@bridge_gpt/mcp-server 0.2.54 → 0.2.55

package/build/commands.generated.js

@@ -2,18 +2,19 @@
 // This file is produced by scripts/bundle-commands.js
 export const COMMANDS = {
     "bridge-research.md": "Run multi-source, fact-checked web research via Bridge API and save a cited report locally.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nExecute all steps in this command as a simple linear sequence of MCP tool calls.\n\n## Step 1 — Parse Arguments\n\nParse `$ARGUMENTS`:\n- The required `query` is the full text of `$ARGUMENTS` after removing any recognized flags.\n- An optional `--ticket <KEY>` flag captures a Jira ticket key (e.g., `BAPI-123`) to associate the research with a specific ticket. If `--ticket` appears, treat the immediately following token as the ticket key and remove both from the query.\n- If `$ARGUMENTS` is empty, or the query (after flag removal) is blank, stop immediately and display:\n\n```\nUsage: /bridge-research <question> [--ticket PROJ-123]\nExample: /bridge-research \"Best practices for rate limiting in FastAPI?\"\n```\n\n## Step 2 — Resolve Docs Directory\n\nCall the `ping` MCP tool (no parameters) and read `docs_dir` from its first (JSON) content item. Store that path as `docs_dir`.\n\n## Step 3 — Run Research\n\nCall `request_deep_research` with:\n- `query`: the parsed question\n- `wait_for_result`: `true`\n- `ticket_number`: the value from `--ticket` if provided; omit this parameter entirely if not present\n\nThis step polls until the research completes (up to 15 minutes) and returns the full cited report directly. The tool appends a literal `Saved to <path>` line to its result — extract that line and store the path as `saved_path`.\n\nIf the tool returns an error, stop immediately and display:\n\n```\nResearch failed: <error message from the tool>\n```\n\n## Step 4 — Confirm\n\nDisplay a confirmation message:\n\n```\nResearch complete.\nSaved to: {saved_path}\n```\n\n## Final Summary\n\nDisplay a summary block:\n\n```\n## Bridge Research Report\n\n- **Query**: <query>\n- **Status**: Completed\n- **Local File**: {saved_path}\n```\n\nOn failure at any step, stop immediately, display which step failed and the error details, and do not proceed.\n",
-    "conduct-epic.md": "---\nschedulable: true\narguments: {\"positionals\":[{\"name\":\"epicKey\",\"type\":\"string\",\"required\":true}],\"flags\":[{\"name\":\"tickets\",\"flag\":\"--tickets\",\"type\":\"string\"},{\"name\":\"baseBranch\",\"flag\":\"--base-branch\",\"type\":\"string\"},{\"name\":\"checkpointPath\",\"flag\":\"--checkpoint-path\",\"type\":\"string\"}]}\n---\n\n# Conduct Epic: $ARGUMENTS\n\n$ARGUMENTS\n\nThis command drives one multi-ticket epic from an approved ticket list to a finished `epic/<EPIC>` branch, one ticket at a time. It is the LLM half of the LLM-as-conductor pivot: there is no server-side reconciler here, no background worker, and no durable run row — the checkpoint file written by the packaged `conduct-epic` CLI plus the live state of GitHub *is* the entire memory of the loop.\n\nCadence is not an argument of this command. `/loop` owns the interval, this command owns exactly one reconcile-then-act step, and the two compose without either one holding state for the other.\n\nIt composes work that already exists rather than reimplementing it: `/review-and-start --auto --base-branch 'epic/<EPIC>' <KEY>` spawns each ticket's worker, the `merge_pull_request` MCP tool merges a green and approved pull request, `parse_repository` (`action: \"start\"`, then `action: \"status\"`) re-indexes the repository after each merge so the next ticket's plan sees its predecessor's code, and the packaged `conduct-epic` CLI (`init`, `status`, `checkpoint set`, `finish`, `spawn`) owns every durable file operation.\n\n---\n\n# Instructions\n\nYou are executing a 5-stage tick. Run the stages in order, take **exactly one** action from the Stage 3 detection table, write **exactly one** checkpoint in Stage 4, then stop. Do not loop internally, do not take a second action because the first one looked cheap, and do not carry assumptions from a previous tick — every tick reconciles from scratch.\n\nThe \"exactly one checkpoint\" rule has **three explicitly documented exemptions** and no others: the two print-only parks, `init_failed` (Stage 1) and `foreign_lock` (Stage 2), which stop before Stage 3; and the `all_done` tick (Row 1), which has no in-flight ticket to name in a `checkpoint set` command. Stage 4 states each one.\n\n## Stage 0 — Arguments and Ping\n\n1. **Parse `$ARGUMENTS`** into exactly one epic positional and the three optional flags. Accept no other input shape.\n\n   - **`<EPIC>`**: exactly one positional token, which must match `[A-Z]+-[0-9]+` (e.g. `BAPI-798`). Zero epic positionals, more than one positional, or a positional that does not match the pattern is malformed input. Extra positionals are rejected rather than ignored.\n   - **`--tickets <K1,K2,…>`** (and the equals form `--tickets=<K1,K2,…>`): a non-empty, comma-separated, **ordered** list of ticket keys. Preserve the caller's order exactly — it is the execution order of the epic. Every entry must match `[A-Z]+-[0-9]+` after trimming surrounding whitespace; reject a malformed key, an empty entry, and a duplicate key. This flag is required **only on the first tick** (see Stage 1); later ticks read the order from the checkpoint.\n   - **`--base-branch <branch>`** (and the equals form `--base-branch=<branch>`): validated with the same rules as `/start-tickets` Stage 0 — after trimming surrounding whitespace it must be non-empty, at most 255 characters, must not start with `-`, and must not contain ASCII control characters (`0x00`–`0x1F` or `0x7F`). It is the branch `epic/<EPIC>` is cut from at `init` time; it is not the pull-request base of a ticket, which is always `epic/<EPIC>`.\n   - **`--checkpoint-path <path>`** (and the equals form `--checkpoint-path=<path>`): must be a non-empty string after trimming, checked **before** it is used as a path or interpolated into a CLI invocation. When omitted, the CLI's own default (`~/.config/bridge/conduct/<repo>/<EPIC>.json`) applies and `status` prints the resolved path.\n\n   Reject malformed input before any side effect: an unsupported flag, a flag given without its value, a `--tickets` list that fails the rules above, a `--base-branch` value that fails validation, an empty `--checkpoint-path`, a missing epic, or an extra positional. On any of these, stop immediately and display:\n\n   ```\n   Invalid arguments.\n   Usage: /conduct-epic [flags] <EPIC>\n     <EPIC>                    required, matches [A-Z]+-[0-9]+ (e.g. BAPI-798)\n     --tickets K1,K2,…         ordered ticket keys; required only on the first tick\n     --base-branch <branch>    branch epic/<EPIC> is cut from (default: the repo base)\n     --checkpoint-path <path>  override the checkpoint file location\n   ```\n\n2. **Connectivity check**: call the `ping` MCP tool with **no parameters**. If the call fails, or does not return `\"status\": \"ok\"`, stop immediately — before Stage 1 initialization, before any CLI invocation, and before any state is written — and display:\n\n   ```\n   Connectivity check failed. Please verify:\n   - Check that the Bridge API MCP server is configured in your editor's MCP settings\n   - Check that BAPI_BASE_URL is set and the server is reachable\n   - Check that BAPI_API_KEY is valid\n   - Check that BAPI_REPO_NAME matches a configured repository\n   ```\n\n3. **Execution model.** This command is one tick; run it with `/loop 5m /conduct-epic <EPIC>`; each tick re-reads the checkpoint and GitHub, takes at most one action, and exits. `/loop` is the external driver that re-invokes this command — it is not an internal loop this command runs, and cadence is never an argument of this command.\n\n4. **Bash interpolation rule (global; applies to every Bash invocation in every stage).** Before interpolating any dynamic value — the epic key, a ticket key, a branch name, a checkpoint path, a prompt-file path, a JSON blob, a journal line — replace every `'` in the value with `'\\''`, then wrap the complete value in single quotes. Never expand a dynamic value unquoted, and never build a command by concatenating an unquoted variable. Credentials must never appear in a command argument, in printed output, in a journal line, or in a prompt file: the CLI and the MCP tools resolve their own credentials from the environment and the user-scoped credential store.\n\n5. **Packaged CLI launcher (`BAPI_MCP_CLI`); global, applies to every packaged-CLI invocation in every stage.** Resolve the launcher **once**, here in Stage 0, and reuse that one resolved value for the rest of the tick. Call it `<launcher>`.\n\n   - Read the `BAPI_MCP_CLI` environment variable.\n   - **Unset, empty, or whitespace-only** — `<launcher>` is exactly `npx -y @bridge_gpt/mcp-server`. This is the default, and the resulting shell command is byte-identical to what it was before this override existed.\n   - **Otherwise** — `<launcher>` is that value, used verbatim as the command prefix. It names a local launcher, such as `node /absolute/path/to/mcp_server/build/index.js`. Use it for local pilots and pre-publish verification.\n\n   When the override is set, apply item 4's single-quote escaping rule to `<launcher>` before interpolating it into a Bash command string, keep every dynamic argument independently quoted rather than concatenated into the launcher value, and never put a credential or a credential-bearing environment assignment into it. A stale local build is exactly as misleading as a stale npm publish: rebuild with `cd mcp_server && npm run build` before relying on the override.\n\nThis stage is **critical** — stop immediately on failure. Do not proceed to Stage 1.\n\n## Stage 1 — Initialize If No Checkpoint\n\nRun the first status probe through the **Bash tool**, forwarding `--checkpoint-path '<path>'` only when the user supplied it:\n\n```\n<launcher> conduct-epic status '<EPIC>' --json\n```\n\nA zero-exit response whose `checkpoint_exists` is `false` is the **only** condition under which initialization is allowed.\n\n- **`checkpoint_exists` is `false`** — this is the first tick. `--tickets` is required here, and **only** here: if it was not supplied, halt with the Stage 0 usage message and initialize nothing. On every later tick `--tickets` is optional and ignored, because the ordered list already lives in the checkpoint. Otherwise run, forwarding `--base-branch '<b>'` and `--checkpoint-path '<p>'` only when supplied:\n\n  ```\n  <launcher> conduct-epic init '<EPIC>' --tickets '<K1,K2,…>'\n  ```\n\n  Print the initialization preflight output **verbatim** — do not summarize it, do not suppress its announcements, and do not reorder it. `init` runs one preflight that lists every failure at once, and that listing is the operator's only diagnostic when it refuses.\n\n  On a **non-zero** exit, `init_failed` is a **print-only park**: emit `NEEDS_HUMAN:init_failed` with the complete secret-free output as evidence, print exactly one bounded, secret-free stdout journal line describing this invocation, and stop the tick. Do **not** call `checkpoint set` and do not otherwise mutate durable state. There is nothing to write to: when initialization failed, no writable checkpoint may exist at all, and any checkpoint that does exist may be the unreadable one that caused the failure. Do not attempt a second initialization in the same tick and do not fall through to Stage 2.\n\n- **`checkpoint_exists` is `true`** — an epic that already has a checkpoint must **never** trigger `init`. The CLI deliberately refuses reinitialization (`already initialized`), so a retry is not a recovery path; it is a bug in the caller. Skip straight to Stage 2.\n\n- **The status command exits non-zero** (a corrupt or wrong-version checkpoint, for example) — treat it exactly like a failed init, including the print-only rule: preserve the secret-free stderr as evidence, emit `NEEDS_HUMAN:init_failed`, print one journal line, call no `checkpoint set`, and stop the tick. `status` never rewrites a checkpoint it could not read, so nothing has been damaged.\n\n## Stage 2 — Reconcile From Status JSON\n\nRun the status probe **again**, with the same conditional `--checkpoint-path '<path>'` forwarding:\n\n```\n<launcher> conduct-epic status '<EPIC>' --json\n```\n\nThis second response is the action snapshot. **This JSON object is the only evidence the tick acts on.** Worker claims are never trusted — a session that says \"CI passed\", \"review approved\", or \"PR merged\" has told you nothing this tick may use. Every one of those facts is re-derived here from GitHub and the server through `status`, and only from there.\n\nThe top-level contract is exactly: `ok`, `epic_key`, `epic_branch`, `checkpoint_path`, `checkpoint_exists`, `all_done`, `ticket`, `worktree_path`, `worktree_exists`, `branch_head`, `worker_commits_since_spawn`, `last_seen_head`, `last_state_change_at`, `stale_for_seconds`, `pr`, `merged_externally`, `ci`, `review`, `parse`, `deadlines`, `scope`, `lock`, `needs_human`, and `probe_errors`.\n\nThe nested objects the detection table reads are:\n\n- `ticket` — the in-flight ticket (the first entry that is not `done`, or `null` when `all_done`): `key`, `status` (`pending`, `in_progress`, `merged`, `done`, `needs_human`), `branch`, `pr_number`, `spawned_at`, `parse_requested_at`, `parse_requested_for_sha`, `review_verdictless_observations`, `review_verdictless_for_sha`, `respawns`, `conflict_attempts`, `counters.sessions_spawned`, `counters.plan_generations_observed`, `counters.merge_attempts`, and `journal`.\n  - `review_verdictless_observations` is a **non-negative integer** and `review_verdictless_for_sha` is **a string or `null`**. They are Row 10's head-bound verdictless counter: the count is only meaningful for the head named beside it, and a count whose head does not equal `pr.head_sha` is spent evidence about code that no longer exists. Read them together or not at all.\n  - `parse_requested_at` and `parse_requested_for_sha` are **each a string or `null`**. They are RETAINED for the audit trail of epics driven before the scope-status contract existed, and **no row reads them any more**: Row 5 asks the `scope` sub-object directly instead of reconstructing causality from a request timestamp. Do not write them and do not decide on them.\n  - `journal` is the ticket's journal lines, **oldest-first, newest last**, exactly as stored. It is a human-readable audit trail and is **never** the source of a decision: it is capped at 50 lines and evicts oldest-first, so a marker searched for in it would silently vanish after roughly fifty wait ticks and the loop would re-request a parse it had already requested.\n- `pr` — `number`, `state` (`OPEN`, `MERGED`, `CLOSED`), `head_sha`, `base`, `mergeable`, `merge_state`, `updated_at`.\n- `ci` — `required`, `complete`, `stable_across_two_polls`, `head_sha`, and `checks` entries of `name`, `status`, `conclusion`, `required`.\n- `review` — `opted_in`, `source`, `available`, `verdict` (`approved`, `changes_requested`, `unknown`), `head_sha`, `verdictless_disposition`, `verdictless_ceiling`, `config_invalid`.\n  - `verdictless_disposition` is `park`, `fail_open`, or `null`. **`null` means `park`** — it is what a condition that configured no disposition reports, and what an unreadable configuration reports. A value is only ever one of those three; the server-side parser refuses every other spelling outright rather than passing a partially honored one through.\n  - `verdictless_ceiling` is the number of head-bound verdictless observations Row 10 makes before it decides. Read it from this snapshot and compare against it; never hard-code a bound.\n  - `config_invalid` is `true` when the repository's `done_gate_config` exists but could not be read — a `malformed` or `invalid: …` parse. It arrives with `opted_in: true` and `available: false`, because an unreadable review policy is **not** an absent one: reading it as \"no review opt-in\" would merge on CI alone on the strength of a typo. There is no readable condition in that state, so `verdictless_disposition` is `null` and Row 10 parks.\n- `parse` — `status` (`idle`, `queued`, `in_progress`, `succeeded`, `failed`), `terminal`, `started_at`, and `finished_at`. The last two are each **a string or `null`** and are the ISO-8601 times of the current or last parse run. A `null` on either is unavailable evidence and **never** permits advancement — in particular, missing timestamps can never satisfy Row 5's causal check. There is no repository-wide index-branch override field: BAPI-847 retired that control plane, and an epic now gets its own index scope instead of taking the repository's index away.\n- `deadlines` — `soft_seconds`, `hard_seconds`, `elapsed_since_spawn_seconds` (defaults 3600 and 10800).\n- `scope` — the epic's index scope, read directly from the server: `scope_id`, `lifecycle_state`, `freshness_status`, `blocked_reason`, `required_commit_sha`, `indexed_commit_sha`, and `last_error`. It is `null` **only** when this epic declares no scope at all; that is not a probe failure and carries no `probe_errors` entry.\n  - `freshness_status` is one of `fresh`, `pending`, `blocked`, `failed`, `unavailable`, `unobserved_advance`. **`fresh` is the only value that means the index covers this epic's merged code.** `pending` is a refresh still running. `blocked` is an epic advance the server REFUSED to index and will never resolve by waiting — `blocked_reason` names which refusal. `failed` is the scope's own generation failing. `unavailable` means the scope could not be read this tick, or its epic branch head could not be resolved, and is never equivalent to `fresh`. `unobserved_advance` means the epic branch has moved PAST `required_commit_sha` with no accepted observation explaining the move — the two watermarks agree, but at a commit the branch has left behind. It is the one non-`fresh` value this loop can repair itself.\n  - `required_commit_sha` is the commit the scope must cover; `indexed_commit_sha` is the commit it actually has. **They are separate fields because they mean different things** — the required SHA moves the moment a merge is accepted, long before anything is indexed, so a required SHA equal to your merge commit is not evidence that your merge was indexed.\n- `lock` — `held_by_me`, `owner_pid`, `host`, `alive`.\n- `needs_human` — `null`, or `reason`, `evidence`, `at`.\n- `probe_errors` — entries of `probe` and `reason`.\n\nA failed probe leaves its sub-object `null` and is listed in `probe_errors`; it never fails the command. **A `null` sub-object is unavailable evidence, not a negative result.** Never infer a merge, an approval, a CI success, or a parse success from a `null` value, from a missing field, or from narrative output of any kind — an unavailable probe means \"wait for the next tick\", never \"proceed\".\n\n**`pr` is the one sub-object whose `null` has two distinct meanings, and `probe_errors` is what tells them apart:**\n\n- **`pr` is `null` and there is no `{probe: \"pr\"}` entry** — confirmed absence. `gh` was asked and answered that this branch has no pull request. This is the **normal** state of every tick between the first spawn and the moment the worker opens its pull request, it is a negative result the rows may act on, and Rows 6 and 7 exist precisely for it.\n- **`pr` is `null` and there IS a `{probe: \"pr\"}` entry** — unavailable evidence. `gh` could not answer: unauthenticated, rate-limited, offline, or output that did not parse. Treat it as \"wait for the next tick\" and never as absence; a pull request that exists but cannot be seen must not be reasoned about as one that does not exist.\n\nDo not collapse these two into \"no PR\". Reading an outage as absence is how the loop would respawn into, or abandon, a pull request that was there all along.\n\nTwo states stop the tick before any action is selected:\n\n- **Already parked.** If `needs_human` is not `null`, print the stable phrase `already parked`, followed by the persisted `reason`, the persisted string `evidence`, and the persisted `at` timestamp — then stop. Take no action this tick and write no checkpoint. A parked epic is a human's to unpark by editing the checkpoint (`needs_human` back to `null`, the ticket `status` back to `pending`/`in_progress`, counters adjusted if a budget is re-granted). Do not select a new recovery action on top of an existing one.\n- **Foreign lock.** If `lock.held_by_me` is `false` and `lock.alive` is `true`, another live process owns this epic. `foreign_lock` is a **print-only park**: emit `NEEDS_HUMAN:foreign_lock` carrying `lock.owner_pid` and `lock.host` as evidence, print one bounded, secret-free stdout journal line for this invocation, and stop. Do **not** call `checkpoint set`, spawn a session, merge a pull request, or start a parse while that lock is alive. The checkpoint belongs to the other live process; writing to it — even to record a park — is the two-authorities corruption the lock exists to prevent, and `checkpoint set` refuses a live foreign lock anyway.\n\n## Stage 3 — Detect and Take Exactly One Action\n\nEvaluate the rows below **strictly in written order, from top to bottom**. Evaluation stops at the first row whose condition matches; that row's action is the only action this tick performs, and control then proceeds directly to Stage 4. A later row is never \"also\" run because it happens to apply.\n\nOne row states a **forward-looking guard** in its own condition: Row 3 (`stalled`) matches only when no later action or fail-closed row would be selectable for this snapshot. That guard is part of Row 3's condition, not a departure from written order — the ordering rule still holds, and Row 3 simply does not match while a real action is available.\n\nEach row is marked **fail-open** (an uncertain or transient condition waits for the next tick) or **fail-closed** (the tick refuses to act and parks rather than guessing).\n\n### Row 1 — `all_done`: finish the epic and open its pull request\n\nWhen `all_done` is `true`, run `<launcher> conduct-epic finish '<EPIC>'` (forwarding `--checkpoint-path '<p>'` when supplied), then call the `create_pull_request` MCP tool with `head_branch` set to `epic/<EPIC>` and `base_branch` set to `main`. Assemble the `body` from the finish summary: the merged ticket pull requests and any skipped tickets. **Open the pull request; never merge it** — a human reviews and merges the epic into `main`. Then stop.\n\n**This tick writes no checkpoint and does not increment `counters.iterations`.** It is the third documented exemption from Stage 4's one-checkpoint-per-tick rule, and unlike the two print-only parks it reaches Stage 3. The reason is mechanical: `all_done` is `true` exactly when `ticket` is `null`, `checkpoint set` requires `--ticket <KEY>`, and there is no in-flight ticket to name. `finish` is this tick's durable act, and it is the last one the epic needs — so do not invent a ticket key to satisfy the rule, and do not write a checkpoint before or after `finish`.\n\n### Row 2 — Wrong base: do not touch a pull request that is not on the epic branch\n\nWhen `pr.base` is present and is not `epic/<EPIC>`, **do not touch the pull request** — no merge, no comment, no respawn. Select `NEEDS_HUMAN:wrong_base`, carrying the observed `pr.base`, `pr.number`, and the expected `epic/<EPIC>`. **Fail-closed**: only pull requests based on `epic/<EPIC>` are ever acted upon, and this row is evaluated before every work and recovery row precisely so a mis-based pull request cannot be merged, respawned into, or advanced by a later row.\n\n### Row 3 — Hard liveness: a stalled epic parks before it waits\n\nWhen `stale_for_seconds >= deadlines.hard_seconds` (default three hours, `10800`) **and no other row below is selectable this tick**, select `NEEDS_HUMAN:stalled`, carrying the observed `stale_for_seconds` and the `deadlines.hard_seconds` it exceeded. **Fail-closed**.\n\n**This row outranks wait rows only.** Before selecting it, check whether any of the following would otherwise be selectable for this snapshot; if any one of them would, take that row instead and do not park:\n\n- pending work (Row 4's first spawn),\n- Row 5's **action** branches only — branch 1's no-scope completion, branch 2's completion, branch 4's `unobserved_advance` guarded catch-up, branch 5's `shadow_stale_deadline` park, and branch 6's causal `parse_failed` park,\n- a targeted respawn (Rows 7, 9, and 11),\n- CI-red handling (Row 9) and review-remediation handling (Row 11),\n- conflict handling (Row 12),\n- ready-to-merge handling (Row 13),\n- a closed, unmerged pull request (Row 13a),\n- Row 10's **action** branch only — a verdictless review at or above `review.verdictless_ceiling`, whichever disposition it then applies. Row 10's below-ceiling branch is a wait and stays subordinate to this row, exactly as the old unbounded wait did.\n\n`stale_for_seconds` counts from the last observed head or status change, not from the last useful event — so an old but green and approved pull request accumulates staleness while being perfectly actionable. Parking that is the exact defect this guard removes. The row remains ahead of every wait row, because without it a wait would match forever and the epic would sit silent instead of asking for a human.\n\n**Row 5's wait branches are deliberately NOT in that list.** Branch 3 (a `pending` or `unavailable` scope) and branch 7 (nothing matched) are waits, and exempting them would mean a merged ticket whose refresh never starts waits forever with no human ever asked. They accumulate staleness like any other wait and park as `stalled` once `deadlines.hard_seconds` is exceeded.\n\n### Row 4 — Pending ticket: spawn the first worker\n\nWhen `ticket.status` is `pending`, spawn the ticket's session:\n\n```\n/review-and-start --auto --base-branch 'epic/<EPIC>' <KEY>\n```\n\nThen prepare the Stage 4 checkpoint values `spawned_at` (now, ISO-8601), `status=in_progress`, and `counters.sessions_spawned` = the Stage 2 value plus one.\n\n**Fail-closed**: refuse this spawn if the lock is foreign (Stage 2 has already parked in that case). The pull-request base of the spawned worker comes from BAPI-801's `BAPI_BASE_BRANCH` export — `/review-and-start --base-branch` forwards it into the spawned worker shell, and the worker's create-PR step resolves the base from it. That export is what makes the first pull request land on `epic/<EPIC>`; this loop never relies on it alone, because Row 2 independently re-checks the observed `pr.base` on every later tick.\n\n### Row 5 — Merged ticket: refresh the scope index, then mark done\n\nWhen `pr.state` is `MERGED`, or `merged_externally` is `true`, or `ticket.status` is `merged`, the ticket's code is on the epic branch. An **external merge is successful reconciliation, not an error** — a human who merged the pull request by hand did the loop's work for it, and `merged_externally` records exactly that.\n\n**The evidence this row acts on is `scope`, and only `scope`.** The epic's index scope is refreshed by the server the moment it observes the merge: it advances its own `required_commit_sha` to the merge commit and re-parses incrementally. So the question \"has this merge been indexed?\" is a question the scope can answer directly, and this row asks it instead of reconstructing an answer.\n\nThat is a deliberate replacement of the older mechanism. This row used to record the time it called `parse_repository` and the head SHA it called it for, then compare that timestamp against a repository-wide parse run's `started_at` / `finished_at` — because `parse.status` is repository-level and stays `succeeded` from any earlier parse of any earlier ticket, so \"succeeded\" alone proved nothing. Timestamp ordering was the only causality available. It is no longer needed, and inference is strictly worse than an answer: **do not call `parse_repository` from this row, and do not read `parse`, `ticket.parse_requested_at`, or `ticket.parse_requested_for_sha` as freshness evidence.** The server owns the refresh; this loop observes it.\n\nThis row is an **ordered state machine**, evaluated top to bottom, and the first matching branch is the tick's action:\n\n1. **`scope` is `null`** — this epic declares no index scope, so there is nothing to refresh and no freshness to establish. Call `update_jira_status` for the ticket with `target_status` set to the Jira `Done` state, and prepare `status=done`. Journal that the ticket completed with no declared scope. **Fail-open.** An epic that never had a scope must not be blocked by one.\n\n2. **`scope.freshness_status` is `fresh`, and `scope.indexed_commit_sha` equals `scope.required_commit_sha`, both non-null** — the scope's index provably covers the commit the server is holding it to. Only then call `update_jira_status` for the ticket with `target_status` set to the Jira `Done` state, and prepare `status=done`. Journal both observed watermarks.\n\n   **Compare the scope's two watermarks against each other — never against `pr.head_sha` or `branch_head`.** Both of those are the *worker's* pre-merge branch tip: `pr.head_sha` is `headRefOid`, and `branch_head` is `git ls-remote` of the ticket's own branch. What lands on `epic/<EPIC>` is the merge commit GitHub creates, and that differs from the worker's tip under every merge strategy — merge, squash, and rebase alike. Comparing an indexed watermark against either one is therefore false essentially always, and a branch that waits on an always-false condition never marks anything done. For the same reason, do not invent a merge-commit field: the `scope` object carries exactly the seven fields named above, and none of them is one.\n\n   The identity that IS causal runs between the scope's own two watermarks, and it is what replaces the old timestamp ordering. The server advances `required_commit_sha` the moment it observes this merge, and **only the parse** writes `indexed_commit_sha`; the two fields are owned by different writers precisely so their agreement means something. So `indexed == required` is the server's own statement that it has finished indexing everything it was asked to cover. A scope that finished refreshing for a **previous** ticket reads `fresh` too — but it reads it at that previous required commit, and the moment this merge is observed `required` moves ahead of `indexed` and `freshness_status` drops to `pending` until the re-parse lands. If either watermark is `null` the comparison cannot be made, so this branch does not match and the tick falls to branch 6 and waits.\n\n   **The gap this used to see straight through is now detected and repaired.** A merge the server never observed at all leaves both watermarks agreeing at the *previous* commit — which reads exactly like a healthy scope from the row alone — and this branch used to mark the ticket done against an index that did not contain its code. That was not a narrow window that closed on its own; it was a permanent wrong answer whenever nothing observed the merge, and it is what made an unregistered merge webhook a silent correctness failure rather than a latency cost.\n\n   The server now compares `required_commit_sha` against the epic branch's ACTUAL head and reports `unobserved_advance` instead of a false `fresh` (branch 4 below repairs it). So this branch's `fresh` is a stronger statement than it was: the pin is current with the branch *and* the index has caught up to the pin. There remains a genuinely narrow window — the seconds between a merge landing and the server observing it — in which the scope reads `pending` rather than `fresh`, which is a wait, not a wrong answer. If you nevertheless see a `done` whose journaled watermarks match the *previous* ticket's, that is a defect worth recording.\n\n3. **`scope.freshness_status` is `pending`, `unavailable`, or missing** — the refresh is still in flight, or the scope could not be read. Wait. Journal the observed `scope.lifecycle_state`, `scope.required_commit_sha`, and `scope.indexed_commit_sha`. Do not spawn anything and do not advance the next ticket. **An unread scope is never a fresh one**, and neither is one whose epic branch head could not be resolved — that also arrives here as `unavailable`, never as `unobserved_advance`.\n\n4. **`scope.freshness_status` is `unobserved_advance`** — the epic branch has moved past `required_commit_sha` and nothing observed the move. The index provably does not cover the epic's merged code, and unlike branch 3 nothing is in flight that will fix it. Attempt the **guarded catch-up exactly once this tick**:\n\n   ```bash\n   node mcp_server/build/index.js conduct-epic catch-up '<EPIC>' --json\n   ```\n\n   The catch-up replays the missed merges through the SAME AC-10 gate a merge webhook would have used. It has no force, override, or skip-guard flag, and it cannot make that gate say yes — it only makes the gate reachable for a merge nobody observed at the time. Then act on its `outcome`:\n\n   - `repaired`, `parse_scheduled`, `already_current`, or `race_lost` — **wait**, exactly as branch 3 does. Do not mark the ticket done on the catch-up's own say-so: completion still requires branch 2's authoritative `fresh` on a later tick, read from `status`. Journal the outcome and the returned `required_commit_sha`.\n   - `blocked`, `history_limit_exceeded`, `refused_lifecycle`, or `invalid_scope` — a **controlled refusal**. Select `NEEDS_HUMAN:shadow_stale_deadline`, carrying the catch-up `outcome` and its bounded `reason` (a controlled `advance_blocked_*` token for `blocked`) in the bounded `evidence` string. **Fail-closed.**\n   - `unavailable` — the branch or the provider could not be read. **Wait**; this is the same \"we could not look\" condition branch 3 handles, and it often clears on its own.\n\n   **Attempt it once per tick, never in a loop.** A second attempt in the same tick reads the same branch state and reaches the same verdict, and the operation is idempotent precisely so that a repeat is harmless rather than useful.\n\n   `shadow_stale_deadline` is reused rather than given a new token, and that is deliberate: its operational meaning — *the index does not cover this epic's branch and waiting will not fix it* — is exactly true here, and the distinguishing detail lives in the bounded `evidence`. **Do not widen the closed `NEEDS_HUMAN` vocabulary.** It is eleven reasons and no others, nine of them persisted, and it is mirrored in TypeScript, so a twelfth token is a disproportionate blast radius for a distinction one evidence string already carries.\n\n   **Scheduling a parse is not the same as moving the pin.** `parse_repository` — including a scope-addressed call — indexes whatever the scope is *already* pinned to; the parse subprocess is handed `required_commit_sha` and verifies against it. So a parse can never substitute for observation or catch-up, and calling one from this row would do nothing at all. Do not.\n\n5. **`scope.freshness_status` is `blocked`** — the server REFUSED to index this advance, and waiting will never change that. Select `NEEDS_HUMAN:shadow_stale_deadline`, with `scope.blocked_reason` as bounded string evidence, and state plainly in the evidence that **no epic advance was indexed**. **Fail-closed.**\n\n   The controlled reasons and what each one means to a human:\n\n   - `advance_blocked_base_merge` — the base branch was merged forward into the epic branch. The epic branch is pinned at its cut point; a base merge would move that pin.\n   - `advance_blocked_unexpected_parent` — the merge commit does not descend directly from the branch head the scope pinned. Something other than a worker pull request landed on the branch.\n   - `advance_blocked_history_changed` — the pinned head is gone from the branch's history. A force-push or rewrite.\n   - `advance_blocked_unverifiable` — the advance could not be verified at all. Doubt blocks; it never indexes.\n\n   **This park is immediate, and that is deliberate** — it is the one place the pilot escalates faster than v2. The v2 reconciler routes a blocked advance through the same `shadow.stale_deadline_seconds` clock it uses for an ordinary refresh hold, because its hold is anchored on a single durable episode timestamp that every hold reason shares. The pilot has no such episode and no typed `RunPolicy` deadline, and none of the four reasons above resolves by waiting, so waiting out a deadline would only delay a human by up to that deadline and change nothing else. Both conductors emit the **same** `shadow_stale_deadline` reason so one grep finds a refused advance either way; only the latency to the park differs. An operator comparing the two should expect the pilot to ask sooner, not to have asked for a different thing.\n\n6. **`scope.freshness_status` is `failed`** — the scope's own generation failed, which is a different problem from a refused advance. Select `NEEDS_HUMAN:parse_failed`, with `scope.lifecycle_state` and `scope.last_error` as bounded string evidence. **Fail-closed.**\n\n7. **None of branches 1–6 matched** — including a `fresh` scope whose indexed commit still trails its required commit, and a tick where either watermark is missing so no comparison can be made. Wait, and journal the observed scope fields. Neither advance nor park: hard liveness (Row 3) is what eventually escalates a wait that never resolves.\n\n**No next ticket is spawned until this one reaches `done`.** A merged ticket stays in flight until its scope is fresh for its own merge commit, so `ticket` still points at it and Row 4 cannot match for its successor — which is the whole point: the next ticket's review and plan must see this ticket's merged code.\n\n### Row 6 — Worktree working: wait\n\nWhen a worktree exists (`worktree_exists` is `true`), the pull request is **confirmed absent** (`pr` is `null` **and** `probe_errors` carries no `{probe: \"pr\"}` entry), and `worker_commits_since_spawn > 0`, the worker is making observable progress. Wait, and journal the observed `branch_head` and commit count. **Fail-open.**\n\nA `pr: null` accompanied by a PR probe error is unavailable evidence, not absence, and does not match this row — it falls through to Row 15 and waits.\n\n### Row 7 — Soft deadline with no progress: one targeted continuation\n\nWhen the pull request is **confirmed absent** (`pr` is `null` **and** no `{probe: \"pr\"}` entry), `worker_commits_since_spawn` is `0`, and `deadlines.elapsed_since_spawn_seconds >= deadlines.soft_seconds` (default one hour, `3600`), spend the single targeted respawn on kind `continue`, with the prompt:\n\n```\nBranch <b> for <KEY>: continue the existing plan; do not regenerate it; push when done\n```\n\nPrepare `respawns` = the Stage 2 value plus one. `respawns` is **one shared per-ticket budget**, not one allowance per row: Rows 7, 9, and 11 all spend the same single counter, so spending it here leaves nothing for a later CI fix or review fix on this ticket. The attempt **counts only if it pushed** — a later tick observing a non-null `branch_head` is the proof. A respawn that produces no push is a no-op, and a no-op respawn stops the loop rather than spinning: once the one targeted respawn is spent and the ticket still shows no pushed head, select `NEEDS_HUMAN:stalled`. **Fail-closed after one attempt**, which is what keeps a dead worker from being respawned without bound.\n\n### Row 8 — Pull request open, CI not settled: wait\n\nWhen a pull request is open and `ci.complete` is `false` **and no required check in `ci.checks` has already reached a terminal unsuccessful conclusion**, wait; or when `ci.complete` is `true` and green but `ci.stable_across_two_polls` is `false`, wait. **Fail-open.**\n\nThe boolean alone is not the condition. `ci.complete` is `false` both while checks are still running and once a required check has definitively failed, and those are opposite situations: the first is worth waiting on and the second never becomes green on its own. This row therefore covers pending and not-yet-stable checks **only** — a required check with a terminal unsuccessful conclusion is **not** consumed here and falls through to Row 9.\n\n### Row 9 — Pull request open, CI red: one targeted fix\n\nWhen a pull request is open, one or more required checks in `ci.checks` have a terminal unsuccessful conclusion, and there has been no new commit for over 60 minutes (`stale_for_seconds > 3600` is the authoritative no-new-commit duration), spend the single targeted respawn on kind `ci_fix`. Take the failing check names from `ci.checks` — the entries whose `required` is `true` — and use the prompt:\n\n```\nPR #N is red on <checks>: read the check annotations, fix, push; do not regenerate the plan\n```\n\nPrepare `respawns` = the Stage 2 value plus one; the attempt counts only if it pushed. A bare `/implement-ticket --auto` is **prohibited** here: it regenerates the plan, costs a full plan generation, and discards the failure detail the annotations already carry.\n\n`respawns` is **one shared per-ticket budget** across Rows 7, 9, and 11. A continuation respawn spent earlier on this ticket therefore leaves **no** CI-fix attempt: with the counter already at its limit, persistent red CI parks immediately as `NEEDS_HUMAN:ci_red` rather than getting a fix session of its own. Once the shared respawn is spent and CI is still red, select `NEEDS_HUMAN:ci_red` with the failing check names as bounded string evidence. **Fail-closed after one attempt.**\n\n### Row 10 — Review opted in and verdictless: count, then decide\n\nWhen `pr.state` is `OPEN`, `review.opted_in` is `true`, and the review is **verdictless for the current head** — that is, `review.available` is `false`, **or** `review.verdict` is neither `approved` nor `changes_requested` at `pr.head_sha` — the review has produced no usable answer for this code. Count the observation, then act on the count.\n\nThis row covers **both** verdictless shapes on purpose. `review.available` is `false` only when the review read itself failed. A reviewer that ran and died before publishing anything is a different shape: the read succeeds, `review.available` is `true`, and `review.verdict` is `unknown`. Both mean the same thing to this loop — no verdict exists for `pr.head_sha` — and a row that covered only the first would leave the second matching nothing at all.\n\n`changes_requested` at the current head is **explicitly excluded**, so Row 11 stays reachable: a reviewer that asked for changes produced a verdict, and that verdict is Row 11's business. A `changes_requested` verdict whose `review.head_sha` does not equal `pr.head_sha` is about code that no longer exists, so it is verdictless for the current head and does match here.\n\nThe `pr.state` is `OPEN` guard is load-bearing: without it a `CLOSED` pull request whose review is verdictless matches here, ahead of Row 13a, and the loop counts tick after tick on abandoned work instead of parking it.\n\n**Prepare the counter, bound to the current head.**\n\n- If `ticket.review_verdictless_for_sha` does **not** equal `pr.head_sha`, prepare `review_verdictless_observations` = `1` and `review_verdictless_for_sha` = `pr.head_sha`. **The counter resets on a new head.** Observations made against an abandoned head must never spend the budget belonging to the head that replaced it — a later push replaces the code the reviewer failed on, and the new code deserves its own full budget.\n- Otherwise prepare `review_verdictless_observations` = the Stage 2 value plus one, absolute, and leave `review_verdictless_for_sha` at `pr.head_sha`.\n\nWrite both prepared fields through the ordinary single `checkpoint set` for this tick, in every direction below — waiting, parking, and the waived merge alike.\n\n**Compare the prepared count with `review.verdictless_ceiling`**, which the Stage 2 snapshot carries. Compare two numbers read from the snapshot; never compare against a bound written into this prose.\n\n- **Below the ceiling** — wait one tick and journal the observation, naming the prepared count, the ceiling, and the observed `review.available` / `review.verdict`. This is today's behaviour, unchanged. This branch is a **wait**, so Row 3's hard-liveness park still outranks it exactly as it does now.\n- **At or above the ceiling** — apply `review.verdictless_disposition`. This branch is an **action**, so it outranks Row 3, and the ceiling is what an operator actually sees instead of a three-hour `stalled` that names the wrong failure.\n\n**At or above the ceiling, the disposition decides:**\n\n- **`park`** — the default, and the value used whenever `review.verdictless_disposition` is `null`, including when `review.config_invalid` is `true` (a review policy that could not be read carries no readable disposition, so it gets the safe one). Select `NEEDS_HUMAN:review_verdictless_ceiling_reached`, carrying the observed count, the ceiling, `pr.head_sha`, and `review.available` / `review.verdict` as bounded string evidence.\n- **`fail_open`** — treat the ticket as **review-opted-out for this tick** and fall through to Row 13's merge conditions. Row 13 still requires everything else it always required: `pr.state` is `OPEN`, `pr.base` is `epic/<EPIC>`, `ci.complete` is `true`, `ci.stable_across_two_polls` is `true` at `pr.head_sha`, and a non-conflicting pull request. **CI, not a verdict, is the whole of the evidence in that case** — journal `review_waived_verdictless_fail_open` and say the merge proceeded on stable CI evidence alone. Never journal, print, or record it as a review that passed or approved anything.\n\nAny value other than exactly `fail_open` resolves to `park`. There is no third direction, and an unreadable disposition is never treated as permission.\n\n### Row 11 — Changes requested for the current head: one targeted review fix\n\nWhen `pr.state` is `OPEN`, `review.verdict` is `changes_requested`, **and** `review.head_sha` equals `pr.head_sha`, spend the single targeted respawn on kind `review_fix`. The `OPEN` guard is what stops a `changes_requested` verdict left on a **closed** pull request's head from spending this ticket's one respawn on work nobody will merge — that snapshot belongs to Row 13a. The prompt carries the authoritative Stage 2 review evidence: the ticket key, the pull-request number, the reviewed head SHA, and the requested changes. A stale `review.head_sha` (one that does not equal `pr.head_sha`) is a verdict about code that no longer exists and never triggers this row. Prepare `respawns` = the Stage 2 value plus one.\n\n`respawns` is **one shared per-ticket budget** across Rows 7, 9, and 11. Any earlier continuation or CI-fix respawn on this ticket therefore leaves **no** review-fix attempt: with the counter already at its limit, requested changes on the current head park immediately. Once the shared respawn is spent and the verdict still stands for the current head, select `NEEDS_HUMAN:review_changes_requested`. **Fail-closed after one attempt.**\n\n### Row 12 — Conflicting pull request: at most two conflict sessions\n\nWhen `pr.state` is `OPEN` **and** either `pr.mergeable` is `CONFLICTING` or `pr.merge_state` is `DIRTY`, spawn a session of kind `conflict` with the prompt:\n\n```\nrebase onto origin/epic/<EPIC>, resolve, run tests, push\n```\n\nPrepare `conflict_attempts` = the Stage 2 value plus one. The conflict budget is **two** sessions and is counted separately from the single targeted respawn of Rows 7, 9, and 11 — a rebase is a different failure mode from a stalled or red worker. After the second conflict session, if the pull request is still `CONFLICTING`/`DIRTY`, select `NEEDS_HUMAN:conflict`. **Fail-closed after two attempts.**\n\nA **closed** pull request is frequently left `CONFLICTING`/`DIRTY` by GitHub, so without the `pr.state` is `OPEN` guard this row would match ahead of Row 13a and spend a rebase session resolving conflicts on a branch nobody will merge.\n\n### Row 13 — Ready to merge\n\nMerge only when **all** of the following hold on the fresh Stage 2 snapshot: the pull request is open (`pr.state` is `OPEN`); `pr.base` is `epic/<EPIC>`; `ci.complete` is `true` and `ci.stable_across_two_polls` is `true` for `ci.head_sha` equal to `pr.head_sha`; the pull request is not conflicting; and review is either opted out (`review.opted_in` is `false`), approved (`review.verdict` is `approved`) with `review.head_sha` equal to `pr.head_sha`, **or** waived by a `fail_open` verdictless ceiling reached in Row 10 this tick.\n\nWhen the waiver path is what reached this row, journal the shared token `review_waived_verdictless_fail_open` alongside the merge line and say plainly that the merge proceeded **on stable CI evidence alone**. That token is the same string the v2 conductor records for the same degradation, so one grep finds every merge that advanced without a verdict whichever conductor drove the epic. Never write it in language that claims the review passed or approved the pull request — it names what was missing, not what was satisfied.\n\nThen call the `merge_pull_request` MCP tool with exactly `pr_number` set to `pr.number` and `expected_head_sha` set to `pr.head_sha`. **The expected SHA is derived only from the fresh Stage 2 status** — never from the checkpoint, never from a worker's report, never from an earlier tick. The checkpoint deliberately stores no expected head; merge identity always comes from a freshly observed `pr.head_sha`. Prepare `counters.merge_attempts` = the Stage 2 value plus one for **every** invocation of the tool, successful or not.\n\nMap the returned envelope:\n\n- **`merged` is `true`** — the only success. It covers `outcome: merged` and `outcome: already_merged`, both of which carry that boolean. Prepare `status=merged` and top-level `counters.merges` = the Stage 2 value plus one.\n- **`outcome: refused` with `reason: head_sha_drift`** — the head moved under the merge. Journal the complete envelope (including `actual_head_sha`) and take a fresh status snapshot on the next tick. Never retry with the stale SHA.\n- **Outcome `lease_held`, `review_not_approved`, or `unknown`, or any envelope carrying `retry_hint: retry_later`** — journal it and wait for the next reconciliation tick.\n- **Outcome `dry_run`, `pending_approval`, `gate_unresolved`, `action_key_mismatch`, `review_unavailable`, `review_source_unsupported`, `error`, or any `refused` result carrying `retry_hint: needs_human`** — select `NEEDS_HUMAN:merge_blocked`. Preserve the **complete** envelope as the evidence, including `hint`, `actual_head_sha`, `ci_summary`, `paths`, and `http_status` whenever those are present; `hint` is usually the exact operator fix. **JSON-stringify that envelope into a bounded, secret-free string** — `evidence` is string data, never an object (see Stage 4).\n\n**Fail-closed**: only `merged: true` is success. A missing, `false`, or malformed `merged` value is never treated as a merge, no matter what `outcome` says alongside it.\n\n### Row 13a — Pull request closed without being merged\n\nWhen `pr.state` is `CLOSED` and the pull request was not merged, the ticket's work has been abandoned on GitHub and nothing this loop does can advance it. Select `NEEDS_HUMAN:merge_blocked`, with bounded string evidence that identifies `pr.state: CLOSED` along with `pr.number`. **Fail-closed** — a closed pull request is never respawned into, reopened, or merged by this loop.\n\n### Row 14 — Local-mode ticket operation refused\n\nWhen a ticket operation returns `409 UNSUPPORTED_IN_LOCAL_MODE`, tolerate it and journal it. The repository is running the local ticket backend, where that response is the documented terminal answer rather than a failure. It introduces **no** new parking reason. **Fail-open.**\n\n### Row 15 — No row matched: journal the snapshot and do nothing else\n\nWhen no row above matches, that is the tick's outcome, not a licence to improvise. Journal a concise summary of the Stage 2 snapshot, take **no** external action — no MCP tool call, no spawn, no merge, no parse — and change **no** row-specific checkpoint field. The single `checkpoint set` this tick writes therefore carries only the universal `counters.iterations` update and its one journal line.\n\nThis row exists because unmatched snapshots are real and reachable: a `stale_for_seconds` or `elapsed_since_spawn_seconds` that is `null` because nothing has been observed yet; a pull request whose CI is complete but not yet stable across two polls. Each of those is a legitimate \"wait for reality to move\" state, and a tick that improvised an action for it would be acting on evidence it does not have. **Fail-open.**\n\nAn open pull request awaiting a review whose `verdict` is still `unknown` is **no longer** one of these. Row 10 now matches that snapshot, counts it, and eventually decides — falling through to here would be the unbounded wait the ceiling exists to end.\n\n### Shared mechanics for every targeted session\n\nRows 7, 9, 11, and 12 spawn a session the same way. The four kinds are exactly `continue`, `ci_fix`, `review_fix`, and `conflict`.\n\n**First, write the prompt file** with the Write tool, at:\n\n```\n~/.config/bridge/conduct/<repo>/<EPIC>/prompts/<KEY>-<kind>-<n>.md\n```\n\nwhere `<EPIC>` and `<KEY>` are the validated keys, `<kind>` is one of the four kinds above, and `<n>` is the applicable absolute attempt number. **`<repo>` is the repository component of the resolved `checkpoint_path` that Stage 2's `status` returned** — read it from there rather than re-deriving it from credentials, from `BAPI_REPO_NAME`, or from anything remembered in conversation. `status` resolves that path itself, including any `--checkpoint-path` override and any `XDG_CONFIG_HOME` redirection, so it is the only value guaranteed to match where the CLI actually keeps this epic's state.\n\n**End every prompt with this exact wording**, so the spawned worker releases its worktree cleanly instead of lingering:\n\n```\nAfter the final pipeline step completes, cleanly end your worker session (for example by issuing the `/exit` command) so the worktree is released and no idle process lingers — but **only when no follow-up remains that you still own**. Do **not** exit while any of the following is true:\n\n- there are unresolved CI failures you are still correcting (the post-PR CI-correction loop in the CI-monitoring step still owns work),\n- review changes were requested and you have not yet addressed them,\n- there is a merge conflict on your PR that you still own,\n- you have unpushed local commits.\n\nExit only after your final branch state is pushed, the done-gate / CI-monitoring workflow required by the recipe has completed, and no CI/review follow-up remains. A clean `SessionEnd` is both the correct terminal lifecycle signal and the point at which the worker should exit.\n```\n\n**Then spawn**, forwarding `--checkpoint-path '<p>'` only when supplied:\n\n```\n<launcher> conduct-epic spawn '<EPIC>' --ticket '<KEY>' --prompt-file '<path>'\n```\n\n`spawn` opens exactly **one** agent tab in the ticket's `worktree_path` running the prompt file's contents. It refuses when the worktree is missing, the prompt file is unreadable, or the lock is held by another live process.\n\n**The budgets are this command's job, not the CLI's.** `spawn` never checks them: it will happily open a fifth tab if asked. One targeted respawn **shared** across Rows 7, 9, and 11 — a single per-ticket `respawns` counter, not one allowance per row — and two conflict sessions in Row 12, are enforced here, by reading the Stage 2 `respawns` and `conflict_attempts` before choosing the row.\n\nAfter a **successful** spawn, prepare `counters.sessions_spawned` = the Stage 2 value plus one. If the spawn command itself fails, do **not** advance `respawns`, `conflict_attempts`, or `counters.sessions_spawned` — a session that never opened has consumed no budget.\n\nKeep credentials, raw environment values, and unrelated command output out of prompt files and out of the spawn command's arguments. The spawned agent resolves its own credentials.\n\n## Stage 4 — Checkpoint and Stop\n\nEvery tick that reaches this stage ends with **exactly one** checkpoint command and **exactly one** journal line. There are **three exemptions**, and they divide into two kinds:\n\n- **Two print-only parks, before Stage 3.** `init_failed` (Stage 1) and `foreign_lock` (Stage 2) stop the tick *before* Stage 3 and write nothing durable at all — they print their `NEEDS_HUMAN:` line and one stdout journal line and stop. Because they never reach Stage 3 they also never increment `counters.iterations`.\n- **The `all_done` tick, inside Stage 3.** Row 1 reaches Stage 3 but has **no in-flight ticket**: `all_done` is `true` exactly when `ticket` is `null`, and `checkpoint set` requires `--ticket <KEY>`. That tick runs `finish`, opens the epic pull request, writes **no** checkpoint, and — as the single stated exception to the rule below — does **not** increment `counters.iterations`.\n\nEvery other tick, including a Row 15 fallthrough, writes here. Run, forwarding `--checkpoint-path '<p>'` whenever the user supplied it:\n\n```\n<launcher> conduct-epic checkpoint set '<EPIC>' --ticket '<KEY>' --field <name> <absolute-value> … --journal '<line>'\n```\n\nRepeat `--field <name> <absolute-value>` once per changed field, and pass `--journal '<line>'` exactly once. Do not issue a second `checkpoint set` in the same tick, and do not split the fields across two invocations — one tick, one auditable write.\n\n**Every value is absolute, computed from the Stage 2 snapshot.** Relative or guessed increments are prohibited: the CLI stores what it is given, so a \"+1\" that was never resolved against a fresh read silently corrupts the count. Compute `n + 1` from the Stage 2 value for `counters.sessions_spawned`, `respawns`, `conflict_attempts`, `counters.merge_attempts`, `counters.iterations`, and `counters.merges`.\n\n`review_verdictless_observations` follows the same absolute rule with one addition: when `ticket.review_verdictless_for_sha` does not equal `pr.head_sha`, the absolute value is `1` rather than `n + 1`, because the counter is bound to a head and resets when the head moves. `review_verdictless_for_sha` is written as the observed `pr.head_sha`. Row 10 owns both fields; no other row writes them.\n\nInclude only the fields the selected row actually affected — typically some of `status`, `spawned_at`, `respawns`, `conflict_attempts`, `review_verdictless_observations`, `review_verdictless_for_sha`, `counters.sessions_spawned`, `counters.merge_attempts`, `counters.iterations`, and `counters.merges`.\n\n**`parse_requested_at` and `parse_requested_for_sha` are no longer written by any row.** The CLI still accepts them so an older checkpoint stays readable, but Row 5 now reads the `scope` sub-object — the server's own answer about whether this merge was indexed — rather than recording a request and timing it. Writing them would record evidence nothing reads.\n\n**`counters.iterations` increments exactly once for every tick that reaches Stage 3**, and it is written in that tick's single `checkpoint set` as the Stage 2 absolute value plus one. It is the one field every such tick updates, including a Row 15 fallthrough — which is why a fallthrough tick's checkpoint contains only `counters.iterations` and its journal line, with no status, retry, merge, or parking mutation. The two print-only parks never reach Stage 3 and so never increment it, and the `all_done` tick reaches Stage 3 but writes no checkpoint, so it does not increment it either.\n\n**Parking** adds two fields to the same single command:\n\n```\n--field status needs_human --field needs_human '{\"reason\":\"<reason>\",\"evidence\":\"<bounded secret-free JSON-stringified envelope or output>\",\"at\":\"<ISO-8601 timestamp>\"}'\n```\n\n**`evidence` is a JSON string, never an object.** The CLI's checkpoint schema accepts only `{reason: string, evidence: string, at: string}` and rejects anything else outright, so an object-valued `evidence` makes `checkpoint set` exit non-zero: the `NEEDS_HUMAN:` line prints, the park never persists, and the next tick repeats the failing action. When the evidence is structured — a merge envelope, a command's output — JSON-stringify it and escape every embedded quote and control character so the result is a single valid JSON string value. Keep it bounded and secret-free.\n\nThe `reason` is one of the closed list below and `at` is an ISO-8601 timestamp. Every `NEEDS_HUMAN:<reason>` line printed by a stage carries the **same** evidence that is persisted here — the printed line and the checkpoint never disagree.\n\nThe parking vocabulary is closed — **eleven reasons** and no others — and it has two partitions:\n\n- **Nine persisted reasons**, each written durably by the single `checkpoint set` above: `stalled`, `ci_red`, `review_changes_requested`, `merge_blocked`, `conflict`, `parse_failed`, `shadow_stale_deadline`, `wrong_base`, and `review_verdictless_ceiling_reached`. A persisted park is what makes the *next* tick report `already parked` and stop.\n  - `shadow_stale_deadline` is the reason for both of Row 5's index refusals — branch 5's `blocked` scope, and branch 4's controlled guarded-catch-up refusal — and it is deliberately **the same token the v2 conductor parks under** for the same condition. One token covers all three because they mean one thing operationally: *the index does not cover this epic's branch and waiting will not fix it*. The distinguishing detail (which refusal, and which controlled `advance_blocked_*` token) travels in the bounded `evidence` string, which is exactly what a bounded evidence field is for and is far cheaper than a twelfth reason mirrored across TypeScript. Both conductors reaching for one string is what lets an operator grep for a refused epic advance without first working out which conductor drove the epic. It is distinct from `parse_failed`: `parse_failed` means the index generation broke, while `shadow_stale_deadline` means the index refused to accept the branch advance at all.\n  - `review_verdictless_ceiling_reached` is Row 10's park, and it is **byte-identical to v2's own token** for the same reason `shadow_stale_deadline` is shared: one grep finds a verdictless ceiling whichever conductor drove the epic. Four alternatives were considered and rejected. `stalled` is the label this row exists to stop emitting — it says the worker died when what actually died was the reviewer. `merge_blocked` is wrong because the merge tool was never called, and its evidence table is built entirely around merge envelopes. `review_changes_requested` is factually false: nobody requested changes, nobody said anything. And a fresh `review_unavailable` token would collide with the merge tool's existing `review_unavailable` *outcome*, which Row 13 already maps to `merge_blocked` — two different conditions answering to one string is exactly the confusion a closed vocabulary exists to prevent.\n- **Two print-only reasons**, which are printed and journaled to stdout for the current invocation only and write nothing durable: `init_failed` and `foreign_lock`. Neither may call `checkpoint set`. A print-only park leaves no durable record, so it does not produce an `already parked` tick — the next tick reconciles from scratch and reports the condition again if it persists.\n\nDo not invent a new reason; a genuinely new failure mode is a change to this command and to the BAPI-805 runbook together.\n\nThe journal line is one line containing the ISO-8601 time, the selected action, and concise evidence. Print it **last**, after the checkpoint command has succeeded, so the operator's final line of output is the tick's durable record.\n\nEvery dynamic value in this stage follows the Stage 0 single-quote rule — the epic key, the ticket key, the checkpoint path, the `needs_human` JSON, and the journal line are each escaped (`'` → `'\\''`) and wrapped in single quotes. Credentials never appear in a checkpoint argument or in journal evidence.\n\n## Operational Guarantees\n\n- **Spec freshness is `/review-and-start`'s job, not a separate check.** Each ticket's review phase runs in a worktree cut from the current `epic/<EPIC>` tip, so its review and its plan already see every predecessor's merged code. This command runs no separate spec-freshness check and needs none.\n- **The checkpoint plus GitHub are the resume point.** Nothing relies on conversation memory. A sleeping laptop merely misses ticks; the next invocation reconciles from scratch and continues where reality actually is.\n- **This command never creates an `epic_run`.** It must never be combined with `setup-epic` on the same epic — the v2 conductor stays active there, and two authorities transitioning one epic is exactly the failure this pivot removes.\n- **`/loop 5m /conduct-epic <EPIC>` is the driver.** The operator runbook is BAPI-805's, not this file's.\n- **Recovery is bounded**: one targeted respawn *shared* across Rows 7, 9, and 11, and two conflict sessions, then park. There is no third chance and no escalating retry.\n- **The first spawn relies on BAPI-801's `BAPI_BASE_BRANCH` contract**, while every tick still independently verifies the observed `pr.base` (Row 2). The export makes the right thing happen; the check catches it when it does not.\n",
+    "conduct-epic.md": "---\nschedulable: true\narguments: {\"positionals\":[{\"name\":\"epicKey\",\"type\":\"string\",\"required\":true}],\"flags\":[{\"name\":\"tickets\",\"flag\":\"--tickets\",\"type\":\"string\"},{\"name\":\"baseBranch\",\"flag\":\"--base-branch\",\"type\":\"string\"},{\"name\":\"checkpointPath\",\"flag\":\"--checkpoint-path\",\"type\":\"string\"}]}\n---\n\n# Conduct Epic: $ARGUMENTS\n\n$ARGUMENTS\n\nThis command drives one multi-ticket epic from an approved ticket list to a finished `epic/<EPIC>` branch, one ticket at a time. It is the LLM half of the LLM-as-conductor pivot: there is no server-side reconciler here, no background worker, and no durable run row — the checkpoint file written by the packaged `implement-epic` CLI plus the live state of GitHub *is* the entire memory of the loop.\n\nCadence is not an argument of this command. `/loop` owns the interval, this command owns exactly one reconcile-then-act step, and the two compose without either one holding state for the other.\n\nIt composes work that already exists rather than reimplementing it: `/review-and-start --auto --base-branch 'epic/<EPIC>' <KEY>` spawns each ticket's worker, the `merge_pull_request` MCP tool merges a green and approved pull request, `parse_repository` (`action: \"start\"`, then `action: \"status\"`) re-indexes the repository after each merge so the next ticket's plan sees its predecessor's code, and the packaged `implement-epic` CLI (`init`, `status`, `checkpoint set`, `finish`, `spawn`) owns every durable file operation.\n\n---\n\n# Instructions\n\nYou are executing a 5-stage tick. Run the stages in order, take **exactly one** action from the Stage 3 detection table, write **exactly one** checkpoint in Stage 4, then stop. Do not loop internally, do not take a second action because the first one looked cheap, and do not carry assumptions from a previous tick — every tick reconciles from scratch.\n\nThe \"exactly one checkpoint\" rule has **three explicitly documented exemptions** and no others: the two print-only parks, `init_failed` (Stage 1) and `foreign_lock` (Stage 2), which stop before Stage 3; and the `all_done` tick (Row 1), which has no in-flight ticket to name in a `checkpoint set` command. Stage 4 states each one.\n\n## Stage 0 — Arguments and Ping\n\n1. **Parse `$ARGUMENTS`** into exactly one epic positional and the three optional flags. Accept no other input shape.\n\n   - **`<EPIC>`**: exactly one positional token, which must match `[A-Z]+-[0-9]+` (e.g. `BAPI-798`). Zero epic positionals, more than one positional, or a positional that does not match the pattern is malformed input. Extra positionals are rejected rather than ignored.\n   - **`--tickets <K1,K2,…>`** (and the equals form `--tickets=<K1,K2,…>`): a non-empty, comma-separated, **ordered** list of ticket keys. Preserve the caller's order exactly — it is the execution order of the epic. Every entry must match `[A-Z]+-[0-9]+` after trimming surrounding whitespace; reject a malformed key, an empty entry, and a duplicate key. This flag is required **only on the first tick** (see Stage 1); later ticks read the order from the checkpoint.\n   - **`--base-branch <branch>`** (and the equals form `--base-branch=<branch>`): validated with the same rules as `/start-tickets` Stage 0 — after trimming surrounding whitespace it must be non-empty, at most 255 characters, must not start with `-`, and must not contain ASCII control characters (`0x00`–`0x1F` or `0x7F`). It is the branch `epic/<EPIC>` is cut from at `init` time; it is not the pull-request base of a ticket, which is always `epic/<EPIC>`.\n   - **`--checkpoint-path <path>`** (and the equals form `--checkpoint-path=<path>`): must be a non-empty string after trimming, checked **before** it is used as a path or interpolated into a CLI invocation. When omitted, the CLI's own default (`~/.config/bridge/conduct/<repo>/<EPIC>.json`) applies and `status` prints the resolved path.\n\n   Reject malformed input before any side effect: an unsupported flag, a flag given without its value, a `--tickets` list that fails the rules above, a `--base-branch` value that fails validation, an empty `--checkpoint-path`, a missing epic, or an extra positional. On any of these, stop immediately and display:\n\n   ```\n   Invalid arguments.\n   Usage: /conduct-epic [flags] <EPIC>\n     <EPIC>                    required, matches [A-Z]+-[0-9]+ (e.g. BAPI-798)\n     --tickets K1,K2,…         ordered ticket keys; required only on the first tick\n     --base-branch <branch>    branch epic/<EPIC> is cut from (default: the repo base)\n     --checkpoint-path <path>  override the checkpoint file location\n   ```\n\n2. **Connectivity check**: call the `ping` MCP tool with **no parameters**. If the call fails, or does not return `\"status\": \"ok\"`, stop immediately — before Stage 1 initialization, before any CLI invocation, and before any state is written — and display:\n\n   ```\n   Connectivity check failed. Please verify:\n   - Check that the Bridge API MCP server is configured in your editor's MCP settings\n   - Check that BAPI_BASE_URL is set and the server is reachable\n   - Check that BAPI_API_KEY is valid\n   - Check that BAPI_REPO_NAME matches a configured repository\n   ```\n\n3. **Execution model.** This command is one tick; run it with `/loop 5m /conduct-epic <EPIC>`; each tick re-reads the checkpoint and GitHub, takes at most one action, and exits. `/loop` is the external driver that re-invokes this command — it is not an internal loop this command runs, and cadence is never an argument of this command.\n\n4. **Bash interpolation rule (global; applies to every Bash invocation in every stage).** Before interpolating any dynamic value — the epic key, a ticket key, a branch name, a checkpoint path, a prompt-file path, a JSON blob, a journal line — replace every `'` in the value with `'\\''`, then wrap the complete value in single quotes. Never expand a dynamic value unquoted, and never build a command by concatenating an unquoted variable. Credentials must never appear in a command argument, in printed output, in a journal line, or in a prompt file: the CLI and the MCP tools resolve their own credentials from the environment and the user-scoped credential store.\n\n5. **Packaged CLI launcher (`BAPI_MCP_CLI`); global, applies to every packaged-CLI invocation in every stage.** Resolve the launcher **once**, here in Stage 0, and reuse that one resolved value for the rest of the tick. Call it `<launcher>`.\n\n   - Read the `BAPI_MCP_CLI` environment variable.\n   - **Unset, empty, or whitespace-only** — `<launcher>` is exactly `npx -y @bridge_gpt/mcp-server`. This is the default, and the resulting shell command is byte-identical to what it was before this override existed.\n   - **Otherwise** — `<launcher>` is that value, used verbatim as the command prefix. It names a local launcher, such as `node /absolute/path/to/mcp_server/build/index.js`. Use it for local pilots and pre-publish verification.\n\n   When the override is set, apply item 4's single-quote escaping rule to `<launcher>` before interpolating it into a Bash command string, keep every dynamic argument independently quoted rather than concatenated into the launcher value, and never put a credential or a credential-bearing environment assignment into it. A stale local build is exactly as misleading as a stale npm publish: rebuild with `cd mcp_server && npm run build` before relying on the override.\n\nThis stage is **critical** — stop immediately on failure. Do not proceed to Stage 1.\n\n## Stage 1 — Initialize If No Checkpoint\n\nRun the first status probe through the **Bash tool**, forwarding `--checkpoint-path '<path>'` only when the user supplied it:\n\n```\n<launcher> implement-epic status '<EPIC>' --json\n```\n\nA zero-exit response whose `checkpoint_exists` is `false` is the **only** condition under which initialization is allowed.\n\n- **`checkpoint_exists` is `false`** — this is the first tick. `--tickets` is required here, and **only** here: if it was not supplied, halt with the Stage 0 usage message and initialize nothing. On every later tick `--tickets` is optional and ignored, because the ordered list already lives in the checkpoint. Otherwise run, forwarding `--base-branch '<b>'` and `--checkpoint-path '<p>'` only when supplied:\n\n  ```\n  <launcher> implement-epic init '<EPIC>' --tickets '<K1,K2,…>'\n  ```\n\n  Print the initialization preflight output **verbatim** — do not summarize it, do not suppress its announcements, and do not reorder it. `init` runs one preflight that lists every failure at once, and that listing is the operator's only diagnostic when it refuses.\n\n  On a **non-zero** exit, `init_failed` is a **print-only park**: emit `NEEDS_HUMAN:init_failed` with the complete secret-free output as evidence, print exactly one bounded, secret-free stdout journal line describing this invocation, and stop the tick. Do **not** call `checkpoint set` and do not otherwise mutate durable state. There is nothing to write to: when initialization failed, no writable checkpoint may exist at all, and any checkpoint that does exist may be the unreadable one that caused the failure. Do not attempt a second initialization in the same tick and do not fall through to Stage 2.\n\n- **`checkpoint_exists` is `true`** — an epic that already has a checkpoint must **never** trigger `init`. The CLI deliberately refuses reinitialization (`already initialized`), so a retry is not a recovery path; it is a bug in the caller. Skip straight to Stage 2.\n\n- **The status command exits non-zero** (a corrupt or wrong-version checkpoint, for example) — treat it exactly like a failed init, including the print-only rule: preserve the secret-free stderr as evidence, emit `NEEDS_HUMAN:init_failed`, print one journal line, call no `checkpoint set`, and stop the tick. `status` never rewrites a checkpoint it could not read, so nothing has been damaged.\n\n## Stage 2 — Reconcile From Status JSON\n\nRun the status probe **again**, with the same conditional `--checkpoint-path '<path>'` forwarding:\n\n```\n<launcher> implement-epic status '<EPIC>' --json\n```\n\nThis second response is the action snapshot. **This JSON object is the only evidence the tick acts on.** Worker claims are never trusted — a session that says \"CI passed\", \"review approved\", or \"PR merged\" has told you nothing this tick may use. Every one of those facts is re-derived here from GitHub and the server through `status`, and only from there.\n\nThe top-level contract is exactly: `ok`, `epic_key`, `epic_branch`, `checkpoint_path`, `checkpoint_exists`, `all_done`, `ticket`, `worktree_path`, `worktree_exists`, `branch_head`, `worker_commits_since_spawn`, `last_seen_head`, `last_state_change_at`, `stale_for_seconds`, `pr`, `merged_externally`, `ci`, `review`, `parse`, `deadlines`, `scope`, `lock`, `needs_human`, and `probe_errors`.\n\nThe nested objects the detection table reads are:\n\n- `ticket` — the in-flight ticket (the first entry that is not `done`, or `null` when `all_done`): `key`, `status` (`pending`, `in_progress`, `merged`, `done`, `needs_human`), `branch`, `pr_number`, `spawned_at`, `parse_requested_at`, `parse_requested_for_sha`, `review_verdictless_observations`, `review_verdictless_for_sha`, `respawns`, `conflict_attempts`, `counters.sessions_spawned`, `counters.plan_generations_observed`, `counters.merge_attempts`, and `journal`.\n  - `review_verdictless_observations` is a **non-negative integer** and `review_verdictless_for_sha` is **a string or `null`**. They are Row 10's head-bound verdictless counter: the count is only meaningful for the head named beside it, and a count whose head does not equal `pr.head_sha` is spent evidence about code that no longer exists. Read them together or not at all.\n  - `parse_requested_at` and `parse_requested_for_sha` are **each a string or `null`**. They are RETAINED for the audit trail of epics driven before the scope-status contract existed, and **no row reads them any more**: Row 5 asks the `scope` sub-object directly instead of reconstructing causality from a request timestamp. Do not write them and do not decide on them.\n  - `journal` is the ticket's journal lines, **oldest-first, newest last**, exactly as stored. It is a human-readable audit trail and is **never** the source of a decision: it is capped at 50 lines and evicts oldest-first, so a marker searched for in it would silently vanish after roughly fifty wait ticks and the loop would re-request a parse it had already requested.\n- `pr` — `number`, `state` (`OPEN`, `MERGED`, `CLOSED`), `head_sha`, `base`, `mergeable`, `merge_state`, `updated_at`.\n- `ci` — `required`, `complete`, `stable_across_two_polls`, `head_sha`, and `checks` entries of `name`, `status`, `conclusion`, `required`.\n- `review` — `opted_in`, `source`, `available`, `verdict` (`approved`, `changes_requested`, `unknown`), `head_sha`, `verdictless_disposition`, `verdictless_ceiling`, `config_invalid`.\n  - `verdictless_disposition` is `park`, `fail_open`, or `null`. **`null` means `park`** — it is what a condition that configured no disposition reports, and what an unreadable configuration reports. A value is only ever one of those three; the server-side parser refuses every other spelling outright rather than passing a partially honored one through.\n  - `verdictless_ceiling` is the number of head-bound verdictless observations Row 10 makes before it decides. Read it from this snapshot and compare against it; never hard-code a bound.\n  - `config_invalid` is `true` when the repository's `done_gate_config` exists but could not be read — a `malformed` or `invalid: …` parse. It arrives with `opted_in: true` and `available: false`, because an unreadable review policy is **not** an absent one: reading it as \"no review opt-in\" would merge on CI alone on the strength of a typo. There is no readable condition in that state, so `verdictless_disposition` is `null` and Row 10 parks.\n- `parse` — `status` (`idle`, `queued`, `in_progress`, `succeeded`, `failed`), `terminal`, `started_at`, and `finished_at`. The last two are each **a string or `null`** and are the ISO-8601 times of the current or last parse run. A `null` on either is unavailable evidence and **never** permits advancement — in particular, missing timestamps can never satisfy Row 5's causal check. There is no repository-wide index-branch override field: BAPI-847 retired that control plane, and an epic now gets its own index scope instead of taking the repository's index away.\n- `deadlines` — `soft_seconds`, `hard_seconds`, `elapsed_since_spawn_seconds` (defaults 3600 and 10800).\n- `scope` — the epic's index scope, read directly from the server: `scope_id`, `lifecycle_state`, `freshness_status`, `blocked_reason`, `required_commit_sha`, `indexed_commit_sha`, and `last_error`. It is `null` **only** when this epic declares no scope at all; that is not a probe failure and carries no `probe_errors` entry.\n  - `freshness_status` is one of `fresh`, `pending`, `blocked`, `failed`, `unavailable`, `unobserved_advance`. **`fresh` is the only value that means the index covers this epic's merged code.** `pending` is a refresh still running. `blocked` is an epic advance the server REFUSED to index and will never resolve by waiting — `blocked_reason` names which refusal. `failed` is the scope's own generation failing. `unavailable` means the scope could not be read this tick, or its epic branch head could not be resolved, and is never equivalent to `fresh`. `unobserved_advance` means the epic branch has moved PAST `required_commit_sha` with no accepted observation explaining the move — the two watermarks agree, but at a commit the branch has left behind. It is the one non-`fresh` value this loop can repair itself.\n  - `required_commit_sha` is the commit the scope must cover; `indexed_commit_sha` is the commit it actually has. **They are separate fields because they mean different things** — the required SHA moves the moment a merge is accepted, long before anything is indexed, so a required SHA equal to your merge commit is not evidence that your merge was indexed.\n- `lock` — `held_by_me`, `owner_pid`, `host`, `alive`.\n- `needs_human` — `null`, or `reason`, `evidence`, `at`.\n- `probe_errors` — entries of `probe` and `reason`.\n\nA failed probe leaves its sub-object `null` and is listed in `probe_errors`; it never fails the command. **A `null` sub-object is unavailable evidence, not a negative result.** Never infer a merge, an approval, a CI success, or a parse success from a `null` value, from a missing field, or from narrative output of any kind — an unavailable probe means \"wait for the next tick\", never \"proceed\".\n\n**`pr` is the one sub-object whose `null` has two distinct meanings, and `probe_errors` is what tells them apart:**\n\n- **`pr` is `null` and there is no `{probe: \"pr\"}` entry** — confirmed absence. `gh` was asked and answered that this branch has no pull request. This is the **normal** state of every tick between the first spawn and the moment the worker opens its pull request, it is a negative result the rows may act on, and Rows 6 and 7 exist precisely for it.\n- **`pr` is `null` and there IS a `{probe: \"pr\"}` entry** — unavailable evidence. `gh` could not answer: unauthenticated, rate-limited, offline, or output that did not parse. Treat it as \"wait for the next tick\" and never as absence; a pull request that exists but cannot be seen must not be reasoned about as one that does not exist.\n\nDo not collapse these two into \"no PR\". Reading an outage as absence is how the loop would respawn into, or abandon, a pull request that was there all along.\n\nTwo states stop the tick before any action is selected:\n\n- **Already parked.** If `needs_human` is not `null`, print the stable phrase `already parked`, followed by the persisted `reason`, the persisted string `evidence`, and the persisted `at` timestamp — then stop. Take no action this tick and write no checkpoint. A parked epic is a human's to unpark by editing the checkpoint (`needs_human` back to `null`, the ticket `status` back to `pending`/`in_progress`, counters adjusted if a budget is re-granted). Do not select a new recovery action on top of an existing one.\n- **Foreign lock.** If `lock.held_by_me` is `false` and `lock.alive` is `true`, another live process owns this epic. `foreign_lock` is a **print-only park**: emit `NEEDS_HUMAN:foreign_lock` carrying `lock.owner_pid` and `lock.host` as evidence, print one bounded, secret-free stdout journal line for this invocation, and stop. Do **not** call `checkpoint set`, spawn a session, merge a pull request, or start a parse while that lock is alive. The checkpoint belongs to the other live process; writing to it — even to record a park — is the two-authorities corruption the lock exists to prevent, and `checkpoint set` refuses a live foreign lock anyway.\n\n## Stage 3 — Detect and Take Exactly One Action\n\nEvaluate the rows below **strictly in written order, from top to bottom**. Evaluation stops at the first row whose condition matches; that row's action is the only action this tick performs, and control then proceeds directly to Stage 4. A later row is never \"also\" run because it happens to apply.\n\nOne row states a **forward-looking guard** in its own condition: Row 3 (`stalled`) matches only when no later action or fail-closed row would be selectable for this snapshot. That guard is part of Row 3's condition, not a departure from written order — the ordering rule still holds, and Row 3 simply does not match while a real action is available.\n\nEach row is marked **fail-open** (an uncertain or transient condition waits for the next tick) or **fail-closed** (the tick refuses to act and parks rather than guessing).\n\n### Row 1 — `all_done`: finish the epic and open its pull request\n\nWhen `all_done` is `true`, run `<launcher> implement-epic finish '<EPIC>'` (forwarding `--checkpoint-path '<p>'` when supplied), then call the `create_pull_request` MCP tool with `head_branch` set to `epic/<EPIC>` and `base_branch` set to `main`. Assemble the `body` from the finish summary: the merged ticket pull requests and any skipped tickets. **Open the pull request; never merge it** — a human reviews and merges the epic into `main`. Then stop.\n\n**This tick writes no checkpoint and does not increment `counters.iterations`.** It is the third documented exemption from Stage 4's one-checkpoint-per-tick rule, and unlike the two print-only parks it reaches Stage 3. The reason is mechanical: `all_done` is `true` exactly when `ticket` is `null`, `checkpoint set` requires `--ticket <KEY>`, and there is no in-flight ticket to name. `finish` is this tick's durable act, and it is the last one the epic needs — so do not invent a ticket key to satisfy the rule, and do not write a checkpoint before or after `finish`.\n\n### Row 2 — Wrong base: do not touch a pull request that is not on the epic branch\n\nWhen `pr.base` is present and is not `epic/<EPIC>`, **do not touch the pull request** — no merge, no comment, no respawn. Select `NEEDS_HUMAN:wrong_base`, carrying the observed `pr.base`, `pr.number`, and the expected `epic/<EPIC>`. **Fail-closed**: only pull requests based on `epic/<EPIC>` are ever acted upon, and this row is evaluated before every work and recovery row precisely so a mis-based pull request cannot be merged, respawned into, or advanced by a later row.\n\n### Row 3 — Hard liveness: a stalled epic parks before it waits\n\nWhen `stale_for_seconds >= deadlines.hard_seconds` (default three hours, `10800`) **and no other row below is selectable this tick**, select `NEEDS_HUMAN:stalled`, carrying the observed `stale_for_seconds` and the `deadlines.hard_seconds` it exceeded. **Fail-closed**.\n\n**This row outranks wait rows only.** Before selecting it, check whether any of the following would otherwise be selectable for this snapshot; if any one of them would, take that row instead and do not park:\n\n- pending work (Row 4's first spawn),\n- Row 5's **action** branches only — branch 1's no-scope completion, branch 2's completion, branch 4's `unobserved_advance` guarded catch-up, branch 5's `shadow_stale_deadline` park, and branch 6's causal `parse_failed` park,\n- a targeted respawn (Rows 7, 9, and 11),\n- CI-red handling (Row 9) and review-remediation handling (Row 11),\n- conflict handling (Row 12),\n- ready-to-merge handling (Row 13),\n- a closed, unmerged pull request (Row 13a),\n- Row 10's **action** branch only — a verdictless review at or above `review.verdictless_ceiling`, whichever disposition it then applies. Row 10's below-ceiling branch is a wait and stays subordinate to this row, exactly as the old unbounded wait did.\n\n`stale_for_seconds` counts from the last observed head or status change, not from the last useful event — so an old but green and approved pull request accumulates staleness while being perfectly actionable. Parking that is the exact defect this guard removes. The row remains ahead of every wait row, because without it a wait would match forever and the epic would sit silent instead of asking for a human.\n\n**Row 5's wait branches are deliberately NOT in that list.** Branch 3 (a `pending` or `unavailable` scope) and branch 7 (nothing matched) are waits, and exempting them would mean a merged ticket whose refresh never starts waits forever with no human ever asked. They accumulate staleness like any other wait and park as `stalled` once `deadlines.hard_seconds` is exceeded.\n\n### Row 4 — Pending ticket: spawn the first worker\n\nWhen `ticket.status` is `pending`, spawn the ticket's session:\n\n```\n/review-and-start --auto --base-branch 'epic/<EPIC>' <KEY>\n```\n\nThen prepare the Stage 4 checkpoint values `spawned_at` (now, ISO-8601), `status=in_progress`, and `counters.sessions_spawned` = the Stage 2 value plus one.\n\n**Fail-closed**: refuse this spawn if the lock is foreign (Stage 2 has already parked in that case). The pull-request base of the spawned worker comes from BAPI-801's `BAPI_BASE_BRANCH` export — `/review-and-start --base-branch` forwards it into the spawned worker shell, and the worker's create-PR step resolves the base from it. That export is what makes the first pull request land on `epic/<EPIC>`; this loop never relies on it alone, because Row 2 independently re-checks the observed `pr.base` on every later tick.\n\n### Row 5 — Merged ticket: refresh the scope index, then mark done\n\nWhen `pr.state` is `MERGED`, or `merged_externally` is `true`, or `ticket.status` is `merged`, the ticket's code is on the epic branch. An **external merge is successful reconciliation, not an error** — a human who merged the pull request by hand did the loop's work for it, and `merged_externally` records exactly that.\n\n**The evidence this row acts on is `scope`, and only `scope`.** The epic's index scope is refreshed by the server the moment it observes the merge: it advances its own `required_commit_sha` to the merge commit and re-parses incrementally. So the question \"has this merge been indexed?\" is a question the scope can answer directly, and this row asks it instead of reconstructing an answer.\n\nThat is a deliberate replacement of the older mechanism. This row used to record the time it called `parse_repository` and the head SHA it called it for, then compare that timestamp against a repository-wide parse run's `started_at` / `finished_at` — because `parse.status` is repository-level and stays `succeeded` from any earlier parse of any earlier ticket, so \"succeeded\" alone proved nothing. Timestamp ordering was the only causality available. It is no longer needed, and inference is strictly worse than an answer: **do not call `parse_repository` from this row, and do not read `parse`, `ticket.parse_requested_at`, or `ticket.parse_requested_for_sha` as freshness evidence.** The server owns the refresh; this loop observes it.\n\nThis row is an **ordered state machine**, evaluated top to bottom, and the first matching branch is the tick's action:\n\n1. **`scope` is `null`** — this epic declares no index scope, so there is nothing to refresh and no freshness to establish. Call `update_jira_status` for the ticket with `target_status` set to the Jira `Done` state, and prepare `status=done`. Journal that the ticket completed with no declared scope. **Fail-open.** An epic that never had a scope must not be blocked by one.\n\n2. **`scope.freshness_status` is `fresh`, and `scope.indexed_commit_sha` equals `scope.required_commit_sha`, both non-null** — the scope's index provably covers the commit the server is holding it to. Only then call `update_jira_status` for the ticket with `target_status` set to the Jira `Done` state, and prepare `status=done`. Journal both observed watermarks.\n\n   **Compare the scope's two watermarks against each other — never against `pr.head_sha` or `branch_head`.** Both of those are the *worker's* pre-merge branch tip: `pr.head_sha` is `headRefOid`, and `branch_head` is `git ls-remote` of the ticket's own branch. What lands on `epic/<EPIC>` is the merge commit GitHub creates, and that differs from the worker's tip under every merge strategy — merge, squash, and rebase alike. Comparing an indexed watermark against either one is therefore false essentially always, and a branch that waits on an always-false condition never marks anything done. For the same reason, do not invent a merge-commit field: the `scope` object carries exactly the seven fields named above, and none of them is one.\n\n   The identity that IS causal runs between the scope's own two watermarks, and it is what replaces the old timestamp ordering. The server advances `required_commit_sha` the moment it observes this merge, and **only the parse** writes `indexed_commit_sha`; the two fields are owned by different writers precisely so their agreement means something. So `indexed == required` is the server's own statement that it has finished indexing everything it was asked to cover. A scope that finished refreshing for a **previous** ticket reads `fresh` too — but it reads it at that previous required commit, and the moment this merge is observed `required` moves ahead of `indexed` and `freshness_status` drops to `pending` until the re-parse lands. If either watermark is `null` the comparison cannot be made, so this branch does not match and the tick falls to branch 6 and waits.\n\n   **The gap this used to see straight through is now detected and repaired.** A merge the server never observed at all leaves both watermarks agreeing at the *previous* commit — which reads exactly like a healthy scope from the row alone — and this branch used to mark the ticket done against an index that did not contain its code. That was not a narrow window that closed on its own; it was a permanent wrong answer whenever nothing observed the merge, and it is what made an unregistered merge webhook a silent correctness failure rather than a latency cost.\n\n   The server now compares `required_commit_sha` against the epic branch's ACTUAL head and reports `unobserved_advance` instead of a false `fresh` (branch 4 below repairs it). So this branch's `fresh` is a stronger statement than it was: the pin is current with the branch *and* the index has caught up to the pin. There remains a genuinely narrow window — the seconds between a merge landing and the server observing it — in which the scope reads `pending` rather than `fresh`, which is a wait, not a wrong answer. If you nevertheless see a `done` whose journaled watermarks match the *previous* ticket's, that is a defect worth recording.\n\n3. **`scope.freshness_status` is `pending`, `unavailable`, or missing** — the refresh is still in flight, or the scope could not be read. Wait. Journal the observed `scope.lifecycle_state`, `scope.required_commit_sha`, and `scope.indexed_commit_sha`. Do not spawn anything and do not advance the next ticket. **An unread scope is never a fresh one**, and neither is one whose epic branch head could not be resolved — that also arrives here as `unavailable`, never as `unobserved_advance`.\n\n4. **`scope.freshness_status` is `unobserved_advance`** — the epic branch has moved past `required_commit_sha` and nothing observed the move. The index provably does not cover the epic's merged code, and unlike branch 3 nothing is in flight that will fix it. Attempt the **guarded catch-up exactly once this tick**:\n\n   ```bash\n   node mcp_server/build/index.js implement-epic catch-up '<EPIC>' --json\n   ```\n\n   The catch-up replays the missed merges through the SAME AC-10 gate a merge webhook would have used. It has no force, override, or skip-guard flag, and it cannot make that gate say yes — it only makes the gate reachable for a merge nobody observed at the time. Then act on its `outcome`:\n\n   - `repaired`, `parse_scheduled`, `already_current`, or `race_lost` — **wait**, exactly as branch 3 does. Do not mark the ticket done on the catch-up's own say-so: completion still requires branch 2's authoritative `fresh` on a later tick, read from `status`. Journal the outcome and the returned `required_commit_sha`.\n   - `blocked`, `history_limit_exceeded`, `refused_lifecycle`, or `invalid_scope` — a **controlled refusal**. Select `NEEDS_HUMAN:shadow_stale_deadline`, carrying the catch-up `outcome` and its bounded `reason` (a controlled `advance_blocked_*` token for `blocked`) in the bounded `evidence` string. **Fail-closed.**\n   - `unavailable` — the branch or the provider could not be read. **Wait**; this is the same \"we could not look\" condition branch 3 handles, and it often clears on its own.\n\n   **Attempt it once per tick, never in a loop.** A second attempt in the same tick reads the same branch state and reaches the same verdict, and the operation is idempotent precisely so that a repeat is harmless rather than useful.\n\n   `shadow_stale_deadline` is reused rather than given a new token, and that is deliberate: its operational meaning — *the index does not cover this epic's branch and waiting will not fix it* — is exactly true here, and the distinguishing detail lives in the bounded `evidence`. **Do not widen the closed `NEEDS_HUMAN` vocabulary.** It is eleven reasons and no others, nine of them persisted, and it is mirrored in TypeScript, so a twelfth token is a disproportionate blast radius for a distinction one evidence string already carries.\n\n   **Scheduling a parse is not the same as moving the pin.** `parse_repository` — including a scope-addressed call — indexes whatever the scope is *already* pinned to; the parse subprocess is handed `required_commit_sha` and verifies against it. So a parse can never substitute for observation or catch-up, and calling one from this row would do nothing at all. Do not.\n\n5. **`scope.freshness_status` is `blocked`** — the server REFUSED to index this advance, and waiting will never change that. Select `NEEDS_HUMAN:shadow_stale_deadline`, with `scope.blocked_reason` as bounded string evidence, and state plainly in the evidence that **no epic advance was indexed**. **Fail-closed.**\n\n   The controlled reasons and what each one means to a human:\n\n   - `advance_blocked_base_merge` — the base branch was merged forward into the epic branch. The epic branch is pinned at its cut point; a base merge would move that pin.\n   - `advance_blocked_unexpected_parent` — the merge commit does not descend directly from the branch head the scope pinned. Something other than a worker pull request landed on the branch.\n   - `advance_blocked_history_changed` — the pinned head is gone from the branch's history. A force-push or rewrite.\n   - `advance_blocked_unverifiable` — the advance could not be verified at all. Doubt blocks; it never indexes.\n\n   **This park is immediate, and that is deliberate** — it is the one place the pilot escalates faster than v2. The v2 reconciler routes a blocked advance through the same `shadow.stale_deadline_seconds` clock it uses for an ordinary refresh hold, because its hold is anchored on a single durable episode timestamp that every hold reason shares. The pilot has no such episode and no typed `RunPolicy` deadline, and none of the four reasons above resolves by waiting, so waiting out a deadline would only delay a human by up to that deadline and change nothing else. Both conductors emit the **same** `shadow_stale_deadline` reason so one grep finds a refused advance either way; only the latency to the park differs. An operator comparing the two should expect the pilot to ask sooner, not to have asked for a different thing.\n\n6. **`scope.freshness_status` is `failed`** — the scope's own generation failed, which is a different problem from a refused advance. Select `NEEDS_HUMAN:parse_failed`, with `scope.lifecycle_state` and `scope.last_error` as bounded string evidence. **Fail-closed.**\n\n7. **None of branches 1–6 matched** — including a `fresh` scope whose indexed commit still trails its required commit, and a tick where either watermark is missing so no comparison can be made. Wait, and journal the observed scope fields. Neither advance nor park: hard liveness (Row 3) is what eventually escalates a wait that never resolves.\n\n**No next ticket is spawned until this one reaches `done`.** A merged ticket stays in flight until its scope is fresh for its own merge commit, so `ticket` still points at it and Row 4 cannot match for its successor — which is the whole point: the next ticket's review and plan must see this ticket's merged code.\n\n### Row 6 — Worktree working: wait\n\nWhen a worktree exists (`worktree_exists` is `true`), the pull request is **confirmed absent** (`pr` is `null` **and** `probe_errors` carries no `{probe: \"pr\"}` entry), and `worker_commits_since_spawn > 0`, the worker is making observable progress. Wait, and journal the observed `branch_head` and commit count. **Fail-open.**\n\nA `pr: null` accompanied by a PR probe error is unavailable evidence, not absence, and does not match this row — it falls through to Row 15 and waits.\n\n### Row 7 — Soft deadline with no progress: one targeted continuation\n\nWhen the pull request is **confirmed absent** (`pr` is `null` **and** no `{probe: \"pr\"}` entry), `worker_commits_since_spawn` is `0`, and `deadlines.elapsed_since_spawn_seconds >= deadlines.soft_seconds` (default one hour, `3600`), spend the single targeted respawn on kind `continue`, with the prompt:\n\n```\nBranch <b> for <KEY>: continue the existing plan; do not regenerate it; push when done\n```\n\nPrepare `respawns` = the Stage 2 value plus one. `respawns` is **one shared per-ticket budget**, not one allowance per row: Rows 7, 9, and 11 all spend the same single counter, so spending it here leaves nothing for a later CI fix or review fix on this ticket. The attempt **counts only if it pushed** — a later tick observing a non-null `branch_head` is the proof. A respawn that produces no push is a no-op, and a no-op respawn stops the loop rather than spinning: once the one targeted respawn is spent and the ticket still shows no pushed head, select `NEEDS_HUMAN:stalled`. **Fail-closed after one attempt**, which is what keeps a dead worker from being respawned without bound.\n\n### Row 8 — Pull request open, CI not settled: wait\n\nWhen a pull request is open and `ci.complete` is `false` **and no required check in `ci.checks` has already reached a terminal unsuccessful conclusion**, wait; or when `ci.complete` is `true` and green but `ci.stable_across_two_polls` is `false`, wait. **Fail-open.**\n\nThe boolean alone is not the condition. `ci.complete` is `false` both while checks are still running and once a required check has definitively failed, and those are opposite situations: the first is worth waiting on and the second never becomes green on its own. This row therefore covers pending and not-yet-stable checks **only** — a required check with a terminal unsuccessful conclusion is **not** consumed here and falls through to Row 9.\n\n### Row 9 — Pull request open, CI red: one targeted fix\n\nWhen a pull request is open, one or more required checks in `ci.checks` have a terminal unsuccessful conclusion, and there has been no new commit for over 60 minutes (`stale_for_seconds > 3600` is the authoritative no-new-commit duration), spend the single targeted respawn on kind `ci_fix`. Take the failing check names from `ci.checks` — the entries whose `required` is `true` — and use the prompt:\n\n```\nPR #N is red on <checks>: read the check annotations, fix, push; do not regenerate the plan\n```\n\nPrepare `respawns` = the Stage 2 value plus one; the attempt counts only if it pushed. A bare `/implement-ticket --auto` is **prohibited** here: it regenerates the plan, costs a full plan generation, and discards the failure detail the annotations already carry.\n\n`respawns` is **one shared per-ticket budget** across Rows 7, 9, and 11. A continuation respawn spent earlier on this ticket therefore leaves **no** CI-fix attempt: with the counter already at its limit, persistent red CI parks immediately as `NEEDS_HUMAN:ci_red` rather than getting a fix session of its own. Once the shared respawn is spent and CI is still red, select `NEEDS_HUMAN:ci_red` with the failing check names as bounded string evidence. **Fail-closed after one attempt.**\n\n### Row 10 — Review opted in and verdictless: count, then decide\n\nWhen `pr.state` is `OPEN`, `review.opted_in` is `true`, and the review is **verdictless for the current head** — that is, `review.available` is `false`, **or** `review.verdict` is neither `approved` nor `changes_requested` at `pr.head_sha` — the review has produced no usable answer for this code. Count the observation, then act on the count.\n\nThis row covers **both** verdictless shapes on purpose. `review.available` is `false` only when the review read itself failed. A reviewer that ran and died before publishing anything is a different shape: the read succeeds, `review.available` is `true`, and `review.verdict` is `unknown`. Both mean the same thing to this loop — no verdict exists for `pr.head_sha` — and a row that covered only the first would leave the second matching nothing at all.\n\n`changes_requested` at the current head is **explicitly excluded**, so Row 11 stays reachable: a reviewer that asked for changes produced a verdict, and that verdict is Row 11's business. A `changes_requested` verdict whose `review.head_sha` does not equal `pr.head_sha` is about code that no longer exists, so it is verdictless for the current head and does match here.\n\nThe `pr.state` is `OPEN` guard is load-bearing: without it a `CLOSED` pull request whose review is verdictless matches here, ahead of Row 13a, and the loop counts tick after tick on abandoned work instead of parking it.\n\n**Prepare the counter, bound to the current head.**\n\n- If `ticket.review_verdictless_for_sha` does **not** equal `pr.head_sha`, prepare `review_verdictless_observations` = `1` and `review_verdictless_for_sha` = `pr.head_sha`. **The counter resets on a new head.** Observations made against an abandoned head must never spend the budget belonging to the head that replaced it — a later push replaces the code the reviewer failed on, and the new code deserves its own full budget.\n- Otherwise prepare `review_verdictless_observations` = the Stage 2 value plus one, absolute, and leave `review_verdictless_for_sha` at `pr.head_sha`.\n\nWrite both prepared fields through the ordinary single `checkpoint set` for this tick, in every direction below — waiting, parking, and the waived merge alike.\n\n**Compare the prepared count with `review.verdictless_ceiling`**, which the Stage 2 snapshot carries. Compare two numbers read from the snapshot; never compare against a bound written into this prose.\n\n- **Below the ceiling** — wait one tick and journal the observation, naming the prepared count, the ceiling, and the observed `review.available` / `review.verdict`. This is today's behaviour, unchanged. This branch is a **wait**, so Row 3's hard-liveness park still outranks it exactly as it does now.\n- **At or above the ceiling** — apply `review.verdictless_disposition`. This branch is an **action**, so it outranks Row 3, and the ceiling is what an operator actually sees instead of a three-hour `stalled` that names the wrong failure.\n\n**At or above the ceiling, the disposition decides:**\n\n- **`park`** — the default, and the value used whenever `review.verdictless_disposition` is `null`, including when `review.config_invalid` is `true` (a review policy that could not be read carries no readable disposition, so it gets the safe one). Select `NEEDS_HUMAN:review_verdictless_ceiling_reached`, carrying the observed count, the ceiling, `pr.head_sha`, and `review.available` / `review.verdict` as bounded string evidence.\n- **`fail_open`** — treat the ticket as **review-opted-out for this tick** and fall through to Row 13's merge conditions. Row 13 still requires everything else it always required: `pr.state` is `OPEN`, `pr.base` is `epic/<EPIC>`, `ci.complete` is `true`, `ci.stable_across_two_polls` is `true` at `pr.head_sha`, and a non-conflicting pull request. **CI, not a verdict, is the whole of the evidence in that case** — journal `review_waived_verdictless_fail_open` and say the merge proceeded on stable CI evidence alone. Never journal, print, or record it as a review that passed or approved anything.\n\nAny value other than exactly `fail_open` resolves to `park`. There is no third direction, and an unreadable disposition is never treated as permission.\n\n### Row 11 — Changes requested for the current head: one targeted review fix\n\nWhen `pr.state` is `OPEN`, `review.verdict` is `changes_requested`, **and** `review.head_sha` equals `pr.head_sha`, spend the single targeted respawn on kind `review_fix`. The `OPEN` guard is what stops a `changes_requested` verdict left on a **closed** pull request's head from spending this ticket's one respawn on work nobody will merge — that snapshot belongs to Row 13a. The prompt carries the authoritative Stage 2 review evidence: the ticket key, the pull-request number, the reviewed head SHA, and the requested changes. A stale `review.head_sha` (one that does not equal `pr.head_sha`) is a verdict about code that no longer exists and never triggers this row. Prepare `respawns` = the Stage 2 value plus one.\n\n`respawns` is **one shared per-ticket budget** across Rows 7, 9, and 11. Any earlier continuation or CI-fix respawn on this ticket therefore leaves **no** review-fix attempt: with the counter already at its limit, requested changes on the current head park immediately. Once the shared respawn is spent and the verdict still stands for the current head, select `NEEDS_HUMAN:review_changes_requested`. **Fail-closed after one attempt.**\n\n### Row 12 — Conflicting pull request: at most two conflict sessions\n\nWhen `pr.state` is `OPEN` **and** either `pr.mergeable` is `CONFLICTING` or `pr.merge_state` is `DIRTY`, spawn a session of kind `conflict` with the prompt:\n\n```\nrebase onto origin/epic/<EPIC>, resolve, run tests, push\n```\n\nPrepare `conflict_attempts` = the Stage 2 value plus one. The conflict budget is **two** sessions and is counted separately from the single targeted respawn of Rows 7, 9, and 11 — a rebase is a different failure mode from a stalled or red worker. After the second conflict session, if the pull request is still `CONFLICTING`/`DIRTY`, select `NEEDS_HUMAN:conflict`. **Fail-closed after two attempts.**\n\nA **closed** pull request is frequently left `CONFLICTING`/`DIRTY` by GitHub, so without the `pr.state` is `OPEN` guard this row would match ahead of Row 13a and spend a rebase session resolving conflicts on a branch nobody will merge.\n\n### Row 13 — Ready to merge\n\nMerge only when **all** of the following hold on the fresh Stage 2 snapshot: the pull request is open (`pr.state` is `OPEN`); `pr.base` is `epic/<EPIC>`; `ci.complete` is `true` and `ci.stable_across_two_polls` is `true` for `ci.head_sha` equal to `pr.head_sha`; the pull request is not conflicting; and review is either opted out (`review.opted_in` is `false`), approved (`review.verdict` is `approved`) with `review.head_sha` equal to `pr.head_sha`, **or** waived by a `fail_open` verdictless ceiling reached in Row 10 this tick.\n\nWhen the waiver path is what reached this row, journal the shared token `review_waived_verdictless_fail_open` alongside the merge line and say plainly that the merge proceeded **on stable CI evidence alone**. That token is the same string the v2 conductor records for the same degradation, so one grep finds every merge that advanced without a verdict whichever conductor drove the epic. Never write it in language that claims the review passed or approved the pull request — it names what was missing, not what was satisfied.\n\nThen call the `merge_pull_request` MCP tool with exactly `pr_number` set to `pr.number` and `expected_head_sha` set to `pr.head_sha`. **The expected SHA is derived only from the fresh Stage 2 status** — never from the checkpoint, never from a worker's report, never from an earlier tick. The checkpoint deliberately stores no expected head; merge identity always comes from a freshly observed `pr.head_sha`. Prepare `counters.merge_attempts` = the Stage 2 value plus one for **every** invocation of the tool, successful or not.\n\nMap the returned envelope:\n\n- **`merged` is `true`** — the only success. It covers `outcome: merged` and `outcome: already_merged`, both of which carry that boolean. Prepare `status=merged` and top-level `counters.merges` = the Stage 2 value plus one.\n- **`outcome: refused` with `reason: head_sha_drift`** — the head moved under the merge. Journal the complete envelope (including `actual_head_sha`) and take a fresh status snapshot on the next tick. Never retry with the stale SHA.\n- **Outcome `lease_held`, `review_not_approved`, or `unknown`, or any envelope carrying `retry_hint: retry_later`** — journal it and wait for the next reconciliation tick.\n- **Outcome `dry_run`, `pending_approval`, `gate_unresolved`, `action_key_mismatch`, `review_unavailable`, `review_source_unsupported`, `error`, or any `refused` result carrying `retry_hint: needs_human`** — select `NEEDS_HUMAN:merge_blocked`. Preserve the **complete** envelope as the evidence, including `hint`, `actual_head_sha`, `ci_summary`, `paths`, and `http_status` whenever those are present; `hint` is usually the exact operator fix. **JSON-stringify that envelope into a bounded, secret-free string** — `evidence` is string data, never an object (see Stage 4).\n\n**Fail-closed**: only `merged: true` is success. A missing, `false`, or malformed `merged` value is never treated as a merge, no matter what `outcome` says alongside it.\n\n### Row 13a — Pull request closed without being merged\n\nWhen `pr.state` is `CLOSED` and the pull request was not merged, the ticket's work has been abandoned on GitHub and nothing this loop does can advance it. Select `NEEDS_HUMAN:merge_blocked`, with bounded string evidence that identifies `pr.state: CLOSED` along with `pr.number`. **Fail-closed** — a closed pull request is never respawned into, reopened, or merged by this loop.\n\n### Row 14 — Local-mode ticket operation refused\n\nWhen a ticket operation returns `409 UNSUPPORTED_IN_LOCAL_MODE`, tolerate it and journal it. The repository is running the local ticket backend, where that response is the documented terminal answer rather than a failure. It introduces **no** new parking reason. **Fail-open.**\n\n### Row 15 — No row matched: journal the snapshot and do nothing else\n\nWhen no row above matches, that is the tick's outcome, not a licence to improvise. Journal a concise summary of the Stage 2 snapshot, take **no** external action — no MCP tool call, no spawn, no merge, no parse — and change **no** row-specific checkpoint field. The single `checkpoint set` this tick writes therefore carries only the universal `counters.iterations` update and its one journal line.\n\nThis row exists because unmatched snapshots are real and reachable: a `stale_for_seconds` or `elapsed_since_spawn_seconds` that is `null` because nothing has been observed yet; a pull request whose CI is complete but not yet stable across two polls. Each of those is a legitimate \"wait for reality to move\" state, and a tick that improvised an action for it would be acting on evidence it does not have. **Fail-open.**\n\nAn open pull request awaiting a review whose `verdict` is still `unknown` is **no longer** one of these. Row 10 now matches that snapshot, counts it, and eventually decides — falling through to here would be the unbounded wait the ceiling exists to end.\n\n### Shared mechanics for every targeted session\n\nRows 7, 9, 11, and 12 spawn a session the same way. The four kinds are exactly `continue`, `ci_fix`, `review_fix`, and `conflict`.\n\n**First, write the prompt file** with the Write tool, at:\n\n```\n~/.config/bridge/conduct/<repo>/<EPIC>/prompts/<KEY>-<kind>-<n>.md\n```\n\nwhere `<EPIC>` and `<KEY>` are the validated keys, `<kind>` is one of the four kinds above, and `<n>` is the applicable absolute attempt number. **`<repo>` is the repository component of the resolved `checkpoint_path` that Stage 2's `status` returned** — read it from there rather than re-deriving it from credentials, from `BAPI_REPO_NAME`, or from anything remembered in conversation. `status` resolves that path itself, including any `--checkpoint-path` override and any `XDG_CONFIG_HOME` redirection, so it is the only value guaranteed to match where the CLI actually keeps this epic's state.\n\n**End every prompt with this exact wording**, so the spawned worker releases its worktree cleanly instead of lingering:\n\n```\nAfter the final pipeline step completes, cleanly end your worker session (for example by issuing the `/exit` command) so the worktree is released and no idle process lingers — but **only when no follow-up remains that you still own**. Do **not** exit while any of the following is true:\n\n- there are unresolved CI failures you are still correcting (the post-PR CI-correction loop in the CI-monitoring step still owns work),\n- review changes were requested and you have not yet addressed them,\n- there is a merge conflict on your PR that you still own,\n- you have unpushed local commits.\n\nExit only after your final branch state is pushed, the done-gate / CI-monitoring workflow required by the recipe has completed, and no CI/review follow-up remains. A clean `SessionEnd` is both the correct terminal lifecycle signal and the point at which the worker should exit.\n```\n\n**Then spawn**, forwarding `--checkpoint-path '<p>'` only when supplied:\n\n```\n<launcher> implement-epic spawn '<EPIC>' --ticket '<KEY>' --prompt-file '<path>'\n```\n\n`spawn` opens exactly **one** agent tab in the ticket's `worktree_path` running the prompt file's contents. It refuses when the worktree is missing, the prompt file is unreadable, or the lock is held by another live process.\n\n**The budgets are this command's job, not the CLI's.** `spawn` never checks them: it will happily open a fifth tab if asked. One targeted respawn **shared** across Rows 7, 9, and 11 — a single per-ticket `respawns` counter, not one allowance per row — and two conflict sessions in Row 12, are enforced here, by reading the Stage 2 `respawns` and `conflict_attempts` before choosing the row.\n\nAfter a **successful** spawn, prepare `counters.sessions_spawned` = the Stage 2 value plus one. If the spawn command itself fails, do **not** advance `respawns`, `conflict_attempts`, or `counters.sessions_spawned` — a session that never opened has consumed no budget.\n\nKeep credentials, raw environment values, and unrelated command output out of prompt files and out of the spawn command's arguments. The spawned agent resolves its own credentials.\n\n## Stage 4 — Checkpoint and Stop\n\nEvery tick that reaches this stage ends with **exactly one** checkpoint command and **exactly one** journal line. There are **three exemptions**, and they divide into two kinds:\n\n- **Two print-only parks, before Stage 3.** `init_failed` (Stage 1) and `foreign_lock` (Stage 2) stop the tick *before* Stage 3 and write nothing durable at all — they print their `NEEDS_HUMAN:` line and one stdout journal line and stop. Because they never reach Stage 3 they also never increment `counters.iterations`.\n- **The `all_done` tick, inside Stage 3.** Row 1 reaches Stage 3 but has **no in-flight ticket**: `all_done` is `true` exactly when `ticket` is `null`, and `checkpoint set` requires `--ticket <KEY>`. That tick runs `finish`, opens the epic pull request, writes **no** checkpoint, and — as the single stated exception to the rule below — does **not** increment `counters.iterations`.\n\nEvery other tick, including a Row 15 fallthrough, writes here. Run, forwarding `--checkpoint-path '<p>'` whenever the user supplied it:\n\n```\n<launcher> implement-epic checkpoint set '<EPIC>' --ticket '<KEY>' --field <name> <absolute-value> … --journal '<line>'\n```\n\nRepeat `--field <name> <absolute-value>` once per changed field, and pass `--journal '<line>'` exactly once. Do not issue a second `checkpoint set` in the same tick, and do not split the fields across two invocations — one tick, one auditable write.\n\n**Every value is absolute, computed from the Stage 2 snapshot.** Relative or guessed increments are prohibited: the CLI stores what it is given, so a \"+1\" that was never resolved against a fresh read silently corrupts the count. Compute `n + 1` from the Stage 2 value for `counters.sessions_spawned`, `respawns`, `conflict_attempts`, `counters.merge_attempts`, `counters.iterations`, and `counters.merges`.\n\n`review_verdictless_observations` follows the same absolute rule with one addition: when `ticket.review_verdictless_for_sha` does not equal `pr.head_sha`, the absolute value is `1` rather than `n + 1`, because the counter is bound to a head and resets when the head moves. `review_verdictless_for_sha` is written as the observed `pr.head_sha`. Row 10 owns both fields; no other row writes them.\n\nInclude only the fields the selected row actually affected — typically some of `status`, `spawned_at`, `respawns`, `conflict_attempts`, `review_verdictless_observations`, `review_verdictless_for_sha`, `counters.sessions_spawned`, `counters.merge_attempts`, `counters.iterations`, and `counters.merges`.\n\n**`parse_requested_at` and `parse_requested_for_sha` are no longer written by any row.** The CLI still accepts them so an older checkpoint stays readable, but Row 5 now reads the `scope` sub-object — the server's own answer about whether this merge was indexed — rather than recording a request and timing it. Writing them would record evidence nothing reads.\n\n**`counters.iterations` increments exactly once for every tick that reaches Stage 3**, and it is written in that tick's single `checkpoint set` as the Stage 2 absolute value plus one. It is the one field every such tick updates, including a Row 15 fallthrough — which is why a fallthrough tick's checkpoint contains only `counters.iterations` and its journal line, with no status, retry, merge, or parking mutation. The two print-only parks never reach Stage 3 and so never increment it, and the `all_done` tick reaches Stage 3 but writes no checkpoint, so it does not increment it either.\n\n**Parking** adds two fields to the same single command:\n\n```\n--field status needs_human --field needs_human '{\"reason\":\"<reason>\",\"evidence\":\"<bounded secret-free JSON-stringified envelope or output>\",\"at\":\"<ISO-8601 timestamp>\"}'\n```\n\n**`evidence` is a JSON string, never an object.** The CLI's checkpoint schema accepts only `{reason: string, evidence: string, at: string}` and rejects anything else outright, so an object-valued `evidence` makes `checkpoint set` exit non-zero: the `NEEDS_HUMAN:` line prints, the park never persists, and the next tick repeats the failing action. When the evidence is structured — a merge envelope, a command's output — JSON-stringify it and escape every embedded quote and control character so the result is a single valid JSON string value. Keep it bounded and secret-free.\n\nThe `reason` is one of the closed list below and `at` is an ISO-8601 timestamp. Every `NEEDS_HUMAN:<reason>` line printed by a stage carries the **same** evidence that is persisted here — the printed line and the checkpoint never disagree.\n\nThe parking vocabulary is closed — **eleven reasons** and no others — and it has two partitions:\n\n- **Nine persisted reasons**, each written durably by the single `checkpoint set` above: `stalled`, `ci_red`, `review_changes_requested`, `merge_blocked`, `conflict`, `parse_failed`, `shadow_stale_deadline`, `wrong_base`, and `review_verdictless_ceiling_reached`. A persisted park is what makes the *next* tick report `already parked` and stop.\n  - `shadow_stale_deadline` is the reason for both of Row 5's index refusals — branch 5's `blocked` scope, and branch 4's controlled guarded-catch-up refusal — and it is deliberately **the same token the v2 conductor parks under** for the same condition. One token covers all three because they mean one thing operationally: *the index does not cover this epic's branch and waiting will not fix it*. The distinguishing detail (which refusal, and which controlled `advance_blocked_*` token) travels in the bounded `evidence` string, which is exactly what a bounded evidence field is for and is far cheaper than a twelfth reason mirrored across TypeScript. Both conductors reaching for one string is what lets an operator grep for a refused epic advance without first working out which conductor drove the epic. It is distinct from `parse_failed`: `parse_failed` means the index generation broke, while `shadow_stale_deadline` means the index refused to accept the branch advance at all.\n  - `review_verdictless_ceiling_reached` is Row 10's park, and it is **byte-identical to v2's own token** for the same reason `shadow_stale_deadline` is shared: one grep finds a verdictless ceiling whichever conductor drove the epic. Four alternatives were considered and rejected. `stalled` is the label this row exists to stop emitting — it says the worker died when what actually died was the reviewer. `merge_blocked` is wrong because the merge tool was never called, and its evidence table is built entirely around merge envelopes. `review_changes_requested` is factually false: nobody requested changes, nobody said anything. And a fresh `review_unavailable` token would collide with the merge tool's existing `review_unavailable` *outcome*, which Row 13 already maps to `merge_blocked` — two different conditions answering to one string is exactly the confusion a closed vocabulary exists to prevent.\n- **Two print-only reasons**, which are printed and journaled to stdout for the current invocation only and write nothing durable: `init_failed` and `foreign_lock`. Neither may call `checkpoint set`. A print-only park leaves no durable record, so it does not produce an `already parked` tick — the next tick reconciles from scratch and reports the condition again if it persists.\n\nDo not invent a new reason; a genuinely new failure mode is a change to this command and to the BAPI-805 runbook together.\n\nThe journal line is one line containing the ISO-8601 time, the selected action, and concise evidence. Print it **last**, after the checkpoint command has succeeded, so the operator's final line of output is the tick's durable record.\n\nEvery dynamic value in this stage follows the Stage 0 single-quote rule — the epic key, the ticket key, the checkpoint path, the `needs_human` JSON, and the journal line are each escaped (`'` → `'\\''`) and wrapped in single quotes. Credentials never appear in a checkpoint argument or in journal evidence.\n\n## Operational Guarantees\n\n- **Spec freshness is `/review-and-start`'s job, not a separate check.** Each ticket's review phase runs in a worktree cut from the current `epic/<EPIC>` tip, so its review and its plan already see every predecessor's merged code. This command runs no separate spec-freshness check and needs none.\n- **The checkpoint plus GitHub are the resume point.** Nothing relies on conversation memory. A sleeping laptop merely misses ticks; the next invocation reconciles from scratch and continues where reality actually is.\n- **This command never creates an `epic_run`.** It must never be combined with `setup-epic` on the same epic — the v2 conductor stays active there, and two authorities transitioning one epic is exactly the failure this pivot removes.\n- **`/loop 5m /conduct-epic <EPIC>` is the driver.** The operator runbook is BAPI-805's, not this file's.\n- **Recovery is bounded**: one targeted respawn *shared* across Rows 7, 9, and 11, and two conflict sessions, then park. There is no third chance and no escalating retry.\n- **The first spawn relies on BAPI-801's `BAPI_BASE_BRANCH` contract**, while every tick still independently verifies the observed `pr.base` (Row 2). The export makes the right thing happen; the check catches it when it does not.\n",
     "council.md": "Convene a multi-perspective council on a task via Bridge API and save the resulting report locally.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nExecute all steps in this command as a simple linear sequence of MCP tool calls.\n\n## Step 1 — Parse Arguments\n\nParse `$ARGUMENTS`. The supported invocation is exactly:\n\n```text\n/council <task description> [--mode technical|design|discovery|general] [--debate] [--lenses a,b] [--ticket PROJ-123]\n```\n\nParsing rules:\n\n- Keep every non-flag token in its original order; the joined result is the required `task_description`. Remove each recognized flag, and the value token that belongs to it, from that text.\n- `--mode <value>` accepts exactly `technical`, `design`, `discovery`, or `general`. When `--mode` is omitted, the selected mode is `technical`.\n- `--debate` is a valueless boolean flag. It takes no following token.\n- `--lenses <a,b>` takes one comma-separated value. Split it on commas and keep the non-empty entries as the `lenses` array.\n- `--ticket <KEY>` captures the immediately following token as the ticket key.\n- A missing value for `--mode`, `--lenses`, or `--ticket` — including a value position occupied by another recognized flag — is a validation failure. Never let the next flag become a flag's value.\n\nValidation must finish before any MCP tool call. Stop immediately, display the usage response below, and make no tool call when `$ARGUMENTS` is empty, when it contains only flags, when a flag that needs a value has none, or when `--mode` is given an unsupported value:\n\n```text\nUsage: /council <task description> [--mode technical|design|discovery|general] [--debate] [--lenses a,b] [--ticket PROJ-123]\nExample: /council \"How should we add rate limiting to the LLM client?\" --mode technical\n```\n\n## Step 2 — Resolve Docs Directory\n\nCall the `ping` MCP tool (no parameters) and read `docs_dir` from its first (JSON) content item. Store that path as `docs_dir`. This is context only — do not slugify it, predict a filename from it, or otherwise construct a report path yourself.\n\n## Step 3 — Convene the Council\n\nBefore calling the tool, tell the user calmly what to expect:\n\n```text\nConvening the council. This commonly takes around 15 minutes, and may continue in the background if the client deadline expires.\n```\n\nThen call `request_council` with:\n\n- `task_description`: the parsed task text\n- `mode`: the selected mode\n- `wait_for_result`: `true`\n- `save_locally`: `true`\n- `debate`: `true` — include this parameter **only** when `--debate` was supplied\n- `lenses`: the parsed array — include this parameter **only** when `--lenses` was supplied\n- `ticket_number`: the captured key — include this parameter **only** when `--ticket` was supplied\n\nOmit an optional parameter entirely rather than sending a placeholder: never send `debate` with a false value, never send an empty `lenses` array, and never send an empty `ticket_number` string. Do not send any other parameter — no `providers`, no `concerns`, no prior `council_id` to refine, and no lens pair of your own. Omitted `lenses` already defaults server-side; do not re-implement that default here.\n\n## Step 4 — Report the Outcome\n\nKeep the report status-first and compact: status, then the next action, then supporting detail such as the saved path, `council_id`, or mode.\n\n**Completed.** The tool appends a `Saved files:` block listing one `- <path>` line per saved report. Collect those lines as `saved_paths`; each entry is a `saved_path` reported by the tool. Display them before any optional task, mode, or `docs_dir` context, and never invent or predict a filename:\n\n```text\nCouncil complete.\nSaved to: {saved_path}\n```\n\n**Backgrounded.** A response that exceeded the client deadline but carries a `council_id` is a successful submission, not a failure. Do not display \"failed\", an error banner, or unrecoverable-error wording for it. Display the exact returned id and the recovery action:\n\n```text\nCouncil submitted and still running in the background.\nRetrieve it with `get_council` using {\"council_id\": \"<the exact id returned>\", \"save_locally\": true}.\n```\n\n**Not indexed.** When a `technical` or `discovery` request reports that the repository is not indexed, say so and name the workaround — those two modes are codebase-grounded and need an indexed repository, while `general` needs no index:\n\n```text\nThis repository is not indexed, and {mode} mode needs an indexed repository.\nRerun the same task with `--mode general`.\n```\n\n**Failed.** A tool error that carries no `council_id` is a genuine failure. Surface the tool's own actionable message, stop, and do not invent a retrieval handle:\n\n```text\nCouncil failed: <error message from the tool>\n```\n\n## Final Summary\n\nDisplay a summary block:\n\n```text\n## Council Report\n\n- **Saved to**: {saved_path}\n- **Task**: <task_description>\n- **Mode**: <selected mode>\n- **Status**: Completed\n```\n\nFor a backgrounded council, replace the saved-path line with the returned `council_id` and the `get_council` recovery action, and set the status to `Submitted — running in the background`.\n",
     "create-decision-page.md": "Turn open decisions from this conversation into an interactive HTML decision page, then fold the answers back in.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\n`$ARGUMENTS` is a free-form description of what needs deciding — a topic (\"how we handle rate limiting\"), a list of specific questions, or empty. It is **not** a Jira ticket key.\n\nThis command exists so a decision page can be reached in ordinary conversation, without running a larger automation. A decision page frames each open choice as a card — the question, why it matters, 2-4 concrete options with the consequence of each, and a recommendation — and renders it as a local HTML page the user submits from their browser. The submitted JSON comes back to you and the decisions become settled.\n\nUse it whenever a conversation has accumulated more open choices than are comfortable to settle in prose. Do not use it to ask one simple question — ask that directly.\n\nRun every stage in the main conversation so the user sees the framing as it happens. If a stage fails, say which one and why.\n\n## Stage 1 — Frame the decisions\n\n1. **Gather the candidates.** Take the decisions named in `$ARGUMENTS` plus any open choice raised earlier in this conversation and not yet settled. If `$ARGUMENTS` is empty, use the conversation alone. If you find nothing genuinely open, say so and stop — do not manufacture cards to fill a page.\n\n2. **Write one card per decision.** Each card needs:\n   - `id`: a short stable id, e.g. `D-1`, `D-2`. Ids must be unique — a duplicate is rejected, because the id is the key the user's answer is reported under.\n   - `question`: the decision itself, phrased as a question.\n   - `options`: 2-4 concrete option labels. Do **not** include \"None of these\" or \"Ask about this\" — the renderer appends both automatically, and passing \"None of these\" yourself is rejected.\n   - `option_consequences`: one consequence per option, **parallel to and the same length as** `options`. Say what actually follows from choosing it, not a restatement of the label.\n   - `why_it_matters`: the concrete impact of getting this wrong.\n   - `recommendation_explanation`: why the recommended option is best.\n   - `recommendation_index`: the 0-based index of the recommended option, within range of `options`.\n   - `codebase_evidence` (optional): your assessment plus `file:line` citations, shown collapsed behind a disclosure.\n\n   Give a real recommendation on every card. If one option is obviously right, still supply the strongest alternative as a second option so the user can see what they are ruling out.\n\n3. **Show the list and let the user correct it.** Present the questions and options in chat before rendering anything. The user may add a decision you missed, drop one that is already settled, or reject your framing of a question. Apply their corrections, then proceed. This check is cheap; a page built on the wrong questions is not.\n\n## Stage 2 — Render the page\n\n1. **Pick a slug.** Derive a kebab-case slug from the topic — a few meaningful words, lowercase, non-alphanumerics stripped, at most 60 characters. It **must** match `/^[A-Za-z][A-Za-z0-9_-]*$/`; if it would start with a digit or hyphen, prefix it with `decisions-`. This slug is the `ticket_key`, which accepts any such slug and does not have to be a Jira key.\n\n2. **Call `generate_decision_page`** with the routing fields at the root and everything else nested under `content`. **The nesting is required** — `actionable_items`, `system_goals`, `clear_improvements`, and `implementation_order` passed at the root are silently dropped by the tool's lean input schema, and a call with no `content` at all is rejected.\n   - `ticket_key`: the slug.\n   - `artifact_type`: `review_decisions` (the default).\n   - `output_subdir`: `decisions`.\n   - `output_filename`: `{slug}-decisions.html`.\n   - `labels`: optional presentation overrides — `title`, `intro`, `section_heading`. Set a `title` that names the topic, and an `intro` that says what agreeing to these choices commits the user to.\n   - `content`: an object holding `actionable_items`.\n\n   ```typescript\n   interface DecisionPageContent {\n     actionable_items: Array<{\n       id: string;                    // e.g. \"D-1\"; must be unique\n       question: string;\n       why_it_matters: string;\n       recommendation_explanation: string;\n       options: string[];             // 2-4 labels (no \"None of these\" / \"Ask about this\")\n       option_consequences: string[]; // same length as options\n       recommendation_index: number;  // 0-based within options\n       codebase_evidence?: string;    // optional: assessment + file:line citations\n     }>;\n   }\n   ```\n\n   Example call:\n   ```json\n   {\n     \"ticket_key\": \"rate-limiting\",\n     \"artifact_type\": \"review_decisions\",\n     \"output_subdir\": \"decisions\",\n     \"output_filename\": \"rate-limiting-decisions.html\",\n     \"labels\": { \"title\": \"Rate Limiting Decisions\", \"section_heading\": \"Open Decisions\" },\n     \"content\": {\n       \"actionable_items\": [\n         {\n           \"id\": \"D-1\",\n           \"question\": \"Where should the limit be enforced?\",\n           \"why_it_matters\": \"Determines whether a burst is rejected before or after it reaches the database.\",\n           \"recommendation_explanation\": \"Middleware keeps the limit in one place and protects every route without per-handler work.\",\n           \"options\": [\"In middleware\", \"Per handler\"],\n           \"option_consequences\": [\"One place to change; blunt for routes that need different budgets.\", \"Precise per route; every new route must remember to opt in.\"],\n           \"recommendation_index\": 0,\n           \"codebase_evidence\": \"api/routes/__init__.py:41 already composes shared dependencies for every router.\"\n         }\n       ]\n     }\n   }\n   ```\n\n3. **When the decisions come with framing worth showing**, use `artifact_type: \"pre_ticket_planning\"` instead and add a `system_goals` object inside `content` (`business_goal`, `desired_end_state`, `system_behavior`, and optionally `acceptance_criteria` and `nfrs`). Those render read-only above the cards, each with its own agree / ask / disagree control. Use this when the user needs to see the goal the decisions serve in order to answer them; the plain `review_decisions` page is the right default otherwise.\n\n4. **Handle the response `status`:**\n   - `decision_page_generated`: surface the returned `file_path` and go to Stage 3.\n   - `no_decisions_needed`: no page was written because there was nothing to render. Tell the user, and do not proceed to Stage 3.\n   - `VALIDATION_ERROR`: the message names the field and restates the expected shape. Fix the payload and retry once. If it fails again, report the message verbatim rather than guessing further.\n\nIf the tool fails outright, **output a highly visible warning** (e.g. **⚠ WARNING: The decision page could not be generated** in bold) and fall back to settling the decisions in chat, one at a time. Do not continue silently — the failure must be visible in your output.\n\n## Stage 3 — Capture the answers (stop and wait)\n\n1. **Direct the user to the page.** Give them the `file_path` and tell them to open it in their browser. Explain that they can accept a recommendation, pick another option, reject them all, or flag a card for discussion, and that they can ask you questions in chat before submitting.\n\n2. **Treat each message as a commit or a discussion turn.**\n   - **Commit:** trim the message and try to parse the whole trimmed message as JSON. Treat it as a commit only when the result is an object with all three top-level fields: `ticket_key` (string), `decisions` (object), and `general_comment` (string). The first valid commit-shaped paste commits — do not over-validate the individual cards.\n   - **Discussion:** anything else. Answer it, then keep waiting. If a JSON-shaped paste is missing one of the three fields, say which one rather than treating it as a freeform question.\n   - **In-flight overrides:** if the user changes an answer in chat (\"go with per-handler for D-1\"), record it as an override. On commit, the submitted JSON is the baseline and your recorded overrides win; acknowledge each overridden card in one line.\n\n3. **Resolve every \"ask\" (hard rule).** After accepting a commit, find every item in `decisions` where `choice === \"ask\"`. For each, present the evidence and keep discussing until the user gives an explicit answer. Do not proceed while any `ask` is unresolved, and do not honor \"just skip those\" — an unanswered card is an unmade decision.\n\n4. **Handle \"None of these\".** A `choice` of `\"none\"` means every option you offered was wrong. Ask what the user would do instead and record their answer as the decision. Do not re-render the page for this.\n\n**You MUST stop and wait for the user here.** Do not assume answers, do not proceed on the recommendations, and do not move to Stage 4 until the user commits or explicitly declines. If they decline, say the decisions are unsettled and stop.\n\n## Stage 4 — Fold the answers back\n\n1. **Review the wider implications, then gate on a decision.** Build the review from the complete settled set: the submitted `decisions`, any in-flight overrides recorded during the conversation (these take precedence over the submission), every `\"none\"` answer together with the reason given for it, `general_comment`, and — where this surface tracks acceptance-criterion or NFR stances — those stances too. Do not start the review until every `ask` has an explicit recorded resolution and every in-flight override has been applied.\n\n   Consider three fixed categories, regardless of whether a decision was framed as technical, user-facing, or business-oriented:\n   - **Program / application** — architecture, code paths, operability, maintenance burden, and requirements imposed on other parts of the software.\n   - **User** — end users, new users performing setup, operators, and developers, including prerequisites, setup friction, and additional steps.\n   - **Business** — cost, adoption, support load, compliance, and reversibility.\n\n   Emit only the categories with material second-order implications. For each included category, write at most four one-line bullets of about 25 words, each naming who or what is affected and how — never a restatement of the selected decision. Close with a line naming every considered category that was omitted, e.g. `Considered, nothing material: business.` — omit this closing line only when all three categories have material implications.\n\n   If the review cannot be produced, report that in one line and continue without stalling the workflow or presenting the gate below.\n\n   This review stays in chat: there is no document for this command to update.\n\n   Then present the gate, verbatim: `Implications reviewed. Proceed, or name a decision to revisit.` Accept only a normalized `proceed`, `yes`, `y`, or `go` as a continuation token. Any other response names a decision to reopen: re-settle it in chat, record the new override, rerun the entire implications review against the changed settled set, and present the gate again.\n\n   Literal `auto_approve = true` emits the review but skips this gate entirely; a missing or non-true `auto_approve` value follows the human-in-the-loop path above.\n\n2. **Restate every decision as settled**, in a short list: the question, the chosen answer, and — where the choice went against your recommendation or came from an override — one line on what changes as a result.\n\n3. **Carry `general_comment` as overarching guidance.** It applies across all the decisions, not to any one card. Say plainly how it changes the picture.\n\n4. **Name what these decisions now constrain.** One or two sentences on what is now fixed for the rest of the conversation. From here on, treat the settled answers as the contract — if later work would contradict one, say so and ask rather than quietly re-deciding.\n\nThere is no document to rewrite. The conversation is where the decisions live, unless the user asks you to record them somewhere.\n",
     "create-doc.md": "Generate a design document (TDD, FSD, or PRD) for a Jira ticket, wait for the result, and save it locally.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nExecute all steps in this command as a simple linear sequence of MCP tool calls.\n\n## Step 1 — Parse Arguments\n\n1. **Parse `$ARGUMENTS`**: Extract a required `ticket_key`, a required `--doc-type` flag, an optional `--second-opinion` flag, and an optional `--provider` flag.\n   - Split `$ARGUMENTS` on whitespace.\n   - If `--doc-type` appears followed by one of `tdd`, `fsd`, or `prd`, capture that as `doc_type`.\n   - If `--doc-type` is absent, or is followed by anything other than `tdd`/`fsd`/`prd` (or is the last token), stop immediately and report: \"Usage error: --doc-type requires a document type (tdd, fsd, or prd).\"\n   - If `--second-opinion` appears followed by a provider name (one of `openai`, `anthropic`, `gemini`), capture that provider as `second_opinion_value`.\n   - If `--second-opinion` appears without a provider name following it (or is the last token), set `second_opinion_value = \"auto\"`.\n   - If `--second-opinion` is absent, set `second_opinion_value = null`.\n   - If `--provider` appears followed by a provider name (one of `openai`, `anthropic`, `gemini`), capture that provider as `provider_value`.\n   - If `--provider` appears without a valid provider name following it (or is the last token), stop immediately and report: \"Usage error: --provider requires a provider name (openai, anthropic, or gemini).\"\n   - If `--provider` is absent, set `provider_value = null`.\n   - If both `--second-opinion` and `--provider` are present, `--second-opinion` takes precedence (set `provider_value = null`).\n   - The remaining token (after removing flags and their arguments) is the `ticket_key`.\n\n2. **Validate ticket key format**: Confirm the ticket key matches the Jira key pattern `[A-Za-z][A-Za-z0-9]+-\\d+`. If it does not match (or `ticket_key` is empty or missing), stop immediately and display:\n\n   ```\n   Usage: /create-doc <ticket_key> --doc-type <tdd|fsd|prd> [--second-opinion [provider]] [--provider <name>] (e.g., /create-doc BAPI-150 --doc-type fsd)\n   ```\n\n## Step 2 — Resolve Docs Directory\n\nCall the `ping` MCP tool (no parameters) and read `docs_dir` from its first (JSON) content item. Store that path as `docs_dir`.\n\n## Step 3 — Generate Design Document\n\nCall the `create_doc` MCP tool with:\n- `ticket_number`: the validated `ticket_key`\n- `doc_type`: the parsed `doc_type` (`tdd`, `fsd`, or `prd`)\n- `wait_for_result`: `true`\n- `save_locally`: `true`\n- `second_opinion`: set to `second_opinion_value` if non-null; omit entirely if null\n- `provider`: set to `provider_value` if non-null; omit entirely if null\n\nThis step may take 2-4 minutes while the backend processes the document.\n\nIf the tool returns an error, stop immediately and display:\n\n```\nDesign document generation failed: <error message from the tool>\n```\n\nIf generation did not finish, the document can be retrieved later with the `get_doc` MCP tool using the same `ticket_number` and `doc_type`.\n\n## Step 4 — Confirm Success\n\nResolve the local file path from `doc_type`:\n- `tdd` → `{docs_dir}/architecture/<ticket_key>-architecture-plan.md`\n- `fsd` → `{docs_dir}/fsd/<ticket_key>-fsd-plan.md`\n- `prd` → `{docs_dir}/prd/<ticket_key>-prd-plan.md`\n\nDisplay a confirmation message:\n\n```\nDesign document generated successfully for <ticket_key>\nSaved to: <local file path>\n```\n\n## Final Summary\n\nDisplay a summary block:\n\n```\n## Design Document Report\n\n- **Ticket**: <ticket_key>\n- **Doc Type**: <doc_type>\n- **Status**: Generated successfully\n- **Local File**: <local file path>\n```\n\nOn failure at any step, stop immediately, display which step failed and the error details, and do not proceed.\n",
     "draft-ticket.md": "Draft a local Markdown Jira ticket via the Jira Ticket Writer — no Jira issue is created or updated.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\n`$ARGUMENTS` is the complete description of what the ticket should cover. This command delegates drafting exclusively to the `jira-ticket-writer` agent and produces only a verified local Markdown file — it never creates, updates, or attaches to a Jira issue. To upload a drafted ticket to Jira, use an approved creation workflow separately.\n\nIf any step fails, stop immediately and report which step failed and why, preserving the original description.\n\n## Step 1 — Validate the Description\n\nTreat the complete, original `$ARGUMENTS` value as the description, verbatim — do not derive a title or select a subset of it.\n\nIf `$ARGUMENTS` is empty or whitespace-only, stop immediately and report:\n```\nUsage: /draft-ticket <description of the problem, feature, or improvement>\n```\nDo not attempt to synthesize ticket content from an empty description.\n\n## Step 2 — Resolve the Output Directory\n\nRead the `BAPI_DOCS_DIR` environment variable. If it is not set, default to `docs/tmp`. Store this value as `docs_dir`. The required output directory is `{docs_dir}/tickets/`.\n\n## Step 3 — Delegate to the Jira Ticket Writer\n\nInvoke the `jira-ticket-writer` agent through the Task tool (or a host equivalent you have confirmed can resolve it). This delegation is the only drafting operation this command performs — do not compose a ticket title, summary, requirements, acceptance criteria, or any other ticket body content here, and do not name or invoke a Jira mutation tool (`create_ticket`, `update_ticket_description`, or similar) anywhere in this command.\n\nPass the subagent exactly:\n- The original description, verbatim (from Step 1).\n- The requirement to create the Markdown draft under `{docs_dir}/tickets/`.\n- The requirement to return the exact final file path of the draft it wrote.\n\nDo not copy the writer agent's sizing, research, epic-threshold, approval-gate, or ticket-composition rules into this prompt — those policies belong to the agent, not this command.\n\n**If the writer agent is unavailable in the current host** (for example, no Task-equivalent capability can resolve `jira-ticket-writer` — this is expected in Cursor, which has no `.cursor/agents/` mirror for this agent), this is an unsupported-environment failure. Stop and report:\n```\n/draft-ticket requires an agent-enabled host that can invoke the jira-ticket-writer agent. This\nenvironment cannot resolve it. Switch to a supported agent-enabled environment (e.g. Claude Code),\nor ask the repository tooling owner for guidance.\n```\nPreserve the original description in this failure report. Do not attempt to draft inline as a substitute.\n\n## Step 4 — Verify the Output\n\n1. **Capture exactly one path** from the agent's response. If the response names no path, or more than one plausible output path, this is an output-verification failure — not an agent-invocation failure. Stop and report:\n   ```\n   The Jira Ticket Writer did not return a single verifiable draft path. No substitute was drafted\n   inline. Original description preserved below.\n   ```\n2. **Confirm the returned path** identifies a Markdown file under the resolved `{docs_dir}/tickets/` directory, and that the exact file exists on disk.\n3. **If the agent claimed success but the file does not exist**, treat this as false success — an output-verification failure, distinct from an agent-invocation failure. Stop and report the same output-verification message as above. Do not generate replacement content.\n\nDistinguish the two failure kinds explicitly in whichever report you give: an agent-invocation failure (Step 3) means the writer could not be reached or run at all; an output-verification failure (this step) means the writer ran but its output could not be verified.\n\n## Final Report\n\n**On success**, display:\n\n> **Draft ready**: {verified exact path}\n>\n> No Jira ticket was created or updated. Review the draft, then use an approved creation workflow\n> when you're ready to upload it.\n\nStop after this instruction. Do not offer or initiate automatic continuation into a Jira creation step.\n\n**On failure**, display the agent-invocation or output-verification failure message from Step 3 or Step 4, preserving the original description and stating plainly that no substitute was drafted inline.\n",
+    "drive-epic-run.md": "---\nschedulable: true\narguments: {\"positionals\":[{\"name\":\"epicKey\",\"type\":\"string\",\"required\":true}]}\n---\n\n# Drive Epic Run: $ARGUMENTS\n\n$ARGUMENTS\n\nThis command is the posture for a session that **drives a live Epic Conductor v2 run** — one that `drive-epic` or `setup-epic` has already created. It does not create a run, cut a branch, or approve a plan. It tells the driving session what it may do without asking, what it must stop and ask about, how much it may intervene, which verbs it may use, and what it must write down.\n\nIt replaces the per-run `docs/tmp/<KEY>-driving-manual.md` files. Those were event allowlists (\"you may, without asking: unpark once per ticket…\"), so every recovery their author did not foresee stopped the driver. This command is a **scope rule** instead: it decides by blast radius, not by event.\n\nThe v2 reconciler stays the single transition authority. The driver observes, applies one sanctioned verb at a time, and records what it did. It never edits gate state, never edits code, and never signals a worker process.\n\n---\n\n# Instructions\n\nRun Stage 0 once, at the start of the driving session. Then run the monitoring loop until the run is terminal (`done`, `stopped`, or `abandoned`) or the session must end. Re-invoking the command later (a new session) runs Stage 0 again before anything else.\n\n## Stage 0 — Pre-flight\n\nRun these six operations **in this order**. Any failure stops the command **before any run state is read or written, before the tally is created, and before any recovery verb is applied.**\n\n1. **Parse `$ARGUMENTS`.** Accept exactly one positional token, the epic key, matching `[A-Z][A-Z0-9]+-[0-9]+` (e.g. `BAPI-1145`). Reject a missing key, a second positional, any flag, and a key that does not match. Reject before any side effect and display:\n\n   ```\n   Invalid arguments.\n   Usage: /drive-epic-run <EPIC>\n     <EPIC>  required, matches [A-Z][A-Z0-9]+-[0-9]+ (e.g. BAPI-1145)\n   ```\n\n2. **Connectivity check.** Call the `ping` MCP tool with **no parameters**. If the call fails, or does not return `\"status\": \"ok\"`, stop immediately and display:\n\n   ```\n   Connectivity check failed. Please verify:\n   - Check that the Bridge API MCP server is configured in your editor's MCP settings\n   - Check that BAPI_BASE_URL is set and the server is reachable\n   - Check that BAPI_API_KEY is valid\n   - Check that BAPI_REPO_NAME matches a configured repository\n   ```\n\n3. **Resolve the packaged CLI launcher (`BAPI_MCP_CLI`) once.** Call it `<launcher>` and reuse the one resolved value for the whole session.\n\n   - Read the `BAPI_MCP_CLI` environment variable.\n   - **Unset, empty, or whitespace-only** — `<launcher>` is exactly `npx -y @bridge_gpt/mcp-server`, so the resulting shell command is byte-identical to the default packaged invocation.\n   - **Otherwise** — `<launcher>` is that value, used verbatim as the command prefix (for example `node /absolute/path/to/mcp_server/build/index.js`). Rebuild with `cd mcp_server && npm run build` before relying on a local launcher; a stale local build misleads exactly as a stale npm publish does.\n\n   The packaged CLI resolves its own credentials through the shared resolver (`BAPI_API_KEY` from the environment, then the user-scoped store `~/.config/bridge/credentials.json`). Keep every dynamic argument independently quoted rather than concatenated into the launcher value. Never put a credential, or a credential-bearing environment assignment, into `<launcher>` or into any command argument.\n\n   **Bash interpolation rule (global; applies to every Bash invocation in this command).** Before interpolating any dynamic value — the epic key, a ticket key, a run id, a scope id, a PR number, a path, a JSON body — replace every `'` in the value with `'\\''`, then wrap the complete value in single quotes. Never expand a dynamic value unquoted, and never build a command by concatenating an unquoted variable. Credentials must never appear in a command argument, in printed output, in a log line, in the tally, in the notes log, or in a context file.\n\n4. **Conductor health check.** From the repository root, run `ENVIRONMENT=TEST python scripts/conductor_health_check.py --json` (the BAPI-1146 script; activate `.venv` first). It proves the deterministic conductor suites pass on this checkout and names every failure class that has no drill. Exit `0` passes — undrilled classes alone do not fail it; record their names in the tally header. Exit `1` (a suite failed), `2` (a prerequisite is missing; nothing was proven), or `3` (the failure-class registry is invalid) stops the command.\n\n5. **Conductor readiness read.** Run `'<launcher>' epic-implementer readiness --json` (read-only; it resolves credentials itself). Its server facts come from `GET /jira/epic-runs/conductor-readiness`, with the meanings in the operator runbook's \"What the readiness endpoint reports\" table: `supervisor.required_checks_empty` (an enabled done gate that gates nothing), `reconciler.stale` (no tick within 900 s), `executor.ready` (`null` = never provisioned, `false` = provisioned but not seen recently). The command always exits `0`, so read the report: any check with status `fail`, or any entry in `unavailableSources`, stops the command. Record `warn` checks in the tally header.\n\n6. **Migration head.** Run `alembic -c alembic.ini heads` and `alembic -c alembic.ini current`. The database must report the **single current migration head**. Two heads, or a current revision behind the head, stops the command — a reconciler whose code is ahead of the database misbehaves in ways no verb repairs.\n\n**Run resolution.** Every run-addressed CLI verb and route in this command takes the epic key directly as its run id (`--epic-run-id <EPIC>`). The server resolves it to the single non-terminal run for the repository. Zero matching runs (`404`) or more than one non-terminal run (`409 CONFLICT`) is a non-mutating stop: record it and ask. Never guess, and never pick a run heuristically. Record the resolved `epic_run_id` from the first state read in the tally header.\n\n## Autonomy\n\n**Scope rule (verbatim):**\n\nThe session driving an epic performs every recovery whose blast radius is inside the run without asking, and stops to ask only when a remediation would touch the base branch or any branch outside `epic/<KEY>` and its children, a `.github/workflows/**` file, the reconciler/executor code path itself, a database migration, a plane or database shared with another run, a credential, a governance-tagged guard, or a billable multi-model run beyond the run's budget.\n\n1. **The base branch, or any branch outside `epic/<KEY>` and its children.**\n   Recognition: a ref the remediation would push to, merge into, or rewrite is neither `epic/<KEY>` nor a `branch_name` bound to a ticket of this run in the state read. Match: `target_refs outside run_branches`\n2. **A `.github/workflows/**` file.**\n   Recognition: the remediation's diff adds, edits, or deletes a path under `.github/workflows/`, or the merge path refused with `workflow_files_modified` (`MERGE_BLOCKED_WORKFLOW_FILES_MODIFIED` in `api/library/epic_conductor/reconciler.py`, the `workflow_files_modified` refusal in `api/library/vcs/epic_implementer_merge_service.py`). Match: `touched_paths under .github/workflows/ or signals has workflow_files_modified`\n3. **The reconciler/executor code path itself.**\n   Recognition: the remediation's diff touches a path in the drills lane's transition-authority prefix set (the `run_drills` regex in `.github/workflows/conductor-ci.yml`) or `mcp_server/src/drive-epic.ts`. Match: `touched_paths under api/library/epic_conductor/ api/library/db/epic_* api/library/vcs/epic_observers.py api/library/index_scope/ src/python/llms/executor_jobs.py mcp_server/src/executor/ tests/integration/conductor_harness/ mcp_server/src/drive-epic.ts`\n4. **A database migration.**\n   Recognition: the remediation's diff adds, edits, or deletes an Alembic revision (the `script_location` in `alembic.ini` is `db/alembic`). Match: `touched_paths under db/alembic/versions/`\n5. **A plane or database shared with another run.**\n   Recognition: the state, plane, or scope read shows the resource the remediation acts on — a plane process, an executor lane, a database, an index scope — is bound to, or also serving, another run. Match: `shared_resources nonempty`\n6. **A credential.**\n   Recognition: the operation would read, create, change, disclose, transmit, or record a credential or authorization value — an API key, a token, a git credential helper, a `gh` login, the credential store file. Match: `credentials nonempty`\n7. **A governance-tagged guard.**\n   Recognition: the remediation would change a region marked `BRIDGE-GOVERNANCE-GUARD-BEGIN` / `-END`, a `reconcile_overlap` job reported `governance_guard_change_required`, the ticket parked `merge_overlap_governance_guard`, or the ticket parked `merge_blocked_governance_guard` (merge admission found a child head that changes a marked region; nothing was merged). Escalate to the repository owner, naming the marker's tag; a tag carries no owner field, so read the owner from the region's own comment when it names one. Match: `markers has BRIDGE-GOVERNANCE-GUARD or signals has governance_guard_change_required merge_overlap_governance_guard merge_blocked_governance_guard`\n8. **A billable multi-model run beyond the run's budget.**\n   Recognition: the operation would call `request_council` or `request_deep_research`. Run policy has no multi-model budget field, so the effective budget is `0` and every such call is beyond it. Match: `tools has request_council request_deep_research`\n\n**This is a scope rule, not an event allowlist.** An unforeseen recovery whose blast radius is inside the run is permitted. The absence of a matching row in any table in this command is never grounds to stop. Only a match against the stop list above is.\n\n**How a `Match:` spec reads.** Each spec tests the facts of one proposed remediation. `<fact> outside <set>` matches when some value of the fact is not in the named set. `<fact> under <pattern>…` matches when some path equals a pattern, lies under a pattern ending in `/`, or fits a pattern whose `*` stands for any characters except `/`. `<fact> has <token>…` matches when some value contains one of the tokens. `<fact> nonempty` matches when the fact lists anything. `or` joins alternatives. The facts are: `target_refs` (refs the remediation writes), `run_branches` (`epic/<KEY>` plus every `branch_name` in the state read), `touched_paths` (files its diff changes), `signals` (park reasons, refusal tokens, and job facts it responds to), `markers` (governance markers in the regions it changes), `tools` (MCP tools it calls), `shared_resources`, and `credentials`.\n\n**Stop-and-ask is non-mutating.** It leaves the run exactly as it is. It does not trigger `stop-run` or `abandon-run`. Record the evidence in the tally and the notes log, apply no further verb to the affected ticket, keep driving every other ticket, and put the question to the operator. An operator's answer authorizes that one remediation only; it does not widen this posture.\n\n**Bookkeeping is not remediation.** Writing the tally and the notes log, and committing the notes log on `docs/<KEY>-conductor-run-notes` at wind-down, are bookkeeping. They are exempt from stop-list entry 1 and authorize no code change on any branch.\n\n## Budget\n\nTwo bounds, per ticket:\n\n- **3 interventions per ticket.**\n- **2 hours of no forward progress per ticket.**\n\nWhen either bound is reached, the driver **parks the ticket**: it ends intervention on that ticket instead of applying a fourth verb or intervening further. Parking needs no asking unless the remediation it would have applied independently matches the stop list.\n\n- **An intervention** is one committed, state-changing verb from the verb table below — one whose response reports `committed: true` or an equivalent durable change. A `409`/CAS conflict, a refusal, a retry of a refused call, an idempotent replay (`committed: false`), and a no-op are not interventions. Ending intervention is not an intervention.\n- **A run-wide verb** (`stop-run`, `abandon-run`, `scope-reparse`, `index-scope catch-up`, `policy PATCH`, `resume`) is charged to **each** blocked ticket that motivated it. Record that attribution in the tally.\n- **Forward progress** is measured only by the reconciler's own clock: the ticket's `last_progress_at` in the state read. Two hours of no forward progress is **now minus `last_progress_at`** reaching two hours. When `last_progress_at` is null — the ticket has never progressed — measure from the ticket's `created_at` in the same state read. The reconciler writes `last_progress_at` only on genuine progress, so a regression, pending-state churn, a lease-owner change, or a park does not reset it. The driver keeps no timer of its own.\n- **To park** is to end intervention: call `second_opinion` first (see below), record the exhausted bound in the tally and the notes log, and apply no further verb to that ticket. The reconciler remains the only writer of parking state; the run keeps driving its other tickets. There is no park verb. Do not use the generic ticket `PATCH` or the run-wide `pause` to park.\n\n## Verb table\n\nThe closed set of state-changing recovery verbs. **No verb outside this table is sanctioned.** Reads, pre-flight, tally and notes writes, and `second_opinion` are supporting operations outside this table.\n\n| Verb | When to use | Sanctioned CLI or route | Mapped reasons / preconditions |\n|---|---|---|---|\n| `unpark` | A parked ticket whose cause is now fixed, or whose park a `second_opinion` judged spurious. | `'<launcher>' epic-implementer unpark --epic-run-id <EPIC> --ticket-key <TICKET>` → `POST /jira/epic-runs/runs/{epic_run_id}/tickets/{ticket_key}/unpark` | `pr_not_attached` with the PR attached at the expected head; `review_verdictless_ceiling_reached` only after `second_opinion`; `merge_overlap_peer_parked` after the peer is recovered; `merge_overlap_wait_expired`, `merge_overlap_observation_expired` after the overlap is resolved; `required_checks_empty_no_mode` after `policy PATCH` is observed. |\n| `adopt-current-head-and-unpark` | A parked ticket whose PR head moved off its anchored commit and the current head is the one to keep. | `'<launcher>' epic-implementer adopt-current-head-and-unpark --epic-run-id <EPIC> --ticket-key <TICKET>` → `POST /jira/epic-runs/runs/{epic_run_id}/tickets/{ticket_key}/adopt-current-head-and-unpark` | `pr_not_attached` when the head drifted; `head_sha_drift`; `ci_fix_contract_incomplete` on a green PR. |\n| `stop-run` | The run must end now and its queued work must not run. Terminal. | `'<launcher>' epic-implementer stop-run --epic-run-id <EPIC>` → `POST /jira/epic-runs/runs/{epic_run_id}/stop` | Never an automatic response to Stop-and-ask. Ends every ticket, so `second_opinion` first. A hung worker is recovered by the executor's per-job timeout, or by this verb if the run must end. |\n| `abandon-run` | A stopped or stale run that blocks a fresh run, or whose stale epic shadow fails child reviews closed. Terminal and irreversible. | `'<launcher>' epic-implementer abandon-run --epic-run-id <EPIC>` → `PATCH /jira/epic-runs/runs/{epic_run_id}` with `{\"status\": \"abandoned\"}` | Never on a live run whose work you mean to keep. `second_opinion` first. |\n| `scope-reparse` | The run's index scope is wedged at its pinned commit: `stale` or `seeding` with no parse behind it, and no unobserved merge. | `'<launcher>' epic-implementer scope-reparse --epic-run-id <EPIC> --scope-id <SCOPE>` → `POST /jira/epic-runs/runs/{epic_run_id}/scope-reparse` | `shadow_stale_deadline` when scope status is not `unobserved_advance`. Refused lifecycles return `409 SCOPE_REPARSE_REFUSED`; it does not wait for `ready`. |\n| `dispatch-remediation` | Any code change on a child branch. The **only** way the driver changes code. | `'<launcher>' epic-implementer dispatch-remediation --epic-run-id <EPIC> --ticket-key <TICKET> --context-file <PATH>` → `POST /jira/epic-runs/runs/{epic_run_id}/tickets/{ticket_key}/dispatch-remediation` | `pr_conflicting`; a `code_review` remediation loop that needs the driver's context. Ticket at `code_review`, or parked with `code_review` as its gate. The server picks the job type. Never the MCP tool, never a worktree edit. |\n| `merge-child` | A green, approved child PR the conductor did not merge on its own. | `'<launcher>' epic-implementer merge-child --epic-run-id <EPIC> --pr <NUMBER>` → `POST /jira/epic-runs/runs/{epic_run_id}/merge-child`; the CLI records a local merge through `POST /jira/epic-runs/runs/{epic_run_id}/merge-child/complete` | Only after the gates permit the merge (CI green at head, review approved). Run from the repository checkout with `gh auth status` passing. Never `gh pr merge`, and never the general `merge_pull_request` MCP tool, which is not scoped to the run. |\n| `index-scope catch-up` | A merge landed on `epic/<KEY>` that the scope did not observe. | `POST /jira/index-scope/catch-up` with `{\"repo_name\": \"<REPO>\", \"scope_id\": \"<SCOPE>\"}` (route-only form below) | `shadow_stale_deadline` when scope status is `unobserved_advance`; a child in `parsing` with the scope behind the epic head; a stranded review worker, once the scope reads `ready`. |\n| `policy PATCH` | The run's policy is wrong for the run (a lost key, a cap that must rise). | `PATCH /jira/epic-runs/runs/{epic_run_id}` with the whole `policy_json` (route-only form below) | `required_checks_empty_no_mode`; a cap pause that needs a higher cap. Whole-object replacement: GET the current `policy_json`, change only the key, send the whole object back. |\n| `resume` | A paused run whose cause is cleared. | `POST /jira/epic-runs/runs/{epic_run_id}/resume` with `{\"repo_name\": \"<REPO>\"}` (route-only form below) | `run_control.cause` pauses such as `cap_wall_clock`; `integration_pr_missing` / `integration_pr_ready_failed` after the integration PR is corrected (the correction itself touches the base branch: Stop-and-ask). Never `clear_quarantine` without `second_opinion`. |\n\n`release-job` is **not** a verb. No route exists for it. BAPI-1147 fixed the stranded-worker defect it was specified for, and the executor's per-job timeout bounds any residual case. **The driver never signals, kills, or restarts a worker process directly.**\n\n**Route-only verb forms.** `index-scope catch-up`, `policy PATCH`, and `resume` have no packaged-CLI verb yet. They need `BAPI_API_KEY` and `BAPI_BASE_URL` exported in the driving shell; if either is unset, stop before mutation. Do not read the credential store to obtain a key — that is stop-list entry 6. Pass the key header on stdin (`curl -H @-`), never in argv, output, the tally, the notes log, or a context file. `printf` is a shell builtin, so the key never appears in a process argument list:\n\n```bash\n# index-scope catch-up\nprintf 'X-API-Key: %s\\n' \"$BAPI_API_KEY\" | curl -sS -X POST -H @- -H 'Content-Type: application/json' \\\n  \"$BAPI_BASE_URL/jira/index-scope/catch-up\" -d '{\"repo_name\": \"<REPO>\", \"scope_id\": \"<SCOPE>\"}'\n\n# policy PATCH — send the WHOLE policy_json from a fresh GET, with only the one key changed\nprintf 'X-API-Key: %s\\n' \"$BAPI_API_KEY\" | curl -sS -X PATCH -H @- -H 'Content-Type: application/json' \\\n  \"$BAPI_BASE_URL/jira/epic-runs/runs/<EPIC>\" -d '{\"repo_name\": \"<REPO>\", \"policy_json\": {...}}'\n\n# resume\nprintf 'X-API-Key: %s\\n' \"$BAPI_API_KEY\" | curl -sS -X POST -H @- -H 'Content-Type: application/json' \\\n  \"$BAPI_BASE_URL/jira/epic-runs/runs/<EPIC>/resume\" -d '{\"repo_name\": \"<REPO>\"}'\n```\n\n**Concurrency.** The CLI verbs resolve the ticket's `row_version` themselves and refuse every version flag — never pass one. For the route-only verbs, send the guard the route accepts from a fresh read (`expected_status` on the PATCH, `expected_generation` on resume). On a `409` or any CAS conflict: call `explain_run`, re-read state, and decide again. Retry only if the verb is still needed, and only against the fresh read (its new `row_version`, status, or generation). The refused call is not an intervention.\n\n## Second opinion\n\n**Before any park and before any change you cannot verify with tests, call `second_opinion` and record its answer in the tally.**\n\n- **Why.** The driver assumes the operator is unavailable. A park can stop a ticket, and everything that depends on it, for hours. A second model answers in seconds.\n- **\"Park\"** means ending intervention on a ticket (Budget, above), including `stop-run` and `abandon-run`, which end every ticket.\n- **\"A change you cannot verify with tests\"** is one with no authoritative postcondition. For code, the postcondition is tests. For an operational verb, it is the route's response plus the gate, status, scope, or merge-ledger state it should produce on the next read.\n- **Not a multi-model run.** `second_opinion` is a single-model, synchronous call. It is **not** the billable multi-model run that stop-list entry 8 names; `request_council` and `request_deep_research` are.\n- Record a sanitized summary of the answer, never the raw model reply or provider payload.\n- If the call fails, times out, or returns something unusable, record the failure and retry once. The retry is not an intervention. If it fails again, apply no further verb to that ticket and escalate through the tally and the notes log. Never skip the precondition.\n\n## The driver never edits code (R-6)\n\nThe driver never edits code on `epic/<KEY>` or on any child branch. A code change on a child branch is made by dispatching a worker job that carries the driver's context: `dispatch-remediation --context-file <PATH>`.\n\n- The context file carries only what the worker needs: the reason code, a bounded failing-test excerpt, the conflict or diff summary, and the remediation the driver wants. The CLI rejects an empty, invalid-UTF-8, or oversized file (16384 bytes) but it **does not scrub it**. The driver must remove credentials, environment values, authorization headers, raw provider payloads, and unrelated data before dispatch.\n- **Counter-example.** During the BAPI-979 run the driver hand-merged `ticket_evaluator.py`, `ticket_type_evaluator.py`, and `test_ticket_evaluator.py` on BAPI-981's branch and pushed the result. Under this posture that work is dispatched, not typed: one `dispatch-remediation` whose context names both children's intent.\n- Writing the tally and the notes log is bookkeeping, not a code edit.\n\n## Records\n\n### The tally\n\nWrite `~/.local/state/bapi-conductor-runs/<KEY>/tally.md` **as events happen**, not at wind-down.\n\n- After Stage 0 succeeds, create the directory with user-only permissions (`mkdir -p` then `chmod 700`) and the file with user-only read/write (`chmod 600`). Open it with a header: the epic key, the resolved `epic_run_id`, the scope id, the health-check undrilled classes, and the readiness warnings.\n- The tally is **append-only**. Never rewrite an earlier entry.\n- One entry per issue, with these five fields: `issue`, `evidence` (ids, SHAs, gate, reason code), `verb applied`, `outcome`, and `time` — plus the run id and an event id (`<KEY>-<n>`) so a duplicate is detectable, and the budget charge (`ticket <TICKET>: intervention 2 of 3`).\n- **Write before you act.** Append the entry's pre-action part (issue, evidence, the verb you are about to apply) before applying the verb. Append the outcome under the same event id after the response. A failed pre-action write blocks that recovery. A failed outcome write blocks every later recovery until recording works again.\n\n### The notes log\n\nThe tally sits alongside the durable notes log at `docs/conductor-runs/v2-engine/<date>-<KEY>-run-issues.md` and does not replace it. The tally is the live, write-as-it-happens record; the notes log is the committed artifact. Write the notes log during the run, not only at wind-down. Keep its shape so the two stay compatible:\n\n```markdown\n# <KEY> conductor run — issues log (v2 engine)\n\nDriver: automated LLM session under /drive-epic-run.\nGoal: <one line>.\n\n## Header / baseline (<date>, pre-launch)\n\n| Fact | Value |\n|---|---|\n| `origin/main` | <sha> |\n| DB | <database>, Alembic head <rev> (single head) |\n| epic_run_id | <id> |\n| scope_id | <id> |\n| cut SHA | <sha> |\n| log paths | `~/.local/state/bapi-conductor-runs/<KEY>/` |\n\n## Notes (dated, one per event)\n\n### <date> <time> · <TICKET or run> · <category>\n<what happened, the evidence, the verb applied, the outcome>\n```\n\nEach entry's category is one of `impediment`, `friction`, `defect`, `decision`.\n\nAt wind-down, add the finished run's row to the v2-engine table in `docs/conductor-runs/README.md` — the tally is not a substitute for that row — and commit the notes log and the README row on `docs/<KEY>-conductor-run-notes`. The pull request into the base branch stays human-gated.\n\n### Scrubbing\n\nOne rule for the tally, the notes log, every context file, and every recorded `second_opinion` answer: never record credentials, environment values, authorization headers, raw provider payloads, or raw model answers. Record bounded, secret-free evidence. The notes log gets a sanitized summary, never a raw copy of the tally.\n\n## Monitoring loop\n\nEach pass reads in this order, then applies at most one verb per ticket:\n\n1. **`explain_run`** (`epic_run_id` = the epic key; add `ticket_key` to narrow) — the fastest \"why is this stuck\" read. Inspect the current gate, the latest observation and its age, queued jobs and their queue position, the lease owner, and the expected versus observed head SHA.\n2. **`get_epic_snapshot`** — the run-state read (`GET /jira/epic-runs/runs/{epic_run_id}/state`): every ticket's `status`, `gates`, `needs_human_reason`, `last_progress_at`, `created_at`, `branch_name`, `pr_number`, `row_version`, and the `executor_jobs` block. Read `executor_jobs.state` before `jobs`.\n3. **Scope status** — `GET /jira/index-scope/status` for the run's scope: `lifecycle_state`, `freshness_status`, `required_commit_sha`, `indexed_commit_sha`.\n4. **Plane** — `'<launcher>' plane status`, and `poll_ci_checks` for a PR whose CI you need to see.\n\nAfter every committed verb, run a fresh pass before choosing another action. Between passes, wait with a bounded wait your host provides; never busy-poll.\n\n### Signals → actions\n\nEvery state-changing action below is a verb from the verb table. `Wait` means no action: the condition resolves itself. `End intervention` means park the ticket as the Budget section defines it. `Stop-and-ask` means a stop-list match.\n\n| Signal | Likely cause | Action |\n|---|---|---|\n| `reconciler.stale`, or no tick for more than 3 minutes | Host slept, or the reconciler died | Check `pmset -g log` for sleep. Slept: `Wait`. Process dead: `Stop-and-ask` (restarting a plane process is not a verb, and a plane may be shared). |\n| `needs_human`, `required_checks_empty_no_mode` | The policy lost `required_checks` | `policy PATCH` → `unpark` |\n| `needs_human`, `merge_blocked_workflow_files_modified` | A child diff edits a workflow file | `Stop-and-ask` |\n| `needs_human`, `merge_overlap_governance_guard` | A reconciliation must change a governance-tagged region | `Stop-and-ask` |\n| `needs_human`, `merge_blocked_governance_guard` | A child head changes a governance-tagged region; merge admission parked it before merge | `Stop-and-ask` |\n| `needs_human`, `pr_not_attached` | The worker produced no PR, or the PR head drifted | `explain_run`, then `unpark` (PR attached at the expected head) or `adopt-current-head-and-unpark` (head drifted and the current head is the one to keep) |\n| `needs_human`, `head_sha_drift` or `ci_fix_contract_incomplete` on a green PR | The head moved after the gate anchored it | `adopt-current-head-and-unpark` |\n| `needs_human`, `pr_conflicting` | The child conflicts with `epic/<KEY>` | `dispatch-remediation` → `merge-child` once the gates permit |\n| `needs_human`, `shadow_stale_deadline`, scope `unobserved_advance` | A merge landed that the scope did not observe | `index-scope catch-up` → `unpark` |\n| `needs_human`, `shadow_stale_deadline`, scope wedged at its pinned commit | The scope stopped with no parse behind it | `scope-reparse` → `unpark` once the scope reads ready |\n| Child in `parsing` for more than 45 minutes, scope behind the epic head | A post-merge observation was missed | `index-scope catch-up` |\n| A review worker idle-waiting on a review that expired, scope now `ready` | The review started before the shadow was ready | `index-scope catch-up`, then `Wait` for the executor's per-job timeout; never a process kill |\n| `needs_human`, `review_verdictless_ceiling_reached` | The reviewer returned no verdict until the ceiling | `second_opinion`, then `unpark` |\n| `needs_human` at `code_review` after the remediation budget | The worker cannot fix the finding alone | `dispatch-remediation` with the driver's context, or `End intervention` if the finding repeats |\n| `needs_human` at `spec_review` (spec gap) | The ticket text is ambiguous | `second_opinion`, then `unpark` (the review was wrong) or `End intervention` (the spec must change; no verb edits it) |\n| `merge_held_behind_peer` | A deterministic hold behind an unmerged peer | `Wait` until its hold deadline passes |\n| `needs_human`, `merge_overlap_peer_parked` | The peer it waits behind is parked | Recover the peer with its own verb, then `unpark` |\n| `needs_human`, `merge_overlap_wait_expired` or `merge_overlap_observation_expired` | The overlap hold ran out | `dispatch-remediation` → `merge-child`, or `unpark` once the peer is readable |\n| `needs_human`, `reconcile_overlap_push_unauthorized` | The executor machine's git credential helper refused | `Stop-and-ask` (a credential) |\n| Green, approved child PR not merging (for example `merge_ci_pending_at_head` on a settled head) | The merge path stalled | `merge-child` |\n| Run `paused`, `run_control.cause` a cap | A wall-clock or job cap tripped | `policy PATCH` → `resume` |\n| Run blocked, `integration_pr_missing` or `integration_pr_ready_failed` | The integration PR into base is missing or not ready | `Stop-and-ask` (base branch), then `resume` after the operator corrects it |\n| Scope `lifecycle_state` `failed` | Seed or verify failed | `second_opinion`, then `abandon-run`, or `Stop-and-ask` when the cause lies outside the run |\n\nA signal with no row here is not a reason to stop. Apply the scope rule: find the sanctioned verb whose blast radius stays inside the run, or stop only on a stop-list match.\n\n## Host coverage\n\nThis command ships to the Claude Code and Cursor command mirrors (`.claude/commands` and `.cursor/commands`) — the only slash-command surfaces this repository installs. Codex and Copilot hosts have no slash-command surface here. They rely on the packaged CLI verbs plus the Autonomy section of the operator runbook (`docs/claude/epic-conductor-v2-operator-runbook.md`).\n\nA missing or drifted CLI equivalent on such a host is filed as a conductor defect ticket naming the host, the missing verb, and the command or package version. The agent never improvises an equivalent — no hand-built route call, MCP write, local process action, or worktree edit stands in for a verb this command sanctions.\n",
     "estimate-epic.md": "Estimate an entire Jira Epic or an explicit ticket-key group via the shared epic estimation orchestrator.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\n`$ARGUMENTS` is either a Jira Epic key (e.g. `BAPI-518`) or an explicit `--tickets` key list — never both. This command calls the `estimate_epic` MCP tool, which delegates to the Bridge API epic estimation orchestrator, and renders the structured result.\n\nIf any step fails, stop immediately and report which step failed and why, preserving the user's originally entered epic key or ticket list in the report.\n\n## Step 1 — Parse Arguments\n\n1. **Parse `$ARGUMENTS`**: Extract exactly one key-source input, plus an optional `--allow-partial` flag.\n   - Split `$ARGUMENTS` on whitespace.\n   - If `--tickets` appears, every token after it (up to the next flag or end of input) is the explicit ticket-key list — this is the `ticket_keys` mode.\n   - Otherwise, the first token matching the Jira key pattern (`[A-Z][A-Z0-9]+-\\d+`) is the `epic_key` — this is the epic mode.\n   - `--allow-partial` may appear anywhere; if present, set `allow_partial_value = true`. If absent, omit `allow_partial` entirely (do not pass `false`).\n   - Never resolve both an `epic_key` and a `ticket_keys` list from the same invocation — usage is one mode or the other.\n\n2. **Validate input**:\n   - Usage forms: `/estimate-epic EPIC-KEY` or `/estimate-epic --tickets KEY-1 KEY-2 ...`, plus optional `--allow-partial`.\n   - If neither an `epic_key` nor a `--tickets` list can be resolved, stop immediately and report:\n     ```\n     Usage: /estimate-epic EPIC-KEY [--allow-partial]\n            /estimate-epic --tickets KEY-1 KEY-2 ... [--allow-partial]\n     ```\n   - If `--tickets` is present but followed by zero keys, stop immediately and report: \"`--tickets` requires at least one ticket key.\"\n   - Do not invent or pass a `mode` parameter — there isn't one; the tool infers the source from whichever of `epic_key`/`ticket_keys` is supplied.\n\n## Step 2 — Call the Tool\n\nCall the `estimate_epic` MCP tool with:\n- `epic_key`: the resolved epic key — **only** when in epic mode. Omit entirely in ticket-key mode.\n- `ticket_keys`: the resolved ticket-key list — **only** when in ticket-key mode. Omit entirely in epic mode.\n- `allow_partial`: `allow_partial_value` if `--allow-partial` was passed; omit entirely otherwise (never pass `null`, an empty string, or an empty array for any absent field).\n\nNever pass both `epic_key` and `ticket_keys` in the same call.\n\nIf the tool returns an error envelope (a JSON object with an `error` field), stop and report the error message, preserving the epic key or ticket list the user originally entered.\n\n## Step 3 — Render the Result\n\nRender the successful result as a structured report — do not dump raw JSON by default:\n\n1. **Top**: the final estimate and its scale label (`estimate_label`) as the primary heading — this is the strongest element of the report.\n2. **Immediately after the summary**: `math_source`.\n3. **Next**: resolved child ticket keys (`child_ticket_keys`) and the per-child breakdown, presented compactly.\n4. **Only if non-empty**: a compact warning section listing `failed_child_keys` and `skipped_child_keys`.\n\nKeep the happy-path report concise and scannable. Use backticks for Jira keys and technical identifiers (e.g. `BAPI-518`).\n\n> Note: this tool does not accept a `recreate` parameter — the underlying epic estimation orchestrator (BAPI-522) always reuses cached child estimates and has no recreate knob to forward to.\n\n## Final Report\n\nOn successful completion, display a structured summary per Step 3 above. On failure, display the error message returned by the tool (or the usage error from Step 1), preserving the user's originally entered epic key or ticket list.\n",
     "estimate-ticket.md": "Estimate a single Jira ticket — a one-ticket shortcut over the shared epic estimation orchestrator.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\n`$ARGUMENTS` is exactly one Jira ticket key (e.g. `BAPI-123`), plus an optional `--allow-partial` flag. This command calls the existing `estimate_epic` MCP tool with a single-key `ticket_keys` payload and renders the result in single-ticket language. It does not register a new MCP tool and does not support an epic key or a multi-key list — use `/estimate-epic` for those.\n\nIf any step fails, stop immediately and report which step failed and why, preserving the user's originally entered input.\n\n## Step 1 — Parse Arguments\n\n1. **Scan every token** in `$ARGUMENTS` for the Jira key pattern `[A-Z][A-Z0-9]+-\\d+`, in input order, counting every match rather than stopping at the first one.\n2. `--allow-partial` may appear anywhere; if present, set `allow_partial_value = true`. If absent, omit `allow_partial` entirely (never pass `false`).\n3. **Zero resolved keys** — stop immediately and report:\n   ```\n   Resolved keys: none\n   Usage: /estimate-ticket KEY [--allow-partial]\n   ```\n4. **More than one resolved key** (including a key repeated more than once) — stop immediately and report every resolved key, including duplicates, and state that `/estimate-ticket` accepts exactly one key:\n   ```\n   Resolved keys: <KEY-1>, <KEY-2>, ...\n   /estimate-ticket accepts exactly one ticket key. For multiple tickets, use /estimate-epic --tickets KEY-1 KEY-2 ...\n   ```\n5. **Exactly one resolved key** — proceed to Step 2 with that key.\n\nDo not invent a `mode`, `epic_key`, or `recreate` parameter — none exist on the underlying tool.\n\n## Step 2 — Call the Tool\n\n1. Call the `estimate_epic` MCP tool with:\n   - Absent `--allow-partial`: exactly `{ticket_keys: [KEY]}`.\n   - Present `--allow-partial`: exactly `{ticket_keys: [KEY], allow_partial: true}`.\n   Never include `epic_key`, `mode`, `recreate`, or an explicit `false`/`null` value for `allow_partial`.\n\n2. **If the tool is unavailable or unregistered** (a missing-tool error, or the host reports no such tool), this is a profile configuration gap, not an estimation failure — there is no fallback estimation path. Report:\n   ```\n   The `estimate_epic` tool is not available. It is registered only when the MCP server's estimation\n   profile is active. Ask whoever manages this project's MCP configuration to restart the server with\n   one of:\n     BRIDGE_MCP_PROFILE=estimation\n     BRIDGE_MCP_PROFILE=core,estimation   (or any comma-separated profile list containing \"estimation\")\n     BRIDGE_MCP_PROFILE=full\n   If you don't have configuration access, contact the project or MCP configuration owner.\n   ```\n\n3. **If the parsed response has a non-empty `error` field**, this is a definitive failure — stop before any success rendering:\n   - Render `message` as the primary explanation, when present.\n   - Show `error` secondarily, as a technical code, when `message` is present.\n   - When `message` is absent or empty, render `error` alone as the explanation.\n\n## Step 3 — Render the Result\n\nOn success, render single-ticket language — do not dump raw JSON by default:\n\n1. **Top**: the requested Jira key, `estimate_label`, and the final estimate (`final_hours` / `final_minutes`) as the primary heading.\n2. **Immediately after**: `math_source`.\n3. **Next**: `child_ticket_keys` and the per-ticket detail, presented compactly using the tool's existing response semantics.\n4. **Only if non-empty**: `failed_child_keys` and `skipped_child_keys`, without epic-oriented headings or empty sections.\n5. **Side-effect disclosure** — always include, directly after the rendered result:\n   ```\n   Note: /estimate-ticket adds no new side effects or override flags of its own, but the underlying\n   estimate_epic tool may write a Jira comment and/or update story points when this repository's\n   configuration enables write-back. That behavior cannot be overridden per MCP call, so this command\n   is not guaranteed to be strictly read-only.\n   ```\n\nUse backticks for Jira keys and technical identifiers (e.g. `BAPI-123`).\n\n## Final Report\n\nOn successful completion, display the structured summary from Step 3, including the side-effect disclosure. On failure, display the error/usage message from the step that failed, preserving the user's originally entered input.\n",
-    "explore-ticket.md": "Explore the codebase for a task, settle its acceptance criteria with the user, then propose a design that meets them.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\n`$ARGUMENTS` is a free-form prompt describing a task you want to accomplish and your goals for it. This is **not** a Jira ticket key — it is plain text describing the work.\n\nExecute all exploration and analysis directly in the main conversation. The user should see exploration progress as it happens.\n\nThis command runs strictly outside-in, and the order is the point:\n\n1. **Requirements first.** Establish what the system must do, how it must behave, and what standards it must meet — then get the user to ratify that on an interactive decision page. The page settles **requirements only**. It never asks the user to pick an implementation.\n2. **Then how.** Only once the criteria are ratified do you consider how to meet them, optionally with a council.\n3. **Then the design.** You describe the final proposed design yourself, in the exploration doc. There is no second decision page.\n4. **Then ticket(s).**\n\nNever invert this. A design proposed against unratified criteria is a guess, and an implementation choice presented before the criteria are settled asks the user to commit to a solution for a problem they have not yet agreed on.\n\nIf any critical stage fails, stop immediately and report which stage failed and why.\n\n## Stage 0 — Setup\n\n1. **Parse prompt**: Extract the prompt text from `$ARGUMENTS`. Trim any surrounding whitespace. If the prompt is empty or whitespace-only, stop immediately and display: `Usage: /explore-ticket <prompt describing your task and goals>`\n\n2. **Resolve docs directory**: Call the `ping` MCP tool (no parameters) and read `docs_dir` from its first (JSON) content item. Store that path as `docs_dir`.\n\n3. **Generate output slug**: Create a kebab-case slug from the prompt — take the first 6-8 meaningful words, strip non-alphanumeric characters, lowercase, and truncate to 60 characters. The slug **must start with a letter** so it is a valid decision-page `ticket_key` in Stage 5 (`/^[A-Za-z][A-Za-z0-9_-]*$/`); if it would start with a digit or hyphen, prefix it with `exploration-`. If `{docs_dir}/explorations/{slug}.md` already exists, append a short timestamp suffix (e.g., `-1710000000`) — and fold that suffix **into the `slug` variable itself**, not just the filename, so that Stage 4 (the doc), Stage 5 (`ticket_key`, `output_filename`), and Stage 8 (the doc rewrite) all reference the same slug. The output file path is `{docs_dir}/explorations/{slug}.md`.\n\n4. **Initialize tracking**: Prepare to track `key_files_examined` (list of files read during exploration), `web_searches` (list of topics searched), and `research_queries` (list of deep research queries).\n\nIf this stage fails, stop immediately and report the error. Do not proceed to Stage 1.\n\n## Stage 1 — Codebase Exploration\n\nThis is the core discovery stage. Take your time — thorough exploration is more valuable than speed.\n\n1. **Analyze the prompt** to identify which areas of the codebase are relevant: route files, agent flows, database models, library utilities, LLM integration, MCP server, unit and E2E suites, etc.\n\n2. **Search for files** matching patterns related to the task (e.g., `api/routes/**/*.py`, `src/python/llms/agents/**/*.py`, `db/models/*.py`).\n\n3. **Search for content** — relevant function names, class names, patterns, and keywords across the codebase.\n\n4. **Read the most relevant files** in detail — understand existing implementations, conventions, and patterns that relate to the task.\n\n5. **Build a mental model** of:\n   - What exists today that relates to the task\n   - What patterns and conventions are used in similar features\n   - What dependencies, data flows, and integration points are involved\n   - What gaps or unknowns remain that need external research\n   - Whether there is an established precedent for this kind of work, or none at all — Stage 7 depends on this judgement\n\nExplore to understand the problem and its constraints. Resist designing a solution while you read — you do not yet know what the system is required to do, and Stage 8 is where the design gets written.\n\n6. **Track all significant files** examined in `key_files_examined`.\n\nDo not rush this stage. When in doubt, read more code rather than less. Continue exploring until you have a solid understanding of the relevant code.\n\nThis stage is non-blocking — always proceed to Stage 2 regardless of what you find, since the exploration informs what research is needed.\n\n## Stage 2 — Research Unknowns\n\nBased on gaps identified in Stage 1, decide what research is needed. Apply these decision rules:\n\n- **No research needed**: The codebase exploration answered all questions. Skip directly to Stage 3.\n- **Web search**: For quick factual lookups — library API signatures, configuration syntax, small \"how to\" questions. Examples: \"FastAPI dependency injection with custom headers\", \"Alembic batch migration syntax\". Do web searches inline and capture relevant findings.\n- **Deep research** (via `request_deep_research` MCP tool): For large, multi-faceted unknowns that require synthesizing information from multiple sources. Examples: \"Best practices for implementing WebSocket connection pooling in Python asyncio\", \"Tradeoffs between different approaches to real-time notification delivery in FastAPI applications\". Only use deep research when the question genuinely needs a multi-source investigation.\n\n**If deep research is needed:**\n\n1. Call `request_deep_research` with `wait_for_result` set to `true`, `save_locally` set to `true`, a descriptive `query`, and `context` describing the Bridge API tech stack and the specific task.\n2. If deep research fails, note the failure and fall back to web searches for the same topic. Do NOT halt the pipeline.\n\nTrack all research performed in `research_queries` and `web_searches`.\n\nThis stage is non-blocking — failures degrade the quality of analysis but do not stop the command. Log a warning for any failed research and continue.\n\n## Stage 3 — Frame Acceptance Criteria\n\nEstablish what \"done and correct\" means. Everything in this stage is about the system's obligations, not its implementation. Do not name a technical approach here — that is Stage 8's job, and it does not happen until the user has ratified this framing.\n\n1. **State the frame plainly (required).**\n   - **Business goal** — the value this work delivers and why it matters.\n   - **Desired end-state** — the concrete state the system should reach once this work is done.\n   - **System behavior** — how the system must behave to complete its task (the quality attributes in prose, not a feature list).\n\n2. **Derive the acceptance criteria — what the system must do (required).** Write 3-8 criteria. Each one gets:\n   - An `id` (`AC-1`, `AC-2`, …).\n   - A `criterion` — a single obligation stated concretely enough to be checked. Write it as observable behavior (\"an operator who revokes a key sees the next request rejected\"), not as a task (\"add a revocation endpoint\").\n   - A `verification` — how we would confirm it holds. Name the observable signal: a response code on a specific route, a row state, a log line, a rendered element, a user-visible outcome. **A criterion nobody can check is not yet a criterion** — sharpen it or drop it.\n   - A `status`, using the rubric in step 4.\n\n   Cover the failure and edge behavior, not just the happy path. If the work changes something that already exists, at least one criterion should pin down what must **not** regress.\n\n3. **Identify the non-functional requirements — the standards the system must meet (required).** Consider every one of these canonical NFR categories and include the ones that genuinely apply (omit the rest): security/privacy, performance/latency, reliability/failure-modes, observability/auditability, accessibility/UX, data-integrity/migration, compatibility, operability/config, compliance/SOC2, rollout/reversibility. For each NFR you include, write its `requirement` and its `implication` (what it changes about the implementation) — an NFR with no concrete implication is boilerplate; drop it.\n\n4. **Classify every acceptance criterion and every NFR** with this rubric: `confirmed` only if explicitly stated or observable in code; `assumed` only if a low-risk, reversible default; `open` if it touches architecture, the data model, security, user-visible behavior, migration, or irreversible creation and is not settled. When a criterion or an NFR is genuinely unclear, prefer marking it `open` and asking. Clear criteria make everything downstream more accurate, so surfacing an unclear one is a success, not a delay.\n\n5. **Frame the open requirement questions.** Where a requirement is unsettled **and** has discrete candidate answers, express it as a question the user can answer by clicking (e.g. \"Must revocation take effect immediately, or is eventual acceptable?\"). These become cards in Stage 5. They are questions about *what the system must do* — never about how to build it. If a question has no discrete answers, leave it as prose in the doc instead.\n\n6. **Sanity-check the frame against itself.** Do any two criteria conflict? Does a criterion conflict with an NFR (e.g. an auditability requirement against a latency budget)? Note every tension you find — Stage 7 treats these as a council trigger, and Stage 8 must resolve them explicitly rather than quietly favouring one side.\n\nThis stage is inline analysis — no tool calls required. This stage is non-blocking — always proceed to Stage 4.\n\n## Stage 4 — Write Requirements Draft\n\nWrite what you know so far to disk, so the user has something to read alongside the decision page. The design is deliberately absent — it does not exist yet.\n\n1. Create the `explorations/` directory under `docs_dir` if it does not exist.\n\n2. Write the exploration document to the slug-based path determined in Stage 0 (`{docs_dir}/explorations/{slug}.md`) with this structure:\n\n```markdown\n# Exploration: {concise summary of the prompt}\n\n**Date**: {current date}\n**Prompt**: {original prompt text}\n**Status**: Requirements drafted — awaiting ratification\n\n## Context\n\n{Brief description of the task and what areas of the codebase are relevant.}\n\n## Acceptance Criteria\n\n{The criteria from Stage 3 — what the system must do. One entry per criterion: its id, the criterion itself, how it is verified, and its status (confirmed / assumed / open).}\n\n## Goals & NFRs\n\n{The business goal, desired end-state, and required system behavior from Stage 3. Then the non-functional requirements — the standards the system must meet: each with its category, requirement, implication, and status (confirmed / assumed / open). Note any tension between criteria or between a criterion and an NFR.}\n\n## Open Questions\n\n{Requirement questions that are still unsettled. Mark which ones are going onto the decision page as cards and which are open-ended prose.}\n\n## Codebase Findings\n\n{Key discoveries from Stage 1. What exists today, what patterns are used, what the relevant code paths look like. Reference specific files and functions with file_path:line_number format.}\n\n## Research Findings\n\n{Findings from web searches and deep research, if any. If no research was performed, state \"No external research was needed.\"}\n\n## Key Files\n\n{Bulleted list of the most important files examined, with one-line descriptions of their relevance.}\n```\n\nDo not add a design, an implementation plan, or a recommendation to this draft. Stage 8 adds those once the criteria are settled.\n\nIf the file cannot be written, stop immediately and report the failure.\n\n## Stage 5 — Generate Requirements Decision Page\n\nTurn the Stage 3 framing into an interactive HTML decision page so the user can ratify the requirements by clicking. **This page settles requirements only.** It must not contain a single implementation option — the user is agreeing on what the system must do, not choosing how to build it.\n\n1. **Map the acceptance criteria to `acceptance_criteria`.** Each entry has `id`, `criterion`, `verification`, and `status`. Ids must be unique — a duplicate id is rejected, because the id is the key the page reports the user's stance under. Every criterion renders with an Agreed / Ask about this / Disagree control, so pass all of them, not only the open ones. Pass the NFRs the same way under `nfrs`.\n\n2. **Map each open requirement question from Stage 3 step 5 to an actionable item.** Each entry has:\n   - `id`: a short stable id, e.g. `R-1`, `R-2`.\n   - `question`: the requirement question.\n   - `options`: the 2-4 candidate answers (string array). Do **not** include \"None of these\" or \"Ask about this\" — the renderer auto-appends both.\n   - `option_consequences`: what each answer would mean for the criteria, **parallel to and the same length as** `options`.\n   - `why_it_matters`: the concrete impact line.\n   - `recommendation_explanation`: why the recommended answer is best.\n   - `recommendation_index`: the 0-based index of the recommended answer (must be within `options`).\n   - `codebase_evidence` (optional): the Assessment paragraph plus `file:line` citations, shown collapsed.\n\n   **These cards are requirement questions, never implementation choices.** \"Must revocation be immediate or is eventual acceptable?\" is a valid card. \"Should we use a short-TTL cache or pub/sub invalidation?\" is not — it is a solution, it belongs to Stage 8, and putting it here defeats the purpose of the page. If you cannot phrase a card without naming a mechanism, it is not a requirement question. When there are no such questions, pass an empty array — a criteria-only page is expected and renders correctly.\n\n3. **Call `generate_decision_page`** with routing fields at the root and all heavy arrays nested under `content`:\n   - `artifact_type`: `pre_ticket_planning` (renders the acceptance-criteria and goals panel above any cards).\n   - `ticket_key`: the Stage 0 `slug` (a non-Jira slug is fine — it must start with a letter and contain only letters, digits, hyphens, or underscores).\n   - `output_subdir`: `explorations` (so the page lands beside the markdown doc).\n   - `output_filename`: `{slug}-requirements.html`.\n   - `labels`: requirements-flavored overrides, e.g. `title` = \"Requirements\", `section_heading` = \"Open Requirement Questions\", and an `intro` that frames the page as agreeing on what the system must do before any design work begins.\n   - `content`: an object containing `system_goals` and `actionable_items`. **`system_goals` MUST ALWAYS be passed** inside `content` so the backend always writes a page. Never omit it, even if every criterion and NFR is confirmed. `acceptance_criteria` and `nfrs` both live inside `system_goals`. (Do not pass `implementation_order` inside `content` — that is for epic surfaces, not a single explored task.)\n\n   ```typescript\n   interface ExploreTicketContent {\n     system_goals?: {\n       business_goal: string;\n       desired_end_state: string;\n       system_behavior: string;\n       acceptance_criteria?: Array<{\n         id: string;                    // e.g. \"AC-1\"; must be unique\n         criterion: string;             // what the system must do\n         verification: string;          // how we would confirm it holds\n         status: \"confirmed\" | \"assumed\" | \"open\";\n       }>;\n       nfrs?: Array<{\n         category: string;\n         requirement: string;\n         implication: string;\n         status: \"confirmed\" | \"assumed\" | \"open\";\n       }>;\n     };\n     actionable_items?: Array<{\n       id: string;                      // e.g. \"R-1\"; a REQUIREMENT question, not a design choice\n       question: string;\n       why_it_matters: string;\n       recommendation_explanation: string;\n       options: string[];               // 2-4 candidate answers (no \"None of these\" or \"Ask about this\")\n       option_consequences: string[];   // same length as options\n       recommendation_index: number;   // 0-based within options\n       codebase_evidence?: string;      // optional: assessment + file:line citations\n       original_question?: string;      // optional: only when item maps to a verbatim question\n     }>;\n     // clear_improvements: not used by this command — it captures requirements, not findings\n     // implementation_order: for epic surfaces only — do NOT include for single task explorations\n     // depends_on: hard prerequisites (titles/keys that must land first)\n     // recommended_after: soft sequencing preferences, not hard blockers\n   }\n   ```\n\n   Example call:\n   ```json\n   {\n     \"ticket_key\": \"revoke-api-keys\",\n     \"artifact_type\": \"pre_ticket_planning\",\n     \"output_subdir\": \"explorations\",\n     \"output_filename\": \"revoke-api-keys-requirements.html\",\n     \"labels\": { \"title\": \"Requirements\", \"section_heading\": \"Open Requirement Questions\" },\n     \"content\": {\n       \"system_goals\": {\n         \"business_goal\": \"Operators can cut off a leaked key immediately.\",\n         \"desired_end_state\": \"Revocation is self-serve and takes effect at once.\",\n         \"system_behavior\": \"Rejects revoked credentials without a restart.\",\n         \"acceptance_criteria\": [\n           { \"id\": \"AC-1\", \"criterion\": \"An operator who revokes a key sees the next request with it rejected.\", \"verification\": \"The following call to the protected route returns 401.\", \"status\": \"confirmed\" },\n           { \"id\": \"AC-2\", \"criterion\": \"Revocation is recorded with actor and timestamp.\", \"verification\": \"An audit row names the operator and the revoked key id.\", \"status\": \"open\" }\n         ],\n         \"nfrs\": [\n           { \"category\": \"security/privacy\", \"requirement\": \"The raw key is never logged on the revoke path.\", \"implication\": \"Log the key id, never the secret.\", \"status\": \"open\" }\n         ]\n       },\n       \"actionable_items\": [\n         {\n           \"id\": \"R-1\",\n           \"question\": \"Must revocation take effect immediately, or is eventual acceptable?\",\n           \"why_it_matters\": \"Sets the hard bound AC-1 has to meet.\",\n           \"recommendation_explanation\": \"A leaked key is an active incident; eventual leaves a usable window.\",\n           \"options\": [\"Immediately (under 5s)\", \"Eventually (under 60s is acceptable)\"],\n           \"option_consequences\": [\"AC-1 gains a 5s bound.\", \"AC-1 gains a 60s bound.\"],\n           \"recommendation_index\": 0\n         }\n       ]\n     }\n   }\n   ```\n\n4. **Handle the response `status`:**\n   - `no_decisions_needed`: no page was written. This should not occur when `system_goals` is always passed. Skip Stage 6 entirely, tell the user there were no open requirements, and proceed to Stage 7 treating the Stage 3 framing as the settled criteria.\n   - `decision_page_generated`: surface the returned `file_path` and proceed to Stage 6. **Always proceed to Stage 6 when `decision_page_generated` is returned**, regardless of `actionable_items_count`. A criteria-only page with zero cards still has stance controls that must be submitted.\n\nThis stage is non-blocking: if `generate_decision_page` fails, do not halt. **You MUST output a highly visible warning** (e.g. **⚠ WARNING: The requirements page could not be generated** in bold) explaining that generation failed and that the user should review the criteria in the markdown doc written in Stage 4 instead. Do not silently continue — the failure must be diagnosable from your output. Then ask the user to confirm the criteria in chat before proceeding to Stage 7.\n\n## Stage 6 — Ratify Requirements\n\nCapture the user's stances, settle the criteria, gate on the implications review, and fold the result into the doc. Nothing downstream — including a council in Stage 7 — may start until the criteria are agreed and the implications gate has accepted a proceed token.\n\n1. **Direct the user to the page.** Provide the `file_path` from Stage 5 and tell them to open it in their browser. Explain that they are agreeing on what the system must do — not how it will be built — that they can accept, question, or reject each criterion, and that they can ask questions in chat before submitting.\n\n2. **Q&A loop and commit signal.** Engage with each user message as either a commit or a discussion turn:\n   - **Commit:** trim the full message and attempt to parse the entire trimmed message as JSON. Treat it as a commit only when the parsed value is an object with all three top-level fields: `ticket_key` (string), `decisions` (object), and `general_comment` (string). The first valid commit-shaped paste commits — do not over-validate the per-card fields. The page also submits `acceptance_criteria_feedback` and `nfr_feedback` objects, each keyed by criterion id or NFR category with a `stance` of `agreed`, `ask`, or `disagree` plus a `comment`.\n   - **Discussion:** anything that is not commit-shaped JSON. Answer from the doc written in Stage 4 and from codebase lookups. If a JSON-shaped paste is missing one of the three required fields, say which field is missing rather than treating it as a freeform question.\n   - **In-flight overrides:** when the user clearly changes an answer in chat (\"AC-2 is wrong\", \"go with eventual for R-1\") or gives new overarching guidance, record it as a working-memory override. On commit, the submitted JSON is the baseline and recorded overrides take precedence; post a one-line acknowledgement naming each overridden item before you rewrite the doc.\n\n3. **Resolve every \"ask\" (hard rule).** After accepting a commit, scan all three: any item in `decisions` where `choice === \"ask\"`, any entry in `acceptance_criteria_feedback` where `stance === \"ask\"`, and any entry in `nfr_feedback` where `stance === \"ask\"`. For each, present the relevant evidence and continue the discussion until the user gives an explicit answer, which you record as an override. Do not proceed while any `ask` remains unresolved — do not honor \"just skip those\".\n\n4. **Resolve every \"disagree\".** A disagree means the criterion is wrong as written. Work out with the user what it should say, restate it back, and get explicit agreement on the corrected wording. A rejected criterion is either rewritten or dropped — never carried forward as-is.\n\n5. This surface's tracked stances are the `acceptance_criteria_feedback` and `nfr_feedback` objects captured above. **Review the wider implications, then gate on a decision.** Build the review from the complete settled set: the submitted `decisions`, any in-flight overrides recorded during the conversation (these take precedence over the submission), every `\"none\"` answer together with the reason given for it, `general_comment`, and — where this surface tracks acceptance-criterion or NFR stances — those stances too. Do not start the review until every `ask` has an explicit recorded resolution and every in-flight override has been applied.\n\n   Consider three fixed categories, regardless of whether a decision was framed as technical, user-facing, or business-oriented:\n   - **Program / application** — architecture, code paths, operability, maintenance burden, and requirements imposed on other parts of the software.\n   - **User** — end users, new users performing setup, operators, and developers, including prerequisites, setup friction, and additional steps.\n   - **Business** — cost, adoption, support load, compliance, and reversibility.\n\n   Emit only the categories with material second-order implications. For each included category, write at most four one-line bullets of about 25 words, each naming who or what is affected and how — never a restatement of the selected decision. Close with a line naming every considered category that was omitted, e.g. `Considered, nothing material: business.` — omit this closing line only when all three categories have material implications.\n\n   If the review cannot be produced, report that in one line and continue without stalling the workflow or presenting the gate below.\n\n   Retain the latest emitted review; once the gate below allows continuation, write it under `## Implications` during the Stage 6 document rewrite.\n\n   Then present the gate, verbatim: `Implications reviewed. Proceed, or name a decision to revisit.` Accept only a normalized `proceed`, `yes`, `y`, or `go` as a continuation token. Any other response names a decision to reopen: re-settle it in chat, record the new override, rerun the entire implications review against the changed settled set, and present the gate again.\n\n   Literal `auto_approve = true` emits the review but skips this gate entirely; a missing or non-true `auto_approve` value follows the human-in-the-loop path above.\n\n6. **Settle the criteria and update the doc.** Rewrite the Acceptance Criteria, Goals & NFRs, and Open Questions sections of `{docs_dir}/explorations/{slug}.md` to the agreed set: fold in every correction, resolve each answered requirement question into the criterion it affects, promote settled criteria out of `open`, and weave `general_comment` in as overarching guidance. Add or update the `## Implications` section with the latest review from item 5 — the final one after any reopen/review loop — or its one-line fail-open notice if generation was unavailable. Set the doc's Status line to \"Requirements ratified\". Preserve all unaffected sections unchanged. **The settled criteria are now the contract** — every later stage is judged against them.\n\nThis stage is non-blocking: if the user never commits, leave the doc as written in Stage 4, tell them the requirements are unratified, and stop without forcing a decision. Do not proceed to a council or a design on unratified criteria.\n\n## Stage 7 — Council Gate\n\nThe criteria are ratified. Now assess honestly whether you know **how** to meet them — and offer to convene a council when you do not.\n\n**Lean toward offering.** A council is cheap relative to committing the user to the wrong design, and this command prioritizes discovery over premature commitment. Do not wait for the user to ask for one.\n\n1. **Check the triggers.** Offer a council when **any** of these hold:\n   - More than one materially different approach could satisfy a criterion, and the codebase evidence you gathered cannot separate them.\n   - A criterion has no obvious implementation path in the existing code.\n   - Meeting one criterion appears to trade off against another criterion or against an NFR (any tension noted in Stage 3 step 6, or created by a correction in Stage 6).\n   - The work touches an area with no established pattern — Stage 1 found no precedent to follow.\n   - Ratification materially changed the problem — the user tightened a bound, rejected a criterion, or added an obligation you had not framed.\n   - Stage 2 research surfaced competing approaches with no clear winner.\n\n   Do **not** offer when every ratified criterion maps cleanly onto a well-trodden pattern already used in this codebase and you can point to the precedent.\n\n2. **Ask for approval.** When a trigger fires, first summarize the uncertainty in 1-3 bullets — name the specific criteria at issue and what you cannot currently decide. Then ask exactly:\n\n   ```\n   Significant uncertainty about how to meet {AC ids}. Run a council before I draft the design? (y/N)\n   ```\n\n   Mention that a council polls for up to ~15 minutes before you ask, so the user is choosing with the cost in view.\n\n   Treat an empty response, any negative response (`n`, `no`, or similar), or any ambiguous/unrecognized response as **decline** — do not guess intent. On decline, note in one line that the council was offered and declined, and proceed to Stage 8 on your own analysis. Never run a council without an explicit affirmative (`y` or `yes`).\n\n3. **Run it on approval.** Call `request_council` with:\n   - `task_description`: the task, the **ratified** acceptance criteria and NFRs from Stage 6, and the specific uncertainty you summarized. Sent verbatim — this tool does not read from a file. State plainly that the criteria are settled and the council's job is to find how to meet them, not to revisit what they are.\n   - `mode`: `technical`.\n   - `wait_for_result`: `true`. `save_locally`: `true`.\n\n   While it runs, tell the user it is polling and roughly how long it may take.\n\n4. **Fold the result into your analysis.** Carry the council's approaches, objections, and any option you had not considered into Stage 8. If the council argues a ratified criterion is unmeetable, do not silently drop it — raise it with the user in Stage 8 as an explicit conflict.\n\nIf the council fails or times out, note the failure visibly and proceed with your own analysis — a missing council degrades the design but does not invalidate it. This stage is non-blocking — always proceed to Stage 8.\n\n## Stage 8 — Propose Final Design\n\nNow describe how you would build it. **Do not generate a decision page for this stage.** The requirements page was the user's decision surface; the design is your proposal, written into the doc and discussed in chat. Generating a second page here would ask the user to ratify a solution, which is not what this command does.\n\n1. **Work out the design against the ratified criteria.** Consider the approaches you know plus anything the council surfaced. For each candidate, establish which criteria it satisfies and at what cost. An approach that cannot meet a ratified criterion is not a candidate — discard it and say why.\n\n2. **Resolve any tension explicitly.** Where meeting one criterion costs another, or costs an NFR, state which obligation your design privileges and what that costs the other. Do not let a tension pass silently.\n\n3. **Commit to a single proposed design.** You are recommending, not offering a menu. Name the approach, describe how it works, list the files to create or modify, and map each ratified criterion to the part of the design that satisfies it. Where you seriously considered an alternative, record it and why you rejected it — as history, not as an open choice.\n\n   **Record the implementation slices with a size band.** The design's sequence of work is also the proposed ticket split, so name each slice and size it with the ladder in the posture block below — `S` / `M` / `L` / `XL`, by file-touch breadth and rough LOC. Size the slices here, in the design, rather than at the handoff gate: the split is a design decision, and deciding it under the pressure of an approval prompt is how a coherent design turns into a swarm of tiny tickets. A slice that will not fit in `L` becomes one `XL` slice — do not split it to fit a band. Split only when it is genuinely two independent pieces of work, or when it runs past roughly 40 files / ~3000 LOC.\n\n4. **Rewrite `{docs_dir}/explorations/{slug}.md`** so it reads as a finished proposal, not a mechanical append. Set the Status line to \"Design proposed\". Keep the ratified Acceptance Criteria, Goals & NFRs, and `## Implications` sections intact — they are the contract and must not drift — and add:\n\n```markdown\n## Approaches Considered\n\n{Each candidate, what it would mean, and why it was or was not chosen. Note which came from the council, if one ran. If no alternatives were seriously considered, state that and why the path was obvious.}\n\n## Proposed Design\n\n{The recommended approach in enough detail to implement: how it works, the files to create or modify, the sequence of work, and the risks. Reference specific files with file_path:line_number format.}\n\n## Criteria Coverage\n\n{Each ratified criterion mapped to the part of the design that satisfies it, and how it will be verified. Any criterion the design only partially meets must say so plainly.}\n```\n\n5. **Present the design in chat and invite pushback.** Summarize the proposal in at most six bullets — the approach, the files it creates or modifies, the order of the work, and any ratified criterion the design only partially meets — and state clearly that it is a proposal. If the user objects, revise the design — but if their objection actually changes what the system must do rather than how it is built, say so: that is a criteria change, and it means reopening the criteria rather than quietly bending the design around it.\n\nThis stage is non-blocking — always proceed to Stage 9 once the design is written and summarized in chat, even if the user has not responded to it.\n\n## Stage 9 — Ticket Handoff\n\n<!-- BEGIN BRIDGE TICKET-AUTHORING POSTURE -->\n<!-- Canonical source: docs/bridge-ticket-authoring.md.\n     This block is duplicated byte-identically onto every carrier. Never edit a\n     copy: edit the canonical source and re-copy it verbatim. A cross-surface\n     byte-equality test fails the build if any copy drifts by a single byte. -->\n\n## Ticket-authoring posture\n\nDeep reference: `docs/bridge-ticket-authoring.md`.\n\n**Draft through the writer.** Every ticket body — an epic parent, an epic child,\nand an ordinary sibling alike — is drafted by the `jira-ticket-writer` agent\nbefore `create_ticket` is called. Do not compose a ticket description inline.\n\n**Size the work.** Size each ticket by file-touch breadth and depth plus rough\nlines of code (LOC) changed:\n\n- `S = 1-2 files / <~80 LOC`\n- `M = ~3-8 files / ~80-400 LOC`\n- `L = ~8-15 files / ~400-900 LOC`\n- `XL = >15 files / >~900 LOC`\n\nTarget size priority: **L (target) -> XL (when the work does not fit in L) -> M\n(third choice) -> S (only when unavoidable)**. This applies equally to a\nstandalone ticket and to an epic child.\n\nAim each slice at L. When one will not fit, grow it to XL rather than splitting\nit — split only when the slice is genuinely two independent pieces of work,\nnever merely to land inside a band. Bridge's grooming and implementation process\nhandles a large vertical slice well and is overkill on small ones: every extra\nticket is another worktree, another PR, another rebase, and another chance for\ntwo workers to touch the same file. Reach for M because the work genuinely is\nthat size, not to avoid an XL.\n\nBeyond roughly 40 files or ~3000 LOC, split anyway. Past that point review\nturnaround and rebase cost dominate the run's budget, and a review that wedges\nholds the gate to its full retry ceiling before anyone notices.\n\n**Group at three.** Three or more implementable tickets is an epic: propose an\nepic parent plus an ordered child manifest, and resolve this surface's own\napproval gate before anything is created. One or two tickets are ordinary\nsiblings — no epic parent, no manifest. The threshold is exactly three.\n\n**Hand off once.** An epic handoff names exactly one conductor entry point,\n`drive-epic`, which selects the runnable path itself. Never present a choice\nbetween conductors.\n\n**Departure is closed-list only.** These three exceptions, and no others, permit\ndeparting from the rules above. Invoking one requires no announcement.\n\n- **E1 External-tracker mirroring** — a recorded upstream identifier exists and\n  its granularity is contractual. Bypasses sizing and the epic threshold.\n- **E2 Discovery-only spike** — no committed production-code deliverable.\n  Bypasses sizing only; does not bypass drafting through the writer.\n- **E3 Authorized incident containment** — tied to an active incident record,\n  not to schedule pressure. Bypasses sizing and the epic threshold.\n\nThe list is closed. Anything outside it is an escalation to the operator, not a\njudgement call. Explicitly refused as grounds for departure: a single-file\ntrivial fix (that is `S` reached through the normal path, not an exception),\ngeneric time pressure, \"already well specified\", \"faster without the writer\",\ndeveloper discretion, minor refactor, unattended mode, context limits, and \"hard\nto decompose\" (XL is the normal overflow, so that is the ordinary path and not a\ndeparture). Writer unavailability escalates; it never silently authorizes inline\ndrafting.\n\n<!-- END BRIDGE TICKET-AUTHORING POSTURE -->\n\n1. **Assess readiness.** The work is ready to become a ticket when the criteria are ratified, the design is proposed, and no criterion is left unresolved or only partially covered. If something is still open, say what it is and recommend the follow-up that would close it rather than creating a ticket on a soft foundation:\n   - A **wider council** (`request_council`) when the design would benefit from a broad review before implementation. If Stage 7 already ran one, only suggest another when something material changed since.\n   - A **second opinion** (`second_opinion`) when a few specific contested points need an independent check.\n   - **Web or deep research** (`request_deep_research`) when the design still rests on technical unknowns that need grounding.\n\n2. **Decompose once, then offer to create.** When the work is ready, freeze the split before you show it, and show the whole thing. Never present the gate bare — the user cannot consent to a plan they have not been shown.\n\n   Count the implementation slices Stage 8 recorded. That count decides the shape, and the threshold is exactly three:\n\n   - **Three or more slices → propose an epic.** Freeze an epic parent plus an ordered child manifest in one pass. Per child: title, scope boundary, size band, `depends_on` (hard prerequisites that must land first), `recommended_after` (soft sequencing preferences that are not blockers), and a one-line order rationale. Keep hard prerequisites strictly separate from soft sequencing. A child may be `XL`; grow one rather than splitting it to fit, and split only past roughly 40 files / ~3000 LOC.\n   - **One or two slices → ordinary siblings.** No epic parent, no manifest, no decomposition pass. Just the ticket or the pair.\n\n   This decomposition is the **single authoritative split** for the rest of the stage. It is decided here, once, with the whole portfolio in view; step 3 renders bodies against it and does not revisit it.\n\n   Output the plan outline, kept to roughly one screen:\n\n   - **Plan** — the approach in one sentence.\n   - **Shape** — `epic` (with the parent's title) or `siblings`. Say which and why the count put it there.\n   - **Tickets ({n})** — one line per ticket: title, the slice of scope it covers, its size band, its ratified criterion ids, and — for an epic — its `depends_on` / `recommended_after` and order rationale. Say plainly when it is a single ticket.\n   - **Files** — the files to create or modify, grouped by area.\n   - **Sequence** — the ordered steps of the work, one line each.\n   - **Risks / not covered** — any criterion the design only partially meets, or `none`.\n\n   Then ask exactly:\n\n   ```\n   Requirements ratified and design proposed. Create the ticket(s) now? (y/N)\n   ```\n\n   Treat an empty response, any negative response, or any ambiguous/unrecognized response as **decline** — do not guess intent. On decline, report that the exploration doc is the artifact, and tell the user they can later ask their agent to use the Jira Ticket Writer to create ticket drafts from it. Never create a ticket without an explicit affirmative (`y` or `yes`) — creation is irreversible.\n\n3. **Render every body through `jira-ticket-writer`, then create.** Nothing calls `create_ticket` before the gate in step 2 resolves with an explicit affirmative.\n\n   On approval, fan out **one `jira-ticket-writer` invocation per entry** in the frozen outline — the epic parent, each child, and each ordinary sibling. Each invocation is bound to exactly its own entry and renders a body against that entry's fixed boundary and size band. A child invocation must not re-split, merge, reorder, renumber, or rescope: the decomposition is already frozen, and an invocation that sees one entry is deciding on strictly less information than the pass that saw them all.\n\n   Pass into every invocation, verbatim: the ratified acceptance criteria that entry satisfies, the proposed design for its slice, and the entry's own manifest fields. Do not restate or reinterpret the criteria — they were ratified in that wording, and they go to the writer unchanged.\n\n   Then create exactly the approved set from the writer's drafts:\n\n   - **Siblings** — `create_ticket` per draft, unparented.\n   - **Epic** — `create_ticket` with `issue_type` set to `Epic` for the parent first, capture the resolved epic key, then `create_ticket` with `parent_key` set to that key for each child in manifest order. The `parent_key` argument is required on every child call; never omit it.\n\n   Create the tickets exactly as the outline named them — the split was decided and shown there, so do not create a different set than the user approved. If you now believe the split is wrong, say so and re-ask rather than changing it silently.\n\n   **Track each ticket as you create it.** Immediately after each successful `create_ticket`, call `track_ticket` with that key and the ticket's description. Bridge's workflow state lives in its own database, not in the ticket backend, and an untracked ticket is invisible to it — `update_ticket_state` and `get_ticket_state` both return 404 until something registers the key. Tracking works against either ticket backend and is a safe no-op when the key is already tracked.\n\n   Tracking is **warn, not halt**: if `track_ticket` returns an error of any kind, record the reason for the final report and keep going. The ticket exists and is the thing that mattered; tracking can be repeated later.\n\n   **On a partial failure, stop and report what exists.** Ticket creation is irreversible and there is no resume protocol here: report the epic key and every child key created so far, name the entry that failed, and tell the user to re-run naming the existing epic rather than repeating this command. Do not retry the failed call in a loop, and do not create the remaining children as though nothing happened — a half-built epic the user cannot see is worse than a stopped one they can.\n\n   Idempotency labels are deliberately not used. They ride on `create_ticket`'s `labels` argument and are read back through `get_tickets(labels=...)`, and **both are Jira-only**: against the local ticket backend each returns a terminal `409 UNSUPPORTED_IN_LOCAL_MODE`, and the create rejects without making a ticket. A resume protocol built on them would work on one backend and silently fail on the other.\n\n4. **Write the plan DAG.** After an epic's tickets are created, write the conductor's plan sidecar to `{docs_dir}/epic-plans/{slug}/epic-plan.dag.json`. You already hold everything it needs: the frozen manifest carries each child's `depends_on`, and Stage 8's design named the files each slice touches. Nothing downstream derives this from an epic key, so if you skip it the handoff dead-ends.\n\n   ```json\n   {\n     \"plan_version\": 1,\n     \"nodes\": [\n       {\n         \"ticket_key\": \"KEY-2\",\n         \"status\": \"planned\",\n         \"depends_on\": [],\n         \"touched_files\": [\"src/example/service.py\", \"src/example/handler.py\"]\n       }\n     ],\n     \"edges\": []\n   }\n   ```\n\n   One node per child, in manifest order, using the **real** keys just created — never a `TBD-N` placeholder, since the tickets already exist. `depends_on` carries the manifest's hard prerequisites only; leave `recommended_after` out, because soft sequencing is not a dependency and encoding it there serializes work that could run in parallel. `edges` may be `[]` when `depends_on` already expresses the graph. The graph must be acyclic and every `depends_on` entry must name a node in the file.\n\n   Populate `touched_files` on every node from the design's file list, as repository-relative POSIX paths. It is what lets the conductor serialize two children that would otherwise collide, and it matters more the larger the slices are. An omitted list is accepted by the server as an explicit opt-out, so leaving it off does not fail — it silently turns overlap protection off. Use `[]` only when a ticket genuinely touches no repository files.\n\n5. **Hand off once.** Then point at exactly one conductor entry point:\n\n   ```\n   npx -y @bridge_gpt/mcp-server drive-epic {EPIC-KEY} --plan-file {docs_dir}/epic-plans/{slug}/epic-plan.dag.json\n   ```\n\n   `drive-epic` reads conductor readiness and selects the runnable path itself. Do not name an underlying conductor, and never present a choice between two of them — two transition authorities on one epic is the failure this single entry point exists to remove. A sibling pair or a lone ticket gets no conductor handoff at all.\n\nThis stage is non-blocking: if the user never answers, leave the doc as written in Stage 8 and stop without forcing a decision.\n\n## Final Report\n\nOn successful completion of all stages, display:\n\n> **Exploration Complete**\n>\n> **Prompt**: {first 80 characters of prompt}...\n> **Output**: {full path to the exploration doc}\n> **Requirements Page**: {full path to the generated requirements.html, or \"not generated\" when generation failed}\n> **Acceptance Criteria**: {count} ratified ({count} corrected by the user, {count} still open)\n> **Files Examined**: {count of key_files_examined}\n> **Research**: {count of web_searches} web searches, {count of research_queries} deep research queries, council {\"run\" | \"offered and declined\" | \"not needed\"}\n>\n> **Requirements**: {\"Ratified\" | \"Unratified — page not submitted\"}\n> **Design**: {\"Proposed\" | \"Not reached\"}\n> **Ticket(s)**: {\"Created: KEY-1, KEY-2\" | \"Declined — doc is the artifact\" | \"Not offered — work not ready\"}\n> **Tracking**: {\"Registered\" | \"Warning: not tracked — {reason}\" | \"n/a — no tickets created\"}\n> **Plan DAG**: {full path to epic-plan.dag.json | \"n/a — not an epic\"}\n\nOn failure at any stage, stop immediately and report:\n- Which stage failed (by number and name)\n- The error details\n- Any partial results that were produced before the failure\n",
+    "explore-ticket.md": "Explore the codebase for a task, settle its acceptance criteria with the user, then propose a design that meets them.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\n`$ARGUMENTS` is a free-form prompt describing a task you want to accomplish and your goals for it. This is **not** a Jira ticket key — it is plain text describing the work.\n\nExecute all exploration and analysis directly in the main conversation. The user should see exploration progress as it happens.\n\nThis command runs strictly outside-in, and the order is the point:\n\n1. **Requirements first.** Establish what the system must do, how it must behave, and what standards it must meet — then get the user to ratify that on an interactive decision page. The page settles **requirements only**. It never asks the user to pick an implementation.\n2. **Then how.** Only once the criteria are ratified do you consider how to meet them, optionally with a council.\n3. **Then the design.** You describe the final proposed design yourself, in the exploration doc. There is no second decision page.\n4. **Then ticket(s).**\n\nNever invert this. A design proposed against unratified criteria is a guess, and an implementation choice presented before the criteria are settled asks the user to commit to a solution for a problem they have not yet agreed on.\n\nIf any critical stage fails, stop immediately and report which stage failed and why.\n\n## Stage 0 — Setup\n\n1. **Parse prompt**: Extract the prompt text from `$ARGUMENTS`. Trim any surrounding whitespace. If the prompt is empty or whitespace-only, stop immediately and display: `Usage: /explore-ticket <prompt describing your task and goals>`\n\n2. **Resolve docs directory**: Call the `ping` MCP tool (no parameters) and read `docs_dir` from its first (JSON) content item. Store that path as `docs_dir`.\n\n3. **Generate output slug**: Create a kebab-case slug from the prompt — take the first 6-8 meaningful words, strip non-alphanumeric characters, lowercase, and truncate to 60 characters. The slug **must start with a letter** so it is a valid decision-page `ticket_key` in Stage 5 (`/^[A-Za-z][A-Za-z0-9_-]*$/`); if it would start with a digit or hyphen, prefix it with `exploration-`. If `{docs_dir}/explorations/{slug}.md` already exists, append a short timestamp suffix (e.g., `-1710000000`) — and fold that suffix **into the `slug` variable itself**, not just the filename, so that Stage 4 (the doc), Stage 5 (`ticket_key`, `output_filename`), and Stage 8 (the doc rewrite) all reference the same slug. The output file path is `{docs_dir}/explorations/{slug}.md`.\n\n4. **Initialize tracking**: Prepare to track `key_files_examined` (list of files read during exploration), `web_searches` (list of topics searched), and `research_queries` (list of deep research queries).\n\nIf this stage fails, stop immediately and report the error. Do not proceed to Stage 1.\n\n## Stage 1 — Codebase Exploration\n\nThis is the core discovery stage. Take your time — thorough exploration is more valuable than speed.\n\n1. **Analyze the prompt** to identify which areas of the codebase are relevant: route files, agent flows, database models, library utilities, LLM integration, MCP server, unit and E2E suites, etc.\n\n2. **Search for files** matching patterns related to the task (e.g., `api/routes/**/*.py`, `src/python/llms/agents/**/*.py`, `db/models/*.py`).\n\n3. **Search for content** — relevant function names, class names, patterns, and keywords across the codebase.\n\n4. **Read the most relevant files** in detail — understand existing implementations, conventions, and patterns that relate to the task.\n\n5. **Build a mental model** of:\n   - What exists today that relates to the task\n   - What patterns and conventions are used in similar features\n   - What dependencies, data flows, and integration points are involved\n   - What gaps or unknowns remain that need external research\n   - Whether there is an established precedent for this kind of work, or none at all — Stage 7 depends on this judgement\n\nExplore to understand the problem and its constraints. Resist designing a solution while you read — you do not yet know what the system is required to do, and Stage 8 is where the design gets written.\n\n6. **Track all significant files** examined in `key_files_examined`.\n\nDo not rush this stage. When in doubt, read more code rather than less. Continue exploring until you have a solid understanding of the relevant code.\n\nThis stage is non-blocking — always proceed to Stage 2 regardless of what you find, since the exploration informs what research is needed.\n\n## Stage 2 — Research Unknowns\n\nBased on gaps identified in Stage 1, decide what research is needed. Apply these decision rules:\n\n- **No research needed**: The codebase exploration answered all questions. Skip directly to Stage 3.\n- **Web search**: For quick factual lookups — library API signatures, configuration syntax, small \"how to\" questions. Examples: \"FastAPI dependency injection with custom headers\", \"Alembic batch migration syntax\". Do web searches inline and capture relevant findings.\n- **Deep research** (via `request_deep_research` MCP tool): For large, multi-faceted unknowns that require synthesizing information from multiple sources. Examples: \"Best practices for implementing WebSocket connection pooling in Python asyncio\", \"Tradeoffs between different approaches to real-time notification delivery in FastAPI applications\". Only use deep research when the question genuinely needs a multi-source investigation.\n\n**If deep research is needed:**\n\n1. Call `request_deep_research` with `wait_for_result` set to `true`, `save_locally` set to `true`, a descriptive `query`, and `context` describing the Bridge API tech stack and the specific task.\n2. If deep research fails, note the failure and fall back to web searches for the same topic. Do NOT halt the pipeline.\n\nTrack all research performed in `research_queries` and `web_searches`.\n\nThis stage is non-blocking — failures degrade the quality of analysis but do not stop the command. Log a warning for any failed research and continue.\n\n## Stage 3 — Frame Acceptance Criteria\n\nEstablish what \"done and correct\" means. Everything in this stage is about the system's obligations, not its implementation. Do not name a technical approach here — that is Stage 8's job, and it does not happen until the user has ratified this framing.\n\n1. **State the frame plainly (required).**\n   - **Business goal** — the value this work delivers and why it matters.\n   - **Desired end-state** — the concrete state the system should reach once this work is done.\n   - **System behavior** — how the system must behave to complete its task (the quality attributes in prose, not a feature list).\n\n2. **Derive the acceptance criteria — what the system must do (required).** Write 3-8 criteria. Each one gets:\n   - An `id` (`AC-1`, `AC-2`, …).\n   - A `criterion` — a single obligation stated concretely enough to be checked. Write it as observable behavior (\"an operator who revokes a key sees the next request rejected\"), not as a task (\"add a revocation endpoint\").\n   - A `verification` — how we would confirm it holds. Name the observable signal: a response code on a specific route, a row state, a log line, a rendered element, a user-visible outcome. **A criterion nobody can check is not yet a criterion** — sharpen it or drop it.\n   - A `status`, using the rubric in step 4.\n\n   Cover the failure and edge behavior, not just the happy path. If the work changes something that already exists, at least one criterion should pin down what must **not** regress.\n\n3. **Identify the non-functional requirements — the standards the system must meet (required).** Consider every one of these canonical NFR categories and include the ones that genuinely apply (omit the rest): security/privacy, performance/latency, reliability/failure-modes, observability/auditability, accessibility/UX, data-integrity/migration, compatibility, operability/config, compliance/SOC2, rollout/reversibility. For each NFR you include, write its `requirement` and its `implication` (what it changes about the implementation) — an NFR with no concrete implication is boilerplate; drop it.\n\n4. **Classify every acceptance criterion and every NFR** with this rubric: `confirmed` only if explicitly stated or observable in code; `assumed` only if a low-risk, reversible default; `open` if it touches architecture, the data model, security, user-visible behavior, migration, or irreversible creation and is not settled. When a criterion or an NFR is genuinely unclear, prefer marking it `open` and asking. Clear criteria make everything downstream more accurate, so surfacing an unclear one is a success, not a delay.\n\n5. **Frame the open requirement questions.** Where a requirement is unsettled **and** has discrete candidate answers, express it as a question the user can answer by clicking (e.g. \"Must revocation take effect immediately, or is eventual acceptable?\"). These become cards in Stage 5. They are questions about *what the system must do* — never about how to build it. If a question has no discrete answers, leave it as prose in the doc instead.\n\n6. **Sanity-check the frame against itself.** Do any two criteria conflict? Does a criterion conflict with an NFR (e.g. an auditability requirement against a latency budget)? Note every tension you find — Stage 7 treats these as a council trigger, and Stage 8 must resolve them explicitly rather than quietly favouring one side.\n\nThis stage is inline analysis — no tool calls required. This stage is non-blocking — always proceed to Stage 4.\n\n## Stage 4 — Write Requirements Draft\n\nWrite what you know so far to disk, so the user has something to read alongside the decision page. The design is deliberately absent — it does not exist yet.\n\n1. Create the `explorations/` directory under `docs_dir` if it does not exist.\n\n2. Write the exploration document to the slug-based path determined in Stage 0 (`{docs_dir}/explorations/{slug}.md`) with this structure:\n\n```markdown\n# Exploration: {concise summary of the prompt}\n\n**Date**: {current date}\n**Prompt**: {original prompt text}\n**Status**: Requirements drafted — awaiting ratification\n\n## Context\n\n{Brief description of the task and what areas of the codebase are relevant.}\n\n## Acceptance Criteria\n\n{The criteria from Stage 3 — what the system must do. One entry per criterion: its id, the criterion itself, how it is verified, and its status (confirmed / assumed / open).}\n\n## Goals & NFRs\n\n{The business goal, desired end-state, and required system behavior from Stage 3. Then the non-functional requirements — the standards the system must meet: each with its category, requirement, implication, and status (confirmed / assumed / open). Note any tension between criteria or between a criterion and an NFR.}\n\n## Open Questions\n\n{Requirement questions that are still unsettled. Mark which ones are going onto the decision page as cards and which are open-ended prose.}\n\n## Codebase Findings\n\n{Key discoveries from Stage 1. What exists today, what patterns are used, what the relevant code paths look like. Reference specific files and functions with file_path:line_number format.}\n\n## Research Findings\n\n{Findings from web searches and deep research, if any. If no research was performed, state \"No external research was needed.\"}\n\n## Key Files\n\n{Bulleted list of the most important files examined, with one-line descriptions of their relevance.}\n```\n\nDo not add a design, an implementation plan, or a recommendation to this draft. Stage 8 adds those once the criteria are settled.\n\nIf the file cannot be written, stop immediately and report the failure.\n\n## Stage 5 — Generate Requirements Decision Page\n\nTurn the Stage 3 framing into an interactive HTML decision page so the user can ratify the requirements by clicking. **This page settles requirements only.** It must not contain a single implementation option — the user is agreeing on what the system must do, not choosing how to build it.\n\n1. **Map the acceptance criteria to `acceptance_criteria`.** Each entry has `id`, `criterion`, `verification`, and `status`. Ids must be unique — a duplicate id is rejected, because the id is the key the page reports the user's stance under. Every criterion renders with an Agreed / Ask about this / Disagree control, so pass all of them, not only the open ones. Pass the NFRs the same way under `nfrs`.\n\n2. **Map each open requirement question from Stage 3 step 5 to an actionable item.** Each entry has:\n   - `id`: a short stable id, e.g. `R-1`, `R-2`.\n   - `question`: the requirement question.\n   - `options`: the 2-4 candidate answers (string array). Do **not** include \"None of these\" or \"Ask about this\" — the renderer auto-appends both.\n   - `option_consequences`: what each answer would mean for the criteria, **parallel to and the same length as** `options`.\n   - `why_it_matters`: the concrete impact line.\n   - `recommendation_explanation`: why the recommended answer is best.\n   - `recommendation_index`: the 0-based index of the recommended answer (must be within `options`).\n   - `codebase_evidence` (optional): the Assessment paragraph plus `file:line` citations, shown collapsed.\n\n   **These cards are requirement questions, never implementation choices.** \"Must revocation be immediate or is eventual acceptable?\" is a valid card. \"Should we use a short-TTL cache or pub/sub invalidation?\" is not — it is a solution, it belongs to Stage 8, and putting it here defeats the purpose of the page. If you cannot phrase a card without naming a mechanism, it is not a requirement question. When there are no such questions, pass an empty array — a criteria-only page is expected and renders correctly.\n\n3. **Call `generate_decision_page`** with routing fields at the root and all heavy arrays nested under `content`:\n   - `artifact_type`: `pre_ticket_planning` (renders the acceptance-criteria and goals panel above any cards).\n   - `ticket_key`: the Stage 0 `slug` (a non-Jira slug is fine — it must start with a letter and contain only letters, digits, hyphens, or underscores).\n   - `output_subdir`: `explorations` (so the page lands beside the markdown doc).\n   - `output_filename`: `{slug}-requirements.html`.\n   - `labels`: requirements-flavored overrides, e.g. `title` = \"Requirements\", `section_heading` = \"Open Requirement Questions\", and an `intro` that frames the page as agreeing on what the system must do before any design work begins.\n   - `content`: an object containing `system_goals` and `actionable_items`. **`system_goals` MUST ALWAYS be passed** inside `content` so the backend always writes a page. Never omit it, even if every criterion and NFR is confirmed. `acceptance_criteria` and `nfrs` both live inside `system_goals`. (Do not pass `implementation_order` inside `content` — that is for epic surfaces, not a single explored task.)\n\n   ```typescript\n   interface ExploreTicketContent {\n     system_goals?: {\n       business_goal: string;\n       desired_end_state: string;\n       system_behavior: string;\n       acceptance_criteria?: Array<{\n         id: string;                    // e.g. \"AC-1\"; must be unique\n         criterion: string;             // what the system must do\n         verification: string;          // how we would confirm it holds\n         status: \"confirmed\" | \"assumed\" | \"open\";\n       }>;\n       nfrs?: Array<{\n         category: string;\n         requirement: string;\n         implication: string;\n         status: \"confirmed\" | \"assumed\" | \"open\";\n       }>;\n     };\n     actionable_items?: Array<{\n       id: string;                      // e.g. \"R-1\"; a REQUIREMENT question, not a design choice\n       question: string;\n       why_it_matters: string;\n       recommendation_explanation: string;\n       options: string[];               // 2-4 candidate answers (no \"None of these\" or \"Ask about this\")\n       option_consequences: string[];   // same length as options\n       recommendation_index: number;   // 0-based within options\n       codebase_evidence?: string;      // optional: assessment + file:line citations\n       original_question?: string;      // optional: only when item maps to a verbatim question\n     }>;\n     // clear_improvements: not used by this command — it captures requirements, not findings\n     // implementation_order: for epic surfaces only — do NOT include for single task explorations\n     // depends_on: hard prerequisites (titles/keys that must land first)\n     // recommended_after: soft sequencing preferences, not hard blockers\n   }\n   ```\n\n   Example call:\n   ```json\n   {\n     \"ticket_key\": \"revoke-api-keys\",\n     \"artifact_type\": \"pre_ticket_planning\",\n     \"output_subdir\": \"explorations\",\n     \"output_filename\": \"revoke-api-keys-requirements.html\",\n     \"labels\": { \"title\": \"Requirements\", \"section_heading\": \"Open Requirement Questions\" },\n     \"content\": {\n       \"system_goals\": {\n         \"business_goal\": \"Operators can cut off a leaked key immediately.\",\n         \"desired_end_state\": \"Revocation is self-serve and takes effect at once.\",\n         \"system_behavior\": \"Rejects revoked credentials without a restart.\",\n         \"acceptance_criteria\": [\n           { \"id\": \"AC-1\", \"criterion\": \"An operator who revokes a key sees the next request with it rejected.\", \"verification\": \"The following call to the protected route returns 401.\", \"status\": \"confirmed\" },\n           { \"id\": \"AC-2\", \"criterion\": \"Revocation is recorded with actor and timestamp.\", \"verification\": \"An audit row names the operator and the revoked key id.\", \"status\": \"open\" }\n         ],\n         \"nfrs\": [\n           { \"category\": \"security/privacy\", \"requirement\": \"The raw key is never logged on the revoke path.\", \"implication\": \"Log the key id, never the secret.\", \"status\": \"open\" }\n         ]\n       },\n       \"actionable_items\": [\n         {\n           \"id\": \"R-1\",\n           \"question\": \"Must revocation take effect immediately, or is eventual acceptable?\",\n           \"why_it_matters\": \"Sets the hard bound AC-1 has to meet.\",\n           \"recommendation_explanation\": \"A leaked key is an active incident; eventual leaves a usable window.\",\n           \"options\": [\"Immediately (under 5s)\", \"Eventually (under 60s is acceptable)\"],\n           \"option_consequences\": [\"AC-1 gains a 5s bound.\", \"AC-1 gains a 60s bound.\"],\n           \"recommendation_index\": 0\n         }\n       ]\n     }\n   }\n   ```\n\n4. **Handle the response `status`:**\n   - `no_decisions_needed`: no page was written. This should not occur when `system_goals` is always passed. Skip Stage 6 entirely, tell the user there were no open requirements, and proceed to Stage 7 treating the Stage 3 framing as the settled criteria.\n   - `decision_page_generated`: surface the returned `file_path` and proceed to Stage 6. **Always proceed to Stage 6 when `decision_page_generated` is returned**, regardless of `actionable_items_count`. A criteria-only page with zero cards still has stance controls that must be submitted.\n\nThis stage is non-blocking: if `generate_decision_page` fails, do not halt. **You MUST output a highly visible warning** (e.g. **⚠ WARNING: The requirements page could not be generated** in bold) explaining that generation failed and that the user should review the criteria in the markdown doc written in Stage 4 instead. Do not silently continue — the failure must be diagnosable from your output. Then ask the user to confirm the criteria in chat before proceeding to Stage 7.\n\n## Stage 6 — Ratify Requirements\n\nCapture the user's stances, settle the criteria, gate on the implications review, and fold the result into the doc. Nothing downstream — including a council in Stage 7 — may start until the criteria are agreed and the implications gate has accepted a proceed token.\n\n1. **Direct the user to the page.** Provide the `file_path` from Stage 5 and tell them to open it in their browser. Explain that they are agreeing on what the system must do — not how it will be built — that they can accept, question, or reject each criterion, and that they can ask questions in chat before submitting.\n\n2. **Q&A loop and commit signal.** Engage with each user message as either a commit or a discussion turn:\n   - **Commit:** trim the full message and attempt to parse the entire trimmed message as JSON. Treat it as a commit only when the parsed value is an object with all three top-level fields: `ticket_key` (string), `decisions` (object), and `general_comment` (string). The first valid commit-shaped paste commits — do not over-validate the per-card fields. The page also submits `acceptance_criteria_feedback` and `nfr_feedback` objects, each keyed by criterion id or NFR category with a `stance` of `agreed`, `ask`, or `disagree` plus a `comment`.\n   - **Discussion:** anything that is not commit-shaped JSON. Answer from the doc written in Stage 4 and from codebase lookups. If a JSON-shaped paste is missing one of the three required fields, say which field is missing rather than treating it as a freeform question.\n   - **In-flight overrides:** when the user clearly changes an answer in chat (\"AC-2 is wrong\", \"go with eventual for R-1\") or gives new overarching guidance, record it as a working-memory override. On commit, the submitted JSON is the baseline and recorded overrides take precedence; post a one-line acknowledgement naming each overridden item before you rewrite the doc.\n\n3. **Resolve every \"ask\" (hard rule).** After accepting a commit, scan all three: any item in `decisions` where `choice === \"ask\"`, any entry in `acceptance_criteria_feedback` where `stance === \"ask\"`, and any entry in `nfr_feedback` where `stance === \"ask\"`. For each, present the relevant evidence and continue the discussion until the user gives an explicit answer, which you record as an override. Do not proceed while any `ask` remains unresolved — do not honor \"just skip those\".\n\n4. **Resolve every \"disagree\".** A disagree means the criterion is wrong as written. Work out with the user what it should say, restate it back, and get explicit agreement on the corrected wording. A rejected criterion is either rewritten or dropped — never carried forward as-is.\n\n5. This surface's tracked stances are the `acceptance_criteria_feedback` and `nfr_feedback` objects captured above. **Review the wider implications, then gate on a decision.** Build the review from the complete settled set: the submitted `decisions`, any in-flight overrides recorded during the conversation (these take precedence over the submission), every `\"none\"` answer together with the reason given for it, `general_comment`, and — where this surface tracks acceptance-criterion or NFR stances — those stances too. Do not start the review until every `ask` has an explicit recorded resolution and every in-flight override has been applied.\n\n   Consider three fixed categories, regardless of whether a decision was framed as technical, user-facing, or business-oriented:\n   - **Program / application** — architecture, code paths, operability, maintenance burden, and requirements imposed on other parts of the software.\n   - **User** — end users, new users performing setup, operators, and developers, including prerequisites, setup friction, and additional steps.\n   - **Business** — cost, adoption, support load, compliance, and reversibility.\n\n   Emit only the categories with material second-order implications. For each included category, write at most four one-line bullets of about 25 words, each naming who or what is affected and how — never a restatement of the selected decision. Close with a line naming every considered category that was omitted, e.g. `Considered, nothing material: business.` — omit this closing line only when all three categories have material implications.\n\n   If the review cannot be produced, report that in one line and continue without stalling the workflow or presenting the gate below.\n\n   Retain the latest emitted review; once the gate below allows continuation, write it under `## Implications` during the Stage 6 document rewrite.\n\n   Then present the gate, verbatim: `Implications reviewed. Proceed, or name a decision to revisit.` Accept only a normalized `proceed`, `yes`, `y`, or `go` as a continuation token. Any other response names a decision to reopen: re-settle it in chat, record the new override, rerun the entire implications review against the changed settled set, and present the gate again.\n\n   Literal `auto_approve = true` emits the review but skips this gate entirely; a missing or non-true `auto_approve` value follows the human-in-the-loop path above.\n\n6. **Settle the criteria and update the doc.** Rewrite the Acceptance Criteria, Goals & NFRs, and Open Questions sections of `{docs_dir}/explorations/{slug}.md` to the agreed set: fold in every correction, resolve each answered requirement question into the criterion it affects, promote settled criteria out of `open`, and weave `general_comment` in as overarching guidance. Add or update the `## Implications` section with the latest review from item 5 — the final one after any reopen/review loop — or its one-line fail-open notice if generation was unavailable. Set the doc's Status line to \"Requirements ratified\". Preserve all unaffected sections unchanged. **The settled criteria are now the contract** — every later stage is judged against them.\n\nThis stage is non-blocking: if the user never commits, leave the doc as written in Stage 4, tell them the requirements are unratified, and stop without forcing a decision. Do not proceed to a council or a design on unratified criteria.\n\n## Stage 7 — Council Gate\n\nThe criteria are ratified. Now assess honestly whether you know **how** to meet them — and offer to convene a council when you do not.\n\n**Lean toward offering.** A council is cheap relative to committing the user to the wrong design, and this command prioritizes discovery over premature commitment. Do not wait for the user to ask for one.\n\n1. **Check the triggers.** Offer a council when **any** of these hold:\n   - More than one materially different approach could satisfy a criterion, and the codebase evidence you gathered cannot separate them.\n   - A criterion has no obvious implementation path in the existing code.\n   - Meeting one criterion appears to trade off against another criterion or against an NFR (any tension noted in Stage 3 step 6, or created by a correction in Stage 6).\n   - The work touches an area with no established pattern — Stage 1 found no precedent to follow.\n   - Ratification materially changed the problem — the user tightened a bound, rejected a criterion, or added an obligation you had not framed.\n   - Stage 2 research surfaced competing approaches with no clear winner.\n\n   Do **not** offer when every ratified criterion maps cleanly onto a well-trodden pattern already used in this codebase and you can point to the precedent.\n\n2. **Ask for approval.** When a trigger fires, first summarize the uncertainty in 1-3 bullets — name the specific criteria at issue and what you cannot currently decide. Then ask exactly:\n\n   ```\n   Significant uncertainty about how to meet {AC ids}. Run a council before I draft the design? (y/N)\n   ```\n\n   Mention that a council polls for up to ~15 minutes before you ask, so the user is choosing with the cost in view.\n\n   Treat an empty response, any negative response (`n`, `no`, or similar), or any ambiguous/unrecognized response as **decline** — do not guess intent. On decline, note in one line that the council was offered and declined, and proceed to Stage 8 on your own analysis. Never run a council without an explicit affirmative (`y` or `yes`).\n\n3. **Run it on approval.** Call `request_council` with:\n   - `task_description`: the task, the **ratified** acceptance criteria and NFRs from Stage 6, and the specific uncertainty you summarized. Sent verbatim — this tool does not read from a file. State plainly that the criteria are settled and the council's job is to find how to meet them, not to revisit what they are.\n   - `mode`: `technical`.\n   - `wait_for_result`: `true`. `save_locally`: `true`.\n\n   While it runs, tell the user it is polling and roughly how long it may take.\n\n4. **Fold the result into your analysis.** Carry the council's approaches, objections, and any option you had not considered into Stage 8. If the council argues a ratified criterion is unmeetable, do not silently drop it — raise it with the user in Stage 8 as an explicit conflict.\n\nIf the council fails or times out, note the failure visibly and proceed with your own analysis — a missing council degrades the design but does not invalidate it. This stage is non-blocking — always proceed to Stage 8.\n\n## Stage 8 — Propose Final Design\n\nNow describe how you would build it. **Do not generate a decision page for this stage.** The requirements page was the user's decision surface; the design is your proposal, written into the doc and discussed in chat. Generating a second page here would ask the user to ratify a solution, which is not what this command does.\n\n1. **Work out the design against the ratified criteria.** Consider the approaches you know plus anything the council surfaced. For each candidate, establish which criteria it satisfies and at what cost. An approach that cannot meet a ratified criterion is not a candidate — discard it and say why.\n\n2. **Resolve any tension explicitly.** Where meeting one criterion costs another, or costs an NFR, state which obligation your design privileges and what that costs the other. Do not let a tension pass silently.\n\n3. **Commit to a single proposed design.** You are recommending, not offering a menu. Name the approach, describe how it works, list the files to create or modify, and map each ratified criterion to the part of the design that satisfies it. Where you seriously considered an alternative, record it and why you rejected it — as history, not as an open choice.\n\n   **Record the implementation slices with a size band.** The design's sequence of work is also the proposed ticket split, so name each slice and size it with the ladder in the posture block below — `S` / `M` / `L` / `XL`, by file-touch breadth and rough LOC. Size the slices here, in the design, rather than at the handoff gate: the split is a design decision, and deciding it under the pressure of an approval prompt is how a coherent design turns into a swarm of tiny tickets. A slice that will not fit in `L` becomes one `XL` slice — do not split it to fit a band. Split only when it is genuinely two independent pieces of work, or when it runs past roughly 40 files / ~3000 LOC.\n\n4. **Rewrite `{docs_dir}/explorations/{slug}.md`** so it reads as a finished proposal, not a mechanical append. Set the Status line to \"Design proposed\". Keep the ratified Acceptance Criteria, Goals & NFRs, and `## Implications` sections intact — they are the contract and must not drift — and add:\n\n```markdown\n## Approaches Considered\n\n{Each candidate, what it would mean, and why it was or was not chosen. Note which came from the council, if one ran. If no alternatives were seriously considered, state that and why the path was obvious.}\n\n## Proposed Design\n\n{The recommended approach in enough detail to implement: how it works, the files to create or modify, the sequence of work, and the risks. Reference specific files with file_path:line_number format.}\n\n## Criteria Coverage\n\n{Each ratified criterion mapped to the part of the design that satisfies it, and how it will be verified. Any criterion the design only partially meets must say so plainly.}\n```\n\n5. **Present the design in chat and invite pushback.** Summarize the proposal in at most six bullets — the approach, the files it creates or modifies, the order of the work, and any ratified criterion the design only partially meets — and state clearly that it is a proposal. If the user objects, revise the design — but if their objection actually changes what the system must do rather than how it is built, say so: that is a criteria change, and it means reopening the criteria rather than quietly bending the design around it.\n\nThis stage is non-blocking — always proceed to Stage 9 once the design is written and summarized in chat, even if the user has not responded to it.\n\n## Stage 9 — Ticket Handoff\n\n<!-- BEGIN BRIDGE TICKET-AUTHORING POSTURE -->\n<!-- Canonical source: docs/bridge-ticket-authoring.md.\n     This block is duplicated byte-identically onto every carrier. Never edit a\n     copy: edit the canonical source and re-copy it verbatim. A cross-surface\n     byte-equality test fails the build if any copy drifts by a single byte. -->\n\n## Ticket-authoring posture\n\nDeep reference: `docs/bridge-ticket-authoring.md`.\n\n**Draft through the writer.** Every ticket body — an epic parent, an epic child,\nand an ordinary sibling alike — is drafted by the `jira-ticket-writer` agent\nbefore `create_ticket` is called. Do not compose a ticket description inline.\n\n**Size the work.** Size each ticket by file-touch breadth and depth plus rough\nlines of code (LOC) changed:\n\n- `S = 1-2 files / <~80 LOC`\n- `M = ~3-8 files / ~80-400 LOC`\n- `L = ~8-15 files / ~400-900 LOC`\n- `XL = >15 files / >~900 LOC`\n\nTarget size priority: **L (target) -> XL (when the work does not fit in L) -> M\n(third choice) -> S (only when unavoidable)**. This applies equally to a\nstandalone ticket and to an epic child.\n\nAim each slice at L. When one will not fit, grow it to XL rather than splitting\nit — split only when the slice is genuinely two independent pieces of work,\nnever merely to land inside a band. Bridge's grooming and implementation process\nhandles a large vertical slice well and is overkill on small ones: every extra\nticket is another worktree, another PR, another rebase, and another chance for\ntwo workers to touch the same file. Reach for M because the work genuinely is\nthat size, not to avoid an XL.\n\nBeyond roughly 40 files or ~3000 LOC, split anyway. Past that point review\nturnaround and rebase cost dominate the run's budget, and a review that wedges\nholds the gate to its full retry ceiling before anyone notices.\n\n**Group at three.** Three or more implementable tickets is an epic: propose an\nepic parent plus an ordered child manifest, and resolve this surface's own\napproval gate before anything is created. One or two tickets are ordinary\nsiblings — no epic parent, no manifest. The threshold is exactly three.\n\n**Hand off once.** An epic handoff names exactly one conductor entry point,\n`drive-epic`, which selects the runnable path itself. Never present a choice\nbetween conductors.\n\n**Departure is closed-list only.** These three exceptions, and no others, permit\ndeparting from the rules above. Invoking one requires no announcement.\n\n- **E1 External-tracker mirroring** — a recorded upstream identifier exists and\n  its granularity is contractual. Bypasses sizing and the epic threshold.\n- **E2 Discovery-only spike** — no committed production-code deliverable.\n  Bypasses sizing only; does not bypass drafting through the writer.\n- **E3 Authorized incident containment** — tied to an active incident record,\n  not to schedule pressure. Bypasses sizing and the epic threshold.\n\nThe list is closed. Anything outside it is an escalation to the operator, not a\njudgement call. Explicitly refused as grounds for departure: a single-file\ntrivial fix (that is `S` reached through the normal path, not an exception),\ngeneric time pressure, \"already well specified\", \"faster without the writer\",\ndeveloper discretion, minor refactor, unattended mode, context limits, and \"hard\nto decompose\" (XL is the normal overflow, so that is the ordinary path and not a\ndeparture). Writer unavailability escalates; it never silently authorizes inline\ndrafting.\n\n<!-- END BRIDGE TICKET-AUTHORING POSTURE -->\n\n1. **Assess readiness.** The work is ready to become a ticket when the criteria are ratified, the design is proposed, and no criterion is left unresolved or only partially covered. If something is still open, say what it is and recommend the follow-up that would close it rather than creating a ticket on a soft foundation:\n   - A **wider council** (`request_council`) when the design would benefit from a broad review before implementation. If Stage 7 already ran one, only suggest another when something material changed since.\n   - A **second opinion** (`second_opinion`) when a few specific contested points need an independent check.\n   - **Web or deep research** (`request_deep_research`) when the design still rests on technical unknowns that need grounding.\n\n2. **Decompose once, then offer to create.** When the work is ready, freeze the split before you show it, and show the whole thing. Never present the gate bare — the user cannot consent to a plan they have not been shown.\n\n   Count the implementation slices Stage 8 recorded. That count decides the shape, and the threshold is exactly three:\n\n   - **Three or more slices → propose an epic.** Freeze an epic parent plus an ordered child manifest in one pass. Per child: title, scope boundary, size band, `depends_on` (hard prerequisites that must land first), `recommended_after` (soft sequencing preferences that are not blockers), and a one-line order rationale. Keep hard prerequisites strictly separate from soft sequencing. A child may be `XL`; grow one rather than splitting it to fit, and split only past roughly 40 files / ~3000 LOC.\n   - **One or two slices → ordinary siblings.** No epic parent, no manifest, no decomposition pass. Just the ticket or the pair.\n\n   This decomposition is the **single authoritative split** for the rest of the stage. It is decided here, once, with the whole portfolio in view; step 3 renders bodies against it and does not revisit it.\n\n   Output the plan outline, kept to roughly one screen:\n\n   - **Plan** — the approach in one sentence.\n   - **Shape** — `epic` (with the parent's title) or `siblings`. Say which and why the count put it there.\n   - **Tickets ({n})** — one line per ticket: title, the slice of scope it covers, its size band, its ratified criterion ids, and — for an epic — its `depends_on` / `recommended_after` and order rationale. Say plainly when it is a single ticket.\n   - **Files** — the files to create or modify, grouped by area.\n   - **Sequence** — the ordered steps of the work, one line each.\n   - **Risks / not covered** — any criterion the design only partially meets, or `none`.\n\n   Then ask exactly:\n\n   ```\n   Requirements ratified and design proposed. Create the ticket(s) now? (y/N)\n   ```\n\n   Treat an empty response, any negative response, or any ambiguous/unrecognized response as **decline** — do not guess intent. On decline, report that the exploration doc is the artifact, and tell the user they can later ask their agent to use the Jira Ticket Writer to create ticket drafts from it. Never create a ticket without an explicit affirmative (`y` or `yes`) — creation is irreversible.\n\n3. **Render every body through `jira-ticket-writer`, then create.** Nothing calls `create_ticket` before the gate in step 2 resolves with an explicit affirmative.\n\n   On approval, fan out **one `jira-ticket-writer` invocation per entry** in the frozen outline — the epic parent, each child, and each ordinary sibling. Each invocation is bound to exactly its own entry and renders a body against that entry's fixed boundary and size band. A child invocation must not re-split, merge, reorder, renumber, or rescope: the decomposition is already frozen, and an invocation that sees one entry is deciding on strictly less information than the pass that saw them all.\n\n   Pass into every invocation, verbatim: the ratified acceptance criteria that entry satisfies, the proposed design for its slice, and the entry's own manifest fields. Do not restate or reinterpret the criteria — they were ratified in that wording, and they go to the writer unchanged.\n\n   Then create exactly the approved set from the writer's drafts:\n\n   - **Siblings** — `create_ticket` per draft, unparented.\n   - **Epic** — `create_ticket` with `issue_type` set to `Epic` for the parent first, capture the resolved epic key, then `create_ticket` with `parent_key` set to that key for each child in manifest order. The `parent_key` argument is required on every child call; never omit it.\n\n   Create the tickets exactly as the outline named them — the split was decided and shown there, so do not create a different set than the user approved. If you now believe the split is wrong, say so and re-ask rather than changing it silently.\n\n   **Track each ticket as you create it.** Immediately after each successful `create_ticket`, call `track_ticket` with that key and the ticket's description. Bridge's workflow state lives in its own database, not in the ticket backend, and an untracked ticket is invisible to it — `update_ticket_state` and `get_ticket_state` both return 404 until something registers the key. Tracking works against either ticket backend and is a safe no-op when the key is already tracked.\n\n   Tracking is **warn, not halt**: if `track_ticket` returns an error of any kind, record the reason for the final report and keep going. The ticket exists and is the thing that mattered; tracking can be repeated later.\n\n   **On a partial failure, stop and report what exists.** Ticket creation is irreversible and there is no resume protocol here: report the epic key and every child key created so far, name the entry that failed, and tell the user to re-run naming the existing epic rather than repeating this command. Do not retry the failed call in a loop, and do not create the remaining children as though nothing happened — a half-built epic the user cannot see is worse than a stopped one they can.\n\n   Idempotency labels are deliberately not used. They ride on `create_ticket`'s `labels` argument and are read back through `get_tickets(labels=...)`, and **both are Jira-only**: against the local ticket backend each returns a terminal `409 UNSUPPORTED_IN_LOCAL_MODE`, and the create rejects without making a ticket. A resume protocol built on them would work on one backend and silently fail on the other.\n\n4. **Write the plan DAG.** After an epic's tickets are created, write the conductor's plan sidecar to `{docs_dir}/epic-plans/{slug}/epic-plan.dag.json`. You already hold everything it needs: the frozen manifest carries each child's `depends_on`, and Stage 8's design named the files each slice touches. Nothing downstream derives this from an epic key, so if you skip it the handoff dead-ends.\n\n   ```json\n   {\n     \"plan_version\": 1,\n     \"nodes\": [\n       {\n         \"ticket_key\": \"KEY-2\",\n         \"status\": \"planned\",\n         \"depends_on\": [],\n         \"touched_files\": [\"src/example/service.py\", \"src/example/handler.py\"]\n       }\n     ],\n     \"edges\": []\n   }\n   ```\n\n   One node per child, in manifest order, using the **real** keys just created — never a `TBD-N` placeholder, since the tickets already exist. `depends_on` carries the manifest's hard prerequisites only; leave `recommended_after` out, because soft sequencing is not a dependency and encoding it there serializes work that could run in parallel. `edges` may be `[]` when `depends_on` already expresses the graph. The graph must be acyclic and every `depends_on` entry must name a node in the file.\n\n   Populate `touched_files` on every node from the design's file list, as repository-relative POSIX paths. It is what lets the conductor serialize two children that would otherwise collide, and it matters more the larger the slices are. An omitted list is accepted by the server as an explicit opt-out, so leaving it off does not fail — it silently turns overlap protection off. Use `[]` only when a ticket genuinely touches no repository files.\n\n   Include the files that are easy to forget, because those are the ones that actually collided: **anticipated tests, fixtures, golden files, guard allowlists, and policy files**, not only the production modules. Both known misses were exactly this class — two children shared `ticket_type_evaluator.py` *and its test*, and another pair shared golden files and a `save_repo_to_db` test — and in each case nobody predicted the shared file, so no ordering edge was synthesized and both PRs merged over each other. Paths are compared as exact strings after normalization, so name concrete files: a directory or a glob matches nothing.\n\n5. **Hand off once.** Then point at exactly one conductor entry point:\n\n   ```\n   npx -y @bridge_gpt/mcp-server drive-epic {EPIC-KEY} --plan-file {docs_dir}/epic-plans/{slug}/epic-plan.dag.json\n   ```\n\n   `drive-epic` reads conductor readiness and selects the runnable path itself. Do not name an underlying conductor, and never present a choice between two of them — two transition authorities on one epic is the failure this single entry point exists to remove. A sibling pair or a lone ticket gets no conductor handoff at all.\n\nThis stage is non-blocking: if the user never answers, leave the doc as written in Stage 8 and stop without forcing a decision.\n\n## Final Report\n\nOn successful completion of all stages, display:\n\n> **Exploration Complete**\n>\n> **Prompt**: {first 80 characters of prompt}...\n> **Output**: {full path to the exploration doc}\n> **Requirements Page**: {full path to the generated requirements.html, or \"not generated\" when generation failed}\n> **Acceptance Criteria**: {count} ratified ({count} corrected by the user, {count} still open)\n> **Files Examined**: {count of key_files_examined}\n> **Research**: {count of web_searches} web searches, {count of research_queries} deep research queries, council {\"run\" | \"offered and declined\" | \"not needed\"}\n>\n> **Requirements**: {\"Ratified\" | \"Unratified — page not submitted\"}\n> **Design**: {\"Proposed\" | \"Not reached\"}\n> **Ticket(s)**: {\"Created: KEY-1, KEY-2\" | \"Declined — doc is the artifact\" | \"Not offered — work not ready\"}\n> **Tracking**: {\"Registered\" | \"Warning: not tracked — {reason}\" | \"n/a — no tickets created\"}\n> **Plan DAG**: {full path to epic-plan.dag.json | \"n/a — not an epic\"}\n\nOn failure at any stage, stop immediately and report:\n- Which stage failed (by number and name)\n- The error details\n- Any partial results that were produced before the failure\n",
     "idea-to-pr.md": "Turn an idea into tickets with the `idea-to-ticket` recipe, then hand the created keys to `/review-and-start --auto`.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nThis command is a thin client-side wrapper around two things that already exist: the\n`idea-to-ticket` recipe, and the `/review-and-start` command. It runs them in that order,\nin this session, and does nothing else.\n\nStage 1 is recipe-driven. Do not call MCP tools directly other than `get_pipeline_recipe`\n— the recipe determines which tools to call and with what parameters. Stage 2 is a single\nslash-command invocation. There is no server-side run to start, poll, resume, or schedule:\nthis command holds no state of its own between the two stages beyond the ticket keys stage 1\nreturned.\n\n## Stage 0 — Parse arguments\n\n1. Tokenize `$ARGUMENTS` on whitespace. Recognize these position-independent flags; every\n   other token is part of the idea:\n   - `--allow-duplicate`\n   - `--max-children=N` where `N` is a positive integer\n   - `--auto`\n\n   Flags are optional. Treat absence of a flag as the default; never pass empty-string or\n   null as a placeholder.\n\n2. Reject **unknown flags** — any token beginning with `--` that is not one of the three\n   above — before calling `get_pipeline_recipe`. Stop and report the offending flag.\n\n3. Reject a **non-positive or non-integer `--max-children`** value before calling\n   `get_pipeline_recipe`. Stop and report the offending value.\n\n4. Everything that is not a recognized flag is the free-form `idea` text. Join those tokens\n   back together preserving order. Trim surrounding whitespace.\n\n5. If the resulting `idea` is empty, stop immediately and display:\n   ```\n   Usage: /idea-to-pr <idea> [--allow-duplicate] [--max-children=N] [--auto]\n   ```\n\n## Stage 1 — Run the `idea-to-ticket` recipe\n\n6. Derive `slug` from the first 6-8 meaningful words of the idea: lowercase, kebab-case,\n   strip non-alphanumeric characters except hyphens, and truncate to roughly 60 characters.\n   Skip stop-words such as \"the\", \"a\", \"an\" when picking the 6-8 meaningful words.\n\n7. Derive `run_id` as `<YYYYMMDD-HHMMSS>-<short-uuid>` using the current UTC time and a\n   short UUID suffix (8 hex chars is enough).\n\n8. Derive the boolean-as-string variables:\n   - `allow_duplicate` is `\"true\"` if `--allow-duplicate` was present, otherwise `\"false\"`.\n   - `auto_approve_external` is `\"true\"` if `--auto` was present, otherwise `\"false\"`.\n   - `max_children` is the integer following `--max-children=` as a string, or `\"10\"` when\n     the flag is absent.\n\n9. Call the `get_pipeline_recipe` MCP tool with:\n   - `pipeline`: `\"idea-to-ticket\"`\n   - `variables`: `{ \"idea\": \"<idea>\", \"slug\": \"<slug>\", \"run_id\": \"<run_id>\", \"allow_duplicate\": \"<allow_duplicate>\", \"auto_approve_external\": \"<auto_approve_external>\", \"max_children\": \"<max_children>\" }`\n\n   Do NOT pass `docs_dir` or `idea_hash` in variables — both are auto-injected by the\n   pipeline system (`docs_dir` from `BAPI_DOCS_DIR`; `idea_hash` is a stable hash derived\n   from the `idea`).\n\n   If the tool returns an error, stop and report the failure.\n\n10. Read and strictly obey the `agent_instructions` field in the response. Execute each\n    returned step in the order returned, announcing each as **Step N of M: <description>**.\n    The recipe owns its internal stages — do not restate them here and do not invoke them\n    directly.\n\n11. **A recipe halt is terminal.** If any step halts — a failed preflight,\n    `too_vague_to_ticket`, duplicate detection, or a screen/resolve halt — stop there.\n    Report the upstream halt reason with enough of its own wording to stay actionable, and\n    state explicitly that no review/start handoff was made. Never fall through a halt to\n    key collection or to Stage 2.\n\n## Stage 2 — Hand off to `/review-and-start`\n\n12. Collect the created Jira keys from the terminal `upload-and-track` step's result, in\n    the exact order that step returned them. Do not query or reconstruct the keys through\n    any other MCP tool.\n\n13. If zero keys were returned, stop with that step's reason. Do not invoke\n    `/review-and-start` with an empty key list.\n\n14. Otherwise invoke, **exactly once, in this same session**:\n    ```\n    /review-and-start --auto <KEY> [KEY ...]\n    ```\n    with the keys in the order stage 1 returned them. `/review-and-start` accepts a variadic\n    key list and threads `--auto` into both the review and the implementation phase of every\n    session it spawns. The handoff always carries `--auto` — this command is hands-off by\n    default, and the command-line `--auto` flag governs stage 1's `auto_approve_external`\n    rather than the handoff.\n\n## Stage 3 — Final report\n\n15. Keep the report compact:\n    ```markdown\n    ## Idea to PR Complete\n\n    Idea: <first 80 chars of idea>...\n    idea-to-ticket: <completed | halted at \"<step description>\">\n    Keys created: <KEY, KEY, ... in returned order, or \"none\">\n    Handoff: <\"/review-and-start --auto <keys>\" issued | not issued — <reason>>\n    ```\n\n    On a halt, `Keys created` is `none` and `Handoff` states `not issued` with the upstream\n    halt reason.\n",
     "idea-to-ticket.md": "Convert a short human idea into a Jira ticket (or Epic plus child tickets) via the server-side idea-to-ticket pipeline.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nThis command is recipe-driven. Do not call MCP tools directly other than `get_pipeline_recipe` — the recipe determines which tools to call and with what parameters.\n\n## Stage 0 — Parse arguments\n\n1. Tokenize `$ARGUMENTS` on whitespace. Recognize the following position-independent flags (any other tokens are part of the idea):\n   - `--allow-duplicate`\n   - `--max-children=N` where `N` is a positive integer\n   - `--auto`\n\n   Flags are optional. Treat absence of a flag as the default; never pass empty-string or null as a placeholder.\n\n2. Everything that is not a recognized flag is the free-form `idea` text. Join those tokens back together preserving order. Trim surrounding whitespace.\n\n3. If the resulting `idea` is empty, stop immediately and display:\n   ```\n   Usage: /idea-to-ticket <idea> [--allow-duplicate] [--max-children=N] [--auto]\n   ```\n\n## Stage 1 — Derive pipeline variables\n\n4. Derive `slug` from the first 6-8 meaningful words of the idea: lowercase, kebab-case, strip non-alphanumeric characters except hyphens, and truncate to roughly 60 characters. Skip stop-words such as \"the\", \"a\", \"an\" when picking the 6-8 meaningful words.\n\n5. Derive `run_id` as `<YYYYMMDD-HHMMSS>-<short-uuid>` using the current UTC time and a short UUID suffix (8 hex chars is enough). The combination of `slug` and `run_id` uniquely identifies this run's artifact directory.\n\n6. Derive the boolean-as-string variables:\n   - `allow_duplicate` is `\"true\"` if `--allow-duplicate` was present, otherwise `\"false\"`.\n   - `auto_approve_external` is `\"true\"` if `--auto` was present, otherwise `\"false\"`.\n   - `max_children` is the integer following `--max-children=` as a string, or `\"10\"` when the flag is absent.\n\n## Stage 2 — Call the recipe\n\n7. Call the `get_pipeline_recipe` MCP tool with:\n   - `pipeline`: `\"idea-to-ticket\"`\n   - `variables`: `{ \"idea\": \"<idea>\", \"slug\": \"<slug>\", \"run_id\": \"<run_id>\", \"allow_duplicate\": \"<allow_duplicate>\", \"auto_approve_external\": \"<auto_approve_external>\", \"max_children\": \"<max_children>\" }`\n\n   Do NOT pass `docs_dir` or `idea_hash` in variables — both are auto-injected by the pipeline system (`docs_dir` from `BAPI_DOCS_DIR`; `idea_hash` is a stable hash derived from the `idea`).\n\n   If the tool returns an error, stop and report the failure.\n\n8. Read and strictly obey the `agent_instructions` field in the response. Execute each step in order, announcing each as **Step N of M: <description>**.\n\n   The recipe drives the ordered stages for you — do not invoke them directly. In order they are: preflight-and-readiness → research-decision → execute-research → duplicate-and-context-scan → screen-and-resolve → frame-goals-and-nfrs → **comp-analysis** (a gated, backend-safe perception step that maps any attached/referenced design comp to existing components, templates, SCSS/CSS tokens, and routes before drafting; it short-circuits for backend-only or no-comp work) → draft-and-critique → upload-and-track.\n\n## Stage 3 — Final summary\n\n9. After all steps complete, display a summary:\n   ```\n   ## Pipeline Complete\n\n   **Idea**: <first 80 characters of idea>...\n   **Slug**: <slug>\n   **Run directory**: <docs_dir>/idea-to-ticket/<slug>-<run_id>/\n   **Steps executed**: N of M\n   **Status**: Success / Failed at step N\n   ```\n",
     "implement-ticket.md": "# Implement Ticket\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nThis command is recipe-driven. Do not call MCP tools directly -- the recipe determines which tools to call and with what parameters.\n\n1. Parse `$ARGUMENTS` to extract:\n   - A required `ticket_key` matching the Jira key pattern (`[A-Z][A-Z0-9]+-\\d+`).\n   - An optional position-independent `--auto` flag.\n\n   Tokenize `$ARGUMENTS` on whitespace. The first token matching the Jira key pattern is the `ticket_key`; ignore any additional ticket-key tokens. The presence of a `--auto` token (anywhere in `$ARGUMENTS`) sets `auto_approve` to `true`.\n\n   If `$ARGUMENTS` is empty or contains no token matching the Jira key pattern, stop immediately and display:\n   ```\n   Invalid ticket key format. Expected: PROJ-123 [--auto]\n   Usage: /implement-ticket <ticket_key> [--auto]\n   ```\n\n2. Call the `get_pipeline_recipe` MCP tool with:\n   - `pipeline`: `\"implement-ticket\"`\n   - `variables`: `{ \"ticket_key\": \"<ticket_key>\" }`\n   - `auto_approve`: `true` — only when `--auto` was passed; otherwise omit this field entirely.\n\n   If the tool returns an error, stop and report the failure.\n\n   Retain the **complete** JSON response — every step, plus `total_steps`, `auto_approve`, and `execution_mode`. `execution_mode` is `\"inline\"` here: you are the executor, there is no server-side orchestrator, and each step's instruction branches on that value.\n\n3. Read and strictly obey the `agent_instructions` field in the response. Execute **every** step the response returned, in its resolved order, announcing each as **Step N of M: <description>** immediately before executing it. Traverse the whole array — never a fixed or remembered subset, which would silently omit the later steps.\n\n   **A phase's durable-recording tool call is not the end of the command.** The `record_phase_result` and `record_checkpoint` tools return a success envelope; that envelope means the *current* `agent_task` step finished, nothing more. Continue immediately with the next step in the same turn. This applies in particular at the 3 → 4, 5 → 6, and 8 → 9 boundaries, which are where the recipe previously stopped: those instruction files used to end in a text envelope addressed to an orchestrator that does not exist inline, so the turn ended there and the run stalled with no error. None of these boundaries introduces an approval pause.\n\n   **Background work is not completion.** A step that submits server-side work is finished\n   only when you have retrieved its terminal result. That includes step 2's\n   `request_plan_generation`: a `GATEWAY_TIMEOUT`/`504` envelope carrying a `recovery_get`\n   field means server-side processing may still be running, so poll `get_plan` (or that URL)\n   until a terminal result — never reissue the step. The same rule covers any ticket review\n   started before the commit handoff: reach a terminal result through a bounded wait tool\n   (`wait_for_ticket_review`, called again on each normal `state: \"pending\"` result) or the\n   artifact's retrieval tool. A shell `sleep`, an unbounded foreground poll, and re-issuing\n   the original generation request are all forbidden — the last abandons the run already in\n   flight and starts a second billable one. This applies to pre-handoff prerequisites only;\n   post-push CI belongs to the CI-monitoring step.\n\n4. After all steps complete, display a summary:\n   ```\n   ## Pipeline Complete\n\n   **Ticket**: <ticket_key>\n   **Steps executed**: N of M\n   **Status**: Success / Failed at step N\n\n   **Branch**: <selected branch>\n   **PR**: <pull request URL>\n   **last_commit_sha**: <latest pushed head SHA>\n   **Verification**: <each bounded command run, with its observed outcome>\n   **Correction commit**: <subject of the correction commit pushed after verification, or \"None\">\n   **Unresolved findings**: <findings reported but not fixed, or \"None\">\n   ```\n\n   Rules for this summary:\n\n   - **`last_commit_sha` is the latest *pushed* head**, not the commit that opened\n     the PR. The pipeline opens the pull request before running its bounded\n     verification, so a correction pushed afterwards moves the head — report the\n     head as it stands after the final push.\n   - **`Status` describes pipeline execution, not a merge verdict.** \"Success\" means\n     the recipe's steps ran to completion; it does not mean CI passed, that the code\n     review approved the change, or that the ticket is mergeable.\n   - **Never self-declare a gate outcome.** Do not label the summary with claims such\n     as \"CI passed\", \"checks green\", \"review approved\", or \"gate met\". Report the\n     exact check and review states you observed instead — the `ci` and `code_review`\n     gates are authoritative and the reconciler observes them independently.\n\n---\n\n# Worker scope discipline and clean exit (Conductor auto mode)\n\nThese rules apply only when you were launched under the Conductor in auto mode (`--auto`); a normal interactive `/implement-ticket` run is unaffected.\n\n## Declared file-scope discipline (N-2)\n\nIf your environment carries `BAPI_CONDUCTOR_DECLARED_TOUCHED_FILES_JSON` (a JSON array of the ticket's declared touched files), stay within that declared file boundary: do **not** create, modify, or delete files outside the declared set. Those out-of-scope files typically belong to a sibling ticket, and editing them risks a merge conflict or a silent clobber of the sibling's merged work. The pre-PR file-scope guard (run at the PR-creation step) will warn about any out-of-scope diff — treat that warning as a signal to re-check your scope, not as a blocker. When the variable is absent, empty, or invalid there is no declared boundary and this rule does not apply.\n\n## Clean session exit (D2)\n\nAfter the final pipeline step completes, cleanly end your worker session (for example by issuing the `/exit` command) so the worktree is released and no idle process lingers — but **only when no follow-up remains that you still own**. Do **not** exit while any of the following is true:\n\n- there are unresolved CI failures you are still correcting (the post-PR CI-correction loop in the CI-monitoring step still owns work),\n- review changes were requested and you have not yet addressed them,\n- there is a merge conflict on your PR that you still own,\n- you have unpushed local commits.\n\nExit only after your final branch state is pushed, the done-gate / CI-monitoring workflow required by the recipe has completed, and no CI/review follow-up remains. A clean `SessionEnd` is both the correct terminal lifecycle signal and the point at which the worker should exit.\n",
-    "install-bridge.md": "Bootstrap Bridge API project configuration from the local codebase via the easy-install manifest.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nThis command takes no arguments; any `$ARGUMENTS` text is ignored.\n\nCommand contract version: **8**. The install manifest returns a `command_contract_version`; Stage 2\ncompares it to this number to detect a stale scaffolded copy of this command.\n\nThis command has three visibly labeled outcomes, chosen by the project's state (Stage 2 decides them\nin a fixed order from the manifest), never by the caller's role:\n\n- **JOIN MODE** (`configured == true`): the project is already set up, so this run proposes and\n  applies ZERO configuration changes for ANY caller. A new teammate — including a non-admin \"member\"\n  key — gets a graceful welcome and the concise capability report instead of an error. An eligible\n  b2b admin is additionally offered the teammate-invite stage (Stage 11).\n- **GREENFIELD REPORT-ONLY** (the manifest's `greenfield_setup_path` field is set): this project\n  already completed the greenfield setup path, so this run asks nothing and writes nothing — it\n  reports and stops. See Stage 2.\n- **CLASSIFY** (neither of the above): Stage 2b decides whether this is a greenfield or an\n  established repository. Established runs the original, unchanged derive → approve → apply → report\n  flow (Stages 3 → 4 → 5 → 6 → 7 → 8 → 9 → 10). Greenfield hands off to the packaged\n  `greenfield-setup` recipe in this same session and then resumes reporting.\n\nThe stale-command rule and the server-authoritative `next_step` behavior are unchanged by all three:\na manifest version higher than this file's still warns, and Stage 8 still renders only what the\nserver put in `next_step.command`.\n\nThis command performs a one-time \"easy install\" of Bridge API project configuration. It reads the\nserver-owned install manifest once, derives values for the **unset** bootstrap fields from the local\ncodebase, asks for human approval of the project description, applies everything in a single atomic\ncall, and then presents a **concise capability report** derived from a fresh read-after-write manifest\nread. The server owns all skip-if-set, conflict, and confirmation semantics — this command never makes\nits own skip-if-set decisions — and the server owns the complete tool catalog and the bounded concise\nprojection over it, their grouping and ordering, and every gate and dependency relationship; this\ncommand formats the server's contract and never recomputes it from prose. Indexing is never a decision\nthis command makes or asks about: it starts automatically, gated entirely by server-side readiness (see\nStage 8).\n\nThis command is portable across Claude Code, Cursor, and Copilot: it only drives the Bridge API MCP\ntools (`ping`, `get_install_manifest`, `apply_install_manifest`, and — in the gated Stage 11\nonly — `invite_member`).\n\nIf any stage fails, stop immediately and report which stage failed and why.\n\n**Install-spawn context.** When the spawning instruction explicitly says this `/install-bridge` run is\nin the \"install-spawn context\" (it was launched by the `install-bridge` CLI's fresh agent session),\nStage 8, Stage 9, and Stage 10 are SKIPPED and the single closing interaction is the concise capability\nreport plus whatever conditional next step the manifest provided — the spawn prompt owns that same\n`next_step.command` rendering there, under the same non-empty rule. When you invoke\n`/install-bridge` directly (manual invocation), Stages 8, 9, and 10 run normally. Stage 11 is NOT part\nof that install-spawn skip set — it is independently gated (admin + b2b + interactive) and best-effort,\nso it may still run in the install-spawn context for an eligible admin.\n\nStage 6 is likewise NOT skipped in the install-spawn context — it still persists the routing\ncredential — but its SUCCESS output is silent there, because the `install-bridge` CLI has already told\nthe user where the credential landed and a second sentence is a duplicate. Silence covers every place\nthat success would otherwise be restated, including the Return contract, and it covers success ONLY: a\nStage 6 failure is still reported loudly and in full. JOIN MODE and manual invocation are outside this\nrule and report Stage 6 success exactly as they do today.\n\n## Packaged CLI launcher (`BAPI_MCP_CLI`)\n\nResolve the packaged-CLI launcher **once**, before the first packaged-CLI command this stage set names, and reuse that one resolved value throughout. Call it `<launcher>`.\n\n- Read the `BAPI_MCP_CLI` environment variable.\n- **Unset, empty, or whitespace-only** — `<launcher>` is exactly `npx -y @bridge_gpt/mcp-server`. This is the default, and the resulting shell command is byte-identical to what it was before this override existed.\n- **Otherwise** — `<launcher>` is that value, used verbatim as the command prefix. It names a local launcher, such as `node /absolute/path/to/mcp_server/build/index.js`. Use it for local pilots and pre-publish verification.\n\nWhen the override is set:\n\n- Apply this command's mandatory single-quote escaping rule (`'` → `'\\''`, then wrap the whole value in single quotes) before interpolating `<launcher>` into a Bash command string. Never expand it unquoted.\n- Keep every dynamic argument — ticket keys, branch names, base branches, file paths — independently quoted. Never concatenate an argument into the launcher value.\n- Never put a credential, an API key, or an environment assignment carrying one into the launcher value, an example, or a dry-run preview.\n\nThe packaged-CLI commands this command *prints for the user to run* (`--init`, `credentials migrate-agent-config`, `doctor`) are written below in their **unset** resolution — the literal `npx -y @bridge_gpt/mcp-server` — because that is the default every operator gets. When `BAPI_MCP_CLI` is set in this environment, print those remediation commands with the resolved `<launcher>` in place of that prefix so the advice matches the operator's actual setup. A stale local build is exactly as misleading as a stale npm publish: rebuild with `cd mcp_server && npm run build` before relying on the override.\n\n## Stage 1 — Admin preflight (defer the permission decision until the manifest is read)\n\n1. Call the `ping` MCP tool (no parameters) and read `role` and `customer_type` from its first\n   (JSON) content item. Retain both — later stages branch on them (Stage 2's mode decision uses\n   `role`; Stage 11's invite gate uses `role` and `customer_type`). Both are best-effort and\n   nullable: a `null` `role` means \"not determined\", which covers a legacy shared key as well as an\n   unconfigured install-join context, and it never fails the ping.\n2. Classify the caller, but do NOT stop here — the member permission decision is DEFERRED until Stage 2\n   has read the manifest and determined whether the project is already `configured`. A member must be\n   allowed to continue at least far enough to read the manifest, because a member CAN join an\n   already-configured project even though a member cannot configure a fresh one:\n   - If `role` is `\"admin\"`, or `role` is `null` (not determined — e.g. a legacy shared key): the\n     caller is configuration-capable (it may run the fresh-configuration flow when the project is\n     unconfigured).\n   - Otherwise (an explicit non-admin role such as `\"member\"`): the caller is join-only. It may proceed\n     into JOIN MODE for a configured project, but must be refused if Stage 2 proves the project is not\n     yet configured (it cannot configure a fresh repo).\n3. Preserve this exact refusal text for later use — it is emitted in Stage 2 ONLY when a non-admin\n   member reaches a `configured == false` project:\n   ```\n   Admin role required to apply install configuration. Ask a project admin to run /install-bridge,\n   or use an admin API key.\n   ```\n\n## Stage 2 — Read the manifest (once, pre-apply)\n\n1. Call the `get_install_manifest` MCP tool exactly once here.\n2. Keep the returned `snapshot_token` verbatim — you must pass the SAME token to the apply call in\n   Stage 5. Do not call `get_install_manifest` again before applying; a fresh read would invalidate\n   the token you are about to use. (Stage 7 performs a SEPARATE, later read-after-write manifest read\n   for the current capability status — that is deliberate and does not reuse this token.)\n3. The manifest contains ordered `groups` of bootstrap fields (each with `field_name`, `is_set`,\n   `current_value`, `guidance`, `examples`, `validation_summary`, `requires_confirmation`), a list of\n   `deferred_fields`, a `next_step`, `done_criteria`, a `command_contract_version`, and the\n   `snapshot_token`. It also carries the additive capability fields (`integrations`, `configured`,\n   `learned`, `indexed`, `tool_capabilities`, `workflows`, `concise_tool_capabilities`, `locked_tools`,\n   `unlocked_tools`) — but ignore those here; the accurate capability status is the post-apply read in\n   Stage 7. `tool_capabilities` is the COMPLETE catalog-backed report field (one entry per registered\n   MCP tool, grouped and ordered by the server); `workflows` is the separate curated collection of\n   non-MCP workflows (slash commands and packaged CLI subcommands, which are not registered MCP tools\n   and never appear in `tool_capabilities`); `concise_tool_capabilities` is the ADDITIVE projection over\n   BOTH of those that Stage 7 actually renders — a server-ordered array of exactly two availability\n   sections, each `{id, name, tools}` holding mixed capability items (see Stage 7); `locked_tools` /\n   `unlocked_tools` are LEGACY compatibility data covering only the VCS/index policy cases and are NOT\n   the tool inventory.\n4. Compare the manifest's `command_contract_version` to this command's contract version (8, stated at\n   the top of this file). If the manifest's version is HIGHER, this scaffolded command copy is stale:\n   warn the user to refresh it by running `npx -y @bridge_gpt/mcp-server --init` in the project root,\n   then proceed conservatively — wherever the manifest's `guidance`, `next_step`, or `done_criteria`\n   differ from these instructions, follow the manifest.\n5. **Decide the outcome from the manifest, in this exact order.** Evaluate these three branches in\n   sequence and take the FIRST one that applies. The order is the contract: a configured project is\n   never classified, and a completed greenfield project is never classified or re-onboarded.\n   1. `configured == true` → **JOIN MODE** (below).\n   2. the manifest's `greenfield_setup_path` field reports `is_set == true` → **GREENFIELD\n      REPORT-ONLY** (below). Read that field from the ordinary `groups[].fields[]` list, the same way\n      you read any other bootstrap field — there is no top-level copy of it. This is the completion\n      marker for the greenfield path: the greenfield recipe writes it exactly once and the server\n      treats it as skip-if-set, so a set value means \"this project already finished greenfield\n      onboarding\".\n   3. neither of the above → **Stage 2b — Classify**.\n\n   The branches:\n   - **`configured == true` → JOIN MODE, for EVERY caller** (admin, legacy, or member). The project is\n     already set up, so this run makes ZERO configuration changes. Emit a concise welcome — for\n     example: \"This Bridge project is already configured. You're joining it as a new teammate; no\n     configuration changes will be proposed or applied.\" Then SKIP Stages 3, 4, and 5 entirely (no\n     field derivation, no project-description approval, no `apply_install_manifest` call, no\n     `config_field` writes), run Stage 6 (persist the routing credential), and render the JOIN MODE\n     branch of Stage 7 (the concise capability report, drawn directly from THIS Stage-2 manifest — no\n     read-after-write). Then EVALUATE Stage 8 conditionally — it renders only when THIS Stage-2\n     manifest's `next_step.command` is a non-empty string, and is silent otherwise — and SKIP Stages 9\n     and 10. Then continue to the independently gated Stage 11.\n     The server has already encoded recommendation eligibility (learned state AND whether this caller\n     can run the command) in `next_step.command`, so do NOT reproduce a role gate or a learned-state\n     gate here: a joining member simply receives an empty command and sees nothing.\n   - **`greenfield_setup_path.is_set == true` → GREENFIELD REPORT-ONLY, for EVERY caller.** This\n     project already completed the greenfield setup path, so this run performs NO classification, NO\n     field derivation, NO interview, NO standards generation, NO `config_field` write, and NO\n     `apply_install_manifest` call. Say so plainly — for example: \"This project already completed\n     greenfield setup. Nothing to configure; here is where it stands.\" Then SKIP Stages 2b, 3, 4, and\n     5 entirely, run Stage 6 (persist the routing credential), and render the GREENFIELD REPORT-ONLY\n     branch of Stage 7. That report is sourced from THIS Stage-2 manifest and there is deliberately no\n     read-after-write call: this branch performs no setup write, so a second read would return the\n     same bytes at the cost of a round trip. Then EVALUATE Stage 8 conditionally against THIS same\n     Stage-2 manifest's `next_step` — the server decides whether to name `connect-github` or nothing —\n     and SKIP Stages 9 and 10. Then continue to the independently gated Stage 11.\n     In the install-spawn context the ordinary Stage 8–10 skip set still applies unchanged, and Stage\n     11 remains independently gated there as it is on every other path.\n     This branch is reached for a member key too: like JOIN MODE it writes nothing, so the \"Admin role\n     required\" refusal does not apply to it.\n   - **`configured == false` and `greenfield_setup_path` is NOT set → apply the deferred Stage 1 role\n     decision:**\n     - `source == \"legacy\"` or `role == \"admin\"`: continue to **Stage 2b — Classify**, which decides\n       between the established flow (Stages 3 → 4 → 5 → 6 → 7 → 8 → 9 → 10, exactly as written and\n       unchanged) and the greenfield handoff.\n     - a non-admin `user_access` \"member\" key: stop immediately and display the exact refusal text\n       preserved in Stage 1 (\"Admin role required to apply install configuration…\"). Do not classify,\n       derive, apply, or persist anything.\n   - **`configured` absent / indeterminate (neither `true` nor `false`) → do NOT treat it as `false`.**\n     `configured` comes from a best-effort capability enrichment that can silently omit the key on a\n     transient server-side probe failure, so a missing value is \"unknown\", not \"unconfigured\". Re-read\n     the manifest ONCE (a fresh `get_install_manifest` call) to try to resolve it, and re-evaluate the\n     ordered branches above against that refreshed response — it is the authoritative manifest from\n     then on, including its `snapshot_token`, its `greenfield_setup_path` field, and its `next_step`.\n     If `configured` is STILL absent:\n     - `source == \"legacy\"` or `role == \"admin\"`: proceed, but NOT silently — first tell the user that\n       configuration status could not be confirmed and that the run will continue anyway (the server\n       owns skip-if-set, so an apply against an already-configured repo is a safe no-op). Then resume\n       the ordered evaluation at branch 2: a set `greenfield_setup_path` still means GREENFIELD\n       REPORT-ONLY, and only an unset one reaches Stage 2b.\n     - a non-admin `user_access` \"member\" key: take the JOIN-MODE-safe path — render the welcome and the\n       Stage 7 capability report (no config writes, no Stage 9/10 offers), evaluate Stage 8 under its\n       normal non-empty-command rule, and note that configuration status could not be confirmed. Do NOT\n       emit the hard \"Admin role required\" STOP: that refusal is reserved for a\n       *definitive* `configured == false`, because treating an unknown state as unconfigured would\n       re-introduce the very member hard-refusal this flow removes.\n\n## Stage 2b — Classify the repository (greenfield or established)\n\n**Run this stage ONLY when neither prior Stage 2 branch applied** — that is, the project is not\n`configured`, its `greenfield_setup_path` is not set, and the Stage 1 permission rules allow this\ncaller to configure the project (`source == \"legacy\"` or `role == \"admin\"`). JOIN MODE and GREENFIELD\nREPORT-ONLY never reach this stage, and a non-admin member was already refused above.\n\nThis stage decides ONE thing: is this a brand-new project that has no code to learn from yet\n(**greenfield**), or a real codebase (**established**)? It writes nothing.\n\n1. **Gather evidence from the local checkout**, excluding Bridge's own files from every count and from\n   the \"is the tree empty\" judgement: `.mcp.json`, `.cursor/mcp.json`, `.claude/commands/`,\n   `.cursor/commands/`, `.bridge/`, and the resolved `{docs_dir}` tree. Those exist because Bridge was\n   installed, so counting them would make every Bridge-installed empty project look established.\n   - **Scaffold evidence** (points toward greenfield): package manifests (`package.json`,\n     `pyproject.toml`, `go.mod`, `Gemfile`, `*.csproj`), lockfiles, a README, generated or template\n     configuration, an empty or near-empty `src/`, and a framework template's untouched entry point.\n   - **Established evidence** (points toward established): substantive implementation files (real\n     modules with real logic, not a generator's placeholder), meaningful tests, and a non-trivial local\n     `git log`.\n2. **Established evidence dominates.** Any credible established evidence decides the answer, no matter\n   how much scaffold evidence sits beside it — a real codebase that also has a `package.json` and a\n   README is established. Scaffold evidence only decides the outcome when there is no credible\n   established evidence at all.\n3. **Missing Git metadata is neutral.** No `.git` directory, no commits, or no remote is neither\n   evidence nor a problem: a brand-new project legitimately has none of them, and so does a directory\n   whose history lives elsewhere. Do not report it as an error or a warning, and do not let it push the\n   decision on its own.\n4. **Print exactly one primary statement**, in this exact form:\n   ```\n   Classification: Greenfield — reply \"established\" to correct, or continue.\n   ```\n   or, for the other result:\n   ```\n   Classification: Established — reply \"greenfield\" to correct, or continue.\n   ```\n   Keep that line and its correction token the primary hierarchy. A compact scaffold-versus-established\n   evidence summary may follow it — one or two short lines — but only when it genuinely explains the\n   decision; never a file listing, a count table, or a diagnostic dump.\n5. **Accept a one-token correction for exactly one response turn.** Accept ONLY the trimmed, lowercased\n   opposite token (`established` when you said Greenfield, `greenfield` when you said Established) and\n   switch to it. Anything else — an unrelated answer, a blank response, an unavailable response,\n   silence — continues with the original classification. Do not re-ask, do not re-explain, and do not\n   enter a correction loop.\n6. **Never wait in a headless session.** When no interactive response can be obtained, do not pause for\n   the correction turn at all, and classify conservatively: choose **greenfield** ONLY for an empty or\n   clearly near-empty tree with no substantive implementation files, no meaningful tests, and no\n   non-trivial history. Everything else is established. A wrong \"established\" costs a skippable\n   derivation pass; a wrong \"greenfield\" would onboard a real codebase as if it had no code.\n7. **Established → run the existing flow unchanged.** Continue to Stage 3 and proceed exactly as\n   written (Stages 3 → 4 → 5 → 6 → 7 → 8 → 9 → 10). This stage changes nothing about derivation,\n   approval, or apply behavior for an established repository.\n8. **Greenfield → hand off to the packaged `greenfield-setup` recipe, in this same session.**\n   1. Call the `get_pipeline_recipe` MCP tool with\n      `{ \"pipeline\": \"greenfield-setup\", \"variables\": { \"repo_name\": \"<this repository's configured\n      name>\" } }`. Pass `repo_name` explicitly — the server auto-supplies `docs_dir` but not\n      `repo_name`. Do not pass `auto_approve`: greenfield onboarding asks the human real questions.\n   2. Follow the returned `agent_instructions` and execute EVERY returned step, in order, to\n      completion, here in this same agent session — exactly as `/learn-repository` runs its recipe.\n      Do not spawn a second agent for it, do not summarize the recipe instead of running it, and do\n      not return to Stage 6 partway through.\n   3. The recipe owns the greenfield apply: its final task makes the one `apply_install_manifest` call\n      (with the Stage-2 snapshot token) and then performs its own read-after-write\n      `get_install_manifest`. **Require that read-after-write manifest as the recipe's final result\n      before resuming.** If the recipe cannot provide it, report that the greenfield setup did not\n      complete and stop — do not fabricate a report from the stale Stage-2 manifest, and do not\n      re-apply.\n   4. Then resume at Stage 6 (persist the routing credential), render the greenfield-completion branch\n      of Stage 7 from that read-after-write manifest, evaluate Stage 8 against that same manifest's\n      `next_step`, SKIP Stages 9 and 10, and continue to the independently gated Stage 11.\n\n## Stage 3 — Derive values for UNSET bootstrap fields only\n\n**Skip this entire stage in JOIN MODE** (Stage 2 selected JOIN MODE because the manifest reported\n`configured == true`). JOIN MODE derives nothing — it proposes and applies zero configuration for every\ncaller. Run this stage only on the `configured == false` fresh-configuration path.\n\n1. Consider ONLY fields whose `is_set` is `false`. Never re-derive or overwrite a field that is\n   already set — the server will also skip already-set fields, but you should not even propose them.\n2. For each unset field, derive a value from the local codebase following that field's `guidance`\n   (e.g. infer `working_in` from frameworks/manifests, `version_control_system` from the git remote,\n   testing stacks from test config and dependencies, directory fields from the repo layout).\n3. Omit any field you are uncertain about or that does not apply — leaving it unset is always\n   acceptable. Do NOT guess.\n4. Specifically, do NOT invent a `version` when the platform has no meaningful version concept; leave\n   it unset.\n5. Before deriving `version_control_system` or `base_branch`, confirm the git repository root is THIS\n   project's root: `git rev-parse --show-toplevel` must resolve to the project directory itself. If\n   the git root is an ancestor directory (the project is nested inside another repository), OMIT both\n   fields and note why in the final report — deriving them would describe the parent repository, not\n   this project.\n6. If you find real evidence for TWO OR MORE plausible values for a field (e.g. configs for multiple\n   test frameworks, multiple frameworks in a monorepo), treat the value as ambiguous and omit it\n   rather than picking one.\n7. Do NOT derive `selected_mcp_slugs` or `greenfield_setup_path`, even though both appear as unset\n   bootstrap-eligible fields in the manifest. MCP validation manual selection is deferred to\n   `/learn-repository`, which derives and confirms it with the codebase already researched. Install\n   neither proposes nor applies this field. `greenfield_setup_path` is not derivable at all: it is the\n   greenfield path's own completion marker and is written ONLY by the `greenfield-setup` recipe's final\n   apply, which records whether the human took the interview or the defaults. Deriving it here would\n   mark an established repository as greenfield-complete and permanently suppress its\n   `/learn-repository` recommendation.\n\n## Stage 4 — Human approval for confirmation-requiring fields\n\n**Skip this entire stage in JOIN MODE** — there is nothing to derive, so there is nothing to approve.\nRun it only on the `configured == false` fresh-configuration path.\n\n`project_description` is the ONLY confirmation-requiring field install proposes. It carries\n`requires_confirmation: true` in the manifest, so it is never applied on derivation alone — it needs\nexplicit human approval. (`selected_mcp_slugs` also requires confirmation, but install does not\nderive it at all; `/learn-repository` asks for it. See Stage 3 step 7.)\n\n1. If `project_description` is unset and you can draft a concise description from the codebase,\n   present the proposed `project_description` text to the human and ask for explicit approval before\n   including it.\n2. Include `project_description` in the apply payload ONLY as\n   `{ \"value\": <approved value>, \"confirmed\": true }`, and only after the human approves it. If the\n   human does not approve it, omit the field entirely.\n3. If no human response can be obtained (e.g. a non-interactive / headless session), do NOT stall and\n   do NOT abandon the install: proceed to Stage 5 with `project_description` omitted, and report it\n   as \"pending human input\" in the final summary. The other derived fields must still be applied — an\n   unapproved description never blocks them.\n\n## Stage 5 — Apply (one call)\n\n**Skip this entire stage in JOIN MODE** — JOIN MODE makes NO `apply_install_manifest` call and writes\nzero fields for every caller. **Skip it in GREENFIELD REPORT-ONLY** for the same reason, and **skip it\nafter a greenfield handoff**: the `greenfield-setup` recipe owns the single final\n`apply_install_manifest` call for that path, made with the Stage-2 snapshot token. Running this stage\nafterwards would be a second apply with a token the recipe already spent. Run it only on the\nestablished fresh-configuration path Stage 2b routed to Stage 3.\n\n1. Make exactly ONE call to the `apply_install_manifest` MCP tool, passing:\n   - `snapshot_token`: the exact token from Stage 2.\n   - `fields`: a map of the derived fields. Scalar fields may be passed directly\n     (e.g. `\"base_branch\": \"main\"`); an approved `project_description` must use the\n     `{ \"value\": ..., \"confirmed\": true }` object form from Stage 4. Never include\n     `selected_mcp_slugs` in this payload — install does not derive or apply it (Stage 3 step 7).\n2. Do not implement any client-side skip-if-set, conflict, or confirmation logic — the server owns\n   all of that and returns the outcome buckets.\n3. The apply is partial-tolerant: fields that fail server-side validation are returned in the\n   `rejected` bucket while the valid fields still commit. A rejected field is therefore reported,\n   not fatal — do NOT retry the whole apply just because one field was rejected.\n\n## Stage 6 — Persist the routing credential\n\n**This stage runs on EVERY path** — JOIN MODE persists the routing credential too, so a joining\nteammate's shell-spawned CLI features (`start-tickets`) can resolve the key, and so do GREENFIELD\nREPORT-ONLY and a completed greenfield handoff, which reach this stage the same way. Its secret\nhandling, fail-open behavior, and install-spawn success-silence rule below are identical on all of\nthem — none of the greenfield paths changes any of it.\n\nThe Bash-spawned `start-tickets` CLI runs in a DIFFERENT runtime surface than this MCP server, so a\nkey that lives only in `.mcp.json` / `.cursor/mcp.json` is invisible to it and difficulty→model\nrouting silently degrades. This final stage persists the already-validated key into the user-scoped\ncredential store (`~/.config/bridge/credentials.json`) so shell-spawned CLI commands can resolve it.\n\n1. Call the `persist_routing_credential` MCP tool with `repo_name` set to the repository name for\n   this install (the configured `BAPI_REPO_NAME`). Pass `repo_name` ONLY — this tool resolves the\n   API key inside the MCP server process. NEVER pass or display `BAPI_API_KEY` (nor `api_key`,\n   `apiKey`, `secret`, or `token`); the key value must not appear in the tool call or in any output.\n2. On success, tell the user the credential was stored, echoing the returned `target` and `path`,\n   e.g. `Stored routing credential for bapi:<repo_name> at <path>`. EXCEPTION — in the install-spawn\n   context (and outside JOIN MODE), persist silently: still call the tool, but print nothing on\n   success and do not restate the `target` or `path` in any later summary or in the Return contract.\n   The CLI already reported the credential destination, so this sentence would be a duplicate. This\n   exception is success-only; step 3's failure reporting is unchanged in every context.\n3. On failure, do NOT block the install — show the rest of the install summary anyway, then tell the\n   user the routing credential could not be stored. They can rerun `/install-bridge` after setting\n   `BAPI_API_KEY` in the environment, or run\n   `npx -y @bridge_gpt/mcp-server credentials migrate-agent-config --write-credentials` to migrate an\n   existing key. Until then, `start-tickets` model routing fails open (assumes a hard ticket and\n   defaults to premium/Opus when available) — this is the MOST EXPENSIVE model tier, so unexplained\n   cost is the symptom of leaving this unfixed. Recommend verifying credential resolution afterwards\n   with `npx -y @bridge_gpt/mcp-server doctor`.\n\n## Stage 7 — Summarize the outcome, then present the concise capability report\n\n**JOIN MODE branch (`configured == true`).** Do NOT print an applied count and do NOT fabricate apply\nbuckets — no apply happened. Instead state plainly that the project was already configured and that\nzero configuration changes were proposed or applied (the welcome from Stage 2). Then render the concise\ncapability report described in \"### Read-after-write\" below, with ONE difference: source it directly\nfrom the `concise_tool_capabilities` field of the Stage-2 manifest you already read — do NOT perform a\nread-after-write `get_install_manifest` call, because no write occurred and there is nothing to\nrefresh. Apply the same server-authority rendering rules (server section order, verbatim item copy,\nmalformed/missing fallback) verbatim. ALSO retain that Stage-2 manifest's `next_step` object: it is the\nauthoritative next-step source for this path, because the Stage-2 read is JOIN MODE's only manifest\nread. After the report, continue to Stage 8, which renders that `next_step` conditionally; an eligible\nadmin then proceeds to Stage 11. The rest of this stage (the applied-count line and six-bucket summary)\napplies ONLY to the `configured == false` fresh-configuration path.\n\n**Greenfield-completion branch (Stage 2b classified greenfield and the recipe ran).** Do NOT print an\napplied count and do NOT reproduce the six apply buckets: this command made no apply call — the recipe\ndid. State plainly that greenfield setup completed, and say which path the human took (the interview\nor the defaults) and what it produced: the standards fields written, and the deferred-VCS guide when\none was written. Then render the concise capability report described in \"### Read-after-write\" below,\nsourced EXCLUSIVELY from the read-after-write manifest the recipe returned as its final result — never\nfrom the stale Stage-2 manifest, and never from a fresh read of your own. Retain that same manifest's\n`next_step` object: it is the authoritative next-step source for this path, and Stage 8 renders it.\nNever mention or recommend `/learn-repository` on this path; the server has already decided what the\nnext step is, and on a greenfield project it is never that.\n\n**GREENFIELD REPORT-ONLY branch (`greenfield_setup_path` was already set).** Behave exactly like JOIN\nMODE's no-apply reporting: no applied count, no fabricated apply buckets, no report of onboarding work\nthat did not happen this run. State that this project already completed greenfield setup and that zero\nconfiguration changes were proposed or applied. Then render the concise capability report sourced\nEXCLUSIVELY from the Stage-2 manifest already read — no read-after-write call, because nothing was\nwritten — and retain that manifest's `next_step` for Stage 8.\n\nBoth greenfield branches use the SAME rendering rules as every other branch: the server's section\norder, verbatim item copy, the malformed/missing fallback line, the empty-section handling, the\nnarrow-terminal wrapping rules, and the closing documentation footer. There is one report format, and\nthese branches only change where the manifest came from.\n\n**Fresh-configuration branch (established path, `configured == false`).**\nFirst, begin with an explicit applied count: \"Applied N of M derivable fields\" (M = the unset\nbootstrap fields you derived values for). An install that applied 0 fields must say so loudly — the\ninstall is NOT complete until the apply call reports applied fields. This applied-count line and the\nsix-bucket summary below remain the PRIMARY install result — the capability report that follows is a\nsecondary close, not a replacement for it.\n\nThen summarize the six buckets from the apply response for the user:\n- `applied` — fields written this run.\n- `skipped` — fields already set (left untouched).\n- `conflict` — fields that changed since the manifest was read (re-run /install-bridge to retry).\n- `needs_confirmation` — fields awaiting explicit confirmation (e.g. project_description).\n- `rejected` — fields that failed validation or were not bootstrap-eligible (the other fields still\n  applied; fix or drop the rejected values before any retry).\n- `deferred` — fields intentionally not auto-written by install (owned by /learn-repository or set\n  deliberately). `selected_mcp_slugs` belongs here: report it as deferred to `/learn-repository`,\n  never as approved, declined, or pending an install-time decision.\n\n### Read-after-write: fetch the concise capability report\n\nAfter the bucket summary, call the `get_install_manifest` MCP tool ONCE MORE. This post-apply read\nreflects the configuration you just wrote (the Stage-2 read was pre-apply and is stale for this\npurpose). This read does not need the snapshot token. Use ONLY this post-write response for the\nreport below.\n\nRetain this post-apply response's `next_step` object as well. On the fresh-configuration path it — not\nthe pre-apply Stage-2 read — is the authoritative source for Stage 8's conditional next step. Neither\nbranch derives a command from `configured`, `learned`, or any role prose: the command is whatever the\nauthoritative manifest returned.\n\nThe response carries `concise_tool_capabilities` — an ADDITIVE projection over the complete\n`tool_capabilities` catalog and the curated `workflows` collection (the response still carries both\nunchanged; this stage simply does not render them). It is a server-ordered array of EXACTLY TWO\navailability sections, in this order:\n\n1. `{\"id\": \"available_now\", \"name\": \"Available now\", \"tools\": [...]}` — what works right now.\n2. `{\"id\": \"needs_setup\", \"name\": \"Needs setup\", \"tools\": [...]}` — what additional setup would unlock.\n\nEach entry in a section's `tools` array is a capability item:\n`{id, kind, display_name, description, how_to_use}`, where `kind` is `\"tool\"` or `\"workflow\"` and\n`how_to_use` may be `null`. Items in the `needs_setup` section carry ONE additional field,\n`availability_text` — the server's plain-language sentence naming what is missing.\n\nServer authority: the server computed this projection's section membership, availability decisions,\nretrieval-pair collapsing, and ordering. Never recompute, re-filter, re-count, re-sort, regroup, infer\navailability, pair a retrieval tool with a primary yourself, invent an item, or re-derive any of it from\n`tool_capabilities`, `workflows`, `docs/mcp-tool-integrations.md`, or any other documentation — render\nexactly what the server sent.\n\nIf the post-write response has no `concise_tool_capabilities` field at all, or it is present but\nmalformed (not the two-section `{id, name, tools}` shape described above), print exactly:\n`No capability status is available yet; configuration can still continue.` and skip the section below\nentirely — never fall back to rendering the complete `tool_capabilities` catalog, the `workflows`\ncollection, or a remembered/hallucinated capability list.\n\nOtherwise, render exactly one section, with this exact heading:\n\n**What Bridge can help with**\n\n- Render the two server sections in the order the server gave them — \"Available now\" first, then\n  \"Needs setup\". Use the server's `name` as each sub-heading. Do not reorder, merge, filter, or drop a\n  section, and do not add a third one.\n- Use semantic Markdown (a short bold or `###` sub-heading per section, one compact list item per\n  capability) with tight vertical spacing — no table, no fixed-width columns, no ASCII box. Long\n  descriptions, commands, and URLs must be free to wrap, so the report stays readable in a narrow\n  terminal.\n- For each item in **Available now**, render its `display_name`, then its `description`, then its\n  `how_to_use` when that value is a non-empty string (omit the invocation line entirely when it is\n  `null` or empty). Render NO availability caveat, asterisk, footnote, or status note in this section —\n  the server already decided these are available, including any tool that works with less codebase\n  context before indexing.\n- For each item in **Needs setup**, render the same three fields, then the server's `availability_text`\n  verbatim as emphasized setup guidance on its own line. Do not reword it, prefix it with a warning, or\n  substitute setup advice of your own.\n- Render a `kind: \"workflow\"` item EXACTLY like a `kind: \"tool\"` item. `kind` is contract metadata for\n  machine consumers, not a visual distinction: no icon, no label, no badge, no separate sub-list, no\n  different indentation, and no client branching of any sort on its value.\n- Keep an empty section VISIBLE rather than dropping it. When **Available now** is empty, render the\n  heading with one short neutral line such as \"Nothing is available yet.\" When **Needs setup** is empty,\n  render the heading with one short line stating that every highlighted capability is ready — for\n  example \"Everything highlighted here is ready to use.\"\n- Close the report with this quiet final line, after BOTH sections and regardless of whether either\n  `tools` array was empty:\n  See the [Bridge MCP server README](https://www.npmjs.com/package/@bridge_gpt/mcp-server) for the\n  complete tool documentation.\n\n## Stage 8 — Render the server's next step (only when there is one)\n\n**Skip this entire stage in the install-spawn context** (the spawning instruction identifies it); the\nspawn prompt owns the same conditional closing recommendation there. On direct manual\n`/install-bridge` invocation, run it normally.\n\n**This stage runs in JOIN MODE too** — it is no longer skipped wholesale for joining teammates. What\ngates it is the server's output, not the mode and not the caller's role.\n\nThis stage renders a decision the SERVER already made. The manifest's `next_step.command` encodes both\nwhether the repository still needs the deeper instruction-tier configuration AND whether this caller\ncan actually run the command; a caller who cannot gets an empty command. On a greenfield project the\nserver may name a `connect-github` command instead, or nothing at all. Never re-derive any of that\nfrom `configured`, `learned`, `greenfield_setup_path`, `role`, or `source`, never assemble a\n`connect-github` invocation locally, and never substitute a command you happen to know.\n\n1. Read `next_step` from the AUTHORITATIVE manifest for this path (Stage 7 names it: the post-apply\n   read-after-write response on the fresh-configuration path; the read-after-write manifest the\n   `greenfield-setup` recipe returned after a greenfield handoff; and the Stage-2 response in JOIN MODE\n   and in GREENFIELD REPORT-ONLY, neither of which performs a read-after-write).\n   Then branch on `next_step.command` alone:\n   - **`next_step.command` is a non-empty string** → render one short heading and the exact command\n     the server returned, formatted as Markdown inline code — for example: ``Next step: `<command>` ``\n     where `<command>` is that returned string, verbatim. Do not rename it, alias it, expand it into\n     a different invocation, or replace it with a locally known command. Do not run it yourself:\n     it remains the human's own explicit next invocation.\n   - **`next_step.command == \"\"`** (or the key is missing/not a string) → be entirely SILENT. Emit no\n     heading, no empty code span, no placeholder or fallback command, no separator or divider, no\n     \"nothing to do\" notice, and no generic success or completion claim. The stage simply produces\n     nothing and the run continues.\n2. Where you give supporting context for a rendered command, use the manifest's own\n   `next_step.post_install_indexing` and `next_step.post_install_scheduler` text as the authoritative\n   source rather than inventing a rationale for whatever command the server named. Keep that context\n   subordinate to the command itself — it is explanation, not a second call to action.\n3. Do NOT ask about repository indexing in any form. There is no consent question, no\n   `parse_repository` tool call, and no `/parse-repository` continuation here — indexing starts\n   automatically once the repository reaches full parse readiness (VCS credentials, the Pinecone\n   index, `working_in` / `project_description`, and SFCC prerequisites where applicable), via the\n   same readiness-gated funnel the GitHub connection-confirm endpoints and the scheduled sweep already\n   use. Do not claim indexing has already started — this command has no visibility into that funnel's\n   outcome.\n\n## Stage 9 — Offer CI follow-up configuration (only when CI is detected)\n\n**Skip this entire stage in the install-spawn context** (the spawning instruction identifies it), so\nthe concise report and the conditional next step the spawn prompt owns remain the sole closing interaction. On direct\nmanual `/install-bridge` invocation, run it normally.\n\n**Also skip this stage in JOIN MODE** (`configured == true`), in **GREENFIELD REPORT-ONLY**, and after\na **greenfield handoff**: none of those paths offers configuration follow-up; they stop after the Stage\n7 capability report and the conditional Stage 8 next step (an eligible admin continues to Stage 11).\n\nThis stage is separate from the single apply call: it writes at most ONE field via the\n`config_field` MCP tool. `ci_followup_config` is REPLACE-NOT-AUGMENT — leaving it NULL already means\nsafe poll-only defaults, so \"skip\" is always a correct answer.\n\n1. Gate: only offer this when Stage 3 found real CI evidence (e.g. `.github/workflows/*.yml`,\n   `bitbucket-pipelines.yml`). With no CI detected, skip this stage and note \"no CI detected — CI\n   follow-up not offered\" in the final report.\n2. Ask ONE question, referencing the discovered CI by name, with exactly these preset profiles\n   (default: skip):\n   - **observe** — poll CI results only, never attempt fixes:\n     `{\"strategy\": \"poll_only\", \"max_iterations\": 1, \"max_minutes\": 10, \"instructions\": \"\"}`\n   - **self-heal** — bounded fix-and-iterate loop on the automation's own PRs:\n     `{\"strategy\": \"fix_and_iterate\", \"max_iterations\": 3, \"max_minutes\": 45, \"instructions\": \"\"}`\n   - **skip** (default) — leave `ci_followup_config` unset (NULL = baseline poll-only behavior).\n3. Do NOT offer or compose a `custom` strategy during install — free-form CI follow-up instructions\n   are a deliberate later act (setup UI or a direct `config_field` update).\n4. When the human picks a profile, call the `config_field` MCP tool once (operation `\"update\"`,\n   `field_name: \"ci_followup_config\"`, `value`: the profile's JSON object). If the server rejects the\n   value, retry ONCE with a corrected payload; if it is rejected again, stop, show the proposed JSON\n   to the human, and leave the field unset.\n5. Never select a profile in a non-interactive session — skip and list it as a pending next step. An\n   unattended install must never enable fix-and-iterate behavior.\n\n## Stage 10 — Offer the speed-vs-quality repository preference\n\n**Skip this entire stage in the install-spawn context** (the spawning instruction identifies it), so\nthe concise report and the conditional next step the spawn prompt owns remain the sole closing interaction. On direct\nmanual `/install-bridge` invocation, run it normally.\n\n**Also skip this stage in JOIN MODE** (`configured == true`), in **GREENFIELD REPORT-ONLY**, and after\na **greenfield handoff**: none of those paths offers configuration follow-up; they stop after the Stage\n7 capability report and the conditional Stage 8 next step (an eligible admin continues to Stage 11).\n\nThis stage is separate from the single apply call: it writes at most ONE field via the `config_field`\nMCP tool, `field_name: \"speed_vs_quality\"`. The column defaults to `5` (max quality) for every\nrepository, so \"skip\" always leaves a safe, valid value in place.\n\n1. Never ask this question in a non-interactive session — skip and list it as a pending next step. An\n   unattended install must never block on or silently answer this preference.\n2. Read the current value first via the `config_field` MCP tool (operation `\"get\"`,\n   `field_name: \"speed_vs_quality\"`) so a reinstall can show the stored value — not always `5` — as\n   the displayed default rather than silently re-asking from scratch.\n3. Ask ONE question with exactly these five numbered presets (default: the value from step 2, or `5`\n   if this is the first install):\n   - **1** — Max speed\n   - **2** — Prefer speed\n   - **3** — Balanced\n   - **4** — Prefer quality\n   - **5** — Max quality (default)\n4. Persist ONLY on an explicit answer — including an explicitly accepted default — by calling the\n   `config_field` MCP tool once (operation `\"update\"`, `field_name: \"speed_vs_quality\"`,\n   `value`: the selected integer 1-5). Do not overwrite an existing value when the human gives no\n   answer at all (e.g. the session cannot obtain one) — leave the stored/default value untouched in\n   that case, distinct from an explicit accepted-default selection of `5`.\n5. After a successful persist, emit one additive structured log line (mirroring the codebase's\n   `logging.info(msg, extra={...})` convention for non-response-body observability signals) with\n   fields: `event=\"install.speed_vs_quality\"`, `repo_name`, `field_name=\"speed_vs_quality\"`,\n   `selected_preset` (the persisted integer), `outcome=\"persisted\"`. Never log this event before the\n   `config_field` tool call has confirmed the write.\n6. When skipped, emit the corresponding structured event with `outcome` set to one of\n   `\"skipped_non_interactive\"` (non-interactive session) or `\"skipped_no_answer\"` (interactive session,\n   no explicit answer obtained) — omit `selected_preset` and any prompt text from this event.\n7. If the `config_field` call is rejected, retry ONCE with a corrected payload; if it is rejected\n   again, stop, show the proposed value to the human, and leave the field at its current stored value.\n   Do not log a persistence-success event for a rejected or failed write.\n\n## Stage 11 — Invite teammates (gated: b2b admins only; best-effort)\n\nThis is a best-effort final stage that lets an eligible admin mint teammate keys after configuration or\nJOIN MODE. It is **independently gated** and is NOT part of the install-spawn skip set (Stages 8–10) —\nit may run in the install-spawn context for an eligible admin. The entire stage is **fail-open**: a\nprompt failure, a declined offer, a non-interactive context, a malformed tool response, or an\n`invite_member` failure must NEVER cause this command to report the (already-completed) install as\nfailed.\n\n1. **Eligibility gate (AND).** Offer this stage ONLY when BOTH hold, using the values retained in\n   Stage 1:\n   - `role == \"admin\"`, AND\n   - `customer_type == \"b2b\"`.\n   Otherwise skip the stage silently with NO prompt: a member, a legacy-source caller whose role is not\n   explicitly `\"admin\"`, and a b2c admin all skip. Reachable from EVERY Stage 2 outcome for an eligible\n   admin: the established fresh-admin close (after Stage 10), the JOIN MODE path (after the Stage 7\n   report), the GREENFIELD REPORT-ONLY path, and a completed greenfield handoff — including an admin\n   re-run against an already-configured or already-greenfielded b2b project.\n2. **Interactive surface required.** This stage needs a human response. If no interactive response can\n   be obtained (a non-TTY / headless / spawn context that cannot prompt), skip the stage silently\n   without changing the completed install result — do not stall.\n3. **Offer prompt.** Ask exactly: `Invite teammates to this project? (y/N)`. Treat a blank answer, `n`,\n   `no`, an unavailable response, or any prompt failure as a non-fatal decline — skip the rest of the\n   stage and report it as declined.\n4. **Collect invitees.** On an affirmative answer, collect one or more teammate email entries using\n   normal **echoed** input (email is PII, not a secret — never use a muted/hidden secret prompt).\n   Optionally collect a display name per entry.\n5. **Per-invite role.** Default each invitation to role `member`. Only set a specific request's role to\n   `admin` after an explicit per-invite opt-up for that entry; never opt up by default.\n6. **Mint.** For each invitee, call the `invite_member` MCP tool exactly once with `{email, name?,\n   role}`. Do NOT pass `repo_name` — the tool resolves the repository from the current session.\n7. **Show each key once.** After each successful call, display that response's plaintext `api_key`\n   exactly once, associated with its intended recipient, followed by the exact warning:\n   `Distribute securely; this key is shown once.` Do NOT repeat a minted key anywhere else — not in the\n   final summary, not in retry guidance, not in diagnostics, not in a later stage.\n8. **Per-invite failure isolation.** Treat each mint failure as local to that invitee: report a\n   sanitized failure for it, continue to any remaining invitees, and never change the already-completed\n   install outcome. Do not surface raw error text, headers, or the caller's key.\n\n## Return\n\nThe Return contract depends on which outcome Stage 2 selected.\n\n**Fresh-configuration return (established path, `configured == false` admin/legacy).**\nReport the admin check result, the \"Applied N of M\" count, the\napplied/skipped/conflict/needs_confirmation/rejected field names, any fields omitted for\nnested-repository or ambiguity reasons, the approval outcome for `project_description` — install's\nonly confirmation-requiring field — (approved / declined / pending human input), the fact that\n`selected_mcp_slugs` is deferred to `/learn-repository` rather than decided here, whether a\nstale-command warning was raised (manifest\n`command_contract_version` higher than this command's), whether the routing credential was persisted\n(the returned `target` and `path`, or the non-blocking failure remediation — but in the install-spawn\ncontext report neither on success, per Stage 6's silence exception; a failure is still reported), the \"What Bridge can help\nwith\" concise capability report from the post-apply read-after-write manifest read, the CI follow-up\noutcome (profile written / skipped / no CI detected / pending / skipped in install-spawn context), the\nspeed-vs-quality preference outcome (persisted with its preset / skipped_non_interactive /\nskipped_no_answer / pending / skipped in install-spawn context), and the Stage 8 outcome — the exact\nserver-provided next-step command when the manifest returned a non-empty one, otherwise the fact that\nno next step was recommended (never a substituted command of your own).\n\n**JOIN MODE return (`configured == true` path).**\nReport the caller's role, the \"you're joining an already-configured project\" welcome and the explicit\nno-change status (zero configuration proposed or applied — do NOT report an applied count or apply\nbuckets), whether the routing credential was persisted (the returned `target` and `path`, or the\nnon-blocking failure remediation), the \"What Bridge can help with\" concise capability report drawn\nfrom the Stage-2 manifest (no read-after-write), and the Stage 8 outcome — the exact server-provided\nnext-step command when that manifest returned a non-empty one, otherwise the fact that no next step\nwas recommended.\n\n**Greenfield-completion return (Stage 2b classified greenfield and the recipe ran).**\nReport the classification that was printed and whether the human corrected it, the setup path the\nrecipe recorded (`interview` or `defaults`), which standards fields were written and which were left\nalone, whether the deferred-VCS guide was written or the connect command was printed, whether the\nrouting credential was persisted (the returned `target` and `path`, or the non-blocking failure\nremediation — with Stage 6's install-spawn silence exception unchanged), the \"What Bridge can help\nwith\" concise capability report drawn from the recipe's read-after-write manifest, and the Stage 8\noutcome — the exact server-provided next-step command when that manifest returned a non-empty one,\notherwise the fact that no next step was recommended. Do NOT report the ordinary derivation buckets\n(applied / skipped / conflict / needs_confirmation / rejected / deferred): no Stage 5 apply happened on\nthis path, so there are no buckets to report, and do NOT report an applied count. Never mention or\nrecommend `/learn-repository`.\n\n**GREENFIELD REPORT-ONLY return (`greenfield_setup_path` already set).**\nReport that this project already completed greenfield setup and that this run made ZERO setup changes,\nasked ZERO onboarding questions, and applied nothing — no applied count, no apply buckets, no\nderivation buckets. Report whether the routing credential was persisted (same rule as above), the\n\"What Bridge can help with\" concise capability report drawn from the Stage-2 manifest (no\nread-after-write), and the Stage 8 outcome under its usual non-empty-command rule. Never mention or\nrecommend `/learn-repository` on this path either.\n\n**Teammate-invitation outcome (Stage 11, all outcomes).**\nReport the Stage 11 outcome as counts/statuses only — one of offered, declined, skipped (not eligible,\nor non-interactive context), partially completed, or completed, plus how many keys were minted. Never\nrepeat teammate email addresses or minted key values in this summary. Stage 11's eligibility gate and\nits one-time key-secrecy rule are identical on all three Stage 2 outcomes — JOIN MODE, GREENFIELD\nREPORT-ONLY, and CLASSIFY (established or greenfield) — and are unchanged by this contract version.\n",
+    "install-bridge.md": "Bootstrap Bridge API project configuration from the local codebase via the easy-install manifest.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\n**How to run this.** Run `/install-bridge` now, with no arguments — that is the entire invocation.\n\"Install-spawn context\" (described below) is a mode identified by the spawning instruction's own\nprose, not a command-line option — never turn it into `--context install-spawn` or any other\nargument. The Bridge MCP tools this command names (`ping`, `get_install_manifest`,\n`apply_install_manifest`, `invite_member`) exist only inside the currently running agent session —\nnever invoke them through Bash or by constructing a direct HTTP request. This command owns and\nautomatically drives its own internal stages; do not reproduce or hand-execute those stages\nyourself. Never place the API-key value in a Bash command, its arguments, printed output, a log\nline, an error message, or a file you write; verify connectivity with the `ping` MCP tool only —\nnever through an environment export, `echo`, or another shell command.\n\nThis command takes no arguments; any `$ARGUMENTS` text is ignored.\n\nCommand contract version: **9**. The install manifest returns a `command_contract_version`; Stage 2\ncompares it to this number to detect a stale scaffolded copy of this command.\n\nThis command has three visibly labeled outcomes, chosen by the project's state (Stage 2 decides them\nin a fixed order from the manifest), never by the caller's role:\n\n- **JOIN MODE** (`configured == true`): the project is already set up, so this run proposes and\n  applies ZERO configuration changes for ANY caller. A new teammate — including a non-admin \"member\"\n  key — gets a graceful welcome and the concise capability report instead of an error. An eligible\n  b2b admin is additionally offered the teammate-invite stage (Stage 11).\n- **GREENFIELD REPORT-ONLY** (the manifest's `greenfield_setup_path` field is set): this project\n  already completed the greenfield setup path, so this run asks nothing and writes nothing — it\n  reports and stops. See Stage 2.\n- **CLASSIFY** (neither of the above): Stage 2b decides whether this is a greenfield or an\n  established repository. Established runs the original, unchanged derive → approve → apply → report\n  flow (Stages 3 → 4 → 5 → 6 → 7 → 8 → 9 → 10). Greenfield hands off to the packaged\n  `greenfield-setup` recipe in this same session and then resumes reporting.\n\nThe stale-command rule and the server-authoritative `next_step` behavior are unchanged by all three:\na manifest version higher than this file's still warns, and Stage 8 still renders only what the\nserver put in `next_step.command`.\n\nThis command performs a one-time \"easy install\" of Bridge API project configuration. It reads the\nserver-owned install manifest once, derives values for the **unset** bootstrap fields from the local\ncodebase, asks for human approval of the project description, applies everything in a single atomic\ncall, and then presents a **concise capability report** derived from a fresh read-after-write manifest\nread. The server owns all skip-if-set, conflict, and confirmation semantics — this command never makes\nits own skip-if-set decisions — and the server owns the complete tool catalog and the bounded concise\nprojection over it, their grouping and ordering, and every gate and dependency relationship; this\ncommand formats the server's contract and never recomputes it from prose. Indexing is never a decision\nthis command makes or asks about: it starts automatically, gated entirely by server-side readiness (see\nStage 8).\n\nThis command is portable across Claude Code, Cursor, and Copilot: it only drives the Bridge API MCP\ntools (`ping`, `get_install_manifest`, `apply_install_manifest`, and — in the gated Stage 11\nonly — `invite_member`).\n\n**Plain-language narration (BAPI-1174).** Every major operation this command performs — reading the\nmanifest, deriving field values, applying the configuration, persisting the routing credential,\nrendering the capability report — gets a short, plain-language explanation of WHAT is happening and WHY\nit matters to the person running the command, in ordinary words rather than internal shorthand.\nUser-facing narration you author yourself must never name an internal stage number, this command's\ncontract version, a raw internal MCP tool identifier (`get_install_manifest`, `apply_install_manifest`,\n`invite_member`, …), or a manifest field key (`command_contract_version`, `next_step`, `configured`,\n…) — say what is actually happening instead (\"checking whether this project is already set up,\" not\n\"reading `configured` from the manifest\"). This restriction is scoped to narration YOU compose: it does\nnot touch this file's own internal procedure, which keeps its stage headings, contract-version checks,\nand tool/field names exactly as written — an agent following these instructions still reads them\nverbatim to decide what to do next; it just never repeats that internal vocabulary back to the person\nrunning the command. Two exceptions are not narration at all and stay exactly as specified wherever\nthis file calls for them: a user-invocable command name (`/install-bridge`, `/learn-repository`,\n`connect-github`, `start-tickets`, …) that the person needs to actually run, and any value a stage\nexplicitly marks as server-authored and rendered verbatim (the concise capability report, `human_setup`\ndisplay text, a `next_step.command` string) — those are never reworded, curated, or filtered by you\neither way.\n\nIf any stage fails, stop immediately and explain in plain language what could not be completed and why\n— describe the actual problem (a failed manifest read, a rejected field, a connectivity error) rather\nthan naming the internal stage label.\n\n**Install-spawn context.** When the spawning instruction explicitly says this `/install-bridge` run is\nin the \"install-spawn context\" (it was launched by the `install-bridge` CLI's fresh agent session),\nStage 8, Stage 9, and Stage 10 are SKIPPED and the single closing interaction is the concise capability\nreport plus whatever conditional next step the manifest provided — the spawn prompt owns that same\n`next_step.command` rendering there, under the same non-empty rule. When you invoke\n`/install-bridge` directly (manual invocation), Stages 8, 9, and 10 run normally. Stage 11 is NOT part\nof that install-spawn skip set — it is independently gated (admin + b2b + interactive) and best-effort,\nso it may still run in the install-spawn context for an eligible admin.\n\nStage 6 is likewise NOT skipped in the install-spawn context — it still persists the routing\ncredential — but its SUCCESS output is silent there, because the `install-bridge` CLI has already told\nthe user where the credential landed and a second sentence is a duplicate. Silence covers every place\nthat success would otherwise be restated, including the Return contract, and it covers success ONLY: a\nStage 6 failure is still reported loudly and in full. JOIN MODE and manual invocation are outside this\nrule and report Stage 6 success exactly as they do today.\n\n## Packaged CLI launcher (`BAPI_MCP_CLI`)\n\nResolve the packaged-CLI launcher **once**, before the first packaged-CLI command this stage set names, and reuse that one resolved value throughout. Call it `<launcher>`.\n\n- Read the `BAPI_MCP_CLI` environment variable.\n- **Unset, empty, or whitespace-only** — `<launcher>` is exactly `npx -y @bridge_gpt/mcp-server`. This is the default, and the resulting shell command is byte-identical to what it was before this override existed.\n- **Otherwise** — `<launcher>` is that value, used verbatim as the command prefix. It names a local launcher, such as `node /absolute/path/to/mcp_server/build/index.js`. Use it for local pilots and pre-publish verification.\n\nWhen the override is set:\n\n- Apply this command's mandatory single-quote escaping rule (`'` → `'\\''`, then wrap the whole value in single quotes) before interpolating `<launcher>` into a Bash command string. Never expand it unquoted.\n- Keep every dynamic argument — ticket keys, branch names, base branches, file paths — independently quoted. Never concatenate an argument into the launcher value.\n- Never put a credential, an API key, or an environment assignment carrying one into the launcher value, an example, or a dry-run preview.\n\nThe packaged-CLI commands this command *prints for the user to run* (`--init`, `credentials migrate-agent-config`, `doctor`) are written below in their **unset** resolution — the literal `npx -y @bridge_gpt/mcp-server` — because that is the default every operator gets. When `BAPI_MCP_CLI` is set in this environment, print those remediation commands with the resolved `<launcher>` in place of that prefix so the advice matches the operator's actual setup. A stale local build is exactly as misleading as a stale npm publish: rebuild with `cd mcp_server && npm run build` before relying on the override.\n\n## Stage 1 — Admin preflight (defer the permission decision until the manifest is read)\n\n1. Call the `ping` MCP tool (no parameters) and read `role` and `customer_type` from its first\n   (JSON) content item. Retain both — later stages branch on them (Stage 2's mode decision uses\n   `role`; Stage 11's invite gate uses `role` and `customer_type`). Both are best-effort and\n   nullable: a `null` `role` means \"not determined\", which covers a legacy shared key as well as an\n   unconfigured install-join context, and it never fails the ping.\n2. Classify the caller, but do NOT stop here — the member permission decision is DEFERRED until Stage 2\n   has read the manifest and determined whether the project is already `configured`. A member must be\n   allowed to continue at least far enough to read the manifest, because a member CAN join an\n   already-configured project even though a member cannot configure a fresh one:\n   - If `role` is `\"admin\"`, or `role` is `null` (not determined — e.g. a legacy shared key): the\n     caller is configuration-capable (it may run the fresh-configuration flow when the project is\n     unconfigured).\n   - Otherwise (an explicit non-admin role such as `\"member\"`): the caller is join-only. It may proceed\n     into JOIN MODE for a configured project, but must be refused if Stage 2 proves the project is not\n     yet configured (it cannot configure a fresh repo).\n3. Preserve this exact refusal text for later use — it is emitted in Stage 2 ONLY when a non-admin\n   member reaches a `configured == false` project:\n   ```\n   Admin role required to apply install configuration. Ask a project admin to run /install-bridge,\n   or use an admin API key.\n   ```\n\n## Stage 2 — Read the manifest (once, pre-apply)\n\n1. Call the `get_install_manifest` MCP tool exactly once here.\n2. Keep the returned `snapshot_token` verbatim — you must pass the SAME token to the apply call in\n   Stage 5. Do not call `get_install_manifest` again before applying; a fresh read would invalidate\n   the token you are about to use. (Stage 7 performs a SEPARATE, later read-after-write manifest read\n   for the current capability status — that is deliberate and does not reuse this token.)\n3. The manifest contains ordered `groups` of bootstrap fields (each with `field_name`, `is_set`,\n   `current_value`, `guidance`, `examples`, `validation_summary`, `requires_confirmation`), a list of\n   `deferred_fields`, a `next_step`, `done_criteria`, a `command_contract_version`, and the\n   `snapshot_token`. It also carries the additive capability fields (`integrations`, `configured`,\n   `learned`, `indexed`, `tool_capabilities`, `workflows`, `concise_tool_capabilities`, `human_setup`,\n   `locked_tools`, `unlocked_tools`) — but ignore those here; the accurate capability status is the\n   post-apply read in Stage 7. `tool_capabilities` is the COMPLETE catalog-backed report field (one\n   entry per registered MCP tool, grouped and ordered by the server); `workflows` is the separate\n   curated collection of non-MCP workflows (slash commands and packaged CLI subcommands, which are not\n   registered MCP tools and never appear in `tool_capabilities`); `concise_tool_capabilities` is the\n   ADDITIVE projection over BOTH of those that Stage 7 actually renders — a server-ordered array of\n   exactly two availability sections, each `{id, name, tools}` holding mixed capability items (see\n   Stage 7); `human_setup` is a SEPARATE additive top-level list (contract v9) naming each remaining\n   unconfigured integration and what finishing it unlocks — also rendered by Stage 7, and never a third\n   section of `concise_tool_capabilities`; `locked_tools` / `unlocked_tools` are LEGACY compatibility\n   data covering only the VCS/index policy cases and are NOT the tool inventory.\n4. Compare the manifest's `command_contract_version` to this command's contract version (9, stated at\n   the top of this file). If the manifest's version is HIGHER, this scaffolded command copy is stale:\n   warn the user to refresh it by running `npx -y @bridge_gpt/mcp-server --init` in the project root,\n   then proceed conservatively — wherever the manifest's `guidance`, `next_step`, or `done_criteria`\n   differ from these instructions, follow the manifest.\n5. **Decide the outcome from the manifest, in this exact order.** Evaluate these three branches in\n   sequence and take the FIRST one that applies. The order is the contract: a configured project is\n   never classified, and a completed greenfield project is never classified or re-onboarded.\n   1. `configured == true` → **JOIN MODE** (below).\n   2. the manifest's `greenfield_setup_path` field reports `is_set == true` → **GREENFIELD\n      REPORT-ONLY** (below). Read that field from the ordinary `groups[].fields[]` list, the same way\n      you read any other bootstrap field — there is no top-level copy of it. This is the completion\n      marker for the greenfield path: the greenfield recipe writes it exactly once and the server\n      treats it as skip-if-set, so a set value means \"this project already finished greenfield\n      onboarding\".\n   3. neither of the above → **Stage 2b — Classify**.\n\n   The branches:\n   - **`configured == true` → JOIN MODE, for EVERY caller** (admin, legacy, or member). The project is\n     already set up, so this run makes ZERO configuration changes. Emit a concise welcome — for\n     example: \"This Bridge project is already configured. You're joining it as a new teammate; no\n     configuration changes will be proposed or applied.\" Then SKIP Stages 3, 4, and 5 entirely (no\n     field derivation, no project-description approval, no `apply_install_manifest` call, no\n     `config_field` writes), run Stage 6 (persist the routing credential), and render the JOIN MODE\n     branch of Stage 7 (the concise capability report, drawn directly from THIS Stage-2 manifest — no\n     read-after-write). Then EVALUATE Stage 8 conditionally — it renders only when THIS Stage-2\n     manifest's `next_step.command` is a non-empty string, and is silent otherwise — and SKIP Stages 9\n     and 10. Then continue to the independently gated Stage 11.\n     The server has already encoded recommendation eligibility (learned state AND whether this caller\n     can run the command) in `next_step.command`, so do NOT reproduce a role gate or a learned-state\n     gate here: a joining member simply receives an empty command and sees nothing.\n   - **`greenfield_setup_path.is_set == true` → GREENFIELD REPORT-ONLY, for EVERY caller.** This\n     project already completed the greenfield setup path, so this run performs NO classification, NO\n     field derivation, NO interview, NO standards generation, NO `config_field` write, and NO\n     `apply_install_manifest` call. Say so plainly — for example: \"This project already completed\n     greenfield setup. Nothing to configure; here is where it stands.\" Then SKIP Stages 2b, 3, 4, and\n     5 entirely, run Stage 6 (persist the routing credential), and render the GREENFIELD REPORT-ONLY\n     branch of Stage 7. That report is sourced from THIS Stage-2 manifest and there is deliberately no\n     read-after-write call: this branch performs no setup write, so a second read would return the\n     same bytes at the cost of a round trip. Then EVALUATE Stage 8 conditionally against THIS same\n     Stage-2 manifest's `next_step` — the server decides whether to name `connect-github` or nothing —\n     and SKIP Stages 9 and 10. Then continue to the independently gated Stage 11.\n     In the install-spawn context the ordinary Stage 8–10 skip set still applies unchanged, and Stage\n     11 remains independently gated there as it is on every other path.\n     This branch is reached for a member key too: like JOIN MODE it writes nothing, so the \"Admin role\n     required\" refusal does not apply to it.\n   - **`configured == false` and `greenfield_setup_path` is NOT set → apply the deferred Stage 1 role\n     decision:**\n     - `source == \"legacy\"` or `role == \"admin\"`: continue to **Stage 2b — Classify**, which decides\n       between the established flow (Stages 3 → 4 → 5 → 6 → 7 → 8 → 9 → 10, exactly as written and\n       unchanged) and the greenfield handoff.\n     - a non-admin `user_access` \"member\" key: stop immediately and display the exact refusal text\n       preserved in Stage 1 (\"Admin role required to apply install configuration…\"). Do not classify,\n       derive, apply, or persist anything.\n   - **`configured` absent / indeterminate (neither `true` nor `false`) → do NOT treat it as `false`.**\n     `configured` comes from a best-effort capability enrichment that can silently omit the key on a\n     transient server-side probe failure, so a missing value is \"unknown\", not \"unconfigured\". Re-read\n     the manifest ONCE (a fresh `get_install_manifest` call) to try to resolve it, and re-evaluate the\n     ordered branches above against that refreshed response — it is the authoritative manifest from\n     then on, including its `snapshot_token`, its `greenfield_setup_path` field, and its `next_step`.\n     If `configured` is STILL absent:\n     - `source == \"legacy\"` or `role == \"admin\"`: proceed, but NOT silently — first tell the user that\n       configuration status could not be confirmed and that the run will continue anyway (the server\n       owns skip-if-set, so an apply against an already-configured repo is a safe no-op). Then resume\n       the ordered evaluation at branch 2: a set `greenfield_setup_path` still means GREENFIELD\n       REPORT-ONLY, and only an unset one reaches Stage 2b.\n     - a non-admin `user_access` \"member\" key: take the JOIN-MODE-safe path — render the welcome and the\n       Stage 7 capability report (no config writes, no Stage 9/10 offers), evaluate Stage 8 under its\n       normal non-empty-command rule, and note that configuration status could not be confirmed. Do NOT\n       emit the hard \"Admin role required\" STOP: that refusal is reserved for a\n       *definitive* `configured == false`, because treating an unknown state as unconfigured would\n       re-introduce the very member hard-refusal this flow removes.\n\n## Stage 2b — Classify the repository (greenfield or established)\n\n**Run this stage ONLY when neither prior Stage 2 branch applied** — that is, the project is not\n`configured`, its `greenfield_setup_path` is not set, and the Stage 1 permission rules allow this\ncaller to configure the project (`source == \"legacy\"` or `role == \"admin\"`). JOIN MODE and GREENFIELD\nREPORT-ONLY never reach this stage, and a non-admin member was already refused above.\n\nThis stage decides ONE thing: is this a brand-new project that has no code to learn from yet\n(**greenfield**), or a real codebase (**established**)? It writes nothing.\n\n1. **Gather evidence from the local checkout**, excluding Bridge's own files from every count and from\n   the \"is the tree empty\" judgement: `.mcp.json`, `.cursor/mcp.json`, `.claude/commands/`,\n   `.cursor/commands/`, `.bridge/`, and the resolved `{docs_dir}` tree. Those exist because Bridge was\n   installed, so counting them would make every Bridge-installed empty project look established.\n   - **Scaffold evidence** (points toward greenfield): package manifests (`package.json`,\n     `pyproject.toml`, `go.mod`, `Gemfile`, `*.csproj`), lockfiles, a README, generated or template\n     configuration, an empty or near-empty `src/`, and a framework template's untouched entry point.\n   - **Established evidence** (points toward established): substantive implementation files (real\n     modules with real logic, not a generator's placeholder), meaningful tests, and a non-trivial local\n     `git log`.\n2. **Established evidence dominates.** Any credible established evidence decides the answer, no matter\n   how much scaffold evidence sits beside it — a real codebase that also has a `package.json` and a\n   README is established. Scaffold evidence only decides the outcome when there is no credible\n   established evidence at all.\n3. **Missing Git metadata is neutral.** No `.git` directory, no commits, or no remote is neither\n   evidence nor a problem: a brand-new project legitimately has none of them, and so does a directory\n   whose history lives elsewhere. Do not report it as an error or a warning, and do not let it push the\n   decision on its own.\n4. **Print exactly one primary statement**, in this exact form:\n   ```\n   Classification: Greenfield — reply \"established\" to correct, or continue.\n   ```\n   or, for the other result:\n   ```\n   Classification: Established — reply \"greenfield\" to correct, or continue.\n   ```\n   Keep that line and its correction token the primary hierarchy. A compact scaffold-versus-established\n   evidence summary may follow it — one or two short lines — but only when it genuinely explains the\n   decision; never a file listing, a count table, or a diagnostic dump.\n5. **Accept a one-token correction for exactly one response turn.** Accept ONLY the trimmed, lowercased\n   opposite token (`established` when you said Greenfield, `greenfield` when you said Established) and\n   switch to it. Anything else — an unrelated answer, a blank response, an unavailable response,\n   silence — continues with the original classification. Do not re-ask, do not re-explain, and do not\n   enter a correction loop.\n6. **Never wait in a headless session.** When no interactive response can be obtained, do not pause for\n   the correction turn at all, and classify conservatively: choose **greenfield** ONLY for an empty or\n   clearly near-empty tree with no substantive implementation files, no meaningful tests, and no\n   non-trivial history. Everything else is established. A wrong \"established\" costs a skippable\n   derivation pass; a wrong \"greenfield\" would onboard a real codebase as if it had no code.\n7. **Established → run the existing flow unchanged.** Continue to Stage 3 and proceed exactly as\n   written (Stages 3 → 4 → 5 → 6 → 7 → 8 → 9 → 10). This stage changes nothing about derivation,\n   approval, or apply behavior for an established repository.\n8. **Greenfield → hand off to the packaged `greenfield-setup` recipe, in this same session.**\n   1. Call the `get_pipeline_recipe` MCP tool with\n      `{ \"pipeline\": \"greenfield-setup\", \"variables\": { \"repo_name\": \"<this repository's configured\n      name>\" } }`. Pass `repo_name` explicitly — the server auto-supplies `docs_dir` but not\n      `repo_name`. Do not pass `auto_approve`: greenfield onboarding asks the human real questions.\n   2. Follow the returned `agent_instructions` and execute EVERY returned step, in order, to\n      completion, here in this same agent session — exactly as `/learn-repository` runs its recipe.\n      Do not spawn a second agent for it, do not summarize the recipe instead of running it, and do\n      not return to Stage 6 partway through.\n   3. The recipe owns the greenfield apply: its final task makes the one `apply_install_manifest` call\n      (with the Stage-2 snapshot token) and then performs its own read-after-write\n      `get_install_manifest`. **Require that read-after-write manifest as the recipe's final result\n      before resuming.** If the recipe cannot provide it, report that the greenfield setup did not\n      complete and stop — do not fabricate a report from the stale Stage-2 manifest, and do not\n      re-apply.\n   4. Then resume at Stage 6 (persist the routing credential), render the greenfield-completion branch\n      of Stage 7 from that read-after-write manifest, evaluate Stage 8 against that same manifest's\n      `next_step`, SKIP Stages 9 and 10, and continue to the independently gated Stage 11.\n\n## Stage 3 — Derive values for UNSET bootstrap fields only\n\n**Skip this entire stage in JOIN MODE** (Stage 2 selected JOIN MODE because the manifest reported\n`configured == true`). JOIN MODE derives nothing — it proposes and applies zero configuration for every\ncaller. Run this stage only on the `configured == false` fresh-configuration path.\n\n1. Consider ONLY fields whose `is_set` is `false`. Never re-derive or overwrite a field that is\n   already set — the server will also skip already-set fields, but you should not even propose them.\n2. For each unset field, derive a value from the local codebase following that field's `guidance`\n   (e.g. infer `working_in` from frameworks/manifests, `version_control_system` from the git remote,\n   testing stacks from test config and dependencies, directory fields from the repo layout).\n3. Omit any field you are uncertain about or that does not apply — leaving it unset is always\n   acceptable. Do NOT guess.\n4. Specifically, do NOT invent a `version` when the platform has no meaningful version concept; leave\n   it unset.\n5. Before deriving `version_control_system` or `base_branch`, confirm the git repository root is THIS\n   project's root: `git rev-parse --show-toplevel` must resolve to the project directory itself. If\n   the git root is an ancestor directory (the project is nested inside another repository), OMIT both\n   fields and note why in the final report — deriving them would describe the parent repository, not\n   this project.\n6. If you find real evidence for TWO OR MORE plausible values for a field (e.g. configs for multiple\n   test frameworks, multiple frameworks in a monorepo), treat the value as ambiguous and omit it\n   rather than picking one.\n7. Do NOT derive `selected_mcp_slugs` or `greenfield_setup_path`, even though both appear as unset\n   bootstrap-eligible fields in the manifest. MCP validation manual selection is deferred to\n   `/learn-repository`, which derives and confirms it with the codebase already researched. Install\n   neither proposes nor applies this field. `greenfield_setup_path` is not derivable at all: it is the\n   greenfield path's own completion marker and is written ONLY by the `greenfield-setup` recipe's final\n   apply, which records whether the human took the interview or the defaults. Deriving it here would\n   mark an established repository as greenfield-complete and permanently suppress its\n   `/learn-repository` recommendation.\n\n## Stage 4 — Human approval for confirmation-requiring fields\n\n**Skip this entire stage in JOIN MODE** — there is nothing to derive, so there is nothing to approve.\nRun it only on the `configured == false` fresh-configuration path.\n\n`project_description` is the ONLY confirmation-requiring field install proposes. It carries\n`requires_confirmation: true` in the manifest, so it is never applied on derivation alone — it needs\nexplicit human approval. (`selected_mcp_slugs` also requires confirmation, but install does not\nderive it at all; `/learn-repository` asks for it. See Stage 3 step 7.)\n\n1. If `project_description` is unset and you can draft a concise description from the codebase,\n   present the proposed `project_description` text to the human and ask for explicit approval before\n   including it.\n2. Include `project_description` in the apply payload ONLY as\n   `{ \"value\": <approved value>, \"confirmed\": true }`, and only after the human approves it. If the\n   human does not approve it, omit the field entirely.\n3. If no human response can be obtained (e.g. a non-interactive / headless session), do NOT stall and\n   do NOT abandon the install: proceed to Stage 5 with `project_description` omitted, and report it\n   as \"pending human input\" in the final summary. The other derived fields must still be applied — an\n   unapproved description never blocks them.\n\n## Stage 5 — Apply (one call)\n\n**Skip this entire stage in JOIN MODE** — JOIN MODE makes NO `apply_install_manifest` call and writes\nzero fields for every caller. **Skip it in GREENFIELD REPORT-ONLY** for the same reason, and **skip it\nafter a greenfield handoff**: the `greenfield-setup` recipe owns the single final\n`apply_install_manifest` call for that path, made with the Stage-2 snapshot token. Running this stage\nafterwards would be a second apply with a token the recipe already spent. Run it only on the\nestablished fresh-configuration path Stage 2b routed to Stage 3.\n\n1. Make exactly ONE call to the `apply_install_manifest` MCP tool, passing:\n   - `snapshot_token`: the exact token from Stage 2.\n   - `fields`: a map of the derived fields. Scalar fields may be passed directly\n     (e.g. `\"base_branch\": \"main\"`); an approved `project_description` must use the\n     `{ \"value\": ..., \"confirmed\": true }` object form from Stage 4. Never include\n     `selected_mcp_slugs` in this payload — install does not derive or apply it (Stage 3 step 7).\n2. Do not implement any client-side skip-if-set, conflict, or confirmation logic — the server owns\n   all of that and returns the outcome buckets.\n3. The apply is partial-tolerant: fields that fail server-side validation are returned in the\n   `rejected` bucket while the valid fields still commit. A rejected field is therefore reported,\n   not fatal — do NOT retry the whole apply just because one field was rejected.\n\n## Stage 6 — Persist the routing credential\n\n**This stage runs on EVERY path** — JOIN MODE persists the routing credential too, so a joining\nteammate's shell-spawned CLI features (`start-tickets`) can resolve the key, and so do GREENFIELD\nREPORT-ONLY and a completed greenfield handoff, which reach this stage the same way. Its secret\nhandling, fail-open behavior, and install-spawn success-silence rule below are identical on all of\nthem — none of the greenfield paths changes any of it.\n\nThe Bash-spawned `start-tickets` CLI runs in a DIFFERENT runtime surface than this MCP server, so a\nkey that lives only in `.mcp.json` / `.cursor/mcp.json` is invisible to it and difficulty→model\nrouting silently degrades. This final stage persists the already-validated key into the user-scoped\ncredential store (`~/.config/bridge/credentials.json`) so shell-spawned CLI commands can resolve it.\n\n1. Call the `persist_routing_credential` MCP tool with `repo_name` set to the repository name for\n   this install (the configured `BAPI_REPO_NAME`). Pass `repo_name` ONLY — this tool resolves the\n   API key inside the MCP server process. NEVER pass or display `BAPI_API_KEY` (nor `api_key`,\n   `apiKey`, `secret`, or `token`); the key value must not appear in the tool call or in any output.\n2. On success, tell the user the credential was stored, echoing the returned `target` and `path`,\n   e.g. `Stored routing credential for bapi:<repo_name> at <path>`. EXCEPTION — in the install-spawn\n   context (and outside JOIN MODE), persist silently: still call the tool, but print nothing on\n   success and do not restate the `target` or `path` in any later summary or in the Return contract.\n   The CLI already reported the credential destination, so this sentence would be a duplicate. This\n   exception is success-only; step 3's failure reporting is unchanged in every context.\n3. On failure, do NOT block the install — show the rest of the install summary anyway, then tell the\n   user the routing credential could not be stored. They can rerun `/install-bridge` after setting\n   `BAPI_API_KEY` in the environment, or run\n   `npx -y @bridge_gpt/mcp-server credentials migrate-agent-config --write-credentials` to migrate an\n   existing key. Until then, `start-tickets` model routing fails open (assumes a hard ticket and\n   defaults to premium/Opus when available) — this is the MOST EXPENSIVE model tier, so unexplained\n   cost is the symptom of leaving this unfixed. Recommend verifying credential resolution afterwards\n   with `npx -y @bridge_gpt/mcp-server doctor`.\n\n## Stage 7 — Summarize the outcome, then present the concise capability report\n\n**JOIN MODE branch (`configured == true`).** Do NOT print an applied count and do NOT fabricate apply\nbuckets — no apply happened. Instead state plainly that the project was already configured and that\nzero configuration changes were proposed or applied (the welcome from Stage 2). Then render the concise\ncapability report described in \"### Read-after-write\" below, with ONE difference: source it directly\nfrom the `concise_tool_capabilities` field of the Stage-2 manifest you already read — do NOT perform a\nread-after-write `get_install_manifest` call, because no write occurred and there is nothing to\nrefresh. Apply the same server-authority rendering rules (server section order, verbatim item copy,\nmalformed/missing fallback) verbatim. ALSO retain that Stage-2 manifest's `next_step` object: it is the\nauthoritative next-step source for this path, because the Stage-2 read is JOIN MODE's only manifest\nread. After the report, continue to Stage 8, which renders that `next_step` conditionally; an eligible\nadmin then proceeds to Stage 11. The rest of this stage (the applied-count line and six-bucket summary)\napplies ONLY to the `configured == false` fresh-configuration path.\n\n**Greenfield-completion branch (Stage 2b classified greenfield and the recipe ran).** Do NOT print an\napplied count and do NOT reproduce the six apply buckets: this command made no apply call — the recipe\ndid. State plainly that greenfield setup completed, and say which path the human took (the interview\nor the defaults) and what it produced: the standards fields written, and the deferred-VCS guide when\none was written. Then render the concise capability report described in \"### Read-after-write\" below,\nsourced EXCLUSIVELY from the read-after-write manifest the recipe returned as its final result — never\nfrom the stale Stage-2 manifest, and never from a fresh read of your own. Retain that same manifest's\n`next_step` object: it is the authoritative next-step source for this path, and Stage 8 renders it.\nNever mention or recommend `/learn-repository` on this path; the server has already decided what the\nnext step is, and on a greenfield project it is never that.\n\n**GREENFIELD REPORT-ONLY branch (`greenfield_setup_path` was already set).** Behave exactly like JOIN\nMODE's no-apply reporting: no applied count, no fabricated apply buckets, no report of onboarding work\nthat did not happen this run. State that this project already completed greenfield setup and that zero\nconfiguration changes were proposed or applied. Then render the concise capability report sourced\nEXCLUSIVELY from the Stage-2 manifest already read — no read-after-write call, because nothing was\nwritten — and retain that manifest's `next_step` for Stage 8.\n\nBoth greenfield branches use the SAME rendering rules as every other branch: the server's section\norder, verbatim item copy, the malformed/missing fallback line, the empty-section handling, the\nnarrow-terminal wrapping rules, and the closing documentation footer. There is one report format, and\nthese branches only change where the manifest came from.\n\n**Fresh-configuration branch (established path, `configured == false`).**\nFirst, begin with an explicit applied count: \"Applied N of M derivable fields\" (M = the unset\nbootstrap fields you derived values for). An install that applied 0 fields must say so loudly — the\ninstall is NOT complete until the apply call reports applied fields. This applied-count line and the\nsix-bucket summary below remain the PRIMARY install result — the capability report that follows is a\nsecondary close, not a replacement for it.\n\nThen summarize the six buckets from the apply response for the user:\n- `applied` — fields written this run.\n- `skipped` — fields already set (left untouched).\n- `conflict` — fields that changed since the manifest was read (re-run /install-bridge to retry).\n- `needs_confirmation` — fields awaiting explicit confirmation (e.g. project_description).\n- `rejected` — fields that failed validation or were not bootstrap-eligible (the other fields still\n  applied; fix or drop the rejected values before any retry).\n- `deferred` — fields intentionally not auto-written by install (owned by /learn-repository or set\n  deliberately). `selected_mcp_slugs` belongs here: report it as deferred to `/learn-repository`,\n  never as approved, declined, or pending an install-time decision.\n\n### Read-after-write: fetch the concise capability report\n\nAfter the bucket summary, call the `get_install_manifest` MCP tool ONCE MORE. This post-apply read\nreflects the configuration you just wrote (the Stage-2 read was pre-apply and is stale for this\npurpose). This read does not need the snapshot token. Use ONLY this post-write response for the\nreport below.\n\nRetain this post-apply response's `next_step` object as well. On the fresh-configuration path it — not\nthe pre-apply Stage-2 read — is the authoritative source for Stage 8's conditional next step. Neither\nbranch derives a command from `configured`, `learned`, or any role prose: the command is whatever the\nauthoritative manifest returned.\n\nThe response carries `concise_tool_capabilities` — an ADDITIVE projection over the complete\n`tool_capabilities` catalog and the curated `workflows` collection (the response still carries both\nunchanged; this stage simply does not render them). It is a server-ordered array of EXACTLY TWO\navailability sections, in this order:\n\n1. `{\"id\": \"available_now\", \"name\": \"Available now\", \"tools\": [...]}` — what works right now.\n2. `{\"id\": \"needs_setup\", \"name\": \"Needs setup\", \"tools\": [...]}` — what additional setup would unlock.\n\nEach entry in a section's `tools` array is a capability item:\n`{id, kind, display_name, description, how_to_use}`, where `kind` is `\"tool\"` or `\"workflow\"` and\n`how_to_use` may be `null`. Items in the `needs_setup` section carry ONE additional field,\n`availability_text` — the server's plain-language sentence naming what is missing.\n\nServer authority: the server computed this projection's section membership, availability decisions,\nretrieval-pair collapsing, and ordering. Never recompute, re-filter, re-count, re-sort, regroup, infer\navailability, pair a retrieval tool with a primary yourself, invent an item, or re-derive any of it from\n`tool_capabilities`, `workflows`, `docs/mcp-tool-integrations.md`, or any other documentation — render\nexactly what the server sent.\n\nIf the post-write response has no `concise_tool_capabilities` field at all, or it is present but\nmalformed (not the two-section `{id, name, tools}` shape described above), print exactly:\n`No capability status is available yet; configuration can still continue.` and skip the section below\nentirely — never fall back to rendering the complete `tool_capabilities` catalog, the `workflows`\ncollection, or a remembered/hallucinated capability list.\n\nOtherwise, render exactly one section, with this exact heading:\n\n**What Bridge can help with**\n\n- Render the two server sections in the order the server gave them — \"Available now\" first, then\n  \"Needs setup\". Use the server's `name` as each sub-heading. Do not reorder, merge, filter, or drop a\n  section, and do not add a third one.\n- Use semantic Markdown (a short bold or `###` sub-heading per section, one compact list item per\n  capability) with tight vertical spacing — no table, no fixed-width columns, no ASCII box. Long\n  descriptions, commands, and URLs must be free to wrap, so the report stays readable in a narrow\n  terminal.\n- For each item in **Available now**, render its `display_name`, then its `description`, then its\n  `how_to_use` when that value is a non-empty string (omit the invocation line entirely when it is\n  `null` or empty). Render NO availability caveat, asterisk, footnote, or status note in this section —\n  the server already decided these are available, including any tool that works with less codebase\n  context before indexing.\n- For each item in **Needs setup**, render the same three fields, then the server's `availability_text`\n  verbatim as emphasized setup guidance on its own line. Do not reword it, prefix it with a warning, or\n  substitute setup advice of your own.\n- Render a `kind: \"workflow\"` item EXACTLY like a `kind: \"tool\"` item. `kind` is contract metadata for\n  machine consumers, not a visual distinction: no icon, no label, no badge, no separate sub-list, no\n  different indentation, and no client branching of any sort on its value.\n- Keep an empty section VISIBLE rather than dropping it. When **Available now** is empty, render the\n  heading with one short neutral line such as \"Nothing is available yet.\" When **Needs setup** is empty,\n  render the heading with one short line stating that every highlighted capability is ready — for\n  example \"Everything highlighted here is ready to use.\"\n- Close the report with this quiet final line, after BOTH sections and regardless of whether either\n  `tools` array was empty:\n  See the [Bridge MCP server README](https://www.npmjs.com/package/@bridge_gpt/mcp-server) for the\n  complete tool documentation.\n\n### What finishing setup unlocks: rendering `human_setup` (contract v9)\n\nThe same manifest response this stage already read for `concise_tool_capabilities` also carries a\nSEPARATE top-level field, `human_setup` — an ordered list naming each remaining unconfigured\nintegration and what finishing it unlocks, computed server-side from the same checklist data\n`integrations` already carries. It is additive: a `human_setup`-unaware reader of this manifest sees no\nchange to `integrations`, `tool_capabilities`, `workflows`, or the two-section\n`concise_tool_capabilities` shape above. It is never a third section of that report — keep the two\nsections above exactly as they are and render this as its own, separate block.\n\nCompatibility and validity, in this order:\n\n- **`human_setup` is absent from the response entirely** (an older, still-compatible v8 manifest) →\n  render nothing for this block. Do not fabricate a heading, a completion line, or fallback setup\n  advice, and do not treat the absence as an error.\n- **`human_setup` is present but malformed** (not a list, or an entry missing its `display_text`) →\n  render nothing for this block, the same as if it were absent. The two-section capability report above\n  is unaffected either way.\n- **`human_setup` is a valid, empty list (`[]`)** → render nothing: no heading, no \"everything is set\n  up\" line, no completion claim. An empty list means every integration is already configured, and the\n  Available now / Needs setup sections above already say what that state unlocks.\n- **`human_setup` is a valid, non-empty list** → render it under this exact heading:\n\n  **What finishing setup unlocks**\n\n  Render one compact Markdown list item per entry, in the server's given order, using\n  `human_setup[].display_text` as that item's ENTIRE substantive content, byte-for-byte. Do not\n  summarize, reword, sort, filter, merge, or otherwise compose your own setup advice from\n  `configure_in`, `setup_instructions`, `required_for`, `integrations`, or any tool/workflow catalog —\n  the same server-authority rule Stage 7 already applies to `concise_tool_capabilities` extends to this\n  field. `configure_in` and `setup_instructions` travel on each entry for completeness but are not\n  separately rendered here; `display_text` already carries the one line this block shows.\n\n## Stage 8 — Render the server's next step (only when there is one)\n\n**Skip this entire stage in the install-spawn context** (the spawning instruction identifies it); the\nspawn prompt owns the same conditional closing recommendation there. On direct manual\n`/install-bridge` invocation, run it normally.\n\n**This stage runs in JOIN MODE too** — it is no longer skipped wholesale for joining teammates. What\ngates it is the server's output, not the mode and not the caller's role.\n\nThis stage renders a decision the SERVER already made. The manifest's `next_step.command` encodes both\nwhether the repository still needs the deeper instruction-tier configuration AND whether this caller\ncan actually run the command; a caller who cannot gets an empty command. On a greenfield project the\nserver may name a `connect-github` command instead, or nothing at all. Never re-derive any of that\nfrom `configured`, `learned`, `greenfield_setup_path`, `role`, or `source`, never assemble a\n`connect-github` invocation locally, and never substitute a command you happen to know.\n\n1. Read `next_step` from the AUTHORITATIVE manifest for this path (Stage 7 names it: the post-apply\n   read-after-write response on the fresh-configuration path; the read-after-write manifest the\n   `greenfield-setup` recipe returned after a greenfield handoff; and the Stage-2 response in JOIN MODE\n   and in GREENFIELD REPORT-ONLY, neither of which performs a read-after-write).\n   Then branch on `next_step.command` alone:\n   - **`next_step.command` is a non-empty string** → render one short heading and the exact command\n     the server returned, formatted as Markdown inline code — for example: ``Next step: `<command>` ``\n     where `<command>` is that returned string, verbatim. Do not rename it, alias it, expand it into\n     a different invocation, or replace it with a locally known command. Do not run it yourself:\n     it remains the human's own explicit next invocation.\n   - **`next_step.command == \"\"`** (or the key is missing/not a string) → be entirely SILENT. Emit no\n     heading, no empty code span, no placeholder or fallback command, no separator or divider, no\n     \"nothing to do\" notice, and no generic success or completion claim. The stage simply produces\n     nothing and the run continues.\n2. Where you give supporting context for a rendered command, use the manifest's own\n   `next_step.post_install_indexing` and `next_step.post_install_scheduler` text as the authoritative\n   source rather than inventing a rationale for whatever command the server named. Keep that context\n   subordinate to the command itself — it is explanation, not a second call to action.\n3. Do NOT ask about repository indexing in any form. There is no consent question, no\n   `parse_repository` tool call, and no `/parse-repository` continuation here — indexing starts\n   automatically once the repository reaches full parse readiness (VCS credentials, the Pinecone\n   index, `working_in` / `project_description`, and SFCC prerequisites where applicable), via the\n   same readiness-gated funnel the GitHub connection-confirm endpoints and the scheduled sweep already\n   use. Do not claim indexing has already started — this command has no visibility into that funnel's\n   outcome.\n\n## Stage 9 — Offer CI follow-up configuration (only when CI is detected)\n\n**Skip this entire stage in the install-spawn context** (the spawning instruction identifies it), so\nthe concise report and the conditional next step the spawn prompt owns remain the sole closing interaction. On direct\nmanual `/install-bridge` invocation, run it normally.\n\n**Also skip this stage in JOIN MODE** (`configured == true`), in **GREENFIELD REPORT-ONLY**, and after\na **greenfield handoff**: none of those paths offers configuration follow-up; they stop after the Stage\n7 capability report and the conditional Stage 8 next step (an eligible admin continues to Stage 11).\n\nThis stage is separate from the single apply call: it writes at most ONE field via the\n`config_field` MCP tool. `ci_followup_config` is REPLACE-NOT-AUGMENT — leaving it NULL already means\nsafe poll-only defaults, so \"skip\" is always a correct answer.\n\n1. Gate: only offer this when Stage 3 found real CI evidence (e.g. `.github/workflows/*.yml`,\n   `bitbucket-pipelines.yml`). With no CI detected, skip this stage and note \"no CI detected — CI\n   follow-up not offered\" in the final report.\n2. Ask ONE question, referencing the discovered CI by name, with exactly these preset profiles\n   (default: skip):\n   - **observe** — poll CI results only, never attempt fixes:\n     `{\"strategy\": \"poll_only\", \"max_iterations\": 1, \"max_minutes\": 10, \"instructions\": \"\"}`\n   - **self-heal** — bounded fix-and-iterate loop on the automation's own PRs:\n     `{\"strategy\": \"fix_and_iterate\", \"max_iterations\": 3, \"max_minutes\": 45, \"instructions\": \"\"}`\n   - **skip** (default) — leave `ci_followup_config` unset (NULL = baseline poll-only behavior).\n3. Do NOT offer or compose a `custom` strategy during install — free-form CI follow-up instructions\n   are a deliberate later act (setup UI or a direct `config_field` update).\n4. When the human picks a profile, call the `config_field` MCP tool once (operation `\"update\"`,\n   `field_name: \"ci_followup_config\"`, `value`: the profile's JSON object). If the server rejects the\n   value, retry ONCE with a corrected payload; if it is rejected again, stop, show the proposed JSON\n   to the human, and leave the field unset.\n5. Never select a profile in a non-interactive session — skip and list it as a pending next step. An\n   unattended install must never enable fix-and-iterate behavior.\n\n## Stage 10 — Offer the speed-vs-quality repository preference\n\n**Skip this entire stage in the install-spawn context** (the spawning instruction identifies it), so\nthe concise report and the conditional next step the spawn prompt owns remain the sole closing interaction. On direct\nmanual `/install-bridge` invocation, run it normally.\n\n**Also skip this stage in JOIN MODE** (`configured == true`), in **GREENFIELD REPORT-ONLY**, and after\na **greenfield handoff**: none of those paths offers configuration follow-up; they stop after the Stage\n7 capability report and the conditional Stage 8 next step (an eligible admin continues to Stage 11).\n\nThis stage is separate from the single apply call: it writes at most ONE field via the `config_field`\nMCP tool, `field_name: \"speed_vs_quality\"`. The column defaults to `5` (max quality) for every\nrepository, so \"skip\" always leaves a safe, valid value in place.\n\n1. Never ask this question in a non-interactive session — skip and list it as a pending next step. An\n   unattended install must never block on or silently answer this preference.\n2. Read the current value first via the `config_field` MCP tool (operation `\"get\"`,\n   `field_name: \"speed_vs_quality\"`) so a reinstall can show the stored value — not always `5` — as\n   the displayed default rather than silently re-asking from scratch.\n3. Ask ONE question with exactly these five numbered presets (default: the value from step 2, or `5`\n   if this is the first install):\n   - **1** — Max speed\n   - **2** — Prefer speed\n   - **3** — Balanced\n   - **4** — Prefer quality\n   - **5** — Max quality (default)\n4. Persist ONLY on an explicit answer — including an explicitly accepted default — by calling the\n   `config_field` MCP tool once (operation `\"update\"`, `field_name: \"speed_vs_quality\"`,\n   `value`: the selected integer 1-5). Do not overwrite an existing value when the human gives no\n   answer at all (e.g. the session cannot obtain one) — leave the stored/default value untouched in\n   that case, distinct from an explicit accepted-default selection of `5`.\n5. After a successful persist, emit one additive structured log line (mirroring the codebase's\n   `logging.info(msg, extra={...})` convention for non-response-body observability signals) with\n   fields: `event=\"install.speed_vs_quality\"`, `repo_name`, `field_name=\"speed_vs_quality\"`,\n   `selected_preset` (the persisted integer), `outcome=\"persisted\"`. Never log this event before the\n   `config_field` tool call has confirmed the write.\n6. When skipped, emit the corresponding structured event with `outcome` set to one of\n   `\"skipped_non_interactive\"` (non-interactive session) or `\"skipped_no_answer\"` (interactive session,\n   no explicit answer obtained) — omit `selected_preset` and any prompt text from this event.\n7. If the `config_field` call is rejected, retry ONCE with a corrected payload; if it is rejected\n   again, stop, show the proposed value to the human, and leave the field at its current stored value.\n   Do not log a persistence-success event for a rejected or failed write.\n\n## Stage 11 — Invite teammates (gated: b2b admins only; best-effort)\n\nThis is a best-effort final stage that lets an eligible admin mint teammate keys after configuration or\nJOIN MODE. It is **independently gated** and is NOT part of the install-spawn skip set (Stages 8–10) —\nit may run in the install-spawn context for an eligible admin. The entire stage is **fail-open**: a\nprompt failure, a declined offer, a non-interactive context, a malformed tool response, or an\n`invite_member` failure must NEVER cause this command to report the (already-completed) install as\nfailed.\n\n1. **Eligibility gate (AND).** Offer this stage ONLY when BOTH hold, using the values retained in\n   Stage 1:\n   - `role == \"admin\"`, AND\n   - `customer_type == \"b2b\"`.\n   Otherwise skip the stage silently with NO prompt: a member, a legacy-source caller whose role is not\n   explicitly `\"admin\"`, and a b2c admin all skip. Reachable from EVERY Stage 2 outcome for an eligible\n   admin: the established fresh-admin close (after Stage 10), the JOIN MODE path (after the Stage 7\n   report), the GREENFIELD REPORT-ONLY path, and a completed greenfield handoff — including an admin\n   re-run against an already-configured or already-greenfielded b2b project.\n2. **Interactive surface required.** This stage needs a human response. If no interactive response can\n   be obtained (a non-TTY / headless / spawn context that cannot prompt), skip the stage silently\n   without changing the completed install result — do not stall.\n3. **Offer prompt.** Ask exactly: `Invite teammates to this project? (y/N)`. Treat a blank answer, `n`,\n   `no`, an unavailable response, or any prompt failure as a non-fatal decline — skip the rest of the\n   stage and report it as declined.\n4. **Collect invitees.** On an affirmative answer, collect one or more teammate email entries using\n   normal **echoed** input (email is PII, not a secret — never use a muted/hidden secret prompt).\n   Optionally collect a display name per entry.\n5. **Per-invite role.** Default each invitation to role `member`. Only set a specific request's role to\n   `admin` after an explicit per-invite opt-up for that entry; never opt up by default.\n6. **Mint.** For each invitee, call the `invite_member` MCP tool exactly once with `{email, name?,\n   role}`. Do NOT pass `repo_name` — the tool resolves the repository from the current session.\n7. **Show each key once.** After each successful call, display that response's plaintext `api_key`\n   exactly once, associated with its intended recipient, followed by the exact warning:\n   `Distribute securely; this key is shown once.` Do NOT repeat a minted key anywhere else — not in the\n   final summary, not in retry guidance, not in diagnostics, not in a later stage.\n8. **Per-invite failure isolation.** Treat each mint failure as local to that invitee: report a\n   sanitized failure for it, continue to any remaining invitees, and never change the already-completed\n   install outcome. Do not surface raw error text, headers, or the caller's key.\n\n## Return\n\nThe Return contract depends on which outcome Stage 2 selected.\n\n**Fresh-configuration return (established path, `configured == false` admin/legacy).**\nReport the admin check result, the \"Applied N of M\" count, the\napplied/skipped/conflict/needs_confirmation/rejected field names, any fields omitted for\nnested-repository or ambiguity reasons, the approval outcome for `project_description` — install's\nonly confirmation-requiring field — (approved / declined / pending human input), the fact that\n`selected_mcp_slugs` is deferred to `/learn-repository` rather than decided here, whether a\nstale-command warning was raised (manifest\n`command_contract_version` higher than this command's), whether the routing credential was persisted\n(the returned `target` and `path`, or the non-blocking failure remediation — but in the install-spawn\ncontext report neither on success, per Stage 6's silence exception; a failure is still reported), the \"What Bridge can help\nwith\" concise capability report from the post-apply read-after-write manifest read, the CI follow-up\noutcome (profile written / skipped / no CI detected / pending / skipped in install-spawn context), the\nspeed-vs-quality preference outcome (persisted with its preset / skipped_non_interactive /\nskipped_no_answer / pending / skipped in install-spawn context), and the Stage 8 outcome — the exact\nserver-provided next-step command when the manifest returned a non-empty one, otherwise the fact that\nno next step was recommended (never a substituted command of your own).\n\n**JOIN MODE return (`configured == true` path).**\nReport the caller's role, the \"you're joining an already-configured project\" welcome and the explicit\nno-change status (zero configuration proposed or applied — do NOT report an applied count or apply\nbuckets), whether the routing credential was persisted (the returned `target` and `path`, or the\nnon-blocking failure remediation), the \"What Bridge can help with\" concise capability report drawn\nfrom the Stage-2 manifest (no read-after-write), and the Stage 8 outcome — the exact server-provided\nnext-step command when that manifest returned a non-empty one, otherwise the fact that no next step\nwas recommended.\n\n**Greenfield-completion return (Stage 2b classified greenfield and the recipe ran).**\nReport the classification that was printed and whether the human corrected it, the setup path the\nrecipe recorded (`interview` or `defaults`), which standards fields were written and which were left\nalone, whether the deferred-VCS guide was written or the connect command was printed, whether the\nrouting credential was persisted (the returned `target` and `path`, or the non-blocking failure\nremediation — with Stage 6's install-spawn silence exception unchanged), the \"What Bridge can help\nwith\" concise capability report drawn from the recipe's read-after-write manifest, and the Stage 8\noutcome — the exact server-provided next-step command when that manifest returned a non-empty one,\notherwise the fact that no next step was recommended. Do NOT report the ordinary derivation buckets\n(applied / skipped / conflict / needs_confirmation / rejected / deferred): no Stage 5 apply happened on\nthis path, so there are no buckets to report, and do NOT report an applied count. Never mention or\nrecommend `/learn-repository`.\n\n**GREENFIELD REPORT-ONLY return (`greenfield_setup_path` already set).**\nReport that this project already completed greenfield setup and that this run made ZERO setup changes,\nasked ZERO onboarding questions, and applied nothing — no applied count, no apply buckets, no\nderivation buckets. Report whether the routing credential was persisted (same rule as above), the\n\"What Bridge can help with\" concise capability report drawn from the Stage-2 manifest (no\nread-after-write), and the Stage 8 outcome under its usual non-empty-command rule. Never mention or\nrecommend `/learn-repository` on this path either.\n\n**Teammate-invitation outcome (Stage 11, all outcomes).**\nReport the Stage 11 outcome as counts/statuses only — one of offered, declined, skipped (not eligible,\nor non-interactive context), partially completed, or completed, plus how many keys were minted. Never\nrepeat teammate email addresses or minted key values in this summary. Stage 11's eligibility gate and\nits one-time key-secrecy rule are identical on all three Stage 2 outcomes — JOIN MODE, GREENFIELD\nREPORT-ONLY, and CLASSIFY (established or greenfield) — and are unchanged by this contract version.\n",
     "learn-repository.md": "Learn and document all configuration fields for the repository by running parallel research agents.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nThis command is recipe-driven. Do not call MCP tools directly -- the recipe determines which tools to call and with what parameters. There are exactly two narrow exceptions, both invoked directly by this command and never through the recipe: the `ping` admin preflight at entry (immediately below) and the closing `get_install_manifest` capability report (step 6).\n\n## Admin preflight (UX guard — run this first)\n\nBefore the onboarding guidance below and before fetching the recipe, call the `ping` MCP tool\nonce (the first of the two narrow exceptions above) and read `role` from its first (JSON) content item:\n\n- If `role` is `\"admin\"`, or `role` is `null` (not determined — e.g. a legacy shared key): continue to\n  the guidance and recipe below.\n- Otherwise (any non-admin `user_access` result — e.g. a \"member\" key): stop immediately, run nothing\n  else (no recipe fetch, no research, no configuration write), and display exactly one concise message:\n  ```\n  Admin role required to learn this repository. Learning writes shared Bridge project configuration,\n  which only an admin may change. Ask a project admin to run /learn-repository, or use an admin API key.\n  ```\n- If the `ping` call fails or returns a malformed / unrecognized response, treat it as a\n  preflight failure: stop with the same admin-required message rather than proceeding into\n  configuration writes.\n\nThis preflight is a UX guard ONLY — it fails fast with one clear message instead of the cascade of\nper-field denials a non-admin would otherwise hit. It is NOT the security control: the authoritative\nenforcement is the existing server-side admin gate on the `config_field` update and\n`apply_install_manifest` routes, which already reject non-admin keys regardless of this client-side\ncheck.\n\n1. This command takes no arguments.\n\n2. **Before executing any recipe step**, tell the user what they are about to sit through and why it\n   is worth it:\n\n   ```\n   Learning this repository. This takes a while — the research agents read the actual codebase, and\n   all the unlearned fields are researched in parallel, so the wait is roughly the slowest single\n   field rather than the sum of all of them. Fields that are already populated are skipped entirely.\n\n   What this buys you: these fields are what ground Bridge's agents in THIS codebase. Planning,\n   reviewing, and code generation all read them, so they follow your repository's actual\n   architecture, testing, documentation, and correctness conventions instead of generic defaults.\n\n   It runs unattended — there are no approval prompts during the run. You may be asked one batched\n   question at the very end.\n   ```\n\n   Do not invent a specific number of minutes; the honest statement is the parallel-wait shape above.\n\n3. Call the `get_pipeline_recipe` MCP tool with:\n   - `pipeline`: `\"learn-repository\"`\n\n   If the tool returns an error, stop and report the failure.\n\n4. Read and strictly obey the `agent_instructions` field in the response. Execute each step in order, announcing each as **Step N of M: <description>**.\n\n   Retain, for the closing summary: the consolidated research task's structured result (its per-field\n   `status`, `character_count`, `condensed`, and `condensation_reason`), each upload step's result,\n   and the final confirmation task's structured result. You will need all three — do not discard them\n   as you go.\n\n5. After all steps complete, display a summary built from the results you retained:\n\n   ```\n   ## Learn Complete\n\n   **Status**: Success / Completed with gaps / Failed at step N\n\n   **Learned and applied**: <fields drafted this run and written to config>\n   **Already populated (skipped)**: <fields skipped because they already had a value>\n   **Condensed to fit the field limit**: <field — reason it was condensed, per field>\n   **Gaps**: <fields whose research or upload failed, each named with its reason>\n   **Confirmation**: <approved / applied / declined / pending human input / not applicable, per field>\n\n   Review or edit any of these on the **Project Configuration** page, under **Code Writer Settings**\n   for the learned instructions and **MCP Validation Manuals** for the manual selection. Bridge's\n   agents read whatever is stored there, so correcting a wrong conclusion there changes their\n   behavior.\n   ```\n\n   Rules for the summary:\n\n   - A run where some fields failed but others applied is **`Completed with gaps`**, not `Failed`.\n     Name every gap explicitly — an unnamed gap is worse than a failed run, because the user believes\n     the field was learned.\n   - Report confirmation candidates that could not be presented in a headless session with the exact\n     phrase `pending human input`.\n   - Every field that was condensed must appear with the reason it was condensed.\n\n6. **After** the `## Learn Complete` summary above is fully displayed, close with the same concise\n   capability report `/install-bridge` renders (BAPI-658, AC-9). This is the recipe's ONE exception to\n   \"do not call MCP tools directly\": call the `get_install_manifest` MCP tool EXACTLY ONCE here,\n   directly, with no arguments beyond what it requires — never through the recipe, never a second time,\n   and never to apply or change any configuration.\n\n   The report is structurally and visually SUBORDINATE to `## Learn Complete` above it — it is a\n   closing addendum, not a replacement for or a distraction from the learn summary's own status,\n   fields, gaps, and confirmation outcome.\n\n   If the `get_install_manifest` call errors, or its response has no `concise_tool_capabilities` field,\n   or that field is present but malformed (not the server's two-section `{id, name, tools}` availability\n   shape), print exactly this line and stop — do not attempt the report in any other form:\n\n   ```\n   capability report unavailable — run /install-bridge to see it\n   ```\n\n   Never substitute documentation, the complete `tool_capabilities` catalog, a remembered tool list, or\n   an inferred capability category for a missing or malformed concise field — a hallucinated report\n   during this trust-critical first run is worse than no report at all.\n\n   Otherwise, render exactly one section, with this exact heading, from `concise_tool_capabilities`\n   only:\n\n   **What Bridge can help with**\n\n   - Render the server's two availability sections in the order it gave them — \"Available now\" first,\n     then \"Needs setup\" — using each section's `name` as its sub-heading. Do not reorder, merge, drop a\n     section the server included even when its `tools` array is empty, or add a third one.\n   - Within a section, render each item's `display_name`, then its `description`, then its `how_to_use`\n     when that value is a non-empty string, in server order. Items in \"Needs setup\" additionally carry\n     the server's `availability_text`, rendered verbatim as the setup guidance; items in \"Available now\"\n     carry no availability caveat at all and must not be given one.\n   - Render a `kind: \"workflow\"` item exactly like a `kind: \"tool\"` item. `kind` is contract metadata,\n     not a visual distinction — no label, icon, separate list, or client branching on its value.\n   - Keep an empty section visible with one short neutral line rather than dropping it, and close the\n     report by pointing at the Bridge MCP server README for the complete tool documentation. Name the\n     README in prose only — this command never hardcodes a URL.\n   - Do not locally filter, count, regroup, sort, infer availability, pair a retrieval tool with a\n     primary, fabricate an item, write configuration, or fall back to the complete `tool_capabilities`\n     or `workflows` collections for this section under any circumstance.\n",
     "parse-repository.md": "Queue a background job to parse and index the repository for Bridge API's AI agents.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nExecute all steps in this command as a simple linear sequence of MCP tool calls.\n\n## Step 1 — Parse Arguments\n\nParse `$ARGUMENTS` for an optional `directory_path` argument (a subdirectory path to scope the parse to, e.g., `src/python`). If no argument is provided, the entire repository will be parsed. If `$ARGUMENTS` is provided but invalid (e.g., contains special characters that suggest it's not a path), report an error.\n\n## Step 2 — Queue Parse Job\n\nCall the `parse_repository` MCP tool with:\n- `directory_path`: set to the parsed `directory_path` from Step 1 if provided, otherwise omit the parameter\n\nIf the response indicates parsing is already in progress, display:\n\n```\nRepository parsing is already in progress. A previous parse job has not yet completed.\n\nAsk your agent to check progress with the `parse_repository` tool (`action` set to `\"status\"`), or wait a few minutes and try again.\n```\n\nStop and do not proceed to the summary.\n\nIf the call fails or returns an error, stop immediately and display:\n\n```\nFailed to queue parse job: <error message from the tool>\n```\n\n## Summary\n\nOn successful queuing, display:\n\n```\nRepository parse job queued successfully.\n\nScope: <entire repository or directory_path if provided>\n\nProcessing typically takes several minutes for large repositories.\nAsk your agent to check progress with the `parse_repository` tool (`action` set to `\"status\"`).\n```\n\nAfter the parse completes, AI-generated plans and clarifying questions will reflect the latest code changes.\n",
     "plan-epic.md": "Plan an epic by decomposing it into sub-tasks with structured exploration documents.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nThis command is recipe-driven. Do not call MCP tools directly -- the recipe determines which tools to call and with what parameters.\n\n## Stage 0 — Setup\n\n1. **Parse arguments**: Extract the input from `$ARGUMENTS`. Trim any surrounding whitespace. If the input is empty or whitespace-only, stop immediately and display:\n   ```\n   Usage: /plan-epic <description of the epic or Jira key>\n   ```\n\n2. **Jira key detection**: If the input matches a Jira key pattern (`[A-Z]+-\\d+`), call the `get_ticket` MCP tool with that key to fetch the epic description. Use the ticket's description as the `epic_description`, and set `epic_key` to that Jira key. If the input does not match a Jira key, use the free-form text directly as the `epic_description` and set `epic_key` to an empty string `\"\"` (there is no Jira epic to update). The recipe uses `epic_key` to decide whether to post the goals/NFRs + recommended implementation order as a comment on the epic.\n\n3. **Generate slug**: Create a kebab-case slug from the epic description — take the first 6-8 meaningful words, strip non-alphanumeric characters (except hyphens), lowercase, and truncate to 60 characters. This becomes the `epic_slug`.\n\n4. **Directory existence check**: Call the `ping` MCP tool (no parameters) and read `docs_dir` from its first (JSON) content item. Then run a terminal command to check if the directory `{docs_dir}/epic-plans/{epic_slug}` already exists:\n   ```\n   test -d {docs_dir}/epic-plans/{epic_slug} && echo \"exists\" || echo \"not_found\"\n   ```\n   If the directory exists, append `-{unix_timestamp}` to the `epic_slug` (e.g., `add-auth-provider-support-1710000000`).\n\n## Stage 1 — Execution\n\n5. Call the `get_pipeline_recipe` MCP tool with:\n   - `pipeline`: `\"plan-epic\"`\n   - `variables`: `{ \"epic_description\": \"<resolved_description>\", \"epic_slug\": \"<slug>\", \"epic_key\": \"<jira_key_or_empty_string>\" }`\n\n   Note: Do NOT pass `docs_dir` in variables — it is auto-injected by the pipeline system.\n\n   If the tool returns an error, stop and report the failure.\n\n6. Read and strictly obey the `agent_instructions` field in the response. Execute each step in order, announcing each as **Step N of M: <description>**.\n\n7. After all steps complete, display a summary:\n   ```\n   ## Pipeline Complete\n\n   **Epic**: <first 80 characters of epic_description>...\n   **Slug**: <epic_slug>\n   **Output**: <docs_dir>/epic-plans/<epic_slug>/overview.md\n   **Steps executed**: N of M\n   **Status**: Success / Failed at step N\n   ```\n",
@@ -27,7 +28,7 @@ export const COMMANDS = {
     "run-tests.md": "Run the project's full test suite (unit and E2E) using the project-configured test stacks, triage failures, fix test-code issues, and produce a structured health-check report.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nThis command discovers how to run tests by reading per-project configuration from the Bridge API, not from hardcoded paths. Stages run only when the project has the corresponding stack configured.\n\n## Stage 0 — Argument Parsing and Setup\n\n1. **Parse `$ARGUMENTS`** for optional flags. Supported flags:\n   - `--skip-e2e` — skip the E2E test stage even if an E2E stack is configured (e.g., when no local server is running)\n   - `--unit-only` — shorthand that implies `--skip-e2e`\n\n   Resolve flags to boolean variables:\n   - Start with: `run_unit = true`, `run_e2e = true`\n   - If `--unit-only` is present: set `run_e2e = false`\n   - If `--skip-e2e` is present: set `run_e2e = false`\n   - Unknown flags: note them in the final report as \"Unrecognized flag ignored\" but do not fail\n\n2. **Generate a run timestamp** using the current date and time in `YYYY-MM-DD-HH-MM` format (e.g., `2026-03-10-14-35`). Store this as `run_timestamp`. Both output documents will use this value.\n\nThis stage has no failure conditions — proceed to Stage 1.\n\n## Stage 1 — Resolve Project Config via MCP\n\nRead the per-project test setup from the Bridge database. Every subsequent stage is driven by what these calls return.\n\n1. **Resolve docs directory**: Call the `ping` MCP tool (no parameters) and read `docs_dir` from its first (JSON) content item. Store that path as `docs_dir`.\n\n2. **Read unit-test stack**: Call the `config_field` MCP tool with `operation` set to `\"get\"` and `field_name` set to `unit_testing_stack`. Store the returned value as `unit_stack` (may be null/empty).\n\n3. **Read unit-test instructions**: Call the `config_field` MCP tool with `operation` set to `\"get\"` and `field_name` set to `unit_testing_instructions`. Store the returned value as `unit_instructions` (may be null/empty).\n\n4. **Read E2E stack**: Call the `config_field` MCP tool with `operation` set to `\"get\"` and `field_name` set to `e2e_testing_stack`. Store as `e2e_stack`.\n\n5. **Read E2E instructions**: Call the `config_field` MCP tool with `operation` set to `\"get\"` and `field_name` set to `e2e_testing_instructions`. Store as `e2e_instructions`.\n\n6. **Compute configuration booleans**:\n   - `unit_configured` = `true` if either `unit_stack` or `unit_instructions` is a non-empty string; otherwise `false`\n   - `e2e_configured` = `true` if either `e2e_stack` or `e2e_instructions` is a non-empty string; otherwise `false`\n\n7. **Create the output directory**:\n   ```\n   mkdir -p {docs_dir}/testing/\n   ```\n   If this fails, stop immediately and report: `Cannot create output directory {docs_dir}/testing/ — check permissions.`\n\nIf any MCP call fails (e.g., the API is unreachable or returns 4xx/5xx), stop immediately and report which call failed. Do not fall back to hardcoded commands — the whole point of this command is that test setup lives in config.\n\n## Stage 2 — Unit / Standard Tests\n\nIf `run_unit` is `false`, skip this stage and record: `Unit tests: SKIPPED — run_unit was set to false (this should not happen in normal use; report as a bug).`\n\nIf `unit_configured` is `false`, skip and record:\n```\nUnit tests: SKIPPED — no unit_testing_stack or unit_testing_instructions configured for this repo. Configure them with /learn-repository (a project admin must run it) or the project setup UI before running /run-tests.\n```\n\nOtherwise:\n\n1. Read `unit_instructions` carefully. It is the source of truth for **how to run unit tests in this repo** — runner binary, paths, environment activation, sub-suites (if the project distinguishes \"unit\" from \"integration\", both belong in this stage), and any flags. Pair it with `unit_stack` (a short label, e.g., `Pytest`, `Jest + React Testing Library`) for context.\n\n2. **Derive the test command(s)**: Extract the literal shell commands the instructions describe. If the instructions describe multiple sub-suites (e.g., a fast unit batch and a slower integration batch), plan to run each as a **separate batch** in the order described. Do not invent runners or paths that the instructions do not mention.\n\n3. **If the instructions do not specify any runnable command**, skip and record:\n   ```\n   Unit tests: SKIPPED — unit_testing_instructions does not describe how to invoke tests; update that field with /teach-bridge, or re-run /learn-repository (a project admin must run it).\n   ```\n\n4. **Run each batch sequentially** in the terminal. **Continue to the next batch even if the current one has failures.** Capture the full output of each batch, including the runner's summary line (e.g., `47 passed, 3 failed in 12.4s` or `Tests: 5 failed, 22 passed`).\n\n5. For each failing test, apply the **Triage Logic** (below), then record the result.\n\n## Stage 3 — E2E Tests\n\nIf `run_e2e` is `false`, skip this stage and record: `E2E tests: SKIPPED — --skip-e2e or --unit-only flag was set.`\n\nIf `e2e_configured` is `false`, skip and record:\n```\nE2E tests: SKIPPED — no e2e_testing_stack or e2e_testing_instructions configured (the project may not have an E2E suite).\n```\n\nOtherwise:\n\n1. Read `e2e_instructions`. It is the source of truth for the E2E runner, spec paths, browser config, and any prerequisites. Pair with `e2e_stack` for context.\n\n2. **Detect server prerequisites**: If `e2e_instructions` indicates that a local server must be running (look for explicit cues such as \"server\", \"running\", \"localhost\", \"started\", \"dev server\", a URL, or a port number) and describes a readiness check, perform that check exactly as described. If the instructions describe a server prerequisite but do not describe a check, attempt the check the instructions imply (e.g., curl the URL the instructions mention) and skip the stage if it fails:\n   ```\n   E2E tests: SKIPPED — e2e_testing_instructions describe a server prerequisite that wasn't met. Start the server per the instructions and re-run.\n   ```\n\n3. **Derive the test command(s)** from the instructions, including any spec-directory batching the instructions specify.\n\n4. **If the instructions do not specify any runnable command**, skip and record:\n   ```\n   E2E tests: SKIPPED — e2e_testing_instructions does not describe how to invoke tests; update that field with /teach-bridge, or re-run /learn-repository (a project admin must run it).\n   ```\n\n5. **Run each batch sequentially** in the terminal. **Continue to the next batch even if the current one has failures.** Capture the full output and summary line of each batch.\n\n6. For each failing test, apply the **Triage Logic** (below), then record the result.\n\n## Triage Logic\n\nFor every failing test, examine the test file and the code it tests. Classify as ONE of the following:\n\n### TEST-CODE ISSUE — fix it directly\n\nClassify as a test-code issue if ANY of the following applies:\n- The test asserts against a hardcoded value that no longer matches current behavior (outdated mock data)\n- The test imports or calls a function that was renamed, moved, or removed\n- The test asserts on a response field that was restructured\n- The test expects a specific error message string that has since changed\n- A fixture references a removed table column, model field, or schema member\n\n**Action**: Apply a minimal, targeted fix to the test file only. Then re-run just that failing test, using the runner described in the relevant instructions field (`unit_instructions` for unit-test failures, `e2e_instructions` for E2E failures). Adapt the runner invocation that the instructions provide to target a single test, following whatever convention the instructions or stack idiomatically use.\n\nIf the re-run **still fails** after your fix, do not make further edits — escalate to implementation-code issue instead and revert your change.\n\n### IMPLEMENTATION-CODE ISSUE (or UNCERTAIN) — document, do not fix\n\nClassify as an implementation issue if ANY of the following applies:\n- The production function raises an unexpected exception\n- A handler returns the wrong status code or response shape for a documented behavior\n- Business logic produces incorrect output that the test correctly asserts against\n- You are not confident the test is wrong\n\n**Action**: Do NOT modify any file outside the test directories described in `unit_testing_instructions` / `e2e_testing_instructions`. When in doubt about whether a path is test-only, treat it as production code and escalate. Record the failure in the implementation-issues document for the user to triage.\n\n## Stage 4 — Write Output Documents\n\n### Document 1: Test Run Report (always write this)\n\nWrite to: `{docs_dir}/testing/test-run-{run_timestamp}.md`\n\n```markdown\n# Test Run: {run_timestamp}\n\n## Configuration\n- Unit stack: {unit_stack or \"not configured\"}\n- E2E stack: {e2e_stack or \"not configured\"}\n- Unit tests: RUN | SKIPPED — (reason)\n- E2E tests: RUN | SKIPPED — (reason)\n\n## Unit Tests\n**Stack**: {unit_stack or \"not configured\"}\n**Result**: X passed, Y failed (sum across batches)\n\n### Batch 1: `<command>`\n**Result**: X passed, Y failed\n**Fixes applied**:\n- `path/to/test_file`: brief description of what was fixed\n- (or \"none\" if no fixes were needed)\n\n### Batch 2: `<command>`\n...\n\n**Failures escalated as implementation issues**: N\n\n## E2E Tests\n**Stack**: {e2e_stack or \"not configured\"}\n**Result**: X passed, Y failed (sum across batches)\n\n### Batch 1: `<command>`\n**Result**: X passed, Y failed\n**Fixes applied**: ...\n\n### Batch 2: `<command>`\n...\n\n**Failures escalated as implementation issues**: N\n\n## Overall Summary\n- Total test fixes applied: N\n- Suspected implementation issues found: N\n- Implementation issues document: {docs_dir}/testing/implementation-issues-{run_timestamp}.md\n  (or \"not created — no issues found\")\n```\n\n### Document 2: Implementation Issues (only write if issues were found)\n\nIf at least one failure was escalated as an implementation-code issue, write to:\n`{docs_dir}/testing/implementation-issues-{run_timestamp}.md`\n\n```markdown\n# Suspected Implementation Issues: {run_timestamp}\n\nThese test failures were NOT fixed. They may indicate bugs in production code.\nA developer should investigate each item before merging.\n\n## Issue 1\n- **Test**: `path/to/test_file::test_function_name`\n- **Tier**: unit | e2e\n- **Failure message**: (paste the key assertion or exception line)\n- **Why not fixed**: (brief reasoning, e.g., \"production function raises KeyError on valid input\")\n\n## Issue 2\n...\n```\n\nIf no implementation issues were found, do NOT create this file.\n\n## Final Output\n\nAfter writing all documents, print this summary:\n\n```\nTest run complete: {run_timestamp}\nReport saved to: {docs_dir}/testing/test-run-{run_timestamp}.md\nImplementation issues: {docs_dir}/testing/implementation-issues-{run_timestamp}.md (if applicable)\nNo suspected implementation issues found. (if none)\n```\n",
     "scan-test-coverage.md": "Scan recently shipped tickets from git history and report which features have or could gain integration tests, and which can only be smoke tested.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nScan the git history for recently shipped tickets and, for each shipped feature, determine whether it already has an integration test, whether it *could* gain one (per this repo's conventions — a test that genuinely executes the system end-to-end via real database operations, real LLM calls, or real FastAPI routing), and — where integration testing is not possible — how it could be smoke tested so it still genuinely executes the system.\n\nThis is an **investigation and discovery** command. Describe features (citing code) and *how* they would be tested at a high level. Do **not** design tests in detail, build or edit any tests, or modify feature code. Orchestrate this run in the main thread, and **fan out one subagent per shipped feature** for the per-feature investigation.\n\nThe report is written to a **durable, committed** directory (`docs/test-coverage/`), and a marker file records when the analysis last ran so subsequent runs only inspect git history since the last run.\n\n## Stage 0 — Parse Arguments and Resolve Analysis Window\n\n1. Read `$ARGUMENTS`. All flags are optional and default-safe. If a flag is malformed, ignore it and add a warning:\n   - `--since=YYYY-MM-DD` — override the window start date.\n   - `--full` — ignore the marker and use a default lookback of 6 months.\n   - `--limit=N` — cap the number of features investigated (parse `N` as an integer; ignore if not a valid integer).\n   With no arguments, run **incrementally** from the marker.\n\n2. Set the durable directory to `docs/test-coverage/` (relative to the repo root) and the marker file to `docs/test-coverage/STATE.md`. This command deliberately does **not** use the configured docs directory (`ping`'s `docs_dir`) — its default (`docs/tmp`) is ephemeral, and this report must be durable.\n\n3. Read `docs/test-coverage/STATE.md` if it exists. It records two values: `last_run_utc` (an ISO-8601 UTC timestamp) and `last_analyzed_commit` (a git commit SHA).\n\n4. Resolve the analysis window with this precedence:\n   - If `--since=YYYY-MM-DD` was given, use `git log --since=<date>`.\n   - Else if `STATE.md` provides `last_analyzed_commit`, use the commit range `<last_analyzed_commit>..HEAD`.\n   - Else (first run, no marker), default to `git log --since=<3 months ago>` (mirrors the `/scan-tickets` default of 3 months). Format the date as `YYYY-MM-DD`. Example: if today is 2026-07-07, the default `--since` is `2026-04-07`.\n   - `--full` overrides the above and uses a 6-month lookback (`--since=<6 months ago>`).\n\n5. Robustness of the marker: capture `head_sha` by running `git rev-parse HEAD`, and capture the current UTC timestamp now. These become the **new** marker values, but only write them after the report is successfully produced (Stage 4). If a stored `last_analyzed_commit` is not present in history (e.g. a rebase/rewrite), fall back to `git log --since=<the date part of last_run_utc>` and add a warning noting the fallback.\n\n6. Initialize tracking variables:\n   - `features` = [] (one entry per shipped feature)\n   - `warnings` = [] (per-item failures and fallbacks; the run never aborts on these)\n\n7. Display the resolved window, e.g. \"Analyzing shipped features in `<range or --since date>` (HEAD = {head_sha})\".\n\n## Stage 1 — Collect Shipped Features from Git History\n\n1. List merged commits in the resolved window with:\n   ```bash\n   git log <range> --first-parent --pretty=format:\"%H|%h|%ad|%s\" --date=short\n   ```\n   `--first-parent` yields roughly one entry per squashed PR merge.\n\n2. For each commit, extract the ticket key by matching `^BAPI-[0-9]+` against the subject. Group commits by ticket key. Commits with no ticket prefix (e.g. `Fix 500 on ...`) each become a standalone feature labeled as an \"untracked change\".\n\n3. For each group, collect the changed-file footprint across its commit(s) using `git show --stat <sha>` or `git diff --name-only`. This file footprint is the primary input to the per-feature investigation.\n\n4. Best-effort enrichment: for each ticket key, call the `get_ticket` MCP tool to fetch the ticket summary. This is **fail-open** — Jira tokens can be expired — so on any error, add a warning and continue without the summary. Do not abort.\n\n5. Build a `features` entry per group: `{ticket_key, subject, commit_shas, changed_files, jira_summary?}`. If `--limit=N` was given, keep only the first `N` features (most recent first).\n\n6. Display: \"Found {count} shipped features to investigate.\"\n\n7. If `git log` returns no commits, skip to Stage 4 and write a report noting an empty window (and still refresh the marker).\n\n## Stage 2 — Investigate Each Feature (fan out subagents)\n\nFor each feature in `features`, launch an **Explore** subagent (batch several in parallel). Give each subagent the feature's `ticket_key`, `subject`, `changed_files`, and `jira_summary`, and instruct it to do read-only investigation only — no edits, no test design, no solutioning — and to return a structured finding.\n\nEach subagent must:\n\n1. Read the changed files and describe what the feature does in 2–4 sentences, with concrete `file:line` citations.\n\n2. Identify the feature's runtime surface — one or more of: real database operations (`postgres_client` / a DAL in `api/library/db/`), real LLM calls (`src/python/llms/ai_client.py`, `async_send_message_to_ai`), real FastAPI routing (a route handler under `api/routes/`), an MCP tool (`mcp_server/`), a shell-spawned / CLI flow, a frontend / Playwright surface, or pure logic / config / docs / tests.\n\n3. Check whether an **integration test already exists**: search `tests/integration/` for a mirror path or for references to the changed modules/functions. The reliable classifier is a path under `tests/integration/` plus `@pytest.mark.integration` or reliance on the `--run-integration` flag (conventions in `docs/claude/testing-integration.md`). Cite any test found.\n\n4. Classify the feature into exactly one `bucket`:\n   - **`has_integration_test`** — already covered end-to-end; cite the existing integration test file.\n   - **`integration_testable`** — no test yet, but the feature exercises real DB / LLM / routing and fits an existing `tests/integration/<area>/` pattern. Give a **high-level** approach only: which real entrypoint to call, which backend it would exercise, and the relevant cost/guard note (the gpt-5-nano override via `INTEGRATION_TEST_MODEL`; the local-DB `skipif` guard; `save_to_db=False`). Cite the entrypoint in code.\n   - **`smoke_only`** — genuine end-to-end execution is possible but not as an automated integration test (e.g. MCP tool behavior inside a host, cross-platform terminal spawning, a headless agent session, or browser E2E). Describe how to smoke test it so it **genuinely executes the system**, citing the relevant runbook: the MCP smoke-test runbook under `mcp_server/smoke-test/`, `tests/mcp/`, `docs/claude/runbooks/self-install-smoke-test.md`, `docs/claude/runbooks/start-tickets-smoke-test.md`, or Playwright (`tests/playwright/`, which needs a running server plus `npm run build`).\n   - **`not_testable`** — nothing to execute end-to-end (docs-only, a wording/comment change, pure config, or a test-only change); state why.\n\n5. Return a structured finding with these fields: `ticket_key`, `subject`, `description_with_cites`, `surface`, `bucket`, `existing_test`, `approach`, `why_not`.\n\nCollect all findings. If a per-feature subagent fails, add a warning and continue — never abort the whole run.\n\n## Stage 3 — Classify and Synthesize\n\n1. Deduplicate features that span multiple commits (merge by `ticket_key`).\n\n2. Sort each finding into the two required report sections:\n   - **Section 1 — Integration Testing (covered or addable):** findings with `bucket` `has_integration_test` (sub-group \"Already covered\") or `integration_testable` (sub-group \"Could be added\").\n   - **Section 2 — Not Integration-Testable:** findings with `bucket` `smoke_only` (sub-group \"Smoke-testable — how\") or `not_testable` (sub-group \"Not testable — why\").\n\n## Stage 4 — Write the Report and Update the Marker\n\n1. Create the `docs/test-coverage/` directory if it does not exist. Choose the report path `docs/test-coverage/REPORT-<YYYYMMDD>.md`; if a same-day file already exists, append `-<HHMMSS>` to avoid clobbering it.\n\n2. Write the report with this layout:\n   - A title and a metadata block: generated-at UTC timestamp; the analysis window (`<from sha or since-date>` → `HEAD <head_sha>`); the feature count; and per-bucket tallies.\n   - **Section 1 — Integration Testing: Covered or Addable.** One `### BAPI-NNN — <subject>` heading per feature, each with **What shipped** (with `file:line` citations), **Current coverage** (cite the existing integration test, or state \"none\"), and **How it could be integration tested (high level)**.\n   - **Section 2 — Not Integration-Testable.** One heading per feature with the same feature description, plus **Why not integration-testable**, and — for `smoke_only` features — **How to smoke test (genuinely execute the system)** with the runbook citation.\n   - A **Warnings** section listing each warning as a bullet — only if `warnings` is non-empty.\n\n3. **Only after** the report file is written successfully, update the marker `docs/test-coverage/STATE.md` with the new `last_run_utc` (the UTC timestamp captured in Stage 0) and `last_analyzed_commit` set to `head_sha`. This date/commit marker is what makes the next run incremental. If the report write fails, do not touch `STATE.md`.\n\n## Final Report\n\nPrint a short summary to chat:\n\n```\n**Test-coverage scan complete**\n\n* Features analyzed: {count}\n* Already covered by integration tests: {n_has}\n* Integration-testable (could be added): {n_addable}\n* Smoke-only: {n_smoke}\n* Not testable: {n_none}\n\nReport: docs/test-coverage/REPORT-<YYYYMMDD>.md\nMarker updated: last_analyzed_commit = {head_sha}\n```\n\nIf `warnings` is non-empty, add a \"Warnings:\" section listing each warning as a bullet. If there are no warnings, omit that section.\n",
     "scan-tickets.md": "$ARGUMENTS\n\n---\n\n# Instructions\n\nSynchronize recently-updated Jira tickets with the local `tickets` database table and backfill missing workflow state timestamps. Perform all work directly in the main thread.\n\n## Stage 0 — Parse Arguments and Calculate Date\n\n1. Read the value of `$ARGUMENTS`. If it is empty, whitespace-only, or not a valid integer, default `months_back` to `3`. If it contains multiple tokens, extract only the first token and attempt to parse it as an integer. If parsing fails, default to `3`.\n\n2. Calculate `updated_since` by subtracting `months_back` months from today's date. Format the result as `YYYY-MM-DD`. Example: if today is 2026-03-07 and `months_back` is 3, then `updated_since` is 2025-12-07.\n\n3. Display the parsed values: \"Scanning tickets updated since {updated_since} (months_back = {months_back})\"\n\n4. Initialize the following tracking variables:\n   - `tickets_scanned` = 0 (total tickets fetched from Jira)\n   - `newly_tracked` = 0 (tickets inserted into database for the first time)\n   - `state_updated_list` = [] (list of objects with ticket key and fields updated)\n   - `warnings` = [] (list of warning strings for any per-ticket failures)\n\n## Stage 1 — Fetch All Tickets from Jira\n\n1. Initialize an empty list `all_tickets` and set `offset` to `0`.\n\n2. Enter a pagination loop:\n   - Call the `get_tickets` MCP tool with: `updated_since` set to the calculated date, `limit` set to `100`, and `offset` set to the current offset value.\n   - Parse the JSON response. The response contains a `tickets` array of ticket objects. Each ticket object has a `ticket_number` field (the Jira key, e.g., `BAPI-42`), along with `summary`, `status`, `issue_type`, `assignee`, and `updated_at`.\n   - Append all tickets from the response's `tickets` array to `all_tickets`.\n   - If the number of tickets returned in this page equals `100`, increment `offset` by `100` and repeat the loop.\n   - If fewer than `100` tickets are returned, exit the loop.\n\n3. Set `tickets_scanned` to the length of `all_tickets`.\n\n4. Display: \"Fetched {tickets_scanned} tickets from Jira. Processing...\"\n\n5. If the `get_tickets` call fails at any point during pagination, **stop** and report the error. Do not proceed to Stage 2.\n\n## Stage 2 — Track Each Ticket\n\n1. Iterate over each ticket in `all_tickets`. For each ticket:\n   - Call the `track_ticket` MCP tool with `ticket_number` set to the ticket's `ticket_number` field. If the ticket object includes a `summary` field, pass it as the `description` parameter.\n   - Inspect the response message. If the response indicates the ticket was newly created/inserted (look for words like \"created\" or \"inserted\" in the message, as opposed to \"already exists\" or \"updated\"), increment `newly_tracked` by 1.\n   - If the `track_ticket` call fails for this ticket, add a warning to the `warnings` list (e.g., \"Warning: Failed to track ticket {ticket_number}: {error}\") and **continue** to the next ticket. Do not abort the scan.\n\n2. Display a brief progress indicator every 25 tickets, e.g., \"Tracked {N} of {tickets_scanned} tickets...\"\n\n## Stage 3 — Detect and Backfill Workflow State\n\nDisplay: \"Checking workflow state for {tickets_scanned} tickets...\"\n\nIterate over each ticket in `all_tickets`. For each ticket (referenced by its `ticket_number` field), perform the following sub-steps. Wrap the entire per-ticket block in error handling: if the `get_ticket_state` call or the subsequent `update_ticket_state` call fails for a ticket, add a warning to `warnings` and continue to the next ticket.\n\n**Sub-step 4a — Retrieve current state**: Call the `get_ticket_state` MCP tool with `ticket_number` set to the ticket's key. The response contains:\n\n- Five timestamp fields (each is a timestamp string or null): `clarify_called`, `clarify_answered`, `critique_called`, `critique_answered`, `plan_generated`\n- Three boolean artifact flags: `has_clarifying_questions`, `has_critique`, `has_plan`\n\nIf the call returns a 404 or any error, add a warning to `warnings` and continue to the next ticket.\n\n**Sub-step 4b — Build fields_to_update list**: Initialize an empty `fields_to_update` list, then apply the following rules:\n\n- If `has_clarifying_questions` is `true` AND `clarify_called` is null -> add `\"clarify_called\"` to `fields_to_update`\n- If `has_clarifying_questions` is `true` AND `clarify_answered` is null -> add `\"clarify_answered\"` to `fields_to_update`\n- If `has_critique` is `true` AND `critique_called` is null -> add `\"critique_called\"` to `fields_to_update`\n- If `has_critique` is `true` AND `critique_answered` is null -> add `\"critique_answered\"` to `fields_to_update`\n- If `has_plan` is `true` AND `plan_generated` is null -> add `\"plan_generated\"` to `fields_to_update`\n\n**Sub-step 4c — Call update_ticket_state if needed**: If `fields_to_update` is non-empty, call the `update_ticket_state` MCP tool with `ticket_number` set to the ticket's key and `fields` set to the `fields_to_update` array. If this succeeds, add an entry to `state_updated_list` recording the ticket key and the list of fields that were set. If `update_ticket_state` fails, add a warning to `warnings` and continue.\n\nDisplay a progress indicator every 25 tickets that includes the current ticket key, e.g., \"Checked state for {TICKET-KEY} ({N} of {tickets_scanned} tickets)\"\n\n## Stage 4 — Report Summary\n\n1. Calculate `state_updated_count` as the length of `state_updated_list`.\n\n2. Display the summary:\n\n   ```\n   **Scan complete**\n\n   * Tickets scanned: {tickets_scanned}\n   * Newly tracked: {newly_tracked}\n   * State updated: {state_updated_count}\n   ```\n\n3. If `state_updated_list` is non-empty, display a section titled \"Updated tickets:\" with one bullet per ticket showing the ticket key and the comma-separated list of fields that were set. Example:\n\n   ```\n   Updated tickets:\n   * BAPI-101: clarify_called, clarify_answered\n   * BAPI-105: critique_called, critique_answered, plan_generated\n   ```\n\n4. If the `warnings` list is non-empty, display a section titled \"Warnings:\" listing each warning string as a bullet. Example:\n\n   ```\n   Warnings:\n   * Warning: Failed to track ticket BAPI-99: Connection timeout\n   * Warning: State query failed for BAPI-112: SQL error\n   ```\n\n5. If there are no warnings, do not display the \"Warnings:\" section.\n",
-    "start-tickets.md": "---\nschedulable: true\narguments: {\"positionals\":[{\"name\":\"ticketKeys\",\"type\":\"string\",\"required\":true,\"variadic\":true}],\"flags\":[{\"name\":\"auto\",\"flag\":\"--auto\",\"type\":\"boolean\"},{\"name\":\"agent\",\"flag\":\"--agent\",\"type\":\"string\"},{\"name\":\"workflow\",\"flag\":\"--workflow\",\"type\":\"string\"},{\"name\":\"rounds\",\"flag\":\"--rounds\",\"type\":\"string\"},{\"name\":\"baseBranch\",\"flag\":\"--base-branch\",\"type\":\"string\"},{\"name\":\"maxParallel\",\"flag\":\"--max-parallel\",\"type\":\"string\"},{\"name\":\"dryRun\",\"flag\":\"--dry-run\",\"type\":\"boolean\"},{\"name\":\"guardStaleBranch\",\"flag\":\"--guard-stale-branch\",\"type\":\"boolean\"}]}\n---\n\n# Start Tickets: $ARGUMENTS\n\n$ARGUMENTS\n\nThis command takes one or more Jira ticket keys (e.g., `BAPI-248 BAPI-250`) and invokes the packaged `@bridge_gpt/mcp-server` CLI subcommand `start-tickets`, which creates a Worktrunk worktree for each key and opens one tab/session per worktree running the **selected agent** — Claude Code (`claude`) by default, or Cursor Agent (`cursor-agent`) via `--agent` — in a macOS Terminal/iTerm tab, a Windows Terminal tab (or PowerShell fallback window), or a detached Linux tmux session, chosen automatically by platform. It replaces Parts 2–5 of `docs/claude/parallel-worktrees.md` with a single command.\n\nBecause the orchestration ships inside the `@bridge_gpt/mcp-server` npm package (not a repo-local script), this command works for every consumer — including projects that installed the package via `--init`.\n\nFor existing ticket keys, `/review-and-start <KEYS>` is the **recommended front door**: it supplies the same connectivity check and branch enrichment as this command, then drives this same packaged CLI with `--workflow review-and-implement` so each worktree reviews the ticket before implementing it. Using `start-tickets --workflow review-and-implement` directly (documented below) remains available as the lower-level launcher seam.\n\nStage 0 and Stage 1 are critical (stop on failure). Stage 2 is non-critical (per-ticket enrichment failures fall back to the default branch and continue). Stage 3 is critical (propagate the packaged CLI's exit code).\n\n---\n\n# Instructions\n\nYou are executing a 4-stage pipeline that spawns N parallel Worktrunk worktrees and selected-agent sessions (Claude Code by default) via the packaged CLI. Execute all stages in sequence directly in the main thread.\n\n## Stage 0 — Argument Parsing and Connectivity Check\n\n1. **Parse `$ARGUMENTS`** into ticket keys, pass-through flags, and branch overrides:\n   - **Ticket keys**: every whitespace-separated token matching `[A-Z]+-[0-9]+` (e.g., `BAPI-248`). If zero keys are found, stop immediately and display:\n\n     ```\n     No ticket keys found in arguments. Expected one or more keys like BAPI-248.\n     Usage: /start-tickets [flags] <KEY> [KEY ...] (e.g., /start-tickets BAPI-248 BAPI-250)\n     ```\n\n   - **Pass-through flags**: collect any of `--agent <name>` (and the equals form `--agent=<name>`), `--terminal terminal|iterm`, `--dry-run`, `--auto`, `--no-refresh-main`, `--base-branch <branch>` (and the equals form `--base-branch=<branch>`), `--max-parallel N`, and `--guard-stale-branch` that the user supplied. These are forwarded verbatim to the CLI in Stage 3. `--guard-stale-branch` turns on the F7 stale-branch guard: a pre-existing `feature/<KEY>` branch whose tip is **not** an ancestor of the resolved base is refused with a `create-failed` row carrying the stale-worktree remedy, instead of being silently reused on top of another run's commits. It is off by default (a plain run keeps reusing a same-named branch) and is implied by `--conductor`. `--auto` makes each spawned agent run the selected workflow's slash command with `--auto` (hands-off); omit it to keep the spawned agents interactive.\n   - **Selected agent**: track a `selected_agent` variable that defaults to `claude`. If the user passed `--agent <name>` / `--agent=<name>`, validate the value against the supported agents `claude` and `cursor-agent`, set `selected_agent` to it, and reject any other (malformed/unsupported) `--agent` value before proceeding. The agent is not auto-detected from the host editor — the user selects it explicitly (default `claude`).\n   - **Selected workflow**: track a `selected_workflow` variable that defaults to `implement`. If the user passed `--workflow <value>` or `--workflow=<value>`, validate it against the two allowed values `implement` and `review-and-implement`, set `selected_workflow`, and reject any other value with the allowlist in the error. `implement` (the default) preserves today's behavior byte-for-byte — each spawned worktree runs `/implement-ticket <KEY> [--auto]`. `review-and-implement` spawns `/review-and-implement <KEY> [--auto] [--rounds=<n>]` instead, which runs `/review-ticket` then, after a per-ticket halt gate, `/implement-ticket` inside the same session. A single chain-level `--auto` applies to the selected workflow as a whole — under `review-and-implement` it auto-approves both the review and the implementation phase.\n   - **Review rounds**: track a `review_rounds` value that defaults to unset. If the user passed `--rounds <n>` or `--rounds=<n>`, normalize it to `--rounds=1` or `--rounds=2` (reject any other value). `--rounds` is **review-only**: reject it (after parsing all flags, so flag order does not matter) if the final `selected_workflow` is not `review-and-implement`.\n   - **User-supplied base branch**: track a `user_supplied_base_branch` boolean that defaults to `false`. If the user passed `--base-branch <branch>` or `--base-branch=<branch>`, set the boolean to `true` and capture the value. A user-supplied `--base-branch` value **takes precedence** over any value resolved from Bridge API config in Stage 2. Validate the user-supplied value before proceeding: after trimming surrounding whitespace it must be non-empty, at most 255 characters, must not start with `-`, and must not contain ASCII control characters (`0x00`–`0x1F` or `0x7F`); reject any malformed value with a clear error.\n   - **User branch overrides**: collect any user-supplied repeatable `--branch KEY=BRANCH` flags. A user-provided override always takes precedence over Stage 2 enrichment for that key.\n   - Reject malformed input before proceeding: if a token looks like a flag but is not one of the supported flags, or a ticket key does not match `[A-Z]+-[0-9]+`, or a `--branch` value is not `KEY=BRANCH`, or `--agent` names an agent other than `claude`/`cursor-agent`, or `--workflow` names anything other than `implement`/`review-and-implement`, or `--rounds` is used outside `review-and-implement` or names anything other than `1`/`2`, or `--base-branch` fails the validation rules above, stop and report the malformed argument.\n\n2. **Connectivity check**: Call the `ping` MCP tool (no parameters). If the ping fails or does not return `\"status\": \"ok\"`, stop immediately and display:\n\n   ```\n   Connectivity check failed. Please verify:\n   - Check that the Bridge API MCP server is configured in your editor's MCP settings\n   - Check that BAPI_BASE_URL is set and the server is reachable\n   - Check that BAPI_API_KEY is valid\n   - Check that BAPI_REPO_NAME matches a configured repository\n   ```\n\nThis stage is **critical** — stop immediately on failure. Do not proceed to Stage 1.\n\n## Stage 1 — Acknowledge CLI Pre-flight\n\nThe packaged CLI runs its own per-platform pre-flight checks and then fetches `origin` and fast-forwards the local **configured base branch** (the value resolved in Stage 2 below, or `main` when none is configured) from `origin/<base>` so the new worktrees are based on an up-to-date base. The historical flag `--no-refresh-main` still controls this behavior — the flag name is preserved for backward compatibility, but it now skips refresh of whatever base branch resolves (default `main`). The required commands depend on the OS:\n\n- **macOS**: `wt`, `git`, `osascript`.\n- **Windows**: `git-wt`, `git`, Git for Windows / Git Bash (Worktrunk runs its `pre-start` / `post-start` hooks via Git Bash), and Windows Terminal **or** PowerShell.\n- **Linux**: `wt`, `git`, `tmux`.\n\nOn **Windows** the Worktrunk binary is `git-wt` (its winget alias), which is a different tool from Windows Terminal's `wt.exe`: the CLI uses `git-wt` to **create worktrees** and `wt.exe` to **open a tab**, and never conflates the two. On **Linux** the CLI opens one detached `tmux` session per ticket (a window is added if that ticket's session already exists); attach later with `tmux attach -t <session>`. An unsupported OS (not macOS/Windows/Linux) fails fast with a clear \"unsupported platform\" message.\n\nThis stage simply notes that the CLI will fail fast if any prerequisite is missing or if local `main` has diverged from `origin/main` — you do not need to verify anything separately here, and you must not run any pre-flight commands yourself. When the CLI's pre-flight fails it now hints the user to run the read-only diagnostics command `npx -y @bridge_gpt/mcp-server doctor`, which reports found/missing for every prerequisite on the current OS — the pre-flight set plus `uv` plus the selected agent's command — and prints the manual install command for each missing one. `doctor` is strictly read-only and never installs anything; never run install commands automatically on the user's behalf. The CLI does not call any Bridge API tools; all credential-bearing work (branch enrichment in Stage 2) stays in this command. Proceed to Stage 2.\n\nThe packaged CLI also performs **secret-free Bridge API MCP provisioning** inside each created worktree: synchronously after the worktree is created and **before the agent tab/session is opened**, it writes both `.mcp.json` (Claude Code) and `.cursor/mcp.json` (Cursor) pointing at the `mcp-invoke` shim. These registrations are **secret-free** — they contain no `env` block and no API key, because the shim resolves credentials at runtime. If a spawned agent (or difficulty→model routing) reports missing Bridge API credentials, fix it by rerunning `/install-bridge` (its final stage persists the routing credential), by running `npx -y @bridge_gpt/mcp-server credentials migrate-agent-config --write-credentials` to migrate a key that lives only in `.mcp.json` / `.cursor/mcp.json`, or by adding a `bapi:<repo>` entry to the user-scoped credentials file (`~/.config/bridge/credentials.json`) — never by putting `BAPI_API_KEY` into the worktree `.mcp.json` or `.cursor/mcp.json` (that env is invisible to the Bash-spawned CLI).\n\nThis stage is **critical** in the sense that the CLI will abort if its pre-flight fails; you will see the error in Stage 3's output and must surface it.\n\n## Stage 2 — Resolve Base Branch + Enrich Branch Names (best-effort)\n\n### Stage 2a — Resolve configured `base_branch`\n\nThe CLI must be told which branch to cut new worktrees from. Resolution order:\n\n1. If `user_supplied_base_branch` from Stage 0 is `true`, **skip the config-field lookup entirely** and use the user-supplied value. The user's explicit `--base-branch` always wins; never call `config_field` for `base_branch` in that case.\n2. Otherwise, call the `config_field` MCP tool with `operation` set to `\"get\"` and `field_name` set to `base_branch` (do not pass any other parameters; the tool resolves the repository from the MCP server's configured `BAPI_REPO_NAME`).\n3. Parse the response. Treat the result as the **configured base branch** only when the response is a JSON object whose `value` field is a non-empty string after trimming surrounding whitespace.\n4. Treat **all** of the following as \"unset\" — emit a single-line warning like `Warning: base_branch is unset; CLI will default to main` and **omit** the `--base-branch` flag entirely from the Stage 3 command (the CLI's own default is `main`):\n   - `value` is `null`.\n   - `value` is an empty string or a whitespace-only string.\n   - The endpoint returns HTTP `400` (invalid field — happens before the registry includes `base_branch`).\n   - The tool returns a network error, timeout, or non-JSON parse failure.\n   - Any other lookup failure.\n5. When the configured value is usable, capture it in a `resolved_base_branch` variable. **Do not** stop the pipeline on a lookup failure; fall through to the CLI default.\n\nWhen forwarding `resolved_base_branch` into the Bash invocation in Stage 3, **shell-escape it safely**: replace every literal single quote `'` in the value with the four-character sequence `'\\''`, then wrap the entire resulting string in single quotes (so the final argument looks like `'<escaped-value>'`). This is the standard POSIX single-quote escaping rule and is **mandatory** because `base_branch` is admin-configurable data that gets interpolated into a Bash command string; any unescaped single quote would otherwise break out of the surrounding quotes. Pass `--base-branch '<escaped-value>'` to the CLI as a single argv element — never expand the value unquoted into the command line.\n\n### Stage 2b — Enrich Branch Names\n\nBranch enrichment happens here, in the command, **before** invoking the CLI — the `get_ticket` MCP tool runs inside the MCP server process, which holds the Bridge API credentials the shell-spawned CLI does not have. For each parsed ticket key that does **not** already have a user-provided `--branch` override:\n\n1. Call the `get_ticket` MCP tool with `ticket_number` set to the key and `save_locally` set to `false`.\n2. From the response, extract the `summary` field. Slugify it: lowercase the string, replace every run of non-alphanumeric characters (`[^a-z0-9]+`) with a single dash `-`, trim leading and trailing dashes, and truncate to at most `40` characters (cutting at a dash boundary if possible).\n3. The enriched branch name is `feature/<KEY>-<slug>`. Example: `BAPI-248` with summary `\"Add PR rating pre-evaluation step\"` becomes `feature/BAPI-248-add-pr-rating-pre-evaluation-step` (trimmed at 40 chars).\n4. If the `get_ticket` call fails for a particular key (404, network error, missing summary) or produces an empty slug, emit a single-line warning like `Warning: could not enrich BAPI-248, falling back to feature/BAPI-248` and let the CLI apply its default `feature/<KEY>` for that key only. Do NOT stop the pipeline.\n5. Build a list of `--branch <KEY>=<BRANCH>` arguments — one entry per key whose enrichment succeeded — and merge it with any user-provided overrides from Stage 0. **Do not** call `get_ticket` for keys that already have a user-provided override; those overrides win.\n\nThis stage is **non-critical** — warnings are acceptable, the pipeline continues with the fallback default for any key that fails. Do not call the Bridge API from the CLI itself; the CLI never has credentials.\n\n## Packaged CLI launcher (`BAPI_MCP_CLI`)\n\nResolve the packaged-CLI launcher **once**, before the first shell-out below, and reuse that one resolved value for every packaged-CLI invocation in this command. Call it `<launcher>`.\n\n- Read the `BAPI_MCP_CLI` environment variable.\n- **Unset, empty, or whitespace-only** — `<launcher>` is exactly `npx -y @bridge_gpt/mcp-server`. This is the default, and the resulting shell command is byte-identical to what it was before this override existed.\n- **Otherwise** — `<launcher>` is that value, used verbatim as the command prefix. It names a local launcher, such as `node /absolute/path/to/mcp_server/build/index.js`. Use it for local pilots and pre-publish verification.\n\nWhen the override is set:\n\n- Apply this command's mandatory single-quote escaping rule (`'` → `'\\''`, then wrap the whole value in single quotes) before interpolating `<launcher>` into a Bash command string. Never expand it unquoted.\n- Keep every dynamic argument — ticket keys, branch names, base branches, file paths — independently quoted. Never concatenate an argument into the launcher value.\n- Never put a credential, an API key, or an environment assignment carrying one into the launcher value, an example, or a dry-run preview.\n\nWorked examples and printed remediation hints below show the **unset** resolution — the literal `npx -y @bridge_gpt/mcp-server` — because that is the default every operator gets. A stale local build is exactly as misleading as a stale npm publish: rebuild with `cd mcp_server && npm run build` before relying on the override.\n\n## Stage 3 — Invoke the Packaged CLI\n\nUse the **Bash tool** to invoke the packaged CLI. Build the command line as:\n\n```\n<launcher> start-tickets <pass-through-flags> <base-branch-flag> <branch-overrides> <ticket-keys>\n```\n\nWhere:\n- `<pass-through-flags>` are the supported flags collected in Stage 0 (`--agent`, `--terminal`, `--dry-run`, `--auto`, `--no-refresh-main`, `--max-parallel`, `--guard-stale-branch`), forwarded verbatim. Forward `--agent <name>` only if the user supplied it; otherwise omit it and the CLI defaults to `claude`. Forward `--auto` only if the user supplied it.\n- Forward `--workflow <selected_workflow>` only when the user explicitly passed `--workflow`; otherwise omit it and the CLI defaults to `implement`. Forward the normalized `--rounds=<n>` from Stage 0 only when the user supplied it (which Stage 0 already guarantees is only possible under `review-and-implement`).\n- `<base-branch-flag>` is `--base-branch '<escaped-value>'` (single-quoted using the Stage 2a escaping rule) **only when** the user supplied `--base-branch` in Stage 0 **or** Stage 2a's `config_field` lookup returned a non-empty configured value. When the configured value is unset / lookup fails / user did not supply one, **omit this flag entirely** so the CLI's own default (`main`) takes effect.\n- `<branch-overrides>` is the list of `--branch KEY=BRANCH` flags assembled in Stage 2 (enrichment results merged with user overrides; omit any key whose enrichment failed and had no user override).\n- `<ticket-keys>` is the original list of ticket keys parsed in Stage 0, space-separated and in the original order.\n\nExample for two tickets after successful enrichment, throttled to 2 concurrent worktrees:\n\n```\nnpx -y @bridge_gpt/mcp-server start-tickets \\\n  --max-parallel 2 \\\n  --branch BAPI-248=feature/BAPI-248-add-pr-rating-pre-evaluation-step \\\n  --branch BAPI-250=feature/BAPI-250-deep-research-durability \\\n  BAPI-248 BAPI-250\n```\n\nExample launching Cursor Agent instead of the default Claude Code:\n\n```\nnpx -y @bridge_gpt/mcp-server start-tickets --agent cursor-agent BAPI-248\n```\n\nExample cutting worktrees from a non-`main` base (either user-supplied via `--base-branch develop` in Stage 0 or resolved from Bridge API config in Stage 2a):\n\n```\nnpx -y @bridge_gpt/mcp-server start-tickets --base-branch develop BAPI-248\n```\n\nExample using the lower-level review-and-implement workflow directly (the `/review-and-start` command is the recommended front door for this; this form is documented here as the advanced launcher seam it drives):\n\n```\nnpx -y @bridge_gpt/mcp-server start-tickets --workflow review-and-implement --auto --rounds=2 BAPI-248\n```\n\nPass through the CLI's stdout and stderr to the user verbatim. If the CLI exits non-zero, treat that as a critical failure: report the exit code and the CLI's error output, and stop.\n\nThis stage is **critical** — propagate any non-zero exit from the packaged CLI.\n\n## Stage 4 — Final Report\n\nOnce the CLI exits 0, parse its `Summary` section (one stable line per ticket in the form `KEY  branch=BRANCH  status=STATUS`, with an optional trailing `path=PATH`) and reformat it as a markdown table:\n\n```\n| Ticket   | Branch                                              | Status   |\n|----------|-----------------------------------------------------|----------|\n| BAPI-248 | feature/BAPI-248-add-pr-rating-pre-evaluation-step  | spawned  |\n| BAPI-250 | feature/BAPI-250-deep-research-durability           | spawned  |\n```\n\nStatus values are `dry-run`, `spawned`, `create-failed`, and `spawn-failed`. This table (and the report as a whole) describes **worktree/spawn status only** — it must never claim that review or implementation itself has completed; that work happens later, independently, inside each spawned session.\n\nCompute `spawned_command` from `selected_workflow`: `/implement-ticket <KEY>` when `implement` (the default), or `/review-and-implement <KEY>` when `review-and-implement`. Append `--auto` when the user passed it, and (workflow `review-and-implement` only) append the normalized `--rounds=<n>` when the user supplied `--rounds`. End the report with the worktree-first explanation, rendered for the tracked `selected_agent` and `spawned_command`. When `selected_agent` is `claude` (the default):\n\n```\nThe CLI created/switched each Worktrunk worktree first (throttled by --max-parallel),\nthen opened one tab/session per successful worktree (macOS Terminal/iTerm tab, Windows\nTerminal tab or PowerShell window, or Linux tmux session). Each one runs\n`claude '<spawned_command>'` inside its already-created worktree, which launches\nClaude Code with the starter prompt as its first message. Switch to each tab — or on\nLinux run `tmux attach -t <session>` — to monitor.\n```\n\nWhen `selected_agent` is `cursor-agent`, render the same explanation but with the Cursor handoff — do **not** claim it launches Claude Code:\n\n```\nThe CLI created/switched each Worktrunk worktree first (throttled by --max-parallel),\nthen opened one tab/session per successful worktree (macOS Terminal/iTerm tab, Windows\nTerminal tab or PowerShell window, or Linux tmux session). Each one runs\n`cursor-agent '<spawned_command>'` inside its already-created worktree, which\nlaunches Cursor Agent with the starter prompt as its first message. Switch to each\ntab — or on Linux run `tmux attach -t <session>` — to monitor.\n```\n\nThe spawned command is identical for both agents; only the launched agent binary differs. Under `review-and-implement`, each spawned session independently runs `/review-ticket`, pauses at its own per-ticket halt gate (unless chain-level `--auto` was passed), and only then runs `/implement-ticket` — do not report that review or implementation succeeded from this parent session.\n\nIf the CLI reported any `create-failed` or `spawn-failed` statuses, or Stage 2 emitted any enrichment warnings, list them under a `Warnings:` heading at the bottom of the report. If there were none, omit that section.\n\nSee `docs/claude/parallel-worktrees.md` for the deep-dive runbook and the Worktrunk verification result behind this worktree-first model.\n\n## Difficulty-Based Implementation-Model Routing\n\nBefore launching the interactive agent for each ticket, the packaged CLI selects an\nimplementation **model tier** from the ticket's `difficulty` rating (1-10) and injects\nit as a `--model` flag at the agent spawn boundary. This happens entirely inside the\nCLI — it is **not** part of the server-side `/implement-ticket` recipe, because the\nmodel an interactive agent session uses is fixed at the moment the process is launched.\n\n- **Tier ladder (fixed):** `difficulty 1-2 → cheap`, `3-6 → basic`, `7-10 → premium`.\n- **Separation of concerns:** the Python backend returns only the coarse tier\n  (`cheap`/`basic`/`premium`) via `GET /jira/tickets/{KEY}/model-tier`; difficulty is\n  computed on demand and cached when absent. The TypeScript CLI alone maps a tier to\n  the agent-specific model alias (`claude`: `haiku`/`sonnet`/`opus`; `cursor-agent`:\n  version-suffixed strings validated against `cursor-agent --list-models`).\n- **Per-repo config:**\n  - `difficulty_model_routing_enabled` — boolean, **default ON**. Set to `false` to\n    disable routing for a repo (the CLI then omits `--model`).\n  - `difficulty_model_tier_overrides` — a JSON object mapping a tier name to a model\n    alias (e.g. `{\"premium\": \"opus\"}`), **not** raw CLI arguments. Only `cheap`,\n    `basic`, and `premium` keys are accepted; aliases must match `^[A-Za-z0-9._:-]+$`.\n- **Fail-open:** routing never aborts a spawn. Credential, network, config, or\n  no-tier routing failures **assume a hard ticket and default to the premium/Opus\n  tier** when the selected agent supports a valid premium alias; routing being\n  disabled (`difficulty_model_routing_enabled = false`) or an agent that does not\n  support `--model` instead omit `--model` so the agent runs on its own default\n  model. Each degraded case is surfaced as exactly one secret-free, per-ticket\n  routing-diagnostic line, never a hard failure.\n\n### Model routing credential\n\nDifficulty→model routing needs Bridge API credentials, and the shell-spawned\n`start-tickets` CLI is a **different runtime surface** from the MCP server: a\n`BAPI_API_KEY` that lives only in `.mcp.json` / `.cursor/mcp.json` is visible to\nthe MCP server but **not** to the Bash-spawned CLI, so routing silently degrades.\nThe durable source of truth both runtimes can resolve is the user-scoped store\n`~/.config/bridge/credentials.json`, keyed `bapi:<repo>`. If a routing-diagnostic\nline reports the credential is missing (e.g. difficulty resolves as `?`), fix it\nby any one of:\n\n1. Rerun `/install-bridge` — its final stage now persists the validated routing\n   credential into `~/.config/bridge/credentials.json` via the\n   `persist_routing_credential` tool.\n2. Migrate a key that lives **only** in `.mcp.json` / `.cursor/mcp.json` into the\n   user-scoped store with the consent-gated, one-shot command:\n\n   ```\n   npx -y @bridge_gpt/mcp-server credentials migrate-agent-config --write-credentials\n   ```\n\n3. Manually add `BAPI_API_KEY` under the `bapi:<repo>` target in the user-scoped\n   store `~/.config/bridge/credentials.json`.\n\nNever put `BAPI_API_KEY` into a worktree `.mcp.json` / `.cursor/mcp.json` as a fix —\nthat env is invisible to the spawned CLI.\n\n## Conductor observability (opt-in via `--conductor`, BAPI-394)\n\nConductor is **opt-in**. By default `start-tickets` spawns the plain\n`cd <worktree> && <agent> '/implement-ticket <KEY> [--auto]'` — no\n`BAPI_CONDUCTOR_*` env, no supervisor window, and no message-relay instruction.\nPass `--conductor` (e.g. `/start-tickets --conductor BAPI-123`) to enable the\nConductor system below.\n\nWith `--conductor`, a run mints a single conductor `run_id` and attributes each\nworker's lifecycle events by `worker_id`, ticket key, and worktree path, and a\nsupervisor peer tab is opened. When the selected agent is **Claude Code**, the CLI\ninjects a conductor lifecycle hook into each created worktree's\n`.claude/settings.local.json` so the spawned session emits local `run.started` /\n`run.stopped` / `agent.notification` (and, when\n`BAPI_CONDUCTOR_ENABLE_PRE_TOOL_USE=1`, `tool.intent`) events into the local\nconductor ledger. These hooks apply **only** when the selected agent is Claude\nCode; other agents (e.g. `cursor-agent`) still participate in the run-level\n`run.started` event but receive no per-worktree Claude hook. Inspect the ledger\nwith the `conductor` CLI (e.g. `conductor doctor`). Conductor observability is\nbest-effort and never blocks or aborts a spawn.\n\nObservability under `--conductor` is one-directional: workers emit lifecycle events\ninto the local ledger and nothing is passed back into a running session. (Epic-tick\ndispatch always runs with conductor enabled, independent of this user-facing flag.)\n",
+    "start-tickets.md": "---\nschedulable: true\narguments: {\"positionals\":[{\"name\":\"ticketKeys\",\"type\":\"string\",\"required\":true,\"variadic\":true}],\"flags\":[{\"name\":\"auto\",\"flag\":\"--auto\",\"type\":\"boolean\"},{\"name\":\"agent\",\"flag\":\"--agent\",\"type\":\"string\"},{\"name\":\"workflow\",\"flag\":\"--workflow\",\"type\":\"string\"},{\"name\":\"rounds\",\"flag\":\"--rounds\",\"type\":\"string\"},{\"name\":\"baseBranch\",\"flag\":\"--base-branch\",\"type\":\"string\"},{\"name\":\"maxParallel\",\"flag\":\"--max-parallel\",\"type\":\"string\"},{\"name\":\"dryRun\",\"flag\":\"--dry-run\",\"type\":\"boolean\"},{\"name\":\"guardStaleBranch\",\"flag\":\"--guard-stale-branch\",\"type\":\"boolean\"}]}\n---\n\n# Start Tickets: $ARGUMENTS\n\n$ARGUMENTS\n\nThis command takes one or more Jira ticket keys (e.g., `BAPI-248 BAPI-250`) and invokes the packaged `@bridge_gpt/mcp-server` CLI subcommand `start-tickets`, which creates a Worktrunk worktree for each key and opens one tab/session per worktree running the **selected agent** — Claude Code (`claude`) by default, or Cursor Agent (`cursor-agent`) via `--agent` — in a macOS Terminal/iTerm tab, a Windows Terminal tab (or PowerShell fallback window), or a detached Linux tmux session, chosen automatically by platform. It replaces Parts 2–5 of `docs/claude/parallel-worktrees.md` with a single command.\n\nBecause the orchestration ships inside the `@bridge_gpt/mcp-server` npm package (not a repo-local script), this command works for every consumer — including projects that installed the package via `--init`.\n\nFor existing ticket keys, `/review-and-start <KEYS>` is the **recommended front door**: it supplies the same connectivity check and branch enrichment as this command, then drives this same packaged CLI with `--workflow review-and-implement` so each worktree reviews the ticket before implementing it. Using `start-tickets --workflow review-and-implement` directly (documented below) remains available as the lower-level launcher seam.\n\nStage 0 and Stage 1 are critical (stop on failure). Stage 2 is non-critical (per-ticket enrichment failures fall back to the default branch and continue). Stage 3 is critical (propagate the packaged CLI's exit code).\n\n---\n\n# Instructions\n\nYou are executing a 4-stage pipeline that spawns N parallel Worktrunk worktrees and selected-agent sessions (Claude Code by default) via the packaged CLI. Execute all stages in sequence directly in the main thread.\n\n## Stage 0 — Argument Parsing and Connectivity Check\n\n1. **Parse `$ARGUMENTS`** into ticket keys, pass-through flags, and branch overrides:\n   - **Ticket keys**: every whitespace-separated token matching `[A-Z]+-[0-9]+` (e.g., `BAPI-248`). If zero keys are found, stop immediately and display:\n\n     ```\n     No ticket keys found in arguments. Expected one or more keys like BAPI-248.\n     Usage: /start-tickets [flags] <KEY> [KEY ...] (e.g., /start-tickets BAPI-248 BAPI-250)\n     ```\n\n   - **Pass-through flags**: collect any of `--agent <name>` (and the equals form `--agent=<name>`), `--terminal terminal|iterm`, `--dry-run`, `--auto`, `--no-refresh-main`, `--base-branch <branch>` (and the equals form `--base-branch=<branch>`), `--max-parallel N`, and `--guard-stale-branch` that the user supplied. These are forwarded verbatim to the CLI in Stage 3. `--guard-stale-branch` turns on the F7 stale-branch guard: a pre-existing `feature/<KEY>` branch whose tip is **not** an ancestor of the resolved base is refused with a `create-failed` row carrying the stale-worktree remedy, instead of being silently reused on top of another run's commits. It is off by default (a plain run keeps reusing a same-named branch) and is implied by `--conductor`. `--auto` makes each spawned agent run the selected workflow's slash command with `--auto` (hands-off); omit it to keep the spawned agents interactive.\n   - **Selected agent**: track a `selected_agent` variable that defaults to `claude`. If the user passed `--agent <name>` / `--agent=<name>`, validate the value against the supported agents `claude` and `cursor-agent`, set `selected_agent` to it, and reject any other (malformed/unsupported) `--agent` value before proceeding. The agent is not auto-detected from the host editor — the user selects it explicitly (default `claude`).\n   - **Selected workflow**: track a `selected_workflow` variable that defaults to `implement`. If the user passed `--workflow <value>` or `--workflow=<value>`, validate it against the two allowed values `implement` and `review-and-implement`, set `selected_workflow`, and reject any other value with the allowlist in the error. `implement` (the default) preserves today's behavior byte-for-byte — each spawned worktree runs `/implement-ticket <KEY> [--auto]`. `review-and-implement` spawns `/review-and-implement <KEY> [--auto] [--rounds=<n>]` instead, which runs `/review-ticket` then, after a per-ticket halt gate, `/implement-ticket` inside the same session. A single chain-level `--auto` applies to the selected workflow as a whole — under `review-and-implement` it auto-approves both the review and the implementation phase.\n   - **Review rounds**: track a `review_rounds` value that defaults to unset. If the user passed `--rounds <n>` or `--rounds=<n>`, normalize it to `--rounds=1` or `--rounds=2` (reject any other value). `--rounds` is **review-only**: reject it (after parsing all flags, so flag order does not matter) if the final `selected_workflow` is not `review-and-implement`.\n   - **User-supplied base branch**: track a `user_supplied_base_branch` boolean that defaults to `false`. If the user passed `--base-branch <branch>` or `--base-branch=<branch>`, set the boolean to `true` and capture the value. A user-supplied `--base-branch` value **takes precedence** over any value resolved from Bridge API config in Stage 2. Validate the user-supplied value before proceeding: after trimming surrounding whitespace it must be non-empty, at most 255 characters, must not start with `-`, and must not contain ASCII control characters (`0x00`–`0x1F` or `0x7F`); reject any malformed value with a clear error.\n   - **User branch overrides**: collect any user-supplied repeatable `--branch KEY=BRANCH` flags. A user-provided override always takes precedence over Stage 2 enrichment for that key.\n   - Reject malformed input before proceeding: if a token looks like a flag but is not one of the supported flags, or a ticket key does not match `[A-Z]+-[0-9]+`, or a `--branch` value is not `KEY=BRANCH`, or `--agent` names an agent other than `claude`/`cursor-agent`, or `--workflow` names anything other than `implement`/`review-and-implement`, or `--rounds` is used outside `review-and-implement` or names anything other than `1`/`2`, or `--base-branch` fails the validation rules above, stop and report the malformed argument.\n\n2. **Connectivity check**: Call the `ping` MCP tool (no parameters). If the ping fails or does not return `\"status\": \"ok\"`, stop immediately and display:\n\n   ```\n   Connectivity check failed. Please verify:\n   - Check that the Bridge API MCP server is configured in your editor's MCP settings\n   - Check that BAPI_BASE_URL is set and the server is reachable\n   - Check that BAPI_API_KEY is valid\n   - Check that BAPI_REPO_NAME matches a configured repository\n   ```\n\nThis stage is **critical** — stop immediately on failure. Do not proceed to Stage 1.\n\n## Stage 1 — Acknowledge CLI Pre-flight\n\nThe packaged CLI runs its own per-platform pre-flight checks and then fetches `origin` and fast-forwards the local **configured base branch** (the value resolved in Stage 2 below, or `main` when none is configured) from `origin/<base>` so the new worktrees are based on an up-to-date base. The historical flag `--no-refresh-main` still controls this behavior — the flag name is preserved for backward compatibility, but it now skips refresh of whatever base branch resolves (default `main`). The required commands depend on the OS:\n\n- **macOS**: `wt`, `git`, `osascript`.\n- **Windows**: `git-wt`, `git`, Git for Windows / Git Bash (Worktrunk runs its `pre-start` / `post-start` hooks via Git Bash), and Windows Terminal **or** PowerShell.\n- **Linux**: `wt`, `git`, `tmux`.\n\nOn **Windows** the Worktrunk binary is `git-wt` (its winget alias), which is a different tool from Windows Terminal's `wt.exe`: the CLI uses `git-wt` to **create worktrees** and `wt.exe` to **open a tab**, and never conflates the two. On **Linux** the CLI opens one detached `tmux` session per ticket (a window is added if that ticket's session already exists); attach later with `tmux attach -t <session>`. An unsupported OS (not macOS/Windows/Linux) fails fast with a clear \"unsupported platform\" message.\n\nThis stage simply notes that the CLI will fail fast if any prerequisite is missing or if local `main` has diverged from `origin/main` — you do not need to verify anything separately here, and you must not run any pre-flight commands yourself. When the CLI's pre-flight fails it now hints the user to run the read-only diagnostics command `npx -y @bridge_gpt/mcp-server doctor`, which reports found/missing for every prerequisite on the current OS — the pre-flight set plus `uv` plus the selected agent's command — and prints the manual install command for each missing one. `doctor` is strictly read-only and never installs anything; never run install commands automatically on the user's behalf. The CLI does not call any Bridge API tools; all credential-bearing work (branch enrichment in Stage 2) stays in this command. Proceed to Stage 2.\n\nThe packaged CLI also performs **secret-free Bridge API MCP provisioning** inside each created worktree: synchronously after the worktree is created and **before the agent tab/session is opened**, it writes both `.mcp.json` (Claude Code) and `.cursor/mcp.json` (Cursor) pointing at the `mcp-invoke` shim. These registrations are **secret-free** — they contain no `env` block and no API key, because the shim resolves credentials at runtime. If a spawned agent (or difficulty→model routing) reports missing Bridge API credentials, fix it by rerunning `/install-bridge` (its final stage persists the routing credential), by running `npx -y @bridge_gpt/mcp-server credentials migrate-agent-config --write-credentials` to migrate a key that lives only in `.mcp.json` / `.cursor/mcp.json`, or by adding a `bapi:<repo>` entry to the user-scoped credentials file (`~/.config/bridge/credentials.json`) — never by putting `BAPI_API_KEY` into the worktree `.mcp.json` or `.cursor/mcp.json` (that env is invisible to the Bash-spawned CLI).\n\nThis stage is **critical** in the sense that the CLI will abort if its pre-flight fails; you will see the error in Stage 3's output and must surface it.\n\n## Stage 2 — Resolve Base Branch + Enrich Branch Names (best-effort)\n\n### Stage 2a — Resolve configured `base_branch`\n\nThe CLI must be told which branch to cut new worktrees from. Resolution order:\n\n1. If `user_supplied_base_branch` from Stage 0 is `true`, **skip the config-field lookup entirely** and use the user-supplied value. The user's explicit `--base-branch` always wins; never call `config_field` for `base_branch` in that case.\n2. Otherwise, call the `config_field` MCP tool with `operation` set to `\"get\"` and `field_name` set to `base_branch` (do not pass any other parameters; the tool resolves the repository from the MCP server's configured `BAPI_REPO_NAME`).\n3. Parse the response. Treat the result as the **configured base branch** only when the response is a JSON object whose `value` field is a non-empty string after trimming surrounding whitespace.\n4. Treat **all** of the following as \"unset\" — emit a single-line warning like `Warning: base_branch is unset; CLI will default to main` and **omit** the `--base-branch` flag entirely from the Stage 3 command (the CLI's own default is `main`):\n   - `value` is `null`.\n   - `value` is an empty string or a whitespace-only string.\n   - The endpoint returns HTTP `400` (invalid field — happens before the registry includes `base_branch`).\n   - The tool returns a network error, timeout, or non-JSON parse failure.\n   - Any other lookup failure.\n5. When the configured value is usable, capture it in a `resolved_base_branch` variable. **Do not** stop the pipeline on a lookup failure; fall through to the CLI default.\n\nWhen forwarding `resolved_base_branch` into the Bash invocation in Stage 3, **shell-escape it safely**: replace every literal single quote `'` in the value with the four-character sequence `'\\''`, then wrap the entire resulting string in single quotes (so the final argument looks like `'<escaped-value>'`). This is the standard POSIX single-quote escaping rule and is **mandatory** because `base_branch` is admin-configurable data that gets interpolated into a Bash command string; any unescaped single quote would otherwise break out of the surrounding quotes. Pass `--base-branch '<escaped-value>'` to the CLI as a single argv element — never expand the value unquoted into the command line.\n\n### Stage 2b — Enrich Branch Names\n\nBranch enrichment happens here, in the command, **before** invoking the CLI — the `get_ticket` MCP tool runs inside the MCP server process, which holds the Bridge API credentials the shell-spawned CLI does not have. For each parsed ticket key that does **not** already have a user-provided `--branch` override:\n\n1. Call the `get_ticket` MCP tool with `ticket_number` set to the key and `save_locally` set to `false`.\n2. From the response, extract the `summary` field. Slugify it: lowercase the string, replace every run of non-alphanumeric characters (`[^a-z0-9]+`) with a single dash `-`, trim leading and trailing dashes, and truncate to at most `40` characters (cutting at a dash boundary if possible).\n3. The enriched branch name is `feature/<KEY>-<slug>`. Example: `BAPI-248` with summary `\"Add PR rating pre-evaluation step\"` becomes `feature/BAPI-248-add-pr-rating-pre-evaluation-step` (trimmed at 40 chars).\n4. If the `get_ticket` call fails for a particular key (404, network error, missing summary) or produces an empty slug, emit a single-line warning like `Warning: could not enrich BAPI-248, falling back to feature/BAPI-248` and let the CLI apply its default `feature/<KEY>` for that key only. Do NOT stop the pipeline.\n5. Build a list of `--branch <KEY>=<BRANCH>` arguments — one entry per key whose enrichment succeeded — and merge it with any user-provided overrides from Stage 0. **Do not** call `get_ticket` for keys that already have a user-provided override; those overrides win.\n\nThis stage is **non-critical** — warnings are acceptable, the pipeline continues with the fallback default for any key that fails. Do not call the Bridge API from the CLI itself; the CLI never has credentials.\n\n## Packaged CLI launcher (`BAPI_MCP_CLI`)\n\nResolve the packaged-CLI launcher **once**, before the first shell-out below, and reuse that one resolved value for every packaged-CLI invocation in this command. Call it `<launcher>`.\n\n- Read the `BAPI_MCP_CLI` environment variable.\n- **Unset, empty, or whitespace-only** — `<launcher>` is exactly `npx -y @bridge_gpt/mcp-server`. This is the default, and the resulting shell command is byte-identical to what it was before this override existed.\n- **Otherwise** — `<launcher>` is that value, used verbatim as the command prefix. It names a local launcher, such as `node /absolute/path/to/mcp_server/build/index.js`. Use it for local pilots and pre-publish verification.\n\nWhen the override is set:\n\n- Apply this command's mandatory single-quote escaping rule (`'` → `'\\''`, then wrap the whole value in single quotes) before interpolating `<launcher>` into a Bash command string. Never expand it unquoted.\n- Keep every dynamic argument — ticket keys, branch names, base branches, file paths — independently quoted. Never concatenate an argument into the launcher value.\n- Never put a credential, an API key, or an environment assignment carrying one into the launcher value, an example, or a dry-run preview.\n\nWorked examples and printed remediation hints below show the **unset** resolution — the literal `npx -y @bridge_gpt/mcp-server` — because that is the default every operator gets. A stale local build is exactly as misleading as a stale npm publish: rebuild with `cd mcp_server && npm run build` before relying on the override.\n\n## Stage 3 — Invoke the Packaged CLI\n\nUse the **Bash tool** to invoke the packaged CLI. Build the command line as:\n\n```\n<launcher> start-tickets <pass-through-flags> <base-branch-flag> <branch-overrides> <ticket-keys>\n```\n\nWhere:\n- `<pass-through-flags>` are the supported flags collected in Stage 0 (`--agent`, `--terminal`, `--dry-run`, `--auto`, `--no-refresh-main`, `--max-parallel`, `--guard-stale-branch`), forwarded verbatim. Forward `--agent <name>` only if the user supplied it; otherwise omit it and the CLI defaults to `claude`. Forward `--auto` only if the user supplied it.\n- Forward `--workflow <selected_workflow>` only when the user explicitly passed `--workflow`; otherwise omit it and the CLI defaults to `implement`. Forward the normalized `--rounds=<n>` from Stage 0 only when the user supplied it (which Stage 0 already guarantees is only possible under `review-and-implement`).\n- `<base-branch-flag>` is `--base-branch '<escaped-value>'` (single-quoted using the Stage 2a escaping rule) **only when** the user supplied `--base-branch` in Stage 0 **or** Stage 2a's `config_field` lookup returned a non-empty configured value. When the configured value is unset / lookup fails / user did not supply one, **omit this flag entirely** so the CLI's own default (`main`) takes effect.\n- `<branch-overrides>` is the list of `--branch KEY=BRANCH` flags assembled in Stage 2 (enrichment results merged with user overrides; omit any key whose enrichment failed and had no user override).\n- `<ticket-keys>` is the original list of ticket keys parsed in Stage 0, space-separated and in the original order.\n\nExample for two tickets after successful enrichment, throttled to 2 concurrent worktrees:\n\n```\nnpx -y @bridge_gpt/mcp-server start-tickets \\\n  --max-parallel 2 \\\n  --branch BAPI-248=feature/BAPI-248-add-pr-rating-pre-evaluation-step \\\n  --branch BAPI-250=feature/BAPI-250-deep-research-durability \\\n  BAPI-248 BAPI-250\n```\n\nExample launching Cursor Agent instead of the default Claude Code:\n\n```\nnpx -y @bridge_gpt/mcp-server start-tickets --agent cursor-agent BAPI-248\n```\n\nExample cutting worktrees from a non-`main` base (either user-supplied via `--base-branch develop` in Stage 0 or resolved from Bridge API config in Stage 2a):\n\n```\nnpx -y @bridge_gpt/mcp-server start-tickets --base-branch develop BAPI-248\n```\n\nExample using the lower-level review-and-implement workflow directly (the `/review-and-start` command is the recommended front door for this; this form is documented here as the advanced launcher seam it drives):\n\n```\nnpx -y @bridge_gpt/mcp-server start-tickets --workflow review-and-implement --auto --rounds=2 BAPI-248\n```\n\nPass through the CLI's stdout and stderr to the user verbatim. If the CLI exits non-zero, treat that as a critical failure: report the exit code and the CLI's error output, and stop.\n\nThis stage is **critical** — propagate any non-zero exit from the packaged CLI.\n\n## Stage 4 — Final Report\n\nOnce the CLI exits 0, parse its `Summary` section (one stable line per ticket in the form `KEY  branch=BRANCH  status=STATUS`, with an optional trailing `path=PATH`) and reformat it as a markdown table:\n\n```\n| Ticket   | Branch                                              | Status   |\n|----------|-----------------------------------------------------|----------|\n| BAPI-248 | feature/BAPI-248-add-pr-rating-pre-evaluation-step  | spawned  |\n| BAPI-250 | feature/BAPI-250-deep-research-durability           | spawned  |\n```\n\nStatus values are `dry-run`, `spawned`, `create-failed`, and `spawn-failed`. This table (and the report as a whole) describes **worktree/spawn status only** — it must never claim that review or implementation itself has completed; that work happens later, independently, inside each spawned session.\n\nCompute `spawned_command` from `selected_workflow`: `/implement-ticket <KEY>` when `implement` (the default), or `/review-and-implement <KEY>` when `review-and-implement`. Append `--auto` when the user passed it, and (workflow `review-and-implement` only) append the normalized `--rounds=<n>` when the user supplied `--rounds`. End the report with the worktree-first explanation, rendered for the tracked `selected_agent` and `spawned_command`. When `selected_agent` is `claude` (the default):\n\n```\nThe CLI created/switched each Worktrunk worktree first (throttled by --max-parallel),\nthen opened one tab/session per successful worktree (macOS Terminal/iTerm tab, Windows\nTerminal tab or PowerShell window, or Linux tmux session). Each one runs\n`claude '<spawned_command>'` inside its already-created worktree, which launches\nClaude Code with the starter prompt as its first message. Switch to each tab — or on\nLinux run `tmux attach -t <session>` — to monitor.\n```\n\nWhen `selected_agent` is `cursor-agent`, render the same explanation but with the Cursor handoff — do **not** claim it launches Claude Code:\n\n```\nThe CLI created/switched each Worktrunk worktree first (throttled by --max-parallel),\nthen opened one tab/session per successful worktree (macOS Terminal/iTerm tab, Windows\nTerminal tab or PowerShell window, or Linux tmux session). Each one runs\n`cursor-agent '<spawned_command>'` inside its already-created worktree, which\nlaunches Cursor Agent with the starter prompt as its first message. Switch to each\ntab — or on Linux run `tmux attach -t <session>` — to monitor.\n```\n\nThe spawned command is identical for both agents; only the launched agent binary differs. Under `review-and-implement`, each spawned session independently runs `/review-ticket`, pauses at its own per-ticket halt gate (unless chain-level `--auto` was passed), and only then runs `/implement-ticket` — do not report that review or implementation succeeded from this parent session.\n\nIf the CLI reported any `create-failed` or `spawn-failed` statuses, or Stage 2 emitted any enrichment warnings, list them under a `Warnings:` heading at the bottom of the report. If there were none, omit that section.\n\nSee `docs/claude/parallel-worktrees.md` for the deep-dive runbook and the Worktrunk verification result behind this worktree-first model.\n\n## Difficulty-Based Implementation-Model Routing\n\nBefore launching the interactive agent for each ticket, the packaged CLI selects an\nimplementation **model tier** from the ticket's `difficulty` rating (1-10) and injects\nit as a `--model` flag at the agent spawn boundary. This happens entirely inside the\nCLI — it is **not** part of the server-side `/implement-ticket` recipe, because the\nmodel an interactive agent session uses is fixed at the moment the process is launched.\n\n- **Tier ladder (fixed):** `difficulty 1-2 → cheap`, `3-6 → basic`, `7-10 → premium`.\n- **Separation of concerns:** the Python backend returns only the coarse tier\n  (`cheap`/`basic`/`premium`) via `GET /jira/tickets/{KEY}/model-tier`; difficulty is\n  computed on demand and cached when absent. The TypeScript CLI alone maps a tier to\n  the agent-specific model alias (`claude`: `haiku`/`sonnet`/`opus`; `cursor-agent`:\n  version-suffixed strings validated against `cursor-agent --list-models`).\n- **Per-repo config:**\n  - `difficulty_model_routing_enabled` — boolean, **default ON**. Set to `false` to\n    disable routing for a repo (the CLI then omits `--model`).\n  - `difficulty_model_tier_overrides` — a JSON object mapping a tier name to a model\n    alias (e.g. `{\"premium\": \"opus\"}`), **not** raw CLI arguments. Only `cheap`,\n    `basic`, and `premium` keys are accepted; aliases must match `^[A-Za-z0-9._:-]+$`.\n- **Fail-open:** routing never aborts a spawn. Credential, network, config, or\n  no-tier routing failures **assume a hard ticket and default to the premium/Opus\n  tier** when the selected agent supports a valid premium alias; routing being\n  disabled (`difficulty_model_routing_enabled = false`) or an agent that does not\n  support `--model` instead omit `--model` so the agent runs on its own default\n  model. Each degraded case is surfaced as exactly one secret-free, per-ticket\n  routing-diagnostic line, never a hard failure.\n\n### Model routing credential\n\nDifficulty→model routing needs Bridge API credentials, and the shell-spawned\n`start-tickets` CLI is a **different runtime surface** from the MCP server: a\n`BAPI_API_KEY` that lives only in `.mcp.json` / `.cursor/mcp.json` is visible to\nthe MCP server but **not** to the Bash-spawned CLI, so routing silently degrades.\nThe durable source of truth both runtimes can resolve is the user-scoped store\n`~/.config/bridge/credentials.json`, keyed `bapi:<repo>`. If a routing-diagnostic\nline reports the credential is missing (e.g. difficulty resolves as `?`), fix it\nby any one of:\n\n1. Rerun `/install-bridge` — its final stage now persists the validated routing\n   credential into `~/.config/bridge/credentials.json` via the\n   `persist_routing_credential` tool.\n2. Migrate a key that lives **only** in `.mcp.json` / `.cursor/mcp.json` into the\n   user-scoped store with the consent-gated, one-shot command:\n\n   ```\n   npx -y @bridge_gpt/mcp-server credentials migrate-agent-config --write-credentials\n   ```\n\n3. Manually add `BAPI_API_KEY` under the `bapi:<repo>` target in the user-scoped\n   store `~/.config/bridge/credentials.json`.\n\nNever put `BAPI_API_KEY` into a worktree `.mcp.json` / `.cursor/mcp.json` as a fix —\nthat env is invisible to the spawned CLI.\n\n## Conductor observability (opt-in via `--conductor`, BAPI-394)\n\nConductor is **opt-in**. By default `start-tickets` spawns the plain\n`cd <worktree> && <agent> '/implement-ticket <KEY> [--auto]'` — no\n`BAPI_CONDUCTOR_*` env, no supervisor window, and no message-relay instruction.\nPass `--conductor` (e.g. `/start-tickets --conductor BAPI-123`) to enable the\nConductor system below.\n\nWith `--conductor`, a run mints a single conductor `run_id` and attributes each\nworker's lifecycle events by `worker_id`, ticket key, and worktree path, and a\nsupervisor peer tab is opened. When the selected agent is **Claude Code**, the CLI\ninjects a conductor lifecycle hook into each created worktree's\n`.claude/settings.local.json` so the spawned session emits local `run.started` /\n`run.stopped` / `agent.notification` (and, when\n`BAPI_CONDUCTOR_ENABLE_PRE_TOOL_USE=1`, `tool.intent`) events into the local\nconductor ledger. These hooks apply **only** when the selected agent is Claude\nCode; other agents (e.g. `cursor-agent`) still participate in the run-level\n`run.started` event but receive no per-worktree Claude hook. Inspect the ledger\nwith the `epic-implementer` CLI (e.g. `epic-implementer doctor`). Conductor observability is\nbest-effort and never blocks or aborts a spawn.\n\nObservability under `--conductor` is one-directional: workers emit lifecycle events\ninto the local ledger and nothing is passed back into a running session. (Epic-tick\ndispatch always runs with conductor enabled, independent of this user-facing flag.)\n",
     "teach-bridge.md": "Update a Bridge API configuration field via a natural-language teaching.\n\n$ARGUMENTS\n\n---\n\n# Instructions\n\nThis command takes a natural-language teaching (e.g., \"use data-testid selectors in Playwright tests\") and updates the appropriate Bridge API configuration field. The teaching is auto-classified to the correct field, merged with existing content as actionable AI instructions, and uploaded after user confirmation.\n\n`$ARGUMENTS` is required — it is the teaching text. If `$ARGUMENTS` is empty, show:\n\n```\nUsage: /teach-bridge <teaching>\n\nExamples:\n  /teach-bridge use data-testid selectors in Playwright tests\n  /teach-bridge always validate input DTOs with Pydantic before passing to service layer\n  /teach-bridge prefer composition over inheritance for service classes\n```\n\nIf any stage fails, stop immediately and report which stage failed and why.\n\n## Stage 0 — Preflight\n\n1. **Validate arguments**: If `$ARGUMENTS` is empty or contains only whitespace, display the usage instructions above and stop.\n\n2. **Admin check**: Call the `ping` MCP tool (no parameters) and read `role` from its first (JSON)\n   content item. Inspect it:\n   - If `role` is `\"admin\"` OR `role` is `null` (not determined — e.g. a legacy shared key): proceed normally.\n   - Otherwise (an explicit non-admin role such as `\"member\"`): stop immediately and display:\n     ```\n     Admin access required. Your API key has role \"<role>\".\n     Only admin keys and legacy shared keys can update configuration fields.\n     Contact your project administrator to request admin access.\n     ```\n\nIf this stage fails, stop immediately and report the error. Do not proceed to Stage 1.\n\n## Stage 1 — Classify\n\n1. **List available fields**: Call the `config_field` MCP tool with `operation` set to `\"list\"` (no other parameters). This returns all available configuration field names with descriptions.\n\n2. **Evaluate the teaching**: Compare the user's teaching (`$ARGUMENTS`) against each field's description to determine which field it applies to.\n\n3. **Handle classification outcomes**:\n   - **Clear single match**: If one field is clearly the best target, proceed to Stage 2 with that field.\n   - **Multiple plausible matches**: If 2-3 fields are equally plausible, present them to the user with their descriptions and ask which one to update. Wait for user input before proceeding.\n   - **No confident match**: If you cannot confidently map the teaching to any field, ask the user to elaborate or specify which field they intend. Wait for user input before proceeding.\n\n## Stage 2 — Merge\n\n1. **Read current value**: Call the `config_field` MCP tool with `operation` set to `\"get\"` and `field_name` set to the selected field from Stage 1. Capture the current value, description, and examples from the response.\n\n2. **Draft the update**:\n   - **If the field is currently null or empty**: Compose initial content from the teaching. Rephrase the user's input as imperative, agent-facing instructions (e.g., convert \"I want you to use data-testid\" to \"Always use `data-testid` attributes for Playwright element locators\"). Do not use the user's exact conversational text.\n   - **If the field has existing content**: Merge the teaching into the existing value at the most appropriate location. Rephrase as imperative, agent-facing instructions. Preserve the existing structure and formatting.\n\n3. **Handle contradictions**: If the teaching contradicts existing instructions in the field, present both the existing instruction and the new teaching side-by-side and ask the user which should take precedence. Wait for user input before proceeding.\n\n## Stage 3 — Confirm and Upload\n\n1. **Show the proposed update**: Display to the user:\n   - **Field**: The name of the field being updated\n   - **Change summary**: A brief description of what was added or changed\n   - **Full proposed value**: The complete new value for the field (not just the diff)\n\n2. **Wait for confirmation**: Ask the user to confirm, request edits, or abort.\n\n3. **On confirmation**: Call the `config_field` MCP tool with:\n   - `operation`: `\"update\"`\n   - `field_name`: the selected field name\n   - `value`: the full merged value (pass inline, do not use `file_path`)\n\n   Display a success message confirming the update.\n\n4. **On rejection**: Ask the user what they'd like to change. If they provide edits, revise the proposed value and show it again. If they abort, stop without making any changes.\n",
     "upgrade-bridge.md": "# Upgrade Bridge\n\n$ARGUMENTS\n\nUse this command to upgrade (or update) the Bridge API MCP — the\n`@bridge_gpt/mcp-server` package, also called the bridge-api MCP — to the latest\npublished version. This is the action behind the ping tool's advice to \"tell\nyour local agent 'upgrade bridge'\".\n\n---\n\n# Instructions\n\nRun the existing packaged upgrade flow. Do not edit files, install anything by\nhand, or invent a new subcommand — just drive the upgrade CLI and report what it\ndid.\n\n## Step 1 — Run the upgrade command\n\nFrom the **project root**, run exactly:\n\n```\nnpx -y @bridge_gpt/mcp-server@latest --upgrade\n```\n\nThis upgrades/updates the installed `@bridge_gpt/mcp-server` (the bridge-api MCP)\nand re-scaffolds the slash commands.\n\n### `BAPI_MCP_CLI` does not apply here — deliberately\n\nEvery other packaged-CLI command honors the `BAPI_MCP_CLI` local-launcher override, which replaces the `npx -y @bridge_gpt/mcp-server` prefix with a local build for pre-publish verification. **This command is the one exception, and ignores it entirely.**\n\nThe reason is that upgrading is inherently about the *published* package. A local launcher points at a build that is already on disk, so running the upgrade through it would \"upgrade\" the operator using the very build they are trying to replace — silently doing nothing while reporting success. That is the exact failure the `@latest` pin below exists to prevent, so honoring the override here would reintroduce it by another route.\n\nRun the command above verbatim even when `BAPI_MCP_CLI` is set, and do not mention the override as an option for this command.\n\nThe `@latest` pin matters. It applies to the short-lived *upgrader* process only:\nwithout it, npx may reuse a cached older copy of the package and \"upgrade\" you\nusing the very build you are trying to replace. The exact version pin the\nupgrader then writes into each host config is a separate, deliberate thing — host\nconfigs stay pinned to an exact `MAJOR.MINOR.PATCH` release so a project's MCP\nserver is reproducible.\n\n## Step 2 — Report the result\n\nThe CLI reports **per config file**, because a project can have several\n(`.mcp.json`, `.cursor/mcp.json`, `.vscode/mcp.json`) and they can disagree.\nRelay those lines as printed. Do not collapse them into a single global\n`oldVersion -> newVersion` transition and do not infer one yourself.\n\nThe forms the CLI emits are:\n\n- `<path>: 0.2.16 -> 0.2.36` — that config's launcher pin moved.\n- `<path>: already 0.2.36` — that config was already at the target.\n- `<path>: normalized <spec> -> @bridge_gpt/mcp-server@0.2.36` — an unpinned or\n  `@latest` launcher was given an exact pin.\n- `<path>: added @bridge_gpt/mcp-server@0.2.36` — a `bridge-api` entry was added\n  to an existing config.\n- `<path>: created with @bridge_gpt/mcp-server@0.2.36` — the config file was\n  created.\n- `<path>: skipped — worktree mcp-invoke shim preserved` — a worktree\n  registration that intentionally has no published-package pin.\n\nThen report the CLI's closing status verbatim:\n\n- If the CLI prints `Already up-to-date.`, report `Already up-to-date.` exactly.\n  Only the CLI decides this; it means every applicable launcher pin was already\n  at the target. Never infer it from a version transition that reads the same on\n  both sides.\n- Otherwise report the CLI's completion line together with the per-config lines\n  above.\n\n## Step 3 — Handle failures\n\nIf the command fails (non-zero exit or an error in its output), **stop** and\nreport the CLI error verbatim. Do not retry blindly and do not attempt manual\nedits to config files to work around it. A non-zero exit means the upgrade did\nnot converge — for example a config could not be read, a launcher carries a\nversion range the upgrader must not rewrite, a competing local install could not\nbe removed, or a written pin failed post-write verification. Those are reported\nfor a human to resolve, not for you to repair.\n\n## Final Report\n\nReport whether the bridge-api MCP was upgraded (relaying the CLI's per-config\nlines and its completion status), was already current (`Already up-to-date.`),\nor failed (with the CLI error).\n"
 };