@garygentry/feature-forge 0.2.3 → 0.2.5

package/adapters/gemini/references/pipeline-state-schema.json

@@ -80,7 +80,7 @@
         "completedAt": { "type": ["string", "null"], "format": "date-time" },
         "commitHash": {
           "type": ["string", "null"],
-          "description": "Git commit SHA after this stage completed"
+          "description": "Git commit SHA of this stage's artifact commit (Commit 1 of the two-commit Git Commit Protocol in shared-conventions.md). Written by a small follow-up commit so it points at the artifact commit itself, never at an orphaned amend or the hash-recording commit. null until recorded (or when there was no new artifact commit)."
         },
         "basedOnVersions": {
           "type": "object",
@@ -107,7 +107,8 @@
         },
         "verifiedAt": { "type": ["string", "null"], "format": "date-time" },
         "fixedAt": { "type": ["string", "null"], "format": "date-time" },
-        "commitHash": { "type": ["string", "null"] }
+        "commitHash": { "type": ["string", "null"], "description": "Git commit SHA of the verify/fix artifact commit. Recorded via the two-commit Git Commit Protocol (shared-conventions.md) so it points at the artifact commit, never an orphaned amend." },
+        "verifiedStageVersion": { "type": ["integer", "null"], "description": "The production stage's `version` at the moment this verify was resolved. The navigator's freshness ledger compares it to the stage's current `version`: equal means the verify is fresh; a mismatch (artifact revised since) or an absent field (legacy state) means stale, so auto-verify re-fires. Recorded by forge-verify/forge-fix when writing a passed/findings-applied status." }
       }
     }
   }

package/adapters/gemini/references/portable-root.md

@@ -12,7 +12,7 @@ against the fenced block here, byte-for-byte.
 ## Canonical bootstrap prelude
 
 ```bash
-R="$(for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done)"
+R="$(bash -c 'for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/cache/*/feature-forge/* "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done')"
 [ -n "$R" ] || { echo "feature-forge: cannot locate plugin root" >&2; exit 1; }
 ```
 
@@ -24,7 +24,7 @@ makes several calls, add the prelude once and reuse `$R` for each. A fresh block
 prelude (per-block re-resolution). Worked example:
 
 ```bash
-R="$(for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done)"
+R="$(bash -c 'for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/cache/*/feature-forge/* "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done')"
 [ -n "$R" ] || { echo "feature-forge: cannot locate plugin root" >&2; exit 1; }
 python3 "$R/scripts/epic-manifest.py" render-status "{epic}" --specs-dir "{specsDir}" --json
 ```
@@ -43,7 +43,10 @@ python3 "$R/scripts/epic-manifest.py" render-status "{epic}" --specs-dir "{specs
 3. **Prelude candidate set is an agent-neutral bootstrap subset.** The prelude's `for d` list
    exists only to bootstrap-discover `forge-root.sh`; the authoritative multi-root probe lives in
    `forge-root.sh` step 2. The list enumerates install roots across agents — the Claude
-   skill/plugin dirs **and** the agent-neutral `.agents/skills/feature-forge` dirs (`$HOME` and the
+   skill/plugin dirs — including the marketplace-cache layout
+   `~/.claude/plugins/cache/<marketplace>/feature-forge/<version>/`, listed before the
+   single-star plugins glob so a versioned cache install always beats the marketplace clone —
+   **and** the agent-neutral `.agents/skills/feature-forge` dirs (`$HOME` and the
    project-relative `./.agents/...`) — so a non-Claude install (e.g. Codex under `.agents/skills`)
    can still discover the resolver. When adding an install root, update `forge-root.sh` first;
    extend the prelude only if the new root is needed to bootstrap-discover `forge-root.sh` itself.

package/adapters/gemini/references/process-overview.md

@@ -36,6 +36,16 @@ The pipeline compiles a fuzzy feature idea into a machine-executable `backlog.js
 **Output:** `{specsDir}/{feature}/.verification/VERIFY-<stage>-<timestamp>.md` (includes both findings and a Fix Execution Plan)
 **Method:** Clean-context analysis producing actionable findings with an ordered fix plan.
 
+**Manual or automatic.** By default the navigator offers verification after each stage. Because
+it runs in a fresh, read-only subagent (clean-room by construction, never needs a `/clear`), it is
+safe to automate: set `autoVerify: true` (or per-stage via `autoVerifyStages`) in
+`forge.config.json` and the navigator runs `forge-verify` automatically once a stage completes,
+returning only a compact digest to the session. The freshness ledger (verify entries record the
+artifact `version` they ran against) means a stage is re-verified only when its artifact changes,
+and an explicitly `skipped` verify is respected. Fixing stays human-gated unless `autoFix: true` is
+also set, which chains `forge-fix` only when preconditions hold (zero unresolved decisions, clean
+tree, passing re-verify). See `references/shared-conventions.md` for the config keys.
+
 After verification, fixes can be applied via:
 - `/feature-forge:forge-fix <feature>` — reads the Fix Execution Plan from the findings document and applies changes (works in any session)
 - Plan mode workflow — enter plan mode, run verify, review plan, exit and execute

package/adapters/gemini/references/shared-conventions.md

@@ -71,6 +71,9 @@ Extract these config values (use defaults if not present):
 - `autoInvokeNextStage` (default: `true` — the `/feature-forge:forge` navigator auto-invokes the next stage via the `Skill` tool after the user confirms; `false` keeps copy-paste behavior. Navigator-only.)
 - `contextWindowTokens` (default: `null` — context window used by the navigator's context-usage check; `null` infers from the session model and falls back to 200000. Set to the model's window, e.g. `1000000` on a 1M model. Navigator-only.)
 - `contextWarnThreshold` (default: `0.7` — fraction of the window past which the navigator recommends a clean session. Navigator-only.)
+- `autoVerify` (default: `false` — when `true`, `forge-verify` runs automatically after a stage completes, no prompt. **In-stage-primary:** the just-completed authoring stage runs it itself, in-session, before the exit block (honoring the verify-before-clear principle). The navigator runs it only as a **catch-up** when verify is still pending (a host that could not dispatch a clean-room subagent, or a stage run before this behavior landed). Either way it runs in a fresh clean-room subagent, so it never needs a `/clear` and costs only a compact digest.)
+- `autoVerifyStages` (default: `{}` — per-stage overrides for `autoVerify`, e.g. `{"forge-1-prd": false}`. Effective value = `autoVerifyStages[stage]` if present, else `autoVerify`. Keys are constrained to the five verify-capable stages; a typo is a config error surfaced as `invalidAutoVerifyKeys`. Both the in-stage run and the navigator catch-up read this same effective value.)
+- `autoFix` (default: `false` — when `true`, `forge-fix` is chained after an auto-verify that finds issues — by the in-stage run (primary) or the navigator catch-up — but only when auto-verify is on for that stage AND preconditions hold (zero unresolved decisions, clean tree, passing re-verify); otherwise a digest is surfaced and the gate is presented.)
 - `loopRunner` (optional object — the loop runner to drive; **defaults to rauf** when absent, with every command templated. See `references/forge-config-schema.json` and `references/ralph-loop-contract.md`.)
 
 ## Feature Directory Resolution
@@ -78,7 +81,7 @@ Extract these config values (use defaults if not present):
 Before any file I/O against a feature's artifacts, resolve its directory through the deterministic helper rather than hardcoding `{specsDir}/{feature}/`. This makes flat (`{specsDir}/{feature}/`) and nested (`{specsDir}/{epic}/{feature}/`) layouts both resolve from a bare feature name (REQ-DIR-03), with standalone features behaving exactly as today.
 
 ```bash
-R="$(for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done)"
+R="$(bash -c 'for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/cache/*/feature-forge/* "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done')"
 [ -n "$R" ] || { echo "feature-forge: cannot locate plugin root" >&2; exit 1; }
 resolvedFeatureDir=$(python3 "$R/scripts/epic-manifest.py" \
   resolve "<feature>" --specs-dir "<specsDir>")
@@ -90,6 +93,19 @@ resolvedFeatureDir=$(python3 "$R/scripts/epic-manifest.py" \
 
 In both failure cases, do not fall back to a guessed path.
 
+**On `not-found`, check other branches before stopping.** With `branchPerFeature`, the feature's directory (and its `.pipeline-state.json`) may exist only on its topic branch — invisible from the default branch of a fresh clone. Before concluding the pipeline does not exist, run the read-only cross-branch discovery:
+
+```bash
+R="$(bash -c 'for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/cache/*/feature-forge/* "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done')"
+[ -n "$R" ] || { echo "feature-forge: cannot locate plugin root" >&2; exit 1; }
+python3 "$R/scripts/forge-session.py" discover-feature "<feature>" --specs-dir "<specsDir>" --json
+```
+
+- **Candidates found** (`candidates` and/or `remoteCandidates` non-empty): summarize them as text (branch, recorded stage, whether the state's own `branch` field matches), then use `AskUserQuestion`: **Switch to `{branch}` (recommended)** — run the candidate's `switchCommand` · **Fetch + switch** — for a `needsFetch` remote candidate, run its `fetchCommand` then `switchCommand` (note its contents were matched by name only, not inspected) · **Treat `{feature}` as new on this branch** · **Stop**. A checkout is a mutation inside an otherwise read-only flow: perform it ONLY on the user's explicit accept AND with a clean working tree (`git status --porcelain` prints nothing) — never auto-switch, never with uncommitted changes. After a successful switch, re-run this Feature Directory Resolution block from the top.
+- **Nothing found** (both lists empty): the pipeline genuinely does not exist anywhere discoverable — STOP and surface the original `not-found` stderr line verbatim (or, where the caller offers to start a new pipeline, offer that).
+
+**Anti-fabrication guard.** Never describe pipeline state that resolution or discovery did not return: if both come back empty, the pipeline does not exist — say exactly that, and never reconstruct stages, backlogs, or history from conversational memory.
+
 **Resolution algorithm (summary; full spec in `02-manifest-helper-cli.md §4`):**
 1. Reject the name if unsafe (path separator, `..`, absolute, or failing `SAFE_NAME_RE`) — before any filesystem access.
 2. If `{specsDir}/{name}/.pipeline-state.json` exists → return that flat path.
@@ -108,7 +124,7 @@ Whenever a stage creates the specs tree for the first time (the first PRD or epi
 Run this after creating the feature/epic directory, before the stage's git commit:
 
 ```bash
-R="$(for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done)"
+R="$(bash -c 'for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/cache/*/feature-forge/* "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done')"
 [ -n "$R" ] || { echo "feature-forge: cannot locate plugin root" >&2; exit 1; }
 mkdir -p "<specsDir>"
 [ -f "<specsDir>/AGENTS.md" ] || cp "$R/references/templates/specs-hygiene/AGENTS.md" "<specsDir>/AGENTS.md"
@@ -135,7 +151,7 @@ After resolving the feature directory, check the feature's `.pipeline-state.json
 To obtain the manifest contracts and the live completion status of each dependency in one deterministic call, run `render-status` and read the per-feature `status` and the `consumes`/`exposes` arrays rather than re-deriving them:
 
 ```bash
-R="$(for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done)"
+R="$(bash -c 'for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/cache/*/feature-forge/* "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done')"
 [ -n "$R" ] || { echo "feature-forge: cannot locate plugin root" >&2; exit 1; }
 python3 "$R/scripts/epic-manifest.py" \
   render-status "<epic>" --specs-dir "<specsDir>" --json
@@ -180,16 +196,18 @@ Invoke this block at the **very start** of a pipeline entry point — `forge-1-p
 
 ## Git Commit Protocol
 
-When `gitCommitAfterStage` is true, follow this exact order to avoid state inconsistency:
+When `gitCommitAfterStage` is true, follow this exact order to avoid state inconsistency.
+
+**Why two commits.** The stage's `.pipeline-state.json` is itself part of the staged commit, but the stage's `commitHash` cannot be known until *after* that commit is made. Recording it *inside* the same commit is a chicken-and-egg with no single-commit solution. Resolve it with a **deterministic two-commit sequence**, and **never** with `git commit --amend`: amending rewrites HEAD, so a hash captured before the amend points at an orphaned commit that is not in the final history (the exact defect this protocol exists to prevent).
 
 1. **Stage specific files only:** `git add {specsDir}/{feature}/` — never use `git add -A` or `git add .`
-2. **Attempt commit:** `git commit -m "{commitPrefix}({feature}): <action>"`
-3. **If commit succeeds:** capture the commit hash, then update pipeline state with `status: "complete"` and the `commitHash`
-4. **If commit fails:** do NOT update pipeline state to complete. Report the error to the user and leave state as `in-progress` so the stage can be resumed. Common failure causes:
+2. **Commit 1 — artifacts + state, hash not yet known:** In `.pipeline-state.json`, set this stage's `status: "complete"` and `commitHash: null`, then `git commit -m "{commitPrefix}({feature}): <action>"`. This is the stage's **artifact commit**; its hash is the provenance hash callers rely on.
+3. **If Commit 1 succeeds — Commit 2 records the hash:** Capture the hash of Commit 1 (`git rev-parse HEAD`). Write it into this stage's `commitHash` in `.pipeline-state.json`, then commit only that one-line change: `git add {specsDir}/{feature}/.pipeline-state.json && git commit -m "{commitPrefix}({feature}): record stage commit hash"`. The stored `commitHash` now points at the artifact commit (Commit 1) — never at Commit 2, and never at an orphaned amend. The working tree is clean afterward, so the next stage's dirty-tree check passes.
+4. **If Commit 1 fails:** do NOT update pipeline state to complete. Report the error to the user and leave state as `in-progress` so the stage can be resumed. Common failure causes:
    - **Pre-commit hook failure:** Report the hook output. Never use `--no-verify` to bypass. Help the user fix the underlying issue.
    - **Merge conflicts:** Report conflicting files. Suggest resolution steps appropriate to the conflict.
-   - **Nothing to commit:** If all artifacts were already committed, this is fine — proceed with state update but note the absence of a new commit hash.
-5. **Never** use `git add -A`, `--no-verify`, or `--force` flags
+   - **Nothing to commit:** If all artifacts were already committed, this is fine — mark the stage `complete`, leave `commitHash` at its existing value (or `null` if there was never an artifact commit), and skip Commit 2. There is no new artifact commit to record.
+5. **Never** use `git add -A`, `--amend`, `--no-verify`, or `--force` flags
 
 ## Crash Recovery
 

package/adapters/gemini/references/stage-exit-protocol.md

@@ -0,0 +1,206 @@
+# Stage Exit Protocol
+
+The single source of truth for how every forge **authoring** stage closes. It
+replaces the old ad-hoc "Next steps:" bullet lists with one fixed, correctly-ordered
+sequence: **verify (if missing or stale) → `/clear` → run the next command.**
+
+Two principles this protocol encodes (do not relitigate — they are locked product
+decisions):
+
+1. **Clearing is recommended on its own merits at every stage boundary** — a clean
+   start for the next stage — *not* as a proxy for a full context window. Window
+   fullness only changes *how emphatically* the clear is recommended, never *whether*
+   it is.
+2. **Verify happens before the clear, never after** — in the authoring session, whether
+   manual **or** auto. Verify's clean-room subagent is dispatched from the *current*
+   session, so the findings digest and any fix decision land where the context to act on
+   them still exists. This holds for auto-verify too: the stage skill dispatches the
+   clean-room verify (and any autoFix) at stage end, in-session, before the exit — it is
+   **not** deferred to the navigator, which runs *after* the `/clear` with none of the
+   authoring context. Clearing first throws that context away.
+
+## How this file is used
+
+The five authoring stages (`forge-0-epic` … `forge-4-backlog`) close with the
+**Scripted Stage Exit**: a short stamped block (below) that runs
+`forge-session.py stage-exit`, obeys the DIRECTIVES it prints per the **directive
+contract** in this file, and prints the script-emitted NEXT-STEPS block verbatim as the
+absolute last output. All the conditional logic the old prose blocks asked the model to
+compute (effective auto-verify, freshness collapse, gate selection, host wording) now
+lives in the script, deterministically; only genuinely interactive work (clean-room
+subagent dispatch, `AskUserQuestion` gates) remains prose — specified once here, not
+per stage.
+
+The loop (`forge-5-loop`) keeps its bespoke exits: it stamps the **standard block**
+(step-6 epic-member handoff) and the **warm variant** (all-done closing) below,
+verbatim. `forge-6-docs` is **terminal** — it stamps no exit block.
+
+A drift-guard test (`tests/test_stage_exit_protocol.py`) asserts each stamp site still
+contains its block, so an edit here must be mirrored into every stamp site (and
+vice-versa).
+
+## Stamp sites
+
+| Stamp site | Block |
+|---|---|
+| `forge-0-epic` … `forge-4-backlog` | scripted-stage-exit stamp |
+| `forge-5-loop` (step-6 epic-member handoff) | standard |
+| `forge-5-loop` (all-done closing → docs) | warm |
+
+The scripted stamp fills one build-time slot, `{stage-exit-args}` — the per-stage
+argument list (e.g. `--feature "{feature}" --stage forge-2-tech`; the epic stage passes
+`--feature "{epic}" --stage forge-0-epic --next-feature "{first-actionable-feature}"`).
+`{feature}` / `{epic}` / `{specsDir}` / `{first-actionable-feature}` remain runtime
+placeholders the skill resolves before running the command, exactly as elsewhere.
+
+<!-- BEGIN: scripted-stage-exit-stamp -->
+**Close this stage with the Scripted Stage Exit** (contract: `references/stage-exit-protocol.md`; do not improvise a "Next steps" list). Run:
+
+```bash
+R="$(bash -c 'for d in "$HOME"/.claude/skills/feature-forge "$HOME"/.claude/plugins/cache/*/feature-forge/* "$HOME"/.claude/plugins/*/feature-forge "$HOME"/.agents/skills/feature-forge ./.agents/skills/feature-forge; do [ -x "$d/scripts/forge-root.sh" ] && exec "$d/scripts/forge-root.sh"; done')"
+[ -n "$R" ] || { echo "feature-forge: cannot locate plugin root" >&2; exit 1; }
+python3 "$R/scripts/forge-session.py" stage-exit {stage-exit-args} --specs-dir "{specsDir}" --host claude
+```
+
+Obey the DIRECTIVES it prints, in order, per the directive contract: `runInStageVerify: true` → dispatch the in-stage clean-room verify now (honoring `autoFixEligible`); `verifyGate: "standard"` → present the Standard Verify Gate; `verifyGate: "manual-print"` → print the `verifyCommand` for the user; non-empty `invalidAutoVerifyKeys` → print a one-line warning. Then **print the NEXT-STEPS block verbatim as your absolute last output — nothing after its sentinel line.**
+<!-- END: scripted-stage-exit-stamp -->
+
+## Directive contract
+
+`stage-exit` emits a DIRECTIVES object and a NEXT-STEPS block. The skill executes the
+directives **in this order**; the script has already computed every conditional, so a
+directive is an instruction, not a question to re-derive.
+
+### `invalidAutoVerifyKeys` (non-empty)
+
+Print a one-line warning first (e.g. "⚠️ forge.config.json `autoVerifyStages` has
+unknown keys: … — they are ignored; fix the typo").
+
+### `runInStageVerify: true` — in-stage auto-verify {stageNoun}
+
+Auto-verify is effective for this stage and verification is outstanding — verify **now,
+in this session** (principle #2 applied to auto-verify: the digest and any fix decision
+land here, where the authoring context still exists — not deferred to a post-`/clear`
+navigator):
+
+1. **Clean-room verify (require-clean).** Dispatch the clean-room `forge-verifier`
+   subagent from this session in require-clean mode — the same path the navigator uses
+   (`skills/forge-verify/SKILL.md`). It inherits none of this session's context, so no
+   `/clear` is needed and only a compact digest returns. **Clean-room unavailable** (no
+   `Agent` tool, `forge-verifier` not dispatchable, or a sentinel returned): do **not**
+   run inline — leave verify **pending** so the navigator catch-up fires on a later
+   Claude-host `/feature-forge:forge`, print the `verifyCommand` for the user to run,
+   and continue to the NEXT-STEPS block.
+2. **Verify passed / no findings** → the fresh verify state is recorded by the
+   clean-room run; continue to the NEXT-STEPS block.
+3. **Verify found findings** →
+   - **`autoFixEligible: true` AND the findings document has zero unresolved decision
+     points** → chain `feature-forge:forge-fix` in-session (it owns its own commit +
+     step tracking), then run a **mandatory re-verify** in require-clean mode. Continue
+     to the NEXT-STEPS block only if the re-verify passes. On any precondition miss, a
+     forge-fix early stop, or a red re-verify, fall through to the digest gate below —
+     never a silent partial mutation. (`autoFixEligible` already folds in the config
+     `autoFix` flag and the clean-tree precondition; a dirty tree or
+     `gitCommitAfterStage: false` arrives here as `false`.)
+   - **`autoFixEligible: false`, or unresolved decision points** → surface a **compact
+     findings digest** as text, then present the gate via `AskUserQuestion`: **Run
+     `forge-fix` now** *(recommended — you are in-context and the digest is right
+     here)* / **Clear + advance anyway** (leave the findings for later) / **Stop
+     here**. Do **not** hard-stop and do **not** silently walk past. Act on the choice,
+     then continue to the NEXT-STEPS block.
+
+### `verifyGate: "standard"` — the Standard Verify Gate
+
+Auto-verify is off for this stage and verification is outstanding (`verifyState` is
+`never`, `stale`, or `failing`). Verify **now, before clearing**, using
+`AskUserQuestion` with exactly these three options — but only when the host has a
+question mechanism **and** the clean-room path is available (the `Agent` tool plus a
+dispatchable `forge-verifier` subagent); otherwise degrade exactly as `manual-print`
+below:
+
+- **Verify {stageNoun} now** *(recommended)* — dispatch the clean-room `forge-verifier`
+  subagent from this session in require-clean mode; the digest returns here so any fix
+  decision keeps its context. One-time — it does **not** change config.
+- **Verify now + enable auto-verify going forward** — verify now **and** patch
+  `"autoVerify": true` into `forge.config.json` in place (preserve formatting and every
+  other key) so future stages verify automatically, no prompt. This complements the
+  `forge-init` opt-in. **Do not auto-commit this config change** — treat it like
+  `notes`: a user-facing edit the user commits on their own cadence, never folded into
+  a stage's artifact commit.
+- **Skip for now** — go straight to the NEXT-STEPS block without verifying. Record this
+  stage's verify status as `"skipped"` in pipeline state (mirroring the existing skip
+  handling) **only** on an explicit skip — a skip does not go stale.
+
+If verify runs and finds findings, handle them exactly as in the in-stage flow above
+(digest + `AskUserQuestion` gate; `autoFixEligible` applies unchanged).
+
+### `verifyGate: "manual-print"`
+
+Verification is outstanding but the host cannot present the gate or dispatch
+clean-room. Do **not** run verify inline — print the `verifyCommand` for the user to
+run (mirroring `autoInvokeNextStage`), offer the auto-verify enable as plain text only
+if a config write is possible, and continue to the NEXT-STEPS block. Verify state stays
+outstanding, so the navigator catch-up can fire later.
+
+### `verifyGate: "none"`
+
+Verification is already resolved (fresh or explicitly skipped) or the in-stage run
+above covers it. Say so in one line and continue to the NEXT-STEPS block.
+
+### The NEXT-STEPS block (always last)
+
+Print the script's NEXT-STEPS block **verbatim as your absolute last output**. Nothing
+follows its final sentinel line (`─ forge: end of stage ─`) — no caveats, no summary,
+no sign-off. The block already carries the `/clear` recommendation (host-aware wording
+via `--host`) and the exact next command, so trailing prose can only push the user's
+next action out of view.
+
+---
+
+## Standard block
+
+Stamped at the loop's step-6 epic-member handoff (finishing feature A → starting
+feature B's PRD). It self-adapts: step 1's verify gate only fires when verification is
+actually outstanding, so at a boundary where verify already ran (or was explicitly
+skipped, or auto-verify is on) it silently collapses to just the `/clear` →
+next-command steps.
+
+Slots: `{stage}` (a lowercase noun phrase), `{verify-command}`, `{next-command}`.
+
+<!-- BEGIN: standard-exit-block -->
+**This stage is done — walk the user through the Stage Exit Protocol** before moving on. The order is fixed, and step 2 is something only the user can do:
+
+1. **Verify {stage} first — if it isn't already verified.** If verify already ran in this session — via the in-stage auto-verify on the authoring stages, or the interactive impl-verify offered above on the loop — or is already fresh on record, or the stage was explicitly skipped, say so and go straight to step 2. Only when `autoVerify` is off for this stage **and** verify is **missing or stale** do you present the **Standard Verify Gate**: verify **now, before clearing**, using `AskUserQuestion` with exactly these three options — but only when the host has a question mechanism **and** the clean-room path is available (the `Agent` tool plus a dispatchable `forge-verifier` subagent):
+   - **Verify {stage} now** *(recommended)* — dispatch the clean-room `forge-verifier` subagent from this session in require-clean mode; the digest returns here so any fix decision keeps its context. One-time — it does **not** change config.
+   - **Verify now + enable auto-verify going forward** — verify now **and** patch `"autoVerify": true` into `forge.config.json` in place (preserve formatting and every other key) so future stages verify automatically, no prompt. This complements the `forge-init` opt-in. **Do not auto-commit this config change** — treat it like `notes`: a user-facing edit the user commits on their own cadence, never folded into a stage's artifact commit.
+   - **Skip for now** — go straight to `/clear` and the next command without verifying. Record this stage's verify status as `"skipped"` in pipeline state (mirroring the existing skip handling) **only** on an explicit skip — a skip does not go stale.
+
+   **Host / clean-room fallback (not a user-selectable option):** if the question mechanism, the `Agent` tool, or the `forge-verifier` subagent is unavailable, do **not** run clean-room — degrade to printing `{verify-command}` for the user to run inline/manually (mirroring `autoInvokeNextStage`), and offer the auto-verify enable as plain text only if a config write is possible.
+2. **Then `/clear`.** Recommended **unconditionally** at this boundary for a clean start — independent of how full the context window is. Every artifact is on disk, so the work survives the clear. **I can't `/clear` for you — you have to run it yourself.**
+3. **Then run `{next-command}`** in the fresh session — or re-run `/feature-forge:forge` to let the navigator resume from disk.
+<!-- END: standard-exit-block -->
+
+---
+
+## Warm-acceptable variant
+
+Stamp this only at the `forge-5-loop → forge-6-docs` boundary (the all-done result
+report). Here clearing is **optional**: the docs stage benefits from the still-warm
+context of what the loop actually did, and impl-verify is already offered interactively
+by the loop itself, so this block defers rather than re-presenting a gate.
+
+> **Note — no literal `/clear` here.** The warm block lives in `result-reporting.md`, a
+> skill-*own* reference that the adapter build copies **verbatim** (unlike skill bodies,
+> it is not host-term translated), so a literal `/clear` would reach non-Claude adapters
+> undegraded. The warm variant says "clearing is optional" anyway, so it is phrased
+> host-neutrally without the token on purpose — do not reintroduce `/clear` here. (The
+> standard block *does* use `/clear`; that is fine because every standard stamp site is a
+> skill **body**, where `scripts/build-adapters.py` degrades it.)
+
+<!-- BEGIN: warm-exit-block -->
+**The loop is complete — this is the one boundary where clearing before the next stage is optional.**
+
+1. **Verify is already offered above.** Impl-verify is offered interactively right after this report (Step 5b for a standalone feature, Step 6.1 for an epic member) — run it there rather than as a second gate. It runs clean-room, so it needs no fresh session.
+2. **Clearing is optional here — warm is fine.** `forge-6-docs` benefits from the still-warm context of what the loop actually did, so continuing in this same session is the easy default. A cold start also works — every artifact is on disk — but there is no need to force it.
+3. **Then run `{next-command}`** — in this warm session, or a fresh one if you prefer.
+<!-- END: warm-exit-block -->

package/adapters/gemini/scripts/epic-manifest.py

@@ -312,6 +312,16 @@ def atomic_write(path: Path, data: dict) -> None:
             handle.flush()
             os.fsync(handle.fileno())
         os.replace(tmp_path, path)
+        # fsync the parent dir so the rename itself is durable on crash, not just
+        # the file bytes (best-effort — some filesystems reject O_RDONLY dir fsync).
+        try:
+            dir_fd = os.open(parent, os.O_RDONLY)
+            try:
+                os.fsync(dir_fd)
+            finally:
+                os.close(dir_fd)
+        except OSError:
+            pass
     except OSError as exc:
         tmp_path.unlink(missing_ok=True)
         raise UsageError(f"atomic write to {path} failed: {exc}")

package/adapters/gemini/scripts/forge-bootstrap.py

@@ -26,6 +26,8 @@ import argparse
 import json
 import os
 import re
+import shlex
+import shutil
 import subprocess
 import sys
 import tempfile
@@ -281,6 +283,32 @@ def _json_text(obj: object) -> str:
     return json.dumps(obj, indent=2, ensure_ascii=False) + "\n"
 
 
+def contained_path(base: Path, *parts: str) -> Path:
+    """Join ``parts`` onto ``base`` and assert the result stays within ``base``.
+
+    Ports epic-manifest.py's ``contained_path`` guard: canonicalizes (symlink-
+    resolves) both ends and verifies containment so no scaffold write or verify
+    cwd can escape the target repo — defense in depth behind ``_validate_answers``
+    should a crafted member path slip through. Containment violations surface only
+    as exit-2 usage errors.
+
+    Args:
+        base: The containing directory (the target repo root), already known to exist.
+        *parts: Repo-relative segments to append (member path, artifact rel path).
+
+    Returns:
+        The resolved, contained absolute path.
+
+    Raises:
+        UsageError: If the resolved path escapes ``base`` (exit 2).
+    """
+    base_real = base.resolve()
+    resolved = (base_real / Path(*parts)).resolve()
+    if resolved != base_real and base_real not in resolved.parents:
+        raise UsageError(f"path escapes target dir: {os.path.join(*parts)}")
+    return resolved
+
+
 def _atomic_write_text(path: Path, text: str) -> None:
     """Write ``text`` to ``path`` atomically via a same-dir temp file + os.replace.
 
@@ -381,7 +409,7 @@ def run(
 
 
 # --------------------------------------------------------------------------- #
-# Subcommand stubs (filled by later backlog items 003/006/008/009/010)
+# Subcommands: check / scaffold / verify / commit / status (02 §8.2)
 # --------------------------------------------------------------------------- #
 
 
@@ -476,6 +504,7 @@ def _write_artifact(
     """
     if rel_path in sentinel["artifactsWritten"]:
         return
+    contained_path(target, rel_path)  # never write outside the target repo (exit 2)
     dest = target / rel_path
     if dest.exists():
         return
@@ -529,7 +558,7 @@ def _resolve_commands(member: Member) -> tuple[str, str]:
 
 
 def write_config(answers: Answers, target: Path, sentinel: Sentinel) -> None:
-    """Write forge.config.json equivalent to forge-init's output (02 §4.3, 00 §7)."""
+    """Write forge.config.json == forge-init's field set + loopRunner (02 §4.3, 00 §7)."""
     config: dict = {
         "specsDir": "./specs",
         "docsDir": "./docs/architecture",
@@ -540,6 +569,14 @@ def write_config(answers: Answers, target: Path, sentinel: Sentinel) -> None:
         "typeCheckCommand": None,
         "testCommand": None,
         "loopIterationMultiplier": 1.5,
+        # Navigator keys — kept in lockstep with forge-init.sh's emitted field set
+        # (00 §7, REQ-CFG-02) so a bootstrapped config == forge-init's + loopRunner.
+        "autoInvokeNextStage": True,
+        "contextWindowTokens": None,
+        "contextWarnThreshold": 0.7,
+        "autoVerify": False,
+        "autoVerifyStages": {},
+        "autoFix": False,
         "loopRunner": {"name": "rauf", "bin": "rauf"},
     }
     if answers["layout"] == "single":
@@ -697,9 +734,9 @@ def scaffold(target: Path, answers: Answers) -> list[str]:
 def toolchain_present(required: list[str]) -> bool:
     """Return True iff every required tool is on PATH (REQ-LIFE-03).
 
-    Probes each binary with ``command -v`` via the run wrapper (a shell builtin,
-    invoked through ``sh -c``). A single missing tool yields False, driving the
-    distinct missing-toolchain outcome (exit 2, 00 §9): the skill then offers
+    Probes each binary with :func:`shutil.which` (a pure PATH lookup, no shell
+    spawned). A single missing tool yields False, driving the distinct
+    missing-toolchain outcome (exit 2, 00 §9): the skill then offers
     scaffold-anyway-unverified vs abort and marks the baseline unverified
     (REQ-LIFE-04). Bootstrap NEVER installs a toolchain (tech-spec §9).
 
@@ -708,17 +745,10 @@ def toolchain_present(required: list[str]) -> bool:
             already {pm}-substituted.
 
     Returns:
-        True iff ``command -v`` succeeds for every entry.
+        True iff every entry resolves on PATH.
     """
     for tool in required:
-        try:
-            proc = run(["sh", "-c", f"command -v {tool}"], cwd=Path.cwd(), check=False)
-        except UsageError:
-            # The probe itself could not be launched (e.g. an empty PATH leaves no
-            # `sh`): treat that as the tool being absent, the missing-toolchain
-            # outcome, never an internal error (REQ-LIFE-03/04).
-            return False
-        if proc.returncode != 0:
+        if shutil.which(tool) is None:
             return False
     return True
 
@@ -763,9 +793,11 @@ def verify(target: Path, answers: Answers) -> VerifyResult:
     test: list[CommandOutcome] = []
     for member in answers["members"]:
         lint_cmd, test_cmd = _resolve_commands(member)
-        cwd = target / member["path"]
+        cwd = contained_path(target, member["path"])  # never run outside the target
         for bucket, cmd in ((lint, lint_cmd), (test, test_cmd)):
-            proc = run(["sh", "-c", cmd], cwd=cwd, check=False)
+            # STACK_COMMANDS templates with an allow-listed {pm} (see _validate_answers)
+            # — split into an argv token list and run without a shell (no `sh -c`).
+            proc = run(shlex.split(cmd), cwd=cwd, check=False)
             bucket.append(
                 {"command": cmd, "ok": proc.returncode == 0, "member": member["path"]}
             )
@@ -867,9 +899,55 @@ def _parse_answers(raw: str) -> Answers:
         raise UsageError(f"malformed --answers JSON: {exc}")
     if not isinstance(parsed, dict):
         raise UsageError("--answers must be a JSON object")
+    _validate_answers(parsed)
     return parsed
 
 
+def _validate_answers(answers: Answers) -> None:
+    """Reject non-allowlisted stack/packageManager and unsafe member paths (exit 2).
+
+    Runs before any ``{pm}`` substitution, ``STACK_COMMANDS[...]`` subscript, or
+    artifact write so a malformed ``--answers`` payload fails with a stderr
+    diagnostic, never a KeyError traceback or a path escape. Mirrors the
+    input-validation rigor epic-manifest.py applies via ``assert_safe_name`` /
+    ``contained_path``:
+
+    - ``stack`` must be a known :data:`Stack` (a ``STACK_COMMANDS`` key).
+    - ``packageManager`` must be in ``PACKAGE_MANAGERS[stack]`` for a stack that
+      has a choice; stacks with no choice (go/rust/generic) must leave it ``None``.
+    - ``path`` must be a repo-relative string — no absolute path, no ``..``
+      component, no backslash — so it can never resolve outside the target.
+    """
+    members = answers.get("members")
+    if not isinstance(members, list) or not members:
+        raise UsageError("--answers must include a non-empty members[] array")
+    for member in members:
+        stack = member.get("stack")
+        if stack not in STACK_COMMANDS:
+            raise UsageError(
+                f"unknown stack {stack!r}: expected one of {sorted(STACK_COMMANDS)}"
+            )
+        pm = member.get("packageManager")
+        choices = PACKAGE_MANAGERS.get(stack)
+        if choices is not None:
+            if pm not in choices:
+                raise UsageError(
+                    f"invalid packageManager {pm!r} for stack {stack!r}: "
+                    f"expected one of {choices}"
+                )
+        elif pm is not None:
+            raise UsageError(
+                f"stack {stack!r} takes no packageManager, got {pm!r}"
+            )
+        path = member.get("path")
+        if not isinstance(path, str) or not path:
+            raise UsageError(f"member path must be a non-empty string, got {path!r}")
+        if os.path.isabs(path) or "\\" in path or ".." in Path(path).parts:
+            raise UsageError(
+                f"unsafe member path {path!r}: must be repo-relative with no '..'"
+            )
+
+
 def _dispatch(args: argparse.Namespace, target: Path) -> int:
     """Route a parsed command to its handler, translating outcomes into exit codes.
 

package/adapters/gemini/scripts/forge-init.sh

@@ -24,7 +24,10 @@ cat > "$CONFIG_FILE" << 'EOF'
   "loopIterationMultiplier": 1.5,
   "autoInvokeNextStage": true,
   "contextWindowTokens": null,
-  "contextWarnThreshold": 0.7
+  "contextWarnThreshold": 0.7,
+  "autoVerify": false,
+  "autoVerifyStages": {},
+  "autoFix": false
 }
 EOF
 
@@ -43,6 +46,9 @@ echo "  loopIterationMultiplier: 1.5 (multiplier for loop iterations)"
 echo "  autoInvokeNextStage: true (navigator auto-starts the next stage after you confirm)"
 echo "  contextWindowTokens: null (infer; set to 1000000 on a 1M-context model)"
 echo "  contextWarnThreshold: 0.7 (suggest a clean session past this fraction of the window)"
+echo "  autoVerify:          false (set true to run forge-verify automatically after each stage)"
+echo "  autoVerifyStages:    {} (per-stage overrides for autoVerify)"
+echo "  autoFix:             false (set true to chain forge-fix after an auto-verify finds issues)"
 echo ""
 echo "The loop runner defaults to rauf. To target a different ralph-style runner,"
 echo "add a \"loopRunner\" block (see references/forge-config-schema.json)."

package/adapters/gemini/scripts/forge-root.sh

@@ -31,14 +31,33 @@ if is_root "$root"; then
   exit 0
 fi
 
+# ── Step 2a: Claude marketplace-cache installs — ~/.claude/plugins/cache/<mp>/<plugin>/<ver>/.
+# This is where Claude Code actually installs marketplace plugins (three segments below
+# plugins/, so the single-star plugins/*/feature-forge glob below can never match it). Version
+# dirs can coexist after upgrades, so probe newest plugin.json first — a stale version must
+# never shadow the current install. Deliberately ordered BEFORE the plugins/* glob: that glob
+# can match the marketplace *clone* (~/.claude/plugins/marketplaces/<mp>/ when the marketplace
+# repo root is itself a plugin root), which may sit at a different commit than the installed
+# skills — the versioned cache install must always win to prevent that version skew.
+while IFS= read -r manifest; do
+  candidate="${manifest%/.claude-plugin/plugin.json}"
+  if is_root "$candidate"; then
+    printf '%s\n' "$candidate"
+    exit 0
+  fi
+done < <(ls -t "$HOME"/.claude/plugins/cache/*/feature-forge/*/.claude-plugin/plugin.json 2>/dev/null || true)
+
 # ── Step 2: candidate-root probe (authoritative multi-agent root list; extend here first). ─
 # Globs that match nothing expand to themselves; the is_root test rejects such literals. Covers
 # every supported agent's install destination under BOTH global ($HOME) and project ($PWD) scope,
 # matching the installer's per-agent layout: claude .claude/skills, codex .agents/skills, copilot
-# .github/feature-forge, cursor .cursor/rules, gemini .gemini/extensions.
+# .github/feature-forge, cursor .cursor/rules, gemini .gemini/extensions. The cache glob repeats
+# step 2a's path for a cache install that carries only the neutral bundle sentinel (no
+# plugin.json for ls -t to key on).
 for candidate in \
   "$HOME/.claude/skills/feature-forge" \
   "$PWD/.claude/skills/feature-forge" \
+  "$HOME"/.claude/plugins/cache/*/feature-forge/* \
   "$HOME"/.claude/plugins/*/feature-forge \
   "$HOME/.agents/skills/feature-forge" \
   "$PWD/.agents/skills/feature-forge" \