baldart 4.61.0 → 4.62.0

package/CHANGELOG.md

@@ -5,6 +5,47 @@ All notable changes to BALDART will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [4.62.0] - 2026-06-22
+
+**Token-economy + reliability hardening for `/new` and `/prd`, distilled from a real `/new FEAT-0041 -full -auto` run that billed ~364M tokens (~95% cache_read).** A deep multi-agent post-mortem (with adversarial verification that refuted several plausible-but-wrong hypotheses — e.g. the workflow-`<result>`-payload-bloat theory was false by ~430×) traced the cost to a small set of root causes: in-context work that should have been off-context (a hallucinated `/mw` delegation and a killed-final-review recovery, ~36M combined), friction multipliers (a NUL-byte/GNU-grep trap ~8–12M, ~28 lone tracker-Edit turns ~11M), and `/prd` design gaps that pushed work downstream into `/new`'s expensive orchestrator context. Plus a proactive Codex-broker teardown so terminated reviews stop leaking MCP processes.
+
+**MINOR** — adds one CLI verb (`baldart teardown-codex-broker`) + new util export + skill/agent guidance edits; no layout/install change, **no new `baldart.config.yml` key** (so the schema-change propagation rule does NOT apply). All `/new` paths without the `Workflow` tool, and all Codex-less installs, are unaffected.
+
+### Added
+
+- **`baldart teardown-codex-broker --cwd <path>`** (`src/commands/teardown-codex-broker.js` + `bin/baldart.js`) + **`tearDownCodexBroker({cwd, pluginScriptsDir})`** in `src/utils/codex-orphans.js` — a **proactive, cwd-scoped** teardown of the Codex `app-server` broker bound to a specific worktree, freeing its Playwright/Figma/… MCP children. The Codex plugin already ships a clean per-cwd broker shutdown but only fires it on Claude Code `SessionEnd`; a `/new` run that ends (or is killed mid-final-review) hours earlier leaves the broker + MCP subtree burning CPU. Tier 1 reuses the plugin's own `loadBrokerSession`→`sendBrokerShutdown`→`teardownBrokerSession` (graceful); Tier 2 (plugin lib absent / no `broker.json` but the process lingers) matches the broker by its `--cwd` token and SIGTERM-pgroup→SIGKILLs the subtree (reusing `listProcesses`/`collectTree`). Scoped strictly by cwd — never touches a broker for another worktree or the user's interactive session. Validated end-to-end against a real 50-minute-orphaned broker (killed 11 processes). Doctor's `reapOrphans` stays the reactive safety net.
+
+### Fixed (`/new` — bugs surfaced by the post-mortem)
+
+- **`/mw` Phase-6 merge now has a concrete `Skill()` invocation** (`merge-cleanup.md` + `worktree-manager/SKILL.md`). The prose-only "delegate to `/mw`" let the orchestrator hallucinate `Skill({skill:"new:new", args:"__phase6_merge__ …"})` → `Unknown skill` → it ran the **entire** Phase-6 merge inline in its ~590k-token context (~20M cache_read). Now: a literal `Skill({skill:"worktree-manager", args:"/mw worktreePath=… checksAlreadyPassed=true"})` + a HARD "never fall back to manual git" rule (interactive → ask; AUTONOMOUS → follow-up card).
+- **Killed `new-final-review` workflow now recovers off-context** (`final-review.md` F.1.5 + `new-final-review.js`). When the workflow was KILLED mid-Fix-phase (no structured return, no notification), the orchestrator hand-parsed the journal (explicitly forbidden) and re-verified all findings + committed fixes in-context (~16M). New "workflow-killed recovery" sub-protocol: detect non-completion, delegate ONE off-context subagent to validate+commit the worktree's uncommitted Fix diffs, never re-verify in-context, never block on a human under AUTONOMOUS. The workflow now emits a **terminal completion sentinel** (its absence ⇒ killed) and the Fix output is documented as uncommitted-until-F.5.
+- **`rg` over GNU `grep` for source searches** (`code-search-protocol.md` + `coder.md`). A single NUL byte in a large generated file silently flips GNU `grep` to binary mode (empty output + exit 1, indistinguishable from "no match"); three coders independently re-discovered one such byte in a `data.ts` (~8–12M). Now a HARD default: prefer `rg`; treat empty+exit-1 on an expected-match file as a binary-mode trap, never as "symbol absent".
+- **Batch tracker relocated off volatile `/tmp`** (`SKILL.md` + `metrics.md`). The recovery-SSOT tracker lived in `/tmp` and was wiped during a pause, forcing an in-context rebuild. Now `$(git rev-parse --git-common-dir)/baldart/run/batch-tracker-<FIRST-CARD-ID>.md` — repo-local, session-stable, survives the worktree removal, never in `git status` (no partition/`.gitignore` change). Plus a degrade branch that reconstructs run state from commits + backlog statuses if the tracker is ever missing.
+
+### Changed (`/new` — optimizations)
+
+- **A turn whose only tool call is a tracker `Edit` is now an explicit VIOLATION** (`SKILL.md` § Context economy) — a measured run paid ~28 such standalone turns (~11M). Ride the tracker write along with the next real tool call.
+- **Out-of-lane `preValidated` guard + build-claim reconciliation** in `new-final-review.js`. A narrow specialist (doc-reviewer / api-perf-cost-auditor) FP-checking a finding OUTSIDE its lane no longer short-circuits to VERIFIED — it routes to the correct domain specialist. And a VERIFIED BLOCKER/HIGH asserting a build/compile failure that the SAME review's build/tsc gate contradicts (PASS) is auto-demoted to NEEDS_MANUAL_CONFIRMATION instead of gating the merge on a phantom (the false-positive perf BLOCKER the post-mortem caught).
+- **Codex broker teardown wired into the run-end finalizers** — `merge-cleanup.md` Phase 6c (cwd-scoped teardown FIRST, then the cumulative `reap-orphans`), `SKILL.md` terminal-hygiene (HALT-before-6c), `new2/SKILL.md` Step 5, and a `codex-gate.md` note that the broker is shared/warm and must NOT be torn down per-card.
+
+### Changed (`/prd` — design gates that prevent downstream `/new` waste)
+
+- **Consumer enumeration for shared multi-mode functions** (`discovery-phase.md` ISA dimension 9) — when the feature changes a shared list/query/selector/filter fn used in >1 mode/call-site, EVERY consumer is now enumerated as its own ISA touchpoint. This closes the gap that let an orders-wizard catalog picker keep showing hidden products (the exclusion was modeled only at the picker-mode call; the catalog-mode consumer page was unowned → surfaced as 2 HIGH only at the final review, ~28M).
+- **Per-AC seam-file closure + `requirements`-at-write-time** (`card-schema.md` authoring invariants, `audit-phase.md` Codex audit, `validation-phase.md` BLOCKING gates d/e, `prd-card-writer.md`). Every AC's render/seam file (and every file named in requirements/AC/integration_points) must be in `files_likely_touched` (⊆ check); every ISA touchpoint must be owned by some card; and a derivable `requirements` block must be emitted at authoring time — so `/new` no longer pays MAY-EDIT stalls, unplanned fix-coders, or an in-context requirements back-fill.
+- **i18n DoD parity gated on `features.has_i18n`** (`prd-card-writer.md`) — a card introducing user-facing strings now carries a DoD line requiring translation to ALL maintained locales (not just "registered"), shifting the reactive `i18n-translator` spawn author-time-left.
+
+## [4.61.1] - 2026-06-22
+
+**`/new` pre-flight no longer wrongly HALTs/skips well-specified `/prd` cards that merely omit the top-level `requirements` block.** Real-world friction: implementing a `/prd`-designed epic with `/new` repeatedly hit *"N cards lack the top-level `requirements` field (profile CHILD)"* even though those cards had complete `scope`, `scope_boundaries`, `acceptance_criteria`, and `business_rationale` — the requirements were fully derivable, and the user had to hand-back-fill them every time or risk the gate destroying the epic. Root cause: `card-schema.md` listed `requirements` as an **unconditional** HALT (non-derivable), so a derivable-but-omitted block forced an ask/skip. The `prd-card-writer` still mandates emitting `requirements`; this only adds the consumer-side safety net that already exists for `review_profile`/`owner_agent`.
+
+**PATCH** — repairs an over-strict ingestion gate; no new capability, no new `baldart.config.yml` key, no install change. The validator and the matrix `R` classification are unchanged (the writer must still emit `requirements`; the back-fill makes it present before validation runs).
+
+### Changed
+
+- **`framework/agents/card-schema.md`** (SSOT consumer contract) — `requirements` reclassified from unconditional HALT to a **conditional** HALT member: it HALTs **only when its deriving material is itself absent** (`acceptance_criteria` empty OR `scope` empty — a genuinely thin card). When both are present and non-empty, a missing `requirements` block is recovered via a new BACK-FILL **sub-kind (b): faithful derivation** — a bounded restatement/decomposition of the existing AC + scope (faithful, never generative; never invents new scope).
+- **`framework/.claude/skills/new/references/setup.md`** — pre-flight 1b-i drops `requirements` from the unconditional HALT list (now conditional on AC/scope being empty); 1b-ii gains the faithful-derivation back-fill (derive from AC + scope, persist to the card on disk in the main repo, log `[BACKFILL] … requirements=<N items, AC+scope-derived>`).
+- **`framework/.claude/workflows/new2.js`** (parallel location, gate G4) — a card missing `requirements` with AC + scope present is now faithfully derived + written back (F-040 main-repo discipline) instead of excluded from the batch; only excluded when AC or scope is also absent.
+
 ## [4.61.0] - 2026-06-22
 
 **UI design-quality critic — a separate, rigorous reviewer for *"is this good design?"*, distinct from the existing *"does it match the mockup?"* check.** Born from the observation that implemented UIs were often poor despite the full `ui-expert` + `/e2e-review` machinery. Root cause, confirmed by exploration: every existing UI verification checks **fidelity / conformance**, not **intrinsic quality** — `visual-fidelity-verifier` asks only whether the render matches the mockup (and never reads source, never judges aesthetics), `code-reviewer`'s design-system steward is mechanical (token/primitive compliance), and the only agent judging quality (`ui-expert`) is *also the implementer*, so it grades its own work. The generator-grades-itself antipattern is exactly what the critic-in-the-loop research (vision-guided iterative refinement; rubric-based LLM-as-judge with per-dimension calibration) identifies as the failure mode. This release adds the missing layer: a separate critic that judges the rendered output against a scientific rubric, in the existing bounded self-heal loop.

package/VERSION

@@ -1 +1 @@
-4.61.0
+4.62.0

package/bin/baldart.js

@@ -164,6 +164,16 @@ program
     await reapCommand({ json: !!options.json, dryRun: !!options.dryRun });
   });
 
+program
+  .command('teardown-codex-broker')
+  .description('Proactively end the Codex app-server broker bound to a specific worktree cwd (and its MCP children). Used by the /new run-end + abort finalizers. Scoped by --cwd — never touches a broker serving another worktree or your interactive session.')
+  .option('--cwd <path>', 'The workspace cwd whose broker to end (the /new batch worktree path). Defaults to the current directory.')
+  .option('--json', 'Machine-readable output: emit a single JSON result object on stdout')
+  .action(async (options) => {
+    const teardownCommand = require('../src/commands/teardown-codex-broker');
+    await teardownCommand({ cwd: options.cwd, json: !!options.json });
+  });
+
 const overlayGroup = program
   .command('overlay')
   .description('Author and check .baldart/overlays/ — scaffolds, validates, and detects drift on skill/agent/command overlays');

package/framework/.claude/agents/coder.md

@@ -70,7 +70,7 @@ Any requirement of the form "verify X; if found → do A; if missing → do B/le
 is a **binary-outcome item**. BOTH branches have a mandatory artifact — silence is never valid.
 
 Rules:
-1. **Verify actively**: read the relevant code (don't assume). Use Grep/Read.
+1. **Verify actively**: read the relevant code (don't assume). Use Grep/Read. **Prefer `rg` over GNU `grep` on source** — GNU `grep` silently flips to binary mode on the first NUL byte (common in large generated/bundled `data.ts`-style files) and then prints nothing + exits 1, *indistinguishable from "no match"*. An empty result + exit 1 on a file you expect to match is a binary-mode trap, NOT proof the symbol is gone: re-run with `rg` (or `grep -a`) before acting. See `framework/agents/code-search-protocol.md` § Budget.
 2. **Branch A (condition satisfied)**: implement A as specified.
 3. **Branch B (condition not met)**: produce B — either:
    - A code change (add the missing piece), OR

package/framework/.claude/agents/prd-card-writer.md

@@ -433,7 +433,13 @@ Every card MUST include ALL fields from the template:
 - `scope` (summary, users, value)
 - `requirements` (concrete, specific)
 - `acceptance_criteria` (testable `[ ] [AC-N]` items)
-- `definition_of_done` (checklist including doc invariants)
+- `definition_of_done` (checklist including doc invariants). **When `features.has_i18n: true`** AND the
+  card introduces user-facing strings, the DoD MUST carry an i18n-completeness line — NOT merely "strings
+  via `t()` + registered", but: *"new keys authored in the source locale + the i18n context registry, AND
+  translated to ALL maintained target languages (via `i18n-translator`); the i18n parity / no-placeholder
+  gate passes."* Otherwise the coder ships only the source locale and `/new`'s per-card review reactively
+  spawns `i18n-translator` — a ~2M-token round-trip the DoD should have made author-time-left. Cross-ref
+  `framework/agents/i18n-protocol.md`.
 - `depends_on` / `blocks` (card IDs)
 - `estimated_complexity`
 - `review_profile` (`skip`|`light`|`balanced`|`deep` — computed via Rule C, or honored if pre-set)

package/framework/.claude/skills/new/SKILL.md

@@ -100,13 +100,20 @@ precedence caveats: `framework/agents/effort-protocol.md`.
 
 ## Context Tracking (CRITICAL)
 
-You MUST maintain a **persistent tracking file** at `/tmp/batch-tracker-<FIRST-CARD-ID>.md` throughout the entire batch run (e.g., `/tmp/batch-tracker-FEAT-0396.md`). Use the **first card ID** from the batch as the suffix. This ensures multiple `/new` sessions running in parallel terminals (e.g., one per worktree) do NOT conflict.
+You MUST maintain a **persistent tracking file** (`$TRACKER`) throughout the entire batch run.
+`$TRACKER` is deterministic — recompute it any time (even after a compaction) from the first card ID:
 
-This file is your single source of truth — if your context gets compacted or you lose track of what happened, **re-read this file first**.
+```
+$TRACKER = "$(git rev-parse --git-common-dir)/baldart/run/batch-tracker-<FIRST-CARD-ID>.md"
+```
+
+i.e. inside the repo's **shared git dir** (e.g. `<main>/.git/baldart/run/batch-tracker-FEAT-0396.md`). Use the **first card ID** as the suffix so parallel-terminal `/new` sessions do NOT conflict; `mkdir -p "$(dirname "$TRACKER")"` once before the first write. **Why the shared git dir and NOT `/tmp`:** `/tmp` is wiped on idle/reboot — a real run lost its tracker mid-pause when `/tmp` was cleaned and had to rebuild it in-context. `git rev-parse --git-common-dir` resolves to `<main>/.git` **even when `/new` runs from inside a worktree** (unlike `--show-toplevel`), so the tracker is session-stable, survives the Phase 6 worktree removal, and is **never tracked by git** (zero `git status` pollution — no clean-tree-partition or `.gitignore` change needed).
+
+This file is your single source of truth — if your context gets compacted or you lose track of what happened, **recompute `$TRACKER` and re-read it first** (§ "Context recovery protocol").
 
 ### Tracking file format
 
-At batch start — **before Phase 0 runs** (Phase 0 step 1 writes to the `## Worktree` section and Phase 0 step 5 writes the phase log, so the file MUST already exist) — create `/tmp/batch-tracker-<FIRST-CARD-ID>.md` with:
+At batch start — **before Phase 0 runs** (Phase 0 step 1 writes to the `## Worktree` section and Phase 0 step 5 writes the phase log, so the file MUST already exist) — create `$TRACKER` with:
 
 ```markdown
 # Batch Run: [CARD-IDS]
@@ -173,14 +180,14 @@ Trunk branch: [resolved git.trunk_branch — Phase 0 step 0 populates]
   - QA findings file (e.g. `/qa/FEAT-XXXX.md` or "skipped")
   - **card_status: DONE (verified)** — confirms the backlog YAML was updated and re-read to verify
 - **When blocked**: log the blocker in `## Issues & Flags`.
-- **On context recovery**: if you ever feel lost or after context compaction, IMMEDIATELY read your tracker file (`/tmp/batch-tracker-<FIRST-CARD-ID>.md`) to restore your state.
+- **On context recovery**: if you ever feel lost or after context compaction, IMMEDIATELY read your tracker file (`$TRACKER`) to restore your state.
 - **The tracker is internal and is the ONLY state surface** (since v4.47.0 — see § "State surface — the tracker only"): the user does NOT see it, and there is no separate user-visible mirror to keep in sync. Do NOT spend a dedicated turn restating a transition to the user (no TaskUpdate, no Progress-Bar block) — surface progress only as a short prose line folded into a turn you are already taking for real work (§ "Context economy" turn-count rule).
 
 ---
 
 ## State surface — the tracker only
 
-`/new` keeps a **single** state surface: the internal `/tmp/batch-tracker-<FIRST-CARD-ID>.md`
+`/new` keeps a **single** state surface: the internal `$TRACKER`
 recovery SSOT (§ "Context Tracking"). There is **no separate user-visible mirror**. The Progress
 Bar markdown block and the per-transition native Task spine (TaskCreate/TaskUpdate) were **removed
 for context economy (v4.47.0)**: every standalone TaskUpdate / Progress-Bar emission was its own
@@ -267,6 +274,14 @@ baselines. Keep that bulk on disk and pass **paths**, not bodies.
 > the transition happens (a compaction between would lose it — § "Context recovery protocol" re-reads
 > the tracker, not your memory).
 >
+> **A turn whose ONLY tool call is a tracker `Edit` is a VIOLATION.** There is ALWAYS a next real action
+> to co-emit it with — the next phase's first `Read`/`Bash`, the card hand-off's next-card `Read`, the
+> commit. A measured run paid ~28 standalone tracker-`Edit` turns (~490k replay each ≈ **~11M wasted**)
+> despite the batching rule above. So make it reflexive: if you are about to emit a lone tracker `Edit`,
+> STOP and fold it into the next real tool call's message instead. The only turns that may carry the
+> tracker write alone are the unavoidable phase-boundary writes where there is genuinely no next action
+> yet (rare) — and even there, prefer to ride it along with the first action of the next phase.
+>
 
 > **HARD RULE — orchestrator is near-silent / maximally cryptic (since v4.49.0).** `/new` is NOT
 > conversational: the user does not read its running narration and keeps the orchestrator attached ONLY so it
@@ -380,7 +395,8 @@ per-card nei sub-step D.x (mai aggregate). Caricalo quando Pre-flight seleziona
 ## Context recovery protocol
 
 If at ANY point you are unsure where you are in the batch:
-1. Read your tracker file (`/tmp/batch-tracker-<FIRST-CARD-ID>.md`)
+1. **Recompute `$TRACKER`** = `$(git rev-parse --git-common-dir)/baldart/run/batch-tracker-<FIRST-CARD-ID>.md` (the first card ID is in the batch args — the path is deterministic) and Read it.
+   - **If the tracker is MISSING** (a pre-this-release run still on volatile `/tmp`, or the file was deleted during a pause): do **NOT** rebuild it from memory in-context. **Reconstruct from the durable witnesses** — git + the backlog, which are the real SSOT the tracker only caches: the worktree branch's commits (`git -C <worktree> log --oneline "$TRUNK"..HEAD` — each `[CARD-ID]` commit ⇒ that card progressed/completed) and each batch card's backlog YAML `status:` (`DONE` ⇒ completed). Re-create `$TRACKER` from those facts (which cards are DONE, which is current), then continue from there.
 2. Check `## Current Card` — if populated, resume that card at the listed phase.
 2b. **Pre-flight not finished?** If `## Current Card` is empty / `(none — starting pre-flight)` AND `## Worktree` still shows skeleton placeholders (no real `Path:`/`Port:`), then a compaction hit mid-pre-flight (most likely during the step-5 background barrier). Re-enter **Pre-flight at step 4** (Read the `setup.md` module first per § "Routing"): its step-4a2 git pre-check re-detects an already-created worktree via `git worktree list` and resumes-or-resets it safely — so re-entry never double-creates or fails-loud.
 3. **Re-Read the phase module** named in `## Current Card`'s `phase_module_loaded:` field (a compaction may have evacuated it from context — see § "Routing" HARD RULE). If the field is absent, map the listed phase to its module via the § "Routing" table and Read that.
@@ -388,6 +404,17 @@ If at ANY point you are unsure where you are in the batch:
 5. Check `## Completed Cards` — know what's already done (don't redo).
 6. Continue the pipeline from where you left off.
 
+**Terminal hygiene — tear down the Codex broker on ANY exit before Phase 6c.** The normal close runs the
+cwd-scoped Codex broker teardown in Phase 6c step 5b. But if the batch ends **before** reaching Phase 6c —
+an unrecoverable HALT, or resuming a run whose final-review workflow was KILLED (see `final-review.md`
+§ "workflow-killed recovery") and then deciding to stop — the batch's `codex app-server` broker (and its
+Playwright/… MCP children) is still alive. Before exiting, run it once (idempotent, cwd-scoped, NON-blocking):
+```bash
+npx baldart teardown-codex-broker --cwd "$MAIN/<relative worktree Path from tracker ## Worktree>" 2>/dev/null || true
+```
+This NEVER touches a broker for another worktree or the user's interactive session (different cwd). If the
+orchestrator is hard-killed and never resumed, the `baldart doctor` reaper is the final safety net.
+
 ---
 
 

package/framework/.claude/skills/new/references/codex-gate.md

@@ -37,6 +37,13 @@ Log the detector output in the tracker for traceability. **This step never short
 
 #### Step C — Invoke per-card `/codexreview` (ALWAYS)
 
+> **Do NOT tear down the Codex broker per-card.** The `codex app-server` broker is **one per worktree
+> cwd, shared and warm-reused across every per-card + final-review Codex call** in the batch. Ending it
+> after a card would force a cold re-spawn (~2s) for the next card's review. The broker (and its MCP
+> children) is torn down ONCE at run-end — `merge-cleanup.md` Phase 6c step 5b (cwd-scoped
+> `baldart teardown-codex-broker`), with the `baldart doctor` reaper as the safety net. Per-card hygiene
+> is none.
+
 For EVERY card (no conditional skip — the gate ALWAYS runs; only its DEPTH varies):
 
 0. **Select the Codex review profile (deterministic — Review Profile Selector)** by mapping the

package/framework/.claude/skills/new/references/final-review.md

@@ -136,6 +136,36 @@ that is a **gate violation**: log it as
   new-final-review workflow (engine: <codexEngine>, fixes: <fixesApplied.length>+<docFixesApplied.length>, residual: <residual.length>)`
   in the tracker.
 
+- **IF the delegated workflow did NOT return a structured value** — the `Workflow`
+  tool result is `null` / an error, OR no completion `<task-notification>` arrived
+  after a bounded poll window — the run was **killed / aborted mid-flight** (e.g. a
+  user interrupt, an OOM, a rate-limit kill — the same abort class as the team-mode
+  empty-result post-mortem). **Do NOT improvise an in-context recovery.** A measured
+  run that hand-parsed the journal and **re-verified all findings + committed the Fix
+  diffs in the orchestrator's own ~540k-token context cost ~16M cache_read** — the
+  single largest avoidable block in that batch. Recover OFF-context instead:
+  1. **Detect, don't reconstruct.** You MAY read ONLY the journal's *terminal status*
+     to confirm non-completion (`status` ≠ `completed`, or a `fix … started` entry with
+     no matching result). You may **NOT** rebuild the findings/verdicts from the journal —
+     that is the forbidden hand-parse, and the Fix phase may have run **partially**, so
+     the journal is an unreliable, half-written picture.
+  2. **Locate the partial work on disk, hands-off.** `git -C <worktree> status --porcelain`
+     + `git -C <worktree> diff --stat` (name/stat only — never read the diff bodies into
+     your context). Uncommitted changes = the killed Fix phase's unvalidated output.
+  3. **Delegate ONE off-context subagent** (a `qa-sentinel` for the build/lint/test/tsc
+     gate, plus — if the diff is non-trivial — a `code-reviewer`/`coder` to judge the
+     partial fixes against the F.3 findings file) to: verify the uncommitted diffs
+     compile and pass the gates, **commit them if good or revert them if broken**, and
+     return a **COMPACT residual** (the still-unaddressed findings, slim). The
+     orchestrator reads only that residual — it does NOT re-verify findings itself.
+  4. **Carry `residual` into Step F.5** exactly as the happy path does. **AUTONOMOUS**:
+     NEVER stop and wait for a human "riprendi" — if the recovery subagent cannot fully
+     resolve, materialize a follow-up card with the residual and proceed (log
+     `AUTO: final-review-killed → recovered off-context (residual: N) [+ follow-up <id>]`).
+     **Interactive**: surface the residual terse and continue; only `AskUserQuestion` if
+     the recovery subagent itself reports an unsafe state (e.g. a broken build it could
+     not revert cleanly).
+
 - **ELSE** (no `Workflow` tool — older Claude Code, `disableWorkflows`, or a
   Codex install where workflows have no equivalent) → run the **inline F.2–F.4
   below exactly as written**. This prose is the SSOT and the always-available

package/framework/.claude/skills/new/references/merge-cleanup.md

@@ -14,10 +14,16 @@ After the final review passes AND all cards are committed in the worktree, deleg
    git status --porcelain
    ```
    If ANY uncommitted files exist (staged, unstaged, or untracked), commit them NOW with `[safety] Auto-commit remaining files before merge`. Do NOT proceed to `/mw` with a dirty worktree — files WILL be lost during rebase.
-2. Invoke `/mw` with:
-   - The worktree path and branch from the tracker
-   - `checksAlreadyPassed: true` (final review + QA already validated the build)
-   - All card IDs for the commit message
+2. **Invoke the worktree-manager skill via the Skill tool — this is the ONLY merge mechanism.** The skill that owns `/mw` is named **`worktree-manager`** (NOT `new`, NOT `new:new`, NOT `mw`). Call it exactly, with a concrete payload — never describe the delegation in prose and hope:
+
+   ```
+   Skill({ skill: "worktree-manager",
+           args: "/mw worktreePath=<absolute worktree path from tracker ## Worktree> checksAlreadyPassed=true" })
+   ```
+
+   - `worktreePath` — the absolute path from the tracker `## Worktree` section (REQUIRED). The skill resolves the branch + the card IDs for the commit message from the worktree/registry itself — do **NOT** invent `branch` / `card_ids` / `worktree_path` payload keys (the real input contract is `{ worktreePath, checksAlreadyPassed }`, worktree-manager/SKILL.md § "/mw programmatic").
+   - `checksAlreadyPassed=true` — the final review + QA already validated the build, so the skill skips re-running quality gates (the safety commit ALWAYS still runs regardless).
+   - **HARD — never hand-merge in this orchestrator.** If the Skill call errors (unknown skill, bad args) or the skill STOPs on a conflict/build failure, do **NOT** fall back to inline `git merge` / `git rebase` / `git push` here. Hand-merging runs the entire Phase 6 in the cached prefix (a measured **~20M-token** regression) and risks a non-isolation-safe `git checkout` on the shared main repo. Instead: surface the skill's error/STOP report verbatim, then — **interactive** → `AskUserQuestion` how to proceed; **AUTONOMOUS** → leave the worktree intact and materialize a follow-up card describing the merge blocker (log `AUTO: merge-delegation-failed → follow-up <id>`). `new:new` and `__phase6_merge__` are hallucinations — there is no such skill or arg; the only valid call is the `worktree-manager` one above.
 3. The skill handles: safety commit of any remaining uncommitted files (step 3), rebasing onto the latest trunk (`$TRUNK` = `git.trunk_branch`) (step 4b — auto-resolves doc conflicts, stops on code conflicts), merging into the trunk via the configured `git.merge_strategy` (step 4c — `pr` uses `gh pr merge`, `local-push` does a direct FF push to `origin/$TRUNK`; NEITHER runs `git checkout` of the trunk on the main repo, respecting the absolute terminal-isolation rule), post-merge verification, worktree removal, registry cleanup, and remote branch deletion.
 4. **If code merge conflicts** → the skill STOPs and reports. Doc-only conflicts (ssot-registry.md, project-status.md, etc.) are auto-resolved by keeping both sides.
 5. **If post-merge build fails** → the skill STOPs and keeps the worktree intact for investigation.
@@ -186,19 +192,34 @@ The most common failure mode is leaving cards IN_PROGRESS after merge. This crea
      - Options: `[Lascia lo stash + apri istruzioni per merge manuale]` / `[Mostrami il conflitto inline]` / `[Halt]`.
      - **When AUTONOMOUS, apply § AUTONOMOUS RESOLUTION RULE (SKILL.md): category=blocker; recommended=none safe (a stash-pop conflict needs human merge) → leave the stash intact (NEVER drop it) and materialize a follow-up card with the restore instructions; log `AUTO: stash-restore-conflict → follow-up <id> (stash retained)`.**
 
-5b. **Process hygiene — reap orphaned Codex MCP servers (NON-BLOCKING)**. This
-   batch's per-card / final-review Codex calls drive `codex app-server`, whose
-   broker spawns the MCP servers declared in `~/.codex/config.toml` (Playwright,
-   …) as children; when a broker dies the OS reparents those MCP servers to init
-   (ppid 1) and they keep burning CPU. Sweep them now so the batch ends clean:
-   ```bash
-   npx baldart reap-orphans 2>/dev/null || true
-   ```
-   This reaps ONLY orphaned MCP servers (ppid 1 ⇒ their broker is already dead ⇒
-   stdio is broken ⇒ dead weight). It deliberately NEVER kills a live
-   `codex app-server` broker (a shared, detached runtime that may still serve the
-   user's interactive session). Never gate the close on this — any error or a
-   "nothing to reap" result is fine; capture its one-line summary for the log.
+5b. **Process hygiene — end THIS batch's Codex broker, then sweep stragglers (NON-BLOCKING, two steps)**.
+   This batch's per-card / final-review Codex calls drive a `codex app-server` broker — **one per
+   worktree cwd, shared across all the batch's Codex calls** — whose broker spawns the MCP servers
+   declared in `~/.codex/config.toml` (Playwright, …) as children. The Codex plugin only ends that
+   broker on Claude Code `SessionEnd` (which can be hours away), so without this the broker and its MCP
+   children keep burning CPU for the rest of the session.
+   1. **Proactive, cwd-scoped teardown FIRST** — end the broker bound to THIS batch's worktree. Resolve
+      the absolute worktree path from the tracker `## Worktree` `Path:` against `$MAIN`, and tear it down
+      **even though `/mw` already removed the worktree directory** (the broker is matched by its `--cwd`
+      token / state-hash, both of which survive the dir removal):
+      ```bash
+      WORKTREE="$MAIN/<relative Path from tracker ## Worktree>"   # absolute worktree path
+      npx baldart teardown-codex-broker --cwd "$WORKTREE" 2>/dev/null || true
+      ```
+      This is **scoped strictly by cwd** — it ends ONLY this worktree's broker (graceful plugin shutdown
+      when the broker.json is present, else a signal to the cwd-matched broker subtree). It NEVER touches
+      a broker serving another worktree or the user's interactive Codex session on the main repo (a
+      different cwd). Do NOT run this per-card — the broker is shared and warm across the batch; ending it
+      mid-batch would force a cold re-spawn for the next card.
+   2. **Then the cumulative orphan sweep** (second line) — mop up any MCP server already orphaned to init
+      (ppid 1) by an earlier broker death this session:
+      ```bash
+      npx baldart reap-orphans 2>/dev/null || true
+      ```
+      `reap-orphans` reaps ONLY orphaned MCP servers (ppid 1 ⇒ their broker is already dead ⇒ stdio is
+      broken ⇒ dead weight); it NEVER kills a live broker.
+   Never gate the close on either — any error or a "nothing to do" result is fine; capture each one-line
+   summary for the log.
 
 6. **Log and exit**:
    ```
@@ -208,7 +229,8 @@ The most common failure mode is leaving cards IN_PROGRESS after merge. This crea
    Divergence (local…origin/$TRUNK): <0\t0 | resolved: pushed/cherry-picked/ff-pulled/rebased>
    Sync-deferred markers: <none | reconciled | user-retained>
    Phase 0 snapshot restore: <n/a | popped clean | conflict-deferred-to-user>
-   Codex MCP hygiene: <reaped N/M | nothing to reap | skipped (error)>
+   Codex broker teardown: <graceful (pid) | signalled N | none | skipped (error)>
+   Codex MCP hygiene (reap): <reaped N/M | nothing to reap | skipped (error)>
    Completed: <timestamp>
    ```
    If any step ended in HALT, set `Status: HALT` and report — Phase 7 must NOT start with an unclean main repo unless the user explicitly chose `[Lascia così]`.

package/framework/.claude/skills/new/references/metrics.md

@@ -12,7 +12,7 @@ has no blocking consequence the honest label is "monitoring signal".
 
 **Steps:**
 
-1. Read the batch tracker file (`/tmp/batch-tracker-<FIRST-CARD-ID>.md`) to extract — every field below has a NAMED upstream producer; read the structured key, do NOT re-parse prose:
+1. Read the batch tracker file (`$TRACKER` = `$(git rev-parse --git-common-dir)/baldart/run/batch-tracker-<FIRST-CARD-ID>.md`, see new/SKILL.md § "Context Tracking") to extract — every field below has a NAMED upstream producer; read the structured key, do NOT re-parse prose:
    - Start timestamp (from `Started:` field)
    - Worktree-creation timestamp (from `## Worktree` — the time Phase 0/Pre-flight created the worktree; the strategy-independent cycle-time anchor, see step 2)
    - Card list and total count
@@ -49,7 +49,7 @@ has no blocking consequence the honest label is "monitoring signal".
 4. Copy the batch tracker to archive:
 
    ```bash
-   cp /tmp/batch-tracker-<FIRST-CARD-ID>.md "$METRICS/archive/"
+   cp "$(git rev-parse --git-common-dir)/baldart/run/batch-tracker-<FIRST-CARD-ID>.md" "$METRICS/archive/"
    ```
 
 5. Note in tracker: `## Metrics Log: WRITTEN (run_id: batch-<FIRST-CARD-ID>)`

package/framework/.claude/skills/new/references/setup.md

@@ -136,15 +136,17 @@
 1b. **Normalize & validate card baseline (pre-flight gate)** — `/new` is **type-blind**: it runs the same pipeline on `FEAT`/`CHORE`/`BUG`/`DOC`/`PERF`. The universal, profile-aware baseline is the SSOT in [`framework/agents/card-schema.md`](../../../agents/card-schema.md). Detect each card's profile (epic / child / standalone) per that module, then run three stages. (Epic-parent cards are excluded from the per-card loop — skip 1b-i/ii for them; the validator at 1b-iii still checks their EPIC column.)
 
    **1b-i — HALT on non-derivable fields (ask the user).** Verify the fields that cannot be safely synthesized are present and non-empty:
-   - `requirements` (>=1), `acceptance_criteria` (>=1), `files_likely_touched` (>=1 file), and `scope`.
+   - `acceptance_criteria` (>=1), `files_likely_touched` (>=1 file), and `scope`.
+   - **`requirements` is conditional, NOT an unconditional HALT** (card-schema.md consumer contract): it HALTs here **only when `acceptance_criteria` OR `scope` is itself empty** (a genuinely thin card). When `acceptance_criteria` AND `scope` are both present and non-empty, a missing `requirements` block is **faithfully derivable** and is back-filled in 1b-ii — do NOT HALT or skip the card for it. (This is the common `/prd` shape: complete AC + scope + scope_boundaries but no top-level `requirements` — skipping it would destroy the epic.)
    - `scope_boundaries` is **NOT** in this set (it is conditional — legitimately omitted for standalone cards with no siblings; see card-schema.md).
    If any card fails: log the specific missing fields in `## Issues & Flags`, ask the user to fill them in before proceeding with that card, and continue pre-flight for any remaining valid cards. **When AUTONOMOUS, apply § AUTONOMOUS RESOLUTION RULE (SKILL.md): category=blocker; recommended=none safe (non-derivable fields cannot be fabricated — SC-4 "missing fields → SKIP+report, never fabricate") → skip that card from the batch and log `AUTO: card-baseline → skipped <CARD-ID> (missing <fields>)`; continue with the valid cards. NEVER invent requirements/AC/scope.**
 
    **1b-ii — Back-fill deterministically-computable fields (compute → persist → log).** For each non-epic card missing a derivable field, compute it and **write it back to the card on disk in `$MAIN/${paths.backlog_dir}`** (the main repo path resolved at Phase 0 step 1 — NOT the worktree copy; same F-040 discipline `new2` uses for follow-ups):
    - `review_profile` absent → compute via `prd-card-writer.md § Rule C` (the SSOT), write it.
    - `owner_agent` absent/empty/`claude` → default `coder`, write it. (Never back-fill an epic's `""`.)
+   - `requirements` absent **AND** `acceptance_criteria` + `scope` both present and non-empty → **faithfully derive** it (card-schema.md consumer contract, BACK-FILL sub-kind b): write a `requirements` list that is a concrete restatement/decomposition of the card's existing AC + scope. **Faithful, not generative** — restate only what the AC and scope already commit to; NEVER invent new scope, constraints, or behaviour. (If `acceptance_criteria` OR `scope` is also empty, this is NOT derivable — it was already HALTed at 1b-i; never reach here.) Log it as `requirements=<N items, derived from AC+scope>`.
    - **Check-and-skip (idempotent):** write ONLY when the field is genuinely absent. A card that already has the field is untouched — so a pre-flight re-entry after compaction never double-writes.
-   - Log each write on its own line in `## Issues & Flags`: `[BACKFILL] <CARD-ID>: <field>=<value> (Rule C / coder-default)`.
+   - Log each write on its own line in `## Issues & Flags`: `[BACKFILL] <CARD-ID>: <field>=<value> (Rule C / coder-default / AC+scope-derived)`.
    - **Commit discipline:** after all cards are processed, if ANY back-fill was written, commit them together in `$MAIN` with the COMMIT_LOCK + doc-freshness precautions of `merge-cleanup.md` Phase 6b (clear stale `COMMIT_LOCK`; `git add` only the back-filled card YAMLs; if the doc-freshness hook blocks, stage `ssot-registry.md`; ≤2 retries): `git commit -m "chore(backlog): normalise card baseline [BACKFILL]"`. Do NOT leave `$MAIN` dirty for the batch duration — an uncommitted backlog edit would surface as noise in the Phase 3/4 git-diff gates.
    - `canonical_docs` / `links.prd` absent on a standalone/non-PRD card → **WARN** only (log, do not block, do not back-fill).
 

package/framework/.claude/skills/new2/SKILL.md

@@ -316,14 +316,21 @@ returns when the batch is done. It returns:
    `/new` on the hard-case follow-ups (vs leaving them tracked). In autonomous mode it is
    `{ eligible:N, offered:false, ran_new:false }`.
    Do NOT re-summarise the cards — the workflow already did.
-6. **Process hygiene — reap orphaned Codex MCP servers (NON-BLOCKING).** The batch's per-card Codex
-   finder calls drive `codex app-server`, whose broker spawns the `~/.codex/config.toml` MCP servers
-   (Playwright, …) as children; when a broker dies they leak to init (ppid 1) and keep burning CPU.
-   Sweep them in the main context (the workflow sandbox cannot run Bash, so this MUST run here, after
-   the workflow returns) so the run ends clean:
+6. **Process hygiene — end this batch's Codex broker, then sweep stragglers (NON-BLOCKING, two lines).**
+   The batch's per-card Codex finder calls drive a `codex app-server` broker (one per worktree cwd,
+   shared across the batch), whose broker spawns the `~/.codex/config.toml` MCP servers (Playwright, …)
+   as children; the Codex plugin only ends that broker on Claude Code `SessionEnd`, so without this it
+   and its MCP children keep burning CPU for the rest of the session. The workflow sandbox cannot run
+   Bash, so BOTH lines MUST run here in the main context, after the workflow returns:
    ```bash
+   # 1. Proactive, cwd-scoped teardown FIRST — end the broker bound to THIS batch's worktree.
+   #    `worktreePath` is the workflow return's worktreePath (preflight); the broker matches by --cwd /
+   #    state-hash, so this works even though the workflow already merged + removed the worktree dir.
+   npx baldart teardown-codex-broker --cwd "<worktreePath from the workflow result>" 2>/dev/null || true
+   # 2. Then the cumulative orphan sweep (second line) — mop up any MCP already orphaned to init.
    npx baldart reap-orphans 2>/dev/null || true
    ```
-   Reaps ONLY orphaned MCP servers (ppid 1 ⇒ broker dead); NEVER kills a live `codex app-server`
-   broker. Non-blocking — any error / "nothing to reap" is fine; fold its one-line summary into the
-   record (`codex_mcp_reaped`).
+   The teardown is **scoped strictly by cwd** — it NEVER touches a broker serving another worktree or the
+   user's interactive Codex session (a different cwd). `reap-orphans` reaps ONLY orphaned MCP servers
+   (ppid 1 ⇒ broker dead) and NEVER kills a live broker. Non-blocking — any error / "nothing to do" is
+   fine; fold both one-line summaries into the record (`codex_broker_teardown`, `codex_mcp_reaped`).

package/framework/.claude/skills/prd/references/audit-phase.md

@@ -183,6 +183,16 @@ Database-specific (apply the subset matching `stack.database`):
 
 Card structure:
 - files_likely_touched missing entries or conflicting across cards
+- **per-AC seam-file closure**: for EACH acceptance_criteria entry, the render/seam file it
+  mechanically touches MUST be in files_likely_touched — an AC that adds a required prop, changes a
+  shared component's public signature, or wires a value into a page but does NOT list that page /
+  render-site / view-model file is a HIGH gap (it forces a MAY-EDIT stall or an unplanned fix-coder
+  in /new, and lets a coder dodge the AC by making the prop optional). Also flag any file named in a
+  card's requirements / acceptance_criteria / integration_points that is absent from
+  files_likely_touched (the ⊆ coverage rule).
+- **shared multi-mode function consumers**: if a card changes a shared list/query/selector/filter fn
+  used in >1 mode/call-site, every consumer call-site must be owned by some card's
+  files_likely_touched — flag any unowned consumer (the "hidden-still-shown-in-the-other-mode" class).
 - areas field incomplete
 - git_strategy set to TBD
 - acceptance_criteria not binary testable

package/framework/.claude/skills/prd/references/discovery-phase.md

@@ -764,6 +764,18 @@ Instead, run the ISA scan and present findings for validation.
      utilities, API route indexes
    - SHARED STATE: stores, contexts, providers that manage related data,
      session/cache keys
+   - SHARED MULTI-MODE FUNCTIONS (consumer enumeration): when this feature changes
+     the behavior of a shared list/query/selector/filter/mapper function that is
+     called from MORE THAN ONE site or in MORE THAN ONE mode (e.g. a `listProducts`
+     used both in a "picker" mode AND a "catalog" mode), enumerate EVERY call-site /
+     consumer of that function as its OWN touchpoint — not just the definition site
+     or the most obvious caller. A filter/flag added at the definition that one
+     consumer relies on is NOT automatically honoured where another consumer calls
+     the same function in a different mode. (This is the gap that let an orders-wizard
+     catalog picker keep showing hidden products: the exclusion was modeled only at
+     the picker-mode call, while the live catalog-mode consumer page was never
+     enumerated → it surfaced as a HIGH only at the final review.) Grep the function
+     name across the codebase and list each distinct caller file:line.
 
    For each touchpoint found, report:
    - Category (from above)

package/framework/.claude/skills/prd/references/validation-phase.md

@@ -33,7 +33,34 @@ Mark task 5 as `in_progress`.
       - **Consistency WARN (soft)**: a `feature`/`enhancement` card with `review_profile: light` may violate Rule C. Log `"[REVIEW-PROFILE-WARN] Card <ID> is feature/enhancement but review_profile: light — verify against Rule C in prd-card-writer.md."` **Do NOT auto-bump the field** — `review_profile` is computed by `prd-card-writer` (Rule C); see `prd-card-writer.md` for the criteria. This is a soft gate: log the warning and surface it for human review.
       - **Epics**: `review_profile` MUST be `skip` (trackers have no code work). Any other value = WARN, normalize to `skip`.
 
-   **Gate**: every card has `owner_agent` AND `review_profile` in their enums. WARN entries from checks (b)/(c) are noted in the audit log; user may accept them after review.
+   d. **`requirements` presence (BLOCKING — PO1)** — for each CHILD/STANDALONE card whose
+      `acceptance_criteria` AND `scope` are both non-empty, `requirements` MUST be present and
+      non-empty. A freshly authored `/prd` card with derivable material but no `requirements` is an
+      authoring miss that shifts a measured ~3–4M-token back-fill into `/new`'s orchestrator context
+      (see `card-schema.md` § "Authoring completeness invariants"). `/prd` fixes it now for free.
+      Message: `"Card <ID> has acceptance_criteria + scope but no requirements — emit requirements
+      (faithful restatement of AC + scope) before commit."` (The `/new` conditional back-fill remains
+      only as the safety net for legacy / externally-authored cards.)
+
+   e. **files_likely_touched coverage (BLOCKING — P1/P2 closure)** — two sub-checks:
+      - **⊆ coverage (P2):** every file path named in a card's `requirements` /
+        `acceptance_criteria` / `integration_points` MUST also appear in that card's
+        `files_likely_touched`. An AC that adds a required prop / changes a shared component signature
+        / wires a value into a page but does NOT list the render-site / page / view-model file is a
+        BLOCKER (it forces a MAY-EDIT stall or an unplanned fix-coder in `/new`, and lets a coder dodge
+        the AC by making the prop optional). Message: `"Card <ID>: file <path> named in <field> is
+        missing from files_likely_touched."`
+      - **ISA-touchpoint ownership (P1 AC-descent):** every touchpoint in the state file
+        `## Integration Surface Map` — including every consumer enumerated for a SHARED MULTI-MODE
+        function (discovery dimension 9) — MUST be owned by some card's `files_likely_touched`. A PRD
+        surface / ISA file owned by NO card is a BLOCKER (this is the gap that otherwise surfaces only
+        at `/new`'s expensive final review). Message: `"ISA touchpoint <path> (<category>) is owned by
+        no card — assign it to a card's files_likely_touched."`
+      Both are SEMANTIC (path extraction from prose / ISA needs interpretation) — bias toward flagging.
+
+   **Gate**: every card has `owner_agent` AND `review_profile` in their enums, `requirements` present when
+   derivable (d), and full files_likely_touched coverage (e). WARN entries from checks (b)/(c) are noted in
+   the audit log; user may accept them after review. The (d)/(e) BLOCKERs must be fixed before commit.
 
 1. **Field Name Audit (runs after item 0 — only when item 0 passes for a given card; skipped for cards blocked in item 0)** — for each generated card:
 

package/framework/.claude/skills/worktree-manager/SKILL.md

@@ -100,6 +100,14 @@ Output: `{ path: string, branch: string, port: number }`
 Input: `{ worktreePath: string, checksAlreadyPassed?: boolean }`
 Output: `{ merged: boolean, developVerified: boolean, prNumber: number|null, mergeCommit: string, strategy: "pr"|"local-push" }`
 
+**Invocation (via the Skill tool — the `/new` Phase 6 merge gate, and any orchestrator):**
+
+```
+Skill({ skill: "worktree-manager", args: "/mw worktreePath=<absolute path> checksAlreadyPassed=true" })
+```
+
+The skill name is **`worktree-manager`** (the skill that owns the `/mw` command) — callers must NOT guess `new` / `new:new` / `mw`. The `/mw` flow parses `worktreePath=` (the merge target; when absent it falls back to reading the registry to identify the active worktree) and `checksAlreadyPassed=` from the args string. Pass ONLY those two keys — the branch and the card IDs for the commit message are resolved from the worktree/registry, never from invented payload keys.
+
 ## Programmatic API — docs mode (for /prd orchestrator)
 
 The docs mode is the **lean** counterpart to `/nw` and `/mw`. It is intentionally

package/framework/.claude/workflows/new-final-review.js

@@ -448,6 +448,25 @@ raw.forEach((f, i) => { f.finding_id = `${f.source || 'src'}#${i}:${f.finding_id
 phase('Verify')
 const classified = (await parallelCapped(raw.map((f) => () => verifyFinding(f)), MAX_PARALLEL)).filter(Boolean)
 
+// O5(b) — deterministic build-claim reconciliation. A VERIFIED BLOCKER/HIGH asserting a build/compile/
+// type/symbol FAILURE is self-contradicted when THIS same review's qa-sentinel build/tsc gate PASSed:
+// the build is green, so the claim cannot be a real build-breaker. Auto-demote it to
+// NEEDS_MANUAL_CONFIRMATION (never silently drop — it still surfaces to the human) instead of letting it
+// gate the merge or spawn a fixer on a phantom. Skipped when the build gate is itself FAIL or absent.
+const BUILD_CLAIM_RE = /\b(build|compil|tsc|type[-\s]?error|does\s*not\s*compile|exported\s+member|missing\s+(symbol|export|function|import)|cannot\s+find\s+(name|module))\b/i
+const buildGatePass = gateTable.some((g) => /build|tsc|type|compile/i.test(g.gate) && g.status === 'PASS')
+const buildGateFail = gateTable.some((g) => /build|tsc|type|compile/i.test(g.gate) && g.status === 'FAIL')
+if (buildGatePass && !buildGateFail) {
+  for (const f of classified) {
+    if (f.classification === 'VERIFIED' && (f.severity === 'BLOCKER' || f.severity === 'HIGH') &&
+        BUILD_CLAIM_RE.test(`${f.title || ''} ${f.evidence || ''} ${f.minimal_fix_direction || ''}`)) {
+      f.classification = 'NEEDS_MANUAL_CONFIRMATION'
+      f.reconciliation = 'auto-demoted: build/compile failure claim contradicted by a green build/tsc gate'
+      log(`Final review: auto-demoted [${f.finding_id}] ${f.severity} (build-failure claim) → NEEDS_MANUAL_CONFIRMATION — the qa build/tsc gate PASSed, contradicting it.`)
+    }
+  }
+}
+
 // Route a finding to the specialist that OWNS its domain (never a generic code-reviewer for
 // doc/api/security). Kept in sync with new2-resolve.js normDomain() routing buckets.
 function domainVerifier(domain) {
@@ -460,7 +479,16 @@ function domainVerifier(domain) {
 }
 
 async function verifyFinding(f) {
-  if (f.preValidated || (typeof f.confidence === 'number' && f.confidence >= 80)) {
+  // O5(a) — out-of-lane preValidated guard. A NARROW specialist (doc-reviewer / api-perf-cost-auditor)
+  // FP-checks ONLY its own lane in the finding pass; its `preValidated` (and self-reported confidence) do
+  // NOT authoritatively cover a finding it emitted OUTSIDE that lane (e.g. the perf auditor asserting a
+  // `domain: code` build/symbol failure). When the source is such a specialist and the finding's domain
+  // routes to a DIFFERENT verifier, strip the short-circuit so the correct domain specialist validates it.
+  // 'codex' / 'code-reviewer' are broad code reviewers (lane = all code domains) → unchanged. This is what
+  // let a phantom perf "BLOCKER" (a symbol it claimed missing, which existed) reach the skill un-verified.
+  const NARROW_SPECIALISTS = new Set(['doc-reviewer', 'api-perf-cost-auditor'])
+  const outOfLane = NARROW_SPECIALISTS.has(f.source) && domainVerifier(f.domain) !== f.source
+  if (!outOfLane && (f.preValidated || (typeof f.confidence === 'number' && f.confidence >= 80))) {
     return { ...f, classification: 'VERIFIED' }
   }
   const verifier = domainVerifier(f.domain)
@@ -630,6 +658,13 @@ if (applyFixes && editableFiles.length) {
   log(`Final Fix phase done: ${fixesApplied.length} fixed, ${docFixesApplied.length} doc-applied, ${residual.length} residual${checksFailed ? ', post-fix checks FAILED' : ''}.`)
 }
 
+// B2 — terminal completion sentinel. The skill's F.1.5 "workflow-killed recovery" branch treats the
+// ABSENCE of this line (with the run journal's terminal status ≠ completed) as "killed mid-flight".
+// Emitting it ONLY here — after every phase finished and the additive fields are assembled — makes it an
+// unambiguous clean-finish marker. NOTE: when applyFixes, the fixes are applied to the WORKTREE but NOT
+// committed (the orchestrator commits in F.5), so a run killed before this line leaves them as
+// uncommitted worktree diffs that the recovery branch re-validates + commits off-context.
+log(`NEW-FINAL-REVIEW COMPLETE — engine=${codexEngine}, findings=${findings.length}, ${applyFixes ? `fixes=${fixesApplied.length}+${docFixesApplied.length} (UNCOMMITTED in worktree), residual=${residual.length}` : 'review-only'}, gates=${gateTable.map((g) => g.gate + ':' + g.status).join(',') || 'none'}.`)
 const base = { codexEngine, findings, noActionFindings, gateTable, summary }
 // Additive ONLY when the Fix phase ran — otherwise byte-identical to today's shape (new2 untouched).
 return applyFixes ? { ...base, fixesApplied, docFixesApplied, residual } : base

package/framework/.claude/workflows/new2.js

@@ -271,7 +271,7 @@ try {
       `• Worktree (setup.md step 4): the ENTIRE build sequence (worktree add → install → env copy → port → registry → baseline) is the SSOT script \`scripts/setup-worktree.sh\` — the SAME script /nw and classic /new run. Do NOT hand-roll the bash (a re-authored copy is the duplication CLAUDE.md forbids, and the script cannot fabricate a baseline or stall). Resolve the branch from the lead/parent card's git_strategy.branch (fallback feat/<PARENT-ID|lead>-<slug>) + a kebab slug, then derive the git-authoritative idempotency pre-check (existing worktree for the branch → reset the half-built orphan before running, never re-create into a fail-loud collision). Resolve the script: \`SETUP_SH="$(ls .framework/framework/.claude/skills/worktree-manager/scripts/setup-worktree.sh .claude/skills/worktree-manager/scripts/setup-worktree.sh 2>/dev/null | head -1)"\`. Run it (foreground — you ARE the background context here): \`bash "$SETUP_SH" --main ${MAIN} --trunk ${TRUNK} --branch <branch> --slug <slug> --cards "${cardIds.join(',')}" --group-parent "<PARENT|>" --env-files "${ENV_FILES.join(',')}" --manifest /tmp/new2-wt-${firstCard}.txt\`. It installs non-interactively, copies stack.env_files (FILES plain cp / DIRS cp -R / missing FILE → WARN, never abort; never .env.example, never bulk-copy untracked), allocates a locked port, writes the registry, and runs the baseline under timeout 600 (tsc/lint report-but-continue, build STOP-on-fail). Read the manifest block (status/baseline/worktree_path/port/created_at/baseline_log): map baseline → your baseline field (pass|fail|timeout) and baseline_log → baselineLog. E2: a \`baseline:fail\`/\`status:error\` build break is NOT yours to fix (role boundary — the coder repairs it); just return it precisely. Then, INDEPENDENTLY of the manifest, VERIFY the worktree on disk and return EVIDENCE (the workflow E2.5 gate string-matches it, never the script's word): set worktreeVerified:true ONLY after running \`git -C ${MAIN} worktree list --porcelain\` (the worktree path MUST appear) AND \`test -d <wt>/node_modules\` AND confirming the branch; put the LITERAL stdout into worktreeEvidence{ worktreeListPorcelain, artifactsLs:\`ls -la <wt>/node_modules <wt>/.next 2>/dev/null | head\`, baselineLogTail }. If the script is ABSENT (older subtree), fall back to running the sequence inline per setup.md step 4 prose. NEVER report worktreeVerified:true without actually running the verify commands.\n` +
       codexResolveBullet +
       g3Bullet +
-      `• G4 card-field validation (setup.md 1b/1c): card missing requirements/acceptance_criteria/files_likely_touched → EXCLUDE (excluded[] + reason). Never HALT for one bad card.\n` +
+      `• G4 card-field validation (setup.md 1b/1c, card-schema.md consumer contract): card missing acceptance_criteria/files_likely_touched/scope → EXCLUDE (excluded[] + reason). \`requirements\` is CONDITIONAL, not an auto-exclude: if it is missing BUT acceptance_criteria + scope are both present and non-empty, FAITHFULLY DERIVE it (a concrete restatement/decomposition of the existing AC + scope — faithful, never inventing new scope) and WRITE it back to the card YAML on disk in ${MAIN}/${paths.backlog_dir || 'backlog'} (F-040: main repo, not the worktree copy; the card artifact is NOT a source/doc file, so this is allowed despite the role boundary), logging \`[BACKFILL] <id>: requirements=<N items, AC+scope-derived>\`. Only EXCLUDE for missing requirements when acceptance_criteria OR scope is ALSO empty (a genuinely thin card). Never HALT for one bad card.\n` +
       `• G5 depends-on: a card whose depends_on names a non-DONE card NOT in this batch → EXCLUDE it AND every in-batch card that transitively depends on it.\n` +
       `• cardGraph (REQUIRED, F-021): for every runnable card return { id, dependsOn:[IN-BATCH deps only], ownerAgent (the card's owner_agent; G25 unknown→'coder'), reviewProfile (the card's review_profile; default 'balanced'), policyDeferredACs, alreadyCommitted, alreadyCommittedSha, isEpic (implement.md §6b epic guard: id ends '-00' OR filename ends '-epic.yml' OR group.is_epic:true OR review_profile 'skip' with no requirements), filesLikelyTouched (verbatim from the YAML) }.\n` +
       `• B1/F-026 idempotency (per card, AFTER the worktree exists): set alreadyCommitted:true (+ alreadyCommittedSha) IFF ALL hold: (a) a commit referencing the card id exists in ${TRUNK}..HEAD of the worktree; (b) the card's validation_commands re-run GREEN right now; (c) NO open follow-up card for it exists in ${paths.backlog_dir || 'backlog'}. On a FRESH worktree ${TRUNK}..HEAD is empty → all false, zero extra work.\n` +

package/framework/agents/card-schema.md

@@ -55,7 +55,7 @@ Legend: **R** = required, present **and non-empty** · **E** = required key, val
 | `context` | R | R | R | |
 | `scope` | R | R | R | **never omittable** — encodes intent |
 | `scope_boundaries` | R | C | C | "Omit for standalone cards with no siblings" (`prd-card-writer.md`) |
-| `requirements` | C | R | R | EPIC tracks via AC-EPIC, not `requirements` |
+| `requirements` | C | R | R | EPIC tracks via AC-EPIC; CHILD/STANDALONE: faithfully derivable from AC+scope when both present — see consumer contract |
 | `acceptance_criteria` | R (AC-EPIC) | R | R | |
 | `definition_of_done` | R | R | R | |
 | `depends_on` / `blocks` | E (`[]`) | E | E | |
@@ -81,16 +81,52 @@ The three classes are deliberately distinct so a thin card is fixed where it can
 blocks where intent is genuinely missing:
 
 - **HALT (non-derivable — ask the user):** a **R** field that cannot be safely synthesized.
-  The canonical HALT set is `requirements`, `acceptance_criteria`, `files_likely_touched`,
-  and `scope` (CHILD/STANDALONE). `scope_boundaries` is **NOT** in the HALT set — it is **C**.
-- **BACK-FILL (deterministically computable — compute, persist, log):** `review_profile`
-  (compute via `prd-card-writer.md § Rule C`) and `owner_agent` (default `coder` for a
-  non-epic; **never** back-fill an epic's `""`). Write the value to the card on disk in the
-  **main repo** with a `[BACKFILL] <id>: <field>=<value>` log. **Check-and-skip**: only write
-  when the field is genuinely absent (idempotent across pre-flight re-entry).
+  The canonical HALT set is `acceptance_criteria`, `files_likely_touched`, and `scope`
+  (CHILD/STANDALONE). `scope_boundaries` is **NOT** in the HALT set — it is **C**.
+  `requirements` is a **conditional** member: it HALTs **only when its deriving material is
+  itself absent** (`acceptance_criteria` empty OR `scope` empty) — a genuinely thin card. When
+  `acceptance_criteria` AND `scope` are both present and non-empty, a missing `requirements`
+  block is *recoverable by faithful derivation* (next bullet, sub-kind b), not a HALT — a
+  well-specified card must never be skipped just because the writer omitted a derivable block.
+- **BACK-FILL (recoverable without asking — compute/derive, persist, log):** two sub-kinds.
+  *(a) Deterministically computable:* `review_profile` (compute via `prd-card-writer.md § Rule C`)
+  and `owner_agent` (default `coder` for a non-epic; **never** back-fill an epic's `""`).
+  *(b) Faithful derivation (bounded synthesis):* `requirements` — **only when**
+  `acceptance_criteria` AND `scope` are both present and non-empty — synthesize the
+  `requirements` list as a concrete restatement/decomposition of the existing AC + scope. This
+  is **faithful, not generative**: restate what the AC and scope already commit to; NEVER invent
+  new scope, constraints, or behaviour (the same procedure a card author performs by hand). For
+  either sub-kind, write the value to the card on disk in the **main repo** with a
+  `[BACKFILL] <id>: <field>=<value>` log. **Check-and-skip**: only write when the field is
+  genuinely absent (idempotent across pre-flight re-entry).
 - **WARN (expected but tolerable absent):** `canonical_docs` / `links.prd` on a STANDALONE /
   non-PRD card. Log, do not block.
 
+## Authoring completeness invariants (write-side — keep `/new` from paying for it)
+
+These bind the card AUTHOR (prd-card-writer) so a card never reaches `/new` with a gap `/new`
+must then patch in its expensive orchestrator context. Two recurring leaks, each with a measured
+downstream cost in a real batch:
+
+- **Per-AC seam-file closure (P2).** Every `acceptance_criteria` entry must have, in
+  `files_likely_touched`, the **render/seam file it mechanically touches** — not only the
+  definition site. An AC that **adds a required prop, changes a shared component's public
+  signature, or wires a new value into a page** MUST list that component's **render-site / page
+  file** (e.g. an AC "Step-2 row shows a `notOrderable` flag" must list the `page.tsx` /
+  view-model that builds the row, not just the data fn). A coder may NOT satisfy such an AC by
+  making the new prop **optional** to dodge an unlisted file — so if the file is missing the card
+  is under-specified. The MAY-EDIT list `/new` derives from `files_likely_touched` is only as
+  complete as this closure; a missing seam file caused MAY-EDIT stalls + unplanned fix-coders.
+  **Coverage rule:** every file named in a card's `requirements` / `acceptance_criteria` /
+  `integration_points` MUST also appear in `files_likely_touched` (⊆ check).
+- **`requirements` emitted at write time (PO1).** For a CHILD / STANDALONE card whose
+  `acceptance_criteria` AND `scope` are present, **emit `requirements` at authoring time** (the
+  faithful restatement/decomposition of AC + scope — same synthesis the `/new` BACK-FILL would do).
+  Do NOT ship the field absent and rely on the `/new` conditional back-fill above: that back-fill
+  runs in the orchestrator's accumulated context (a measured ~3–4M-token in-context tax per batch
+  when most cards omit it). The back-fill stays as the safety net for legacy cards; a freshly
+  authored `/prd` card must not trigger it.
+
 A back-fill never rescues a card missing a HALT field: a card without `scope` blocks regardless
 of how many derivable fields were filled.
 

package/framework/agents/code-search-protocol.md

@@ -80,6 +80,16 @@ Before picking a tier, classify the query:
   Repeated grep on a large barrel burns the turn budget and risks the agent
   returning *before* it synthesizes its report.
 - **Git**: no hard budget.
+- **Tool default — prefer `rg` over GNU `grep` for source searches.** `rg` searches
+  NUL-containing files as text by default; GNU `grep` silently switches to **binary
+  mode** on the first NUL byte (common in large generated / bundled / data files) and
+  then prints nothing and exits 1 — *indistinguishable from a genuine no-match*. So:
+  treat an empty `grep` result + exit 1 **on a file you EXPECT to match** as a
+  binary-mode trap, not "symbol absent" — re-run with `rg` (or `grep -a` to force
+  text mode) before concluding anything, and never spawn a fix to "add a missing
+  symbol" off a bare GNU-grep miss on a large file. (A measured `/new` run burned
+  ~8–12M tokens with three coders independently re-discovering one NUL byte in a
+  single multi-hundred-KB `data.ts`.)
 
 ## Fallback rules
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "baldart",
-  "version": "4.61.0",
+  "version": "4.62.0",
   "description": "Claude Agent Framework - Reusable framework for coordinating AI agents and humans in software projects",
   "bin": {
     "baldart": "./bin/baldart.js"

package/src/commands/teardown-codex-broker.js

@@ -0,0 +1,70 @@
+/**
+ * `baldart teardown-codex-broker --cwd <path>` — proactively end the Codex
+ * app-server broker bound to a SPECIFIC workspace cwd (since v4.62.0).
+ *
+ * The reactive companion `baldart reap-orphans` only mops up MCP servers AFTER
+ * their broker has already died (ppid 1). This command is the PROACTIVE, cwd-
+ * scoped teardown: at the END of a `/new` batch (or an abort), end the broker
+ * the batch's worktree created so its Playwright/Figma/… MCP children stop
+ * burning CPU for the rest of the session — instead of waiting for the plugin's
+ * SessionEnd hook (which may be hours away) or the next reap sweep.
+ *
+ * SCOPE & SAFETY: scoped strictly by `--cwd`. It ends ONLY the broker whose
+ * `--cwd` matches; a broker serving another worktree, or the user's interactive
+ * Codex session on the main repo (a DIFFERENT cwd), is never touched. The SSOT
+ * for the teardown logic is `src/utils/codex-orphans.js` (`tearDownCodexBroker`);
+ * this command only frames it. Always exits 0 — hygiene, never a blocker.
+ */
+
+const UI = require('../utils/ui');
+const CodexOrphans = require('../utils/codex-orphans');
+
+async function teardownCodexBroker(opts = {}) {
+  const json = !!opts.json;
+  const cwd = opts.cwd || process.cwd();
+
+  const result = {
+    schema: 'baldart.teardown-codex-broker/1',
+    cwd,
+    method: 'none',
+    brokerPid: null,
+    killed: 0,
+    failed: 0,
+    failures: [],
+  };
+
+  try {
+    const r = await CodexOrphans.tearDownCodexBroker({ cwd });
+    result.method = r.method;
+    result.brokerPid = r.brokerPid;
+    result.killed = (r.killed || []).length;
+    result.failed = (r.failed || []).length;
+    result.failures = r.failed || [];
+    if (r.error) result.error = r.error;
+  } catch (err) {
+    result.error = (err && err.message) || String(err);
+  }
+
+  if (json) {
+    process.stdout.write(JSON.stringify(result) + '\n');
+    return result;
+  }
+
+  if (result.error) {
+    UI.warning(`Codex broker teardown skipped (error: ${result.error}).`);
+  } else if (result.method === 'none') {
+    UI.success('Codex broker teardown: no broker bound to this cwd — nothing to do.');
+  } else if (result.method === 'graceful') {
+    UI.success(`Codex broker teardown: gracefully shut down the broker for this cwd (pid ${result.brokerPid ?? 'n/a'}) + its MCP children.`);
+  } else {
+    UI.success(`Codex broker teardown: signalled the broker subtree for this cwd (broker pid ${result.brokerPid ?? 'n/a'}, killed ${result.killed} process(es)).`);
+    if (result.failed > 0) {
+      UI.warning(`${result.failed} could not be killed:`);
+      result.failures.forEach((f) => console.log(`    pid ${f.pid}: ${f.error}`));
+    }
+  }
+
+  return result;
+}
+
+module.exports = teardownCodexBroker;

package/src/utils/codex-orphans.js

@@ -40,6 +40,9 @@
  */
 
 const { execSync } = require('child_process');
+const path = require('path');
+const fs = require('fs');
+const os = require('os');
 
 // Command signatures that identify an MCP server. When such a process is
 // orphaned (ppid 1) it is safe to reap (its stdio parent is gone).
@@ -172,11 +175,175 @@ function reapOrphans(orphans = [], procs = listProcesses()) {
   return { killed, failed };
 }
 
+// ─────────────────────────────────────────────────────────────────────────────
+// PROACTIVE, cwd-SCOPED broker teardown (since v4.62.0)
+// ─────────────────────────────────────────────────────────────────────────────
+//
+// WHY (distinct from reapOrphans above)
+// -------------------------------------
+// reapOrphans is REACTIVE: it can only mop up MCP servers AFTER their broker has
+// already died and the OS reparented them to init (ppid 1). It deliberately never
+// ends a live broker (a detached broker at ppid 1 is indistinguishable from a
+// leaked one by ppid alone).
+//
+// But the Codex plugin ALREADY ships a clean per-cwd broker teardown — it just
+// only fires on Claude Code `SessionEnd` (session-lifecycle-hook.mjs). A `/new`
+// run that ends — or is KILLED mid-final-review — long before the Claude session
+// ends therefore leaves its broker (and the Playwright/Figma/… MCP grandchildren)
+// alive for the rest of the session, burning CPU. `tearDownCodexBroker` closes
+// that gap: at /new run-end / abort, end the broker bound to THAT worktree's cwd.
+//
+// SAFETY: scoped strictly by cwd. The broker is keyed per workspace-root, so a
+// worktree's broker has no legitimate concurrent user — but a broker for a
+// DIFFERENT cwd (another worktree, or the user's interactive session on the main
+// repo) is NEVER matched. We never blanket-kill brokers.
+
+// Broker-process command signatures (the spawned app-server broker + its server).
+const CODEX_BROKER_SIGNATURES = [
+  /app-server-broker\.mjs/,
+  /codex\s+app-server/,
+];
+
+// Best-effort locate the installed Codex plugin `scripts/` dir (for Tier 1).
+function resolveCodexPluginScriptsDir() {
+  const home = process.env.HOME || os.homedir();
+  const base = path.join(home, '.claude', 'plugins');
+  const direct = [
+    path.join(base, 'marketplaces', 'openai-codex', 'plugins', 'codex', 'scripts'),
+  ];
+  for (const c of direct) {
+    try { if (fs.existsSync(path.join(c, 'lib', 'broker-lifecycle.mjs'))) return c; } catch (_) { /* ignore */ }
+  }
+  // Shallow scan: <plugins>/{marketplaces,cache}/<marketplace>/(plugins/)?codex/scripts
+  for (const sub of ['marketplaces', 'cache']) {
+    const d = path.join(base, sub);
+    let entries;
+    try { entries = fs.readdirSync(d); } catch (_) { continue; }
+    for (const mk of entries) {
+      for (const cand of [
+        path.join(d, mk, 'plugins', 'codex', 'scripts'),
+        path.join(d, mk, 'codex', 'scripts'),
+      ]) {
+        try { if (fs.existsSync(path.join(cand, 'lib', 'broker-lifecycle.mjs'))) return cand; } catch (_) { /* ignore */ }
+      }
+    }
+  }
+  return null;
+}
+
+// Extract the `--cwd <value>` token from a broker ps command line.
+function brokerCwdOf(command) {
+  const m = String(command).match(/--cwd\s+(\S+)/);
+  return m ? m[1] : null;
+}
+
+// Canonicalize a path for comparison (resolve symlinks where possible).
+function canonical(p) {
+  if (!p) return null;
+  try { return fs.realpathSync(p); } catch (_) { return path.resolve(p).replace(/\/+$/, ''); }
+}
+
+/**
+ * Proactively tear down the Codex app-server broker bound to a SPECIFIC `cwd`
+ * (typically a /new batch worktree), freeing its MCP-server children. Scoped by
+ * cwd — never touches a broker serving another worktree or the user's
+ * interactive session on the main repo. Never throws (hygiene, not a blocker).
+ *
+ * Tier 1 (graceful) — reuse the Codex plugin's OWN lib: loadBrokerSession(cwd) →
+ *   sendBrokerShutdown(endpoint) (the broker closes its app-client + MCP children)
+ *   → teardownBrokerSession(...) (unlink socket/pid/log + SIGTERM the broker pid)
+ *   → clearBrokerSession(cwd). This is exactly the plugin's own stale-broker path.
+ * Tier 2 (fallback) — when the plugin lib is absent/unimportable OR there is no
+ *   broker.json for this cwd but a broker process still lingers: ps-match the
+ *   broker whose command carries `--cwd <cwd>` + a codex-app-server signature,
+ *   SIGTERM its process group, then SIGKILL survivors in its subtree (reusing
+ *   listProcesses/collectTree).
+ *
+ * @param {{ cwd: string, pluginScriptsDir?: string }} opts
+ * @returns {Promise<{ method:'graceful'|'signal'|'none', brokerPid:number|null,
+ *                     killed:number[], failed:Array<{pid:number,error:string}>, error?:string }>}
+ */
+async function tearDownCodexBroker({ cwd, pluginScriptsDir } = {}) {
+  const none = { method: 'none', brokerPid: null, killed: [], failed: [] };
+  if (!cwd || process.platform === 'win32') return none;
+
+  // ---- Tier 1: graceful, via the plugin's own broker-lifecycle lib -----------
+  try {
+    const dir = pluginScriptsDir || resolveCodexPluginScriptsDir();
+    if (dir) {
+      const libPath = path.join(dir, 'lib', 'broker-lifecycle.mjs');
+      if (fs.existsSync(libPath)) {
+        const lib = await import(require('url').pathToFileURL(libPath).href);
+        if (typeof lib.loadBrokerSession === 'function') {
+          const session = lib.loadBrokerSession(cwd);
+          if (session && (session.endpoint || session.pid)) {
+            if (session.endpoint && typeof lib.sendBrokerShutdown === 'function') {
+              try { await lib.sendBrokerShutdown(session.endpoint); } catch (_) { /* ignore */ }
+            }
+            if (typeof lib.teardownBrokerSession === 'function') {
+              try {
+                lib.teardownBrokerSession({
+                  endpoint: session.endpoint ?? null,
+                  pidFile: session.pidFile ?? null,
+                  logFile: session.logFile ?? null,
+                  sessionDir: session.sessionDir ?? null,
+                  pid: session.pid ?? null,
+                  killProcess: (p) => { try { process.kill(p, 'SIGTERM'); } catch (_) { /* ignore */ } },
+                });
+              } catch (_) { /* ignore */ }
+            }
+            if (typeof lib.clearBrokerSession === 'function') {
+              try { lib.clearBrokerSession(cwd); } catch (_) { /* ignore */ }
+            }
+            return { method: 'graceful', brokerPid: session.pid ?? null, killed: session.pid ? [session.pid] : [], failed: [] };
+          }
+        }
+      }
+    }
+  } catch (_) {
+    // Tier 1 unavailable (plugin moved, import failed, …) → fall through to Tier 2.
+  }
+
+  // ---- Tier 2: signal the cwd-matched broker subtree -------------------------
+  try {
+    const target = canonical(cwd);
+    const procs = listProcesses();
+    const brokers = procs.filter((p) =>
+      matchesAny(CODEX_BROKER_SIGNATURES, p.command) && canonical(brokerCwdOf(p.command)) === target);
+    if (!brokers.length) return none;
+    const self = process.pid;
+    const killed = [];
+    const failed = [];
+    for (const b of brokers) {
+      // SIGTERM the broker's process group first (a detached broker is its own
+      // group leader) so it can run its shutdown handler; fall back to the bare pid.
+      try { process.kill(-b.pid, 'SIGTERM'); } catch (_) {
+        try { process.kill(b.pid, 'SIGTERM'); } catch (_) { /* ignore */ }
+      }
+      // Then SIGKILL any survivors in its subtree (force path — Tier 1 already failed/absent).
+      for (const pid of collectTree(b.pid, procs)) {
+        if (pid === self || !Number.isInteger(pid) || pid <= 1) continue;
+        try { process.kill(pid, 'SIGKILL'); killed.push(pid); }
+        catch (err) {
+          if (err && err.code === 'ESRCH') killed.push(pid);
+          else failed.push({ pid, error: (err && err.message) || String(err) });
+        }
+      }
+    }
+    return { method: 'signal', brokerPid: brokers[0].pid, killed, failed };
+  } catch (err) {
+    return { method: 'none', brokerPid: null, killed: [], failed: [], error: (err && err.message) || String(err) };
+  }
+}
+
 module.exports = {
   MCP_SIGNATURES,
   CODEX_RUNTIME_SIGNATURES,
+  CODEX_BROKER_SIGNATURES,
   listProcesses,
   detectOrphans,
   collectTree,
   reapOrphans,
+  tearDownCodexBroker,
+  resolveCodexPluginScriptsDir,
 };