@ai-dev-methodologies/rlp-desk 0.17.0 → 0.18.1

package/CHANGELOG.md

@@ -11,6 +11,55 @@ For pre-v0.15.4 versions, refer to `git log` and individual GitHub release notes
 - Post-v0.15.6: remove `RLP_LIFECYCLE_METRICS` flag entirely (per plan v3 ADR follow-ups).
 - Phase D.1 (handoff documents) + Phase D.2 (per-stage agent role specialization) — both deferred per `docs/plans/v0.15.4-release-runbook.md` §7.6.
 
+## [0.18.1] — 2026-06-25
+
+### Fixed
+- End-of-campaign resilience: when the last user story passes its per-US verify, the
+  leader now runs the final verification directly instead of dispatching one more
+  worker iteration solely to hand off to the final verify. That extra round-trip was a
+  fragile dependency on a healthy worker at the very end of a campaign (a worker stall
+  or API rate-limit at that moment could block a campaign whose work was already
+  complete and verified). Per-US campaigns now finalize without it.
+
+## [0.18.0] — 2026-06-24
+
+**Leader hardening: campaigns that complete.** The `--mode tmux` leader's decision
+gates used to TERMINATE where they should RECOVER, so campaigns could stall at
+`--max-iter` or block on a single transient slip. This release resolves that
+"never completes" failure class.
+
+### Fixed
+- A single transient `blocked` verdict no longer ends the campaign. A fresh-context
+  Worker/Verifier mis-emitting `blocked` once now gets grace (soft-fail + retry);
+  termination only on a genuine infra failure or a repeated/consecutive block.
+- Verifier phrasing variants (`completed`/`done`, lowercase `all`) no longer strand a
+  genuinely-complete campaign at the iteration cap — recommendation/us_id normalized at read.
+- Mixed-engine dead-pane detection: a dead pane is detected per role when Worker and
+  Verifier run different engines (e.g. claude worker + codex verifier).
+- No double-credit of a re-submitted User Story on the verify pass-path.
+- Final verification now retries a single transient poll failure instead of charging a
+  false fail at the end of the campaign.
+- Durable sentinels: atomic writes detect truncation (disk-full/SIGPIPE) instead of
+  renaming a half-written file into the canonical path.
+- Race-safe locks: the runner-lock and per-slug lock close an ABA/empty-owner race that
+  could let two leaders run on one project root.
+- Done-claim freshness gate: auto-synthesis rejects a stale/wrong-US done-claim, preventing
+  the wrong User Story from being credited.
+- Model-upgrade state survives relaunch: a resumed campaign keeps the upgraded worker model
+  instead of resetting to the CLI default.
+- codex 0.141 worker compatibility: the directory-trust prompt is accepted and MCP disabled
+  so codex workers don't block on startup.
+
+### Changed
+- Stronger final verification: `--final-verifier-model` / `--final-verifier-effort` are now
+  actually used for the final ALL-verify pass (previously every verify ran at the per-US
+  model). "Final = stricter" now holds.
+- Consensus cross-verifier model knobs now take effect: `--consensus-model` (per-US, lighter)
+  and `--final-consensus-model` (final, stricter) are wired into the consensus path —
+  previously documented but inert.
+- Verifier FORMAT meta-gates softened cross-engine while correctness gates stay strict
+  (fewer false fails on phrasing, no weakening of substance).
+
 ## [0.17.0] — 2026-06-19
 
 **BREAKING (ADR-001): `node run.mjs run <slug> --mode agent` (the deprecated Node-leader direct-CLI alpha) now hard-errors.** Per the schedule announced in 0.16.0, the `--mode agent` entry point exits 2 with a redirect, and the Node-CLI default mode flips from `agent` to `tmux` (the canonical production leader). The `src/node/**` engine modules are retained (they back the Native Agent() path and the test suite); only the direct-CLI `--mode agent` *dispatch entry* was removed. **Migration:** replace `node run.mjs run <slug> --mode agent` with `--mode tmux`. The slash command's `--mode native` default is unaffected.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-dev-methodologies/rlp-desk",
-  "version": "0.17.0",
+  "version": "0.18.1",
   "description": "Fresh-context iterative loops for Claude Code — autonomous task completion with independent verification",
   "scripts": {
     "postinstall": "node scripts/postinstall.js",

package/src/governance.md

@@ -66,9 +66,25 @@ Verification layers:
 - L3: E2E Simulation — known input → full pipeline → output comparison. Always required.
 - L4: Deploy Verify — production environment checks. Required when deploying.
 
-Layer requirements per US are determined by risk classification (§1c).
-Non-applicable layers must be explicitly marked "N/A — {reason}" in the test-spec.
-Any required layer section with TODO or blank = automatic Verifier FAIL.
+Layer requirements per US are determined by risk classification (§1c). Layer
+completeness means ACTUAL verification COVERAGE, not test-spec FORMAT: a required
+layer is satisfied when its behavior is genuinely exercised by a concrete, PASSING
+check — e.g. a per-AC command in the Criteria→Verification table counts as L1/L3
+coverage. Explicit "## L1/L2/L3" section headers and "N/A — {reason}" markers are
+recommended but NOT mandatory, and their absence is NOT itself a fail.
+FAIL a required layer ONLY when its verification is genuinely absent, blank, TODO,
+or failing — never merely for a different-but-adequate test-spec format, a missing
+N/A marker, or RED evidence aggregated across the US instead of recorded per-AC.
+The iter-signal.json only identifies WHICH US to verify; whether the Worker wrote
+it or the leader synthesized it does not change the verdict — verify the work, not
+the signal's author. Deliverable COMPLETENESS is NOT exempt: if the work an AC
+requires is absent, uncommitted/untracked, or never actually exercised, the layer
+still FAILs. This rule is IDENTICAL for every verifier engine (claude AND codex):
+do NOT block a PASS on test-spec FORMAT when the acceptance criteria are met and
+their checks are green — but DO fail on missing or unverified substance. This
+preserves strict AC + real layer coverage (IL-4 still governs test sufficiency);
+it removes only format pedantry, never substance — the F-17 cross-engine
+consistency fix (scope narrowed to format-only per the F-18 review).
 See §1d for full layer definitions.
 
 **IL-4: Test Sufficiency**