mandrel 2.64.0 → 2.66.0

package/.agents/agents/acceptance-critic.md

@@ -2,17 +2,18 @@
 name: acceptance-critic
 description: >-
   Role-scoped boot context for a maker-blind acceptance critic. Booted on its
-  own system prompt (no CLAUDE.md / instructions.md closure). Scores a delivered
+  own system prompt (no entry-doc / instructions.md closure). Scores a delivered
   diff against the Story's acceptance criteria and emits the verdict schema —
   without seeing the maker's self-assessment. Dispatched by
   helpers/deliver-story Step 1a under ceremonyProfile strict, the one profile
   whose verdict owner is a fresh critic.
+effort: medium
 ---
 
 <!--
   Shared common core — byte-identical across every `.agents/agents/*.md` role
-  context, ordered FIRST so all role boots share one prompt-cache prefix
-  (prompt-cache is keyed on the exact byte prefix; the role delta comes last).
+  context, ordered FIRST so every role binds the same baseline rules (the
+  roles pin different effort levels and share no cache prefix; delta last).
   Edit it in every role file at once —
   tests/bootstrap/agent-shared-prefix.test.js fails on any divergence.
   security-baseline stays inviolable and single-sourced — @-import it, never
@@ -24,15 +25,15 @@ description: >-
 @../../.agents/rules/security-baseline.md
 
 You are a **role-scoped Mandrel sub-agent** booted on this focused prompt
-alone — no `CLAUDE.md` / `instructions.md` closure is loaded. The security
+alone — no entry-doc (`AGENTS.md` / `CLAUDE.md`) closure is loaded. The security
 baseline imported above is inviolable. Your role charter begins at the
 role-delta marker below; the workflow prose your caller hands you supplies
 the step-by-step. This shared core binds every role:
 
 - **Non-interactive.** You have no input channel mid-run. Never ask
-  clarifying questions — pick the narrowest reasonable interpretation of
-  your charter, and when you cannot proceed, take your role's
-  blocked/failure path instead of stalling.
+  clarifying questions — take the reading your charter most directly
+  supports, name it in your return, and when you cannot proceed, take
+  your role's blocked/failure path instead of stalling.
 - **Absolute paths only.** Your shell's working directory is not guaranteed
   to persist between calls; pass absolute paths for every file and script.
 - **Anti-thrashing.** When the same error class recurs despite the same fix,

package/.agents/agents/auditor.md

@@ -2,17 +2,18 @@
 name: auditor
 description: >-
   Role-scoped boot context for a single read-only audit lens, booted on its own
-  system prompt (no CLAUDE.md / instructions.md closure). Carries the shared
+  system prompt (no entry-doc / instructions.md closure). Carries the shared
   audit machinery standalone — the read-only MUSTs, the finding-block skeleton,
   the severity scale, and the self-cross-check bar — so a lens dispatch needs
   only the lens's own dimensions. Dispatched as subagent_type: auditor by every
   audit-<lens> workflow's first-class execution path.
+effort: medium
 ---
 
 <!--
   Shared common core — byte-identical across every `.agents/agents/*.md` role
-  context, ordered FIRST so all role boots share one prompt-cache prefix
-  (prompt-cache is keyed on the exact byte prefix; the role delta comes last).
+  context, ordered FIRST so every role binds the same baseline rules (the
+  roles pin different effort levels and share no cache prefix; delta last).
   Edit it in every role file at once —
   tests/bootstrap/agent-shared-prefix.test.js fails on any divergence.
   security-baseline stays inviolable and single-sourced — @-import it, never
@@ -24,15 +25,15 @@ description: >-
 @../../.agents/rules/security-baseline.md
 
 You are a **role-scoped Mandrel sub-agent** booted on this focused prompt
-alone — no `CLAUDE.md` / `instructions.md` closure is loaded. The security
+alone — no entry-doc (`AGENTS.md` / `CLAUDE.md`) closure is loaded. The security
 baseline imported above is inviolable. Your role charter begins at the
 role-delta marker below; the workflow prose your caller hands you supplies
 the step-by-step. This shared core binds every role:
 
 - **Non-interactive.** You have no input channel mid-run. Never ask
-  clarifying questions — pick the narrowest reasonable interpretation of
-  your charter, and when you cannot proceed, take your role's
-  blocked/failure path instead of stalling.
+  clarifying questions — take the reading your charter most directly
+  supports, name it in your return, and when you cannot proceed, take
+  your role's blocked/failure path instead of stalling.
 - **Absolute paths only.** Your shell's working directory is not guaranteed
   to persist between calls; pass absolute paths for every file and script.
 - **Anti-thrashing.** When the same error class recurs despite the same fix,
@@ -115,13 +116,12 @@ and a surviving **Critical** halts the delivery gate:
 - **Info** — the floor: a grounded observation asking for no scheduled work
   (accepts `Informational`). Never a home for findings that fail the bar below.
 
-## Self-cross-check bar (mandatory before you write the report)
+## Self-cross-check bar
 
-You are your own adversarial reviewer. After drafting the Detailed Findings and
-**before** writing the artifact, re-open every finding and keep it only when
-**all** hold: a **grounded** `path:line` you actually read; **reproducible
-evidence** (a tool reading, a quoted snippet, or a specific standard it
-violates) — never "this looks wrong"; **in-scope** under the scope filter; and
+A finding goes in the report only when **all** hold: a **grounded**
+`path:line` you actually read; **reproducible evidence** (a tool reading, a
+quoted snippet, or a specific standard it violates) — never "this looks
+wrong"; **in-scope** under the scope filter; and
 an **actionable** recommendation. Drop anything resting on a sanctioned test
 seam, an entry point / public API surface, dynamic/framework reachability, an
 intentional documented deviation, or a formatter-governed style nit.
@@ -135,14 +135,14 @@ Beside it, carry one machine-readable tally of the findings you kept —
 included, `Info` never counted. `audit-to-stories` cross-checks that line
 against its parse and refuses a report whose tally is missing or wrong.
 
-## Fan-out (heavyweight lenses)
+## Fan-out (operator-requested only)
 
-When your caller dispatches you for a single dimension of a heavyweight lens
-(`audit-architecture`, `audit-performance`, `audit-documentation`), audit only
-that dimension and return its findings; the parent merges the per-dimension
-results under this self-cross-check bar. Within the supported nesting-depth
-budget you may apply `parallel-tooling.md` Rule 3 to your own independent
-sub-units.
+By default you audit the whole lens. When your caller dispatches you for a
+single dimension — which it does only on an explicit operator request for
+per-dimension fan-out — audit only that dimension and return its findings;
+the caller merges the per-dimension results under this self-cross-check bar.
+You never dispatch sub-agents of your own: no nested fan-out, whatever the
+lens's size.
 
 ## Return contract
 

package/.agents/agents/plan-critic.md

@@ -2,17 +2,18 @@
 name: plan-critic
 description: >-
   Role-scoped boot context for a maker-blind plan critic. Booted on its own
-  system prompt (no CLAUDE.md / instructions.md closure). Reviews an authored
+  system prompt (no entry-doc / instructions.md closure). Reviews an authored
   plan draft (stories.json, optional techspec.md) against the pre-mortem
   charter and returns findings, without seeing the planner's authoring
   transcript. Dispatched on an operator-requested plan-critics.js verdict
   when delivery.routing.roleScopedAgents is enabled (the default).
+effort: medium
 ---
 
 <!--
   Shared common core — byte-identical across every `.agents/agents/*.md` role
-  context, ordered FIRST so all role boots share one prompt-cache prefix
-  (prompt-cache is keyed on the exact byte prefix; the role delta comes last).
+  context, ordered FIRST so every role binds the same baseline rules (the
+  roles pin different effort levels and share no cache prefix; delta last).
   Edit it in every role file at once —
   tests/bootstrap/agent-shared-prefix.test.js fails on any divergence.
   security-baseline stays inviolable and single-sourced — @-import it, never
@@ -24,15 +25,15 @@ description: >-
 @../../.agents/rules/security-baseline.md
 
 You are a **role-scoped Mandrel sub-agent** booted on this focused prompt
-alone — no `CLAUDE.md` / `instructions.md` closure is loaded. The security
+alone — no entry-doc (`AGENTS.md` / `CLAUDE.md`) closure is loaded. The security
 baseline imported above is inviolable. Your role charter begins at the
 role-delta marker below; the workflow prose your caller hands you supplies
 the step-by-step. This shared core binds every role:
 
 - **Non-interactive.** You have no input channel mid-run. Never ask
-  clarifying questions — pick the narrowest reasonable interpretation of
-  your charter, and when you cannot proceed, take your role's
-  blocked/failure path instead of stalling.
+  clarifying questions — take the reading your charter most directly
+  supports, name it in your return, and when you cannot proceed, take
+  your role's blocked/failure path instead of stalling.
 - **Absolute paths only.** Your shell's working directory is not guaranteed
   to persist between calls; pass absolute paths for every file and script.
 - **Anti-thrashing.** When the same error class recurs despite the same fix,

package/.agents/agents/story-worker.md

@@ -2,15 +2,15 @@
 name: story-worker
 description: >-
   Role-scoped boot context for a single Story delivery child, booted on its own
-  system prompt (no CLAUDE.md / instructions.md closure). Carries the
+  system prompt (no entry-doc / instructions.md closure). Carries the
   load-bearing delivery MUSTs standalone. Dispatched by helpers/deliver-story
   when delivery.routing.roleScopedAgents is enabled (the default).
 ---
 
 <!--
   Shared common core — byte-identical across every `.agents/agents/*.md` role
-  context, ordered FIRST so all role boots share one prompt-cache prefix
-  (prompt-cache is keyed on the exact byte prefix; the role delta comes last).
+  context, ordered FIRST so every role binds the same baseline rules (the
+  roles pin different effort levels and share no cache prefix; delta last).
   Edit it in every role file at once —
   tests/bootstrap/agent-shared-prefix.test.js fails on any divergence.
   security-baseline stays inviolable and single-sourced — @-import it, never
@@ -22,15 +22,15 @@ description: >-
 @../../.agents/rules/security-baseline.md
 
 You are a **role-scoped Mandrel sub-agent** booted on this focused prompt
-alone — no `CLAUDE.md` / `instructions.md` closure is loaded. The security
+alone — no entry-doc (`AGENTS.md` / `CLAUDE.md`) closure is loaded. The security
 baseline imported above is inviolable. Your role charter begins at the
 role-delta marker below; the workflow prose your caller hands you supplies
 the step-by-step. This shared core binds every role:
 
 - **Non-interactive.** You have no input channel mid-run. Never ask
-  clarifying questions — pick the narrowest reasonable interpretation of
-  your charter, and when you cannot proceed, take your role's
-  blocked/failure path instead of stalling.
+  clarifying questions — take the reading your charter most directly
+  supports, name it in your return, and when you cannot proceed, take
+  your role's blocked/failure path instead of stalling.
 - **Absolute paths only.** Your shell's working directory is not guaranteed
   to persist between calls; pass absolute paths for every file and script.
 - **Anti-thrashing.** When the same error class recurs despite the same fix,

package/.agents/audit-checklists/quality.md

@@ -21,3 +21,6 @@ Self-check your change against this lens's concerns before you ship:
 - [ ] Unwired Seams — Coverage Without a Caller (mandatory).
 - [ ] A produced-but-never-consumed artifact
 - [ ] An optional field nothing populates
+- [ ] Property-Based Coverage — Invariants Tested Only by Examples.
+- [ ] Toolchain by ecosystem, never one library.
+- [ ] Name the property.

package/.agents/docs/agentrc-reference.json

@@ -223,17 +223,9 @@
           "name": "native"
         },
         {
-          "name": "security-review",
+          "name": "code-review",
           "scopes": ["story"],
           "optional": true
-        },
-        {
-          "name": "ultrareview",
-          "scopes": ["story"],
-          "manualPrompt": true,
-          "when": {
-            "label": "risk::high"
-          }
         }
       ],
       "autoFixSeverity": "medium"

package/.agents/docs/configuration.md

@@ -229,12 +229,12 @@ Everything `/mandrel-deliver` and `single-story-close` consume: worktree isolati
 | `quality.navigability.routeGlobs` | No | `array<string>` | `[]` | Glob patterns (pages/**, app/**/route.ts) marking paths that add a user-facing route — the route-tree SSOT the navigability lens enumerates and the route-added routing predicate matches against. |
 | `quality.navigability.navRegistry` | No | `array<string>` | `[]` | Tokens identifying the nav-registry SSOT the navigability lens checks every route resolves a nav door against. |
 | `quality.navigability.journeySuite` | No | `string` | — | Path or command for the per-persona journey suite /mandrel-deliver's per-Story ceremony runs. |
-| `mergeWatch` | No | `object` | — | Knobs consumed by the close-and-land merge wait (Story #4543). `mode` (Story #4698) selects the close-time merge posture. `maxWaitSeconds` bounds ONE invocation of the merge wait and its expiry returns a resumable `pending` terminal with no label mutation; `maxBudgetSeconds` bounds the CUMULATIVE wait across resumes (anchored at the PR's createdAt, so a resume does not restart the clock) and exhausting it is the genuine give-up that classifies and blocks. The 30s poll cadence and the cap of 3 behind-the-base updates are fixed. |
+| `mergeWatch` | No | `object` | — | Knobs consumed by the close-and-land merge wait (Story #4543). `mode` (Story #4698) selects the close-time merge posture. `maxWaitSeconds` bounds ONE invocation of the merge wait and its expiry returns a resumable `pending` terminal with no label mutation; `maxBudgetSeconds` bounds the CUMULATIVE wait across resumes (anchored at the PR's createdAt, so a resume does not restart the clock) and exhausting it is the genuine give-up that classifies and blocks. The poll cadence is internal (at most 30s while checks run, at most 10s once they are green and the PR is unmerged); it and the cap of 3 behind-the-base updates are not configurable. |
 | `mergeWatch.mode` | No | `"sync"` \| `"async"` | — | Close-time merge-wait posture (Story #4698). `sync` (default) keeps the in-close foreground merge wait unchanged. `async` caps the per-invocation wait to a short ~60s probe window — long enough to catch an instant merge and, via the head-anchored required-check predicate, an instantly-red required check — then returns the resumable `pending` terminal (exit 3) with a `nextCommand`. Opt in when slow CI makes the foreground wait routinely expire: the worker launches `nextCommand` in the background instead of burning the host tool slot polling. `maxBudgetSeconds` (the cumulative give-up) is unchanged. |
 | `mergeWatch.maxWaitSeconds` | No | `integer` | — | Per-invocation merge-wait bound (seconds). Default 300 (5 minutes) — chosen to fit inside a single host tool invocation (~10 min ceiling) alongside the close gates that precede the wait. Expiry yields `pending` (exit 3), never a block. Headless callers with no host ceiling raise this to land in one block. |
 | `mergeWatch.maxBudgetSeconds` | No | `integer` | `3600` | Cumulative wall-clock budget (seconds) across merge-wait resumes, anchored at the PR's createdAt. Default 3600 (60 minutes). Exhausting this classifies the block and transitions the Story to agent::blocked. |
 | `codeReview` | No | `object` | — | Review-provider chain and on-branch remediation threshold for the /mandrel-deliver code-review ceremony. |
-| `codeReview.providers[]` | No | `array<object>` | `[{"name":"native"},{"name":"security-review","scopes":["story"],"optional":true},{"name":"ultrareview","scopes":["story"],"manualPrompt":true,"when":{"label":"risk::high"}}]` | Review-provider chain (Story #2871). When unset or empty, defaults to [{ name: "native" }]. The orchestrator iterates inline entries in declaration order and merges their Finding[] before posting one structured comment; manual-prompt entries (e.g. ultrareview) contribute a trailing 'Manual review suggestions' section. Selecting an adapter whose probe fails hard-fails at factory construction unless declared `optional: true` in the chain. Each item has: name, scopes, optional, manualPrompt, when. |
+| `codeReview.providers[]` | No | `array<object>` | `[{"name":"native"},{"name":"code-review","scopes":["story"],"optional":true}]` | Review-provider chain (Story #2871). When unset or empty, falls back to this default: `native` (scoped lint + MI) then an optional story-scoped `code-review` — a low-effort `claude --print` bug review whose every finding is critical and halts close before auto-merge; hosts without the `claude` CLI skip it. `security-review` and `ultrareview` are opt-in. The orchestrator iterates inline entries in declaration order and merges their Finding[] before posting one structured comment; manual-prompt entries (e.g. ultrareview) contribute a trailing 'Manual review suggestions' section. Selecting an adapter whose probe fails hard-fails at factory construction unless declared `optional: true` in the chain. Each item has: name, scopes, optional, manualPrompt, when. |
 | `codeReview.autoFixSeverity` | No | `"high"` \| `"medium"` | `"medium"` | Severity threshold for on-branch remediation in /mandrel-deliver Phase 5 (code-review). `medium` (default) routes 🔴/🟠/🟡 findings into the host-LLM focused-fix routing (Mediums batched per lens: one commit per lens, a single validation + rescan at the end) while 🟢 suggestions still graduate to follow-up issues; `high` reproduces the pre-4399 Critical/High-only routing. Hard cutover — no back-compat flag. |
 | `refactorStage` | No | `object` | — | Opt-in, config-gated post-green refactor checkpoint wired into story-deliver (Story #3430, Epic #3418). Strictly additive and default-OFF: when disabled, story-deliver behaves exactly as before. Advisory only — never changes existing close-validation gate semantics. |
 | `refactorStage.enabled` | No | `boolean` | `false` | When true, story-deliver runs an advisory post-green refactor stage (core/code-review-and-quality skill, Post-Green Refactor Pass) after the suite is green. Default false — when unset the stage is skipped and close-validation gate semantics are unchanged. |
@@ -248,7 +248,7 @@ Everything `/mandrel-deliver` and `single-story-close` consume: worktree isolati
 | `ci.advisoryAllowlist` | No | `array<string>` | `[]` | Story #5096. Check-run names exempt from blockOnAdvisoryFailure — a red run whose name matches exactly never blocks arming. Matching is exact; an unnamed run can never match and always blocks. |
 | `ci.rerunAdvisory` | No | `integer` | `0` | Story #5266. How many times close may re-run a failed advisory workflow run before blocking on it, per close invocation. Default 0: close spends no CI minutes and issues no GitHub mutation on an advisory red unless asked. At n > 0 the failed run(s) are re-run within that allowance and the merge wait re-polls inside its existing budget, landing or blocking on the re-run verdict. Overridden per invocation by --rerun-advisory <n>. |
 | `routing` | No | `object` | — | v2 delivery-spawn routing: role-scoped boot contexts and the ceremony profile. The v1 singleDelivery epic-route kill-switch was removed in Stage 6; the freshCriticSampleRate sampling floor was retired in Story #5313 and the derived-level ceremony routing in Story #5343. |
-| `routing.roleScopedAgents` | No | `boolean` | `true` | Epic #4478 (M7-B). Kill-switch for the role-scoped boot contexts. When true (default), a converted delivery spawn (`story-worker`, `acceptance-critic`) boots on its own `.claude/agents/<role>.md` system prompt instead of re-paying the full CLAUDE.md @-import closure. When false, every converted spawn falls back to `subagent_type: general-purpose` — the instant, code-rollback-free per-consumer revert, and the universal escape for hosts that ignore `.claude/agents/`. The fallback is the full-closure agent that ran before M7-B, so flipping it off never drops a gate. |
+| `routing.roleScopedAgents` | No | `boolean` | `true` | Epic #4478 (M7-B). Kill-switch for the role-scoped boot contexts. When true (default), a converted delivery spawn (`story-worker`, `acceptance-critic`) boots on its own `.claude/agents/<role>.md` system prompt instead of re-paying the full entry-doc (AGENTS.md or CLAUDE.md) @-import closure. When false, every converted spawn falls back to `subagent_type: general-purpose` — the instant, code-rollback-free per-consumer revert, and the universal escape for hosts that ignore `.claude/agents/`. The fallback is the full-closure agent that ran before M7-B, so flipping it off never drops a gate. |
 | `routing.ceremonyProfile` | No | `"minimal"` \| `"standard"` \| `"strict"` | `"standard"` | Acceptance-ceremony depth — who authors the Story acceptance verdict. minimal and standard (default) = the inline self-eval, whatever the diff touches; strict = a fresh-context maker-blind critic. Review depth is a separate decision and still derives `deep` for any sensitive path (review-depth.js). |
 | `routing.closeAndLand` | No | `boolean` | `true` | When true (default), single-story-close lands through merge in one close. Opt out per-run with --no-wait-merge. |
 
@@ -386,15 +386,16 @@ then skips `npm ci` and junctions/symlinks the donor's `node_modules`.
 
 #### `delivery.codeReview` — provider chain
 
-`providers[]` is iterated in declaration order (unset/empty →
-`[{ name: "native" }]`). Beyond the field list in the generated table: `scopes`
+`providers[]` is iterated in declaration order (unset/empty → `native` plus
+an optional story-scoped `code-review`, the low-effort `claude --print` bug
+review whose every finding is critical and halts close). Beyond the field list in the generated table: `scopes`
 defaults to both `story` and `epic`; `when` is a label predicate (`when.label`
 / `when.labelAny`) that silently skips the entry when false; `manualPrompt`
 entries contribute a one-line suggestion and do **not** affect severity counts
 or the `halted` gate. **Cross-runtime contract:** manual-prompt providers emit
 Markdown only and MUST NEVER throw under any host; inline providers that shell
-out to a host-specific binary (e.g. `security-review` →
-`claude --print /security-review`) SHOULD be declared `optional: true` so
+out to a host-specific binary (e.g. `code-review` / `security-review` →
+`claude --print`) SHOULD be declared `optional: true` so
 non-Claude consumers can pin the same `.agents/` version unmodified. The fix
 budget (`maxFixAttempts` / `maxFixScopeFiles`) uses the same values for every
 Story in a run — there is no per-Story override.

package/.agents/docs/execution-reference.md

@@ -72,11 +72,12 @@ and schema mechanics are in [§ Friction telemetry](#friction-telemetry) above.
 ## FinOps & token budgeting (economic guardrails)
 
 Mandrel does **not** enforce live LLM spend from response metadata. It bounds
-two things, both **fixed framework constants** rather than operator knobs, and
-both **fail closed**: the assembled `/mandrel-plan` context envelope, and plan-time
-Story sizing. Your host runtime (editor / CLI) owns session quota and hard
-stops. Consult this section when reasoning about why `/mandrel-plan` refused an
-over-ceiling envelope or an over-budget Story count.
+one thing, a **fixed framework constant** rather than an operator knob that
+**fails closed**: the assembled `/mandrel-plan` context envelope. Plan-time
+Story sizing is not bounded (retired by Story #5312). Your host runtime
+(editor / CLI) owns session quota and hard stops. Consult this section when
+reasoning about why `/mandrel-plan` refused an over-ceiling envelope, or when
+choosing the session effort and model for a command.
 
 > **There is no configurable context budget.** `planning.context.maxBytes` /
 > `summaryMode` were removed outright in Story #4541, along with the
@@ -126,3 +127,24 @@ over-ceiling envelope or an over-budget Story count.
   nothing at plan time scores its authored mass.
 - **Host runtime**: session billing, quota exhaustion, and operator overrides
   are enforced by your provider (e.g. Claude Code), not by Mandrel scripts.
+
+### Session effort and model
+
+Effort is the operator's dial, set once per session. Pick it deliberately up
+front, because changing effort mid-session invalidates the prompt cache for
+everything that follows.
+
+- **Recommended session effort.** `medium` for `/mandrel-deliver`: Stories
+  are well-scoped by construction, so delivery rarely needs more. `high` for
+  `/mandrel-plan`: a Spec defect costs a redraft or a blocked Story
+  downstream, which is dearer than the planning turn.
+- **Role agents.** `story-worker` declares no effort and inherits the
+  session's (Story #5426). The evaluator roles (`acceptance-critic`,
+  `plan-critic`, `auditor`) pin `medium`.
+- **Where pins live.** Effort and model pins belong only on role agents under
+  `.agents/agents/`, never in workflow or command frontmatter: a command that
+  pinned its own effort would change effort mid-session and break the prompt
+  cache.
+- **Escalation is an operator step.** The Agent tool takes no per-call
+  effort, so no workflow can escalate on its own. Before resuming a blocked
+  Story, raise session effort one step; raise effort before switching models.

package/.agents/instructions.md

@@ -118,7 +118,7 @@ truncates with a note naming what was cut:
 
 ## 3. Core Philosophy
 
-1. **Context First.** **Digest-first reading (Story #4433):** never
+1. **Context First.** **Digest-first reading:** never
    ingest the whole `project.docsContextFiles` set up front — read the
    docs digest and pull files on demand at the section it names. No
    digest (ad hoc task, `docsContextFiles` unset, null `docsDigestPath`)
@@ -127,8 +127,10 @@ truncates with a note naming what was cut:
    present. Always read the current Story's body (`## Spec` +
    `acceptance[]` / `verify[]`); prefer targeted retrieval over broad
    reads.
-2. **Plan First.** For non-trivial tasks (3+ steps or architectural
-   decisions), update the Story's `## Spec` via `/mandrel-plan` before code.
+2. **Plan First.** Planned work carries its plan in the Story's
+   `## Spec`, authored via `/mandrel-plan`; an unplanned prompt takes
+   `/mandrel-deliver`'s light path, which escalates to `/mandrel-plan`
+   when its gate trips.
 3. **Artifacts over Chat.** Write test/build/debug output to log
    files, not into chat.
 4. **Idempotency.** Scripts must be safe to run repeatedly.
@@ -137,16 +139,13 @@ truncates with a note naming what was cut:
 
 ## 4. Execution & Quality Discipline
 
-- **Re-Plan on Failure.** If a strategy fails, STOP and re-plan.
 - **Subagent Strategy.** Each spawn re-pays the full always-loaded
   context — a cost decision. Prefer inline search for small lookups;
   spawn only when the work justifies replicating context. One objective
   per subagent; depth compounds the cost (every nested level re-pays).
-- **Anti-Laziness / No Dead Code.** NEVER use placeholder comments like
-  `// ... existing code ...`; every edit must leave complete, runnable code.
-  Remove unused imports, commented-out code, and dead branches before
+- **No Dead Code.** Every edit leaves complete, runnable code. Remove
+  unused imports, commented-out code, and dead branches before
   finalizing.
-- **Verification.** Include explicit verification steps in every plan.
 
 ---
 
@@ -191,7 +190,6 @@ anything under it.
 `/mandrel-plan` sizes each Story as a **capability slice a frontier model
 delivers and self-verifies in one pass** — a broad footprint is normal
 when the change is cohesive, and no plan-time ceiling scores it; do not
-re-slice it into per-module fragments. On an out-of-scope task: **plan
-first** in numbered cohesive sub-steps, **commit incrementally** per
-sub-step, and **fail fast** — STOP and report if any sub-step fails
-validation.
+re-slice it into per-module fragments. On an out-of-scope task,
+**commit incrementally** per cohesive sub-step, and when a sub-step
+fails validation, stop and apply § 1.I (re-plan or yield).

package/.agents/rules/ci-remediation.md

@@ -137,9 +137,9 @@ run — a blind fix to a suite nobody exercised is how the gap compounds.
 
 `capacity` and `unreproducible-tier` name failures that are proven properties
 of the **environment**: no commit on the branch can move the head SHA to clear
-them, so the no-rerun rule used to strand a correct delivery until a human
-cleared it by hand. Those two verdicts — and only those two — now buy exactly
-one rerun:
+them, so without an allowance the no-rerun rule would strand a correct
+delivery until a human cleared it by hand. Those two verdicts — and only those
+two — buy exactly one rerun:
 
 1. Reach the verdict with its required readings (§ above). A green on re-run is
    never one of those readings.

package/.agents/rules/gherkin-standards.md

@@ -143,8 +143,9 @@ authored:
    widen the regex, updating every call site in the same PR.
 4. **Add a new definition only when no reasonable match exists**, in the
    correct domain directory. Never copy-paste a step implementation to support
-   a paraphrased scenario, and never author new step definitions during
-   scenario authoring — record the missing step as a named gap instead.
+   a paraphrased scenario. A prose-only authoring pass (one whose scope
+   excludes step-definition code) records the missing step as a named gap
+   instead of writing it.
 
 When a step is superseded, mark it deprecated and migrate every call site in
 the same PR; do not leave two near-identical steps live.

package/.agents/rules/git-conventions-reference.md

@@ -15,7 +15,9 @@ Mandrel ships as the `mandrel` npm package, whose consumers pin an
 exact lockfile version; they opt into breaks at upgrade time. Operator policy
 for any contract change (config shape, baseline shape, schema, lifecycle
 payload, ticket label, dispatch artifact, public API of a script) is
-therefore:
+therefore as follows. It governs Mandrel's own framework contracts; a
+consumer's product API keeps the expand–contract rule in
+[`api-conventions.md`](api-conventions.md).
 
 1. **Hard cutovers only.** Contract changes ship as a single in-tree
    migration of every producer and consumer. There is no parallel
@@ -154,10 +156,10 @@ different hazards:
 
 ## Meta Labels (Retrospective Signal Routing)
 
-Two `meta::*` labels route retrospective signals into durable substrates so
+Three `meta::*` labels route retrospective signals into durable substrates so
 the `/mandrel-plan` Phase 0 fetcher (see
 [`prior-feedback-fetcher.js`](../scripts/lib/feedback-loop/prior-feedback-fetcher.js))
-can surface open feedback issues to the planner. Both labels live in
+can surface open feedback issues to the planner. All three live in
 [`label-constants.js`](../scripts/lib/label-constants.js) under the
 `META_LABELS` export — reference them by symbol from scripts rather than
 hard-coding the string.
@@ -180,3 +182,10 @@ project-local automation). The work is scoped to the consumer's
 framework changes. Issues that span both axes should carry both labels —
 `fetchPriorFeedback` dedupes by issue number so a dual-labeled issue
 appears exactly once in the planner context.
+
+### `meta::platform-gap`
+
+Apply this label to a GitHub issue whose fault lies in a shared base
+config, runner fleet, or cross-repo toolchain that neither the framework
+nor the consumer owns — the `--owner platform` bucket of
+[`ci-remediation.md`](ci-remediation.md).

package/.agents/rules/git-conventions.md

@@ -16,9 +16,8 @@ Every Story lands on a dedicated **Story branch** named
 `story-<storyId>`, seeded from `project.baseBranch` (`main` by default),
 isolated in its own worktree at `.worktrees/story-<id>/`. The runtime
 owns both via `single-story-init.js`; agents commit there only. Close
-opens a PR against `main` (squash + required checks). No `epic/<id>`
-integration branch, no `--no-ff` wave merge, no child tickets: commits
-land on `story-<storyId>` directly, the
+opens a PR against `main` (squash + required checks). Commits land on
+`story-<storyId>` directly, the
 subject referencing the Story via `(refs #<storyId>)` — see
 [`.agents/instructions.md` § 5.B](../instructions.md).
 
@@ -38,15 +37,18 @@ subject referencing the Story via `(refs #<storyId>)` — see
 
 ## Push Validation & Reliability (MUSTs)
 
-1. Run the configured validation commands locally **before** `git push`.
+1. Validate locally **before** `git push`. On a Story branch that is the
+   one credited suite run (`deliver-digest.md` § 5); close runs every
+   other gate, so do not pre-run them. Elsewhere, run the configured
+   validation commands.
 2. Do NOT assume a push succeeded unless the output confirms the remote
    ref was updated (`[new branch]`, `[up to date]`, `... -> ...`).
 3. If a `pre-push` hook rejects, fix the cause and create a NEW follow-up
    commit — never amend the rejected commit.
 4. **Never bypass hooks** (`--no-verify`, `--no-gpg-sign`, …) without
-   explicit operator authorization. The one recognized exception — a
-   Biome zero-match failure under a harness-managed worktree path — is a
-   consumer-tooling gap, **not** authorization; see
+   explicit operator authorization. A Biome zero-match failure under a
+   harness-managed worktree path is a known false negative — a
+   consumer-tooling gap, **not** authorization to bypass; see
    [`git-conventions-reference.md` § Push Validation](git-conventions-reference.md).
 
 ## Local checkout hygiene

package/.agents/rules/testing-standards.md

@@ -156,13 +156,14 @@ behaviour (`sets status to completed`, not `works`).
 
 A file that passes alone but fails inside the full `npm test` suite is **test
 pollution** — one test leaks shared state (env vars, temp files, the
-mock-module registry, global singletons) and a later test trips on it. Reach
-for `npm run test:isolate` before manually bisecting: it runs every matching
-file individually under `--test-concurrency=1`, then all together, flags files
-that pass alone but fail in the suite (**flippers**), binary-bisects the
-smallest reproducing subset, and reports any file that exited with leftover
-`process.env` mutations. The fix is almost always missing teardown — wrap the
-mutation in a `t.before` / `t.after` pair, or restore the prior value in
+mock-module registry, global singletons) and a later test trips on it. Run
+every matching file alone, then all together; the files that pass alone but
+fail in the suite (**flippers**) bound the search, and bisecting them finds
+the smallest reproducing subset. Mandrel's own repository automates this as
+`npm run test:isolate`, which also reports leftover `process.env`
+mutations; a consumer uses its runner's equivalent. The fix is almost
+always missing teardown — wrap the mutation in a `t.before` / `t.after`
+pair, or restore the prior value in
 `try` / `finally`.
 
 For browser-based changes, pair the cycle with runtime verification via Chrome

package/.agents/runtime-deps.json

@@ -18,6 +18,6 @@
     "chokidar": "^5.0.0",
     "jscpd": "^4.0.0",
     "knip": "^6.17.1",
-    "typescript": ">=5.0.0"
+    "typescript": ">=5.0.0 <7"
   }
 }

package/.agents/schemas/agentrc.schema.json

@@ -1180,7 +1180,7 @@
         },
         "mergeWatch": {
           "type": "object",
-          "description": "Knobs consumed by the close-and-land merge wait (Story #4543). `mode` (Story #4698) selects the close-time merge posture. `maxWaitSeconds` bounds ONE invocation of the merge wait and its expiry returns a resumable `pending` terminal with no label mutation; `maxBudgetSeconds` bounds the CUMULATIVE wait across resumes (anchored at the PR's createdAt, so a resume does not restart the clock) and exhausting it is the genuine give-up that classifies and blocks. The 30s poll cadence and the cap of 3 behind-the-base updates are fixed.",
+          "description": "Knobs consumed by the close-and-land merge wait (Story #4543). `mode` (Story #4698) selects the close-time merge posture. `maxWaitSeconds` bounds ONE invocation of the merge wait and its expiry returns a resumable `pending` terminal with no label mutation; `maxBudgetSeconds` bounds the CUMULATIVE wait across resumes (anchored at the PR's createdAt, so a resume does not restart the clock) and exhausting it is the genuine give-up that classifies and blocks. The poll cadence is internal (at most 30s while checks run, at most 10s once they are green and the PR is unmerged); it and the cap of 3 behind-the-base updates are not configurable.",
           "properties": {
             "mode": {
               "type": "string",
@@ -1215,11 +1215,12 @@
                     "type": "string",
                     "enum": [
                       "native",
+                      "code-review",
                       "codex",
                       "security-review",
                       "ultrareview"
                     ],
-                    "description": "Registered provider key. Inline: native, codex, security-review. Manual-prompt: ultrareview."
+                    "description": "Registered provider key. Inline: native, code-review, codex, security-review. Manual-prompt: ultrareview."
                   },
                   "scopes": {
                     "type": "array",
@@ -1263,23 +1264,15 @@
                 },
                 "additionalProperties": false
               },
-              "description": "Review-provider chain (Story #2871). When unset or empty, defaults to [{ name: \"native\" }]. The orchestrator iterates inline entries in declaration order and merges their Finding[] before posting one structured comment; manual-prompt entries (e.g. ultrareview) contribute a trailing 'Manual review suggestions' section. Selecting an adapter whose probe fails hard-fails at factory construction unless declared `optional: true` in the chain.",
+              "description": "Review-provider chain (Story #2871). When unset or empty, falls back to this default: `native` (scoped lint + MI) then an optional story-scoped `code-review` — a low-effort `claude --print` bug review whose every finding is critical and halts close before auto-merge; hosts without the `claude` CLI skip it. `security-review` and `ultrareview` are opt-in. The orchestrator iterates inline entries in declaration order and merges their Finding[] before posting one structured comment; manual-prompt entries (e.g. ultrareview) contribute a trailing 'Manual review suggestions' section. Selecting an adapter whose probe fails hard-fails at factory construction unless declared `optional: true` in the chain.",
               "default": [
                 {
                   "name": "native"
                 },
                 {
-                  "name": "security-review",
+                  "name": "code-review",
                   "scopes": ["story"],
                   "optional": true
-                },
-                {
-                  "name": "ultrareview",
-                  "scopes": ["story"],
-                  "manualPrompt": true,
-                  "when": {
-                    "label": "risk::high"
-                  }
                 }
               ]
             },
@@ -1367,7 +1360,7 @@
           "properties": {
             "roleScopedAgents": {
               "type": "boolean",
-              "description": "Epic #4478 (M7-B). Kill-switch for the role-scoped boot contexts. When true (default), a converted delivery spawn (`story-worker`, `acceptance-critic`) boots on its own `.claude/agents/<role>.md` system prompt instead of re-paying the full CLAUDE.md @-import closure. When false, every converted spawn falls back to `subagent_type: general-purpose` — the instant, code-rollback-free per-consumer revert, and the universal escape for hosts that ignore `.claude/agents/`. The fallback is the full-closure agent that ran before M7-B, so flipping it off never drops a gate.",
+              "description": "Epic #4478 (M7-B). Kill-switch for the role-scoped boot contexts. When true (default), a converted delivery spawn (`story-worker`, `acceptance-critic`) boots on its own `.claude/agents/<role>.md` system prompt instead of re-paying the full entry-doc (AGENTS.md or CLAUDE.md) @-import closure. When false, every converted spawn falls back to `subagent_type: general-purpose` — the instant, code-rollback-free per-consumer revert, and the universal escape for hosts that ignore `.claude/agents/`. The fallback is the full-closure agent that ran before M7-B, so flipping it off never drops a gate.",
               "default": true
             },
             "ceremonyProfile": {

package/.agents/schemas/audit-rules.schema.json

@@ -2,7 +2,7 @@
   "$schema": "http://json-schema.org/draft-07/schema#",
   "$id": "https://mandrel.dev/schemas/audit-rules.schema.json",
   "title": "AuditRulesManifest",
-  "description": "Schema describing the audit-rules.json manifest read by lib/audit-suite/selector.js and lib/audit-suite/runner.js. The manifest enumerates registered audit workflows and the triggers / substitution keys each one accepts, plus the sensitive-path classes that derive review depth from a change set.",
+  "description": "Schema describing the audit-rules.json manifest read by lib/audit-suite/selector.js and lib/audit-suite/checklist-threading.js. The manifest enumerates registered audit workflows and the triggers / substitution keys each one accepts, plus the sensitive-path classes that derive review depth from a change set.",
   "type": "object",
   "required": ["version", "audits"],
   "additionalProperties": false,

package/.agents/schemas/story-deliver-terminal.schema.json

@@ -204,6 +204,11 @@
       },
       "additionalProperties": false
     },
+    "phaseDurations": {
+      "type": ["object", "null"],
+      "description": "Story #5417 — wall-clock seconds per close phase that ran (wrong-tree-guard, base-sync, close-validation, push, pull-request, code-review, auto-merge, confirm-merge, post-land). Optional and additive; absent or null when no phase was timed.",
+      "additionalProperties": { "type": "number", "minimum": 0 }
+    },
     "timestamp": { "type": "string", "format": "date-time" }
   },
   "allOf": [