fix(workflow): complete trusted empty-scope tasks

[zaxbysauce]

Closes #2763

Summary

• Allow an explicitly trusted, generation-zero no-mutation settlement with files_touched: [] to complete without inventing a pre_check_batch receipt.
• Bind the empty-scope exception to the matching coder-settlement WAL and transition; derive applicable gates consistently for completion, status, the evidence bridge, and phase completion.
• Pass ledger-replayed task scope into phase-completion inference. Missing or expanded scope revokes stale empty-scope proof; legacy completion diagnostics now report the same required, satisfied, and missing gates as check_gate_status.
• Preserve fail-closed behavior for malformed or missing settlement scope, accepted mutation, repair generations, stale Stage A evidence, and failing secretscan evidence.

Invariant audit

• 1 (plugin init): not touched — no initialization or startup-await path changed.
• 2 (runtime portability): touched — plain TypeScript runtime logic only; bun run build, Node ESM import of dist/index.js, and the gate-portability check passed.
• 3 (subprocesses): not touched — no production subprocess caller changed.
• 4 (.swarm containment): touched — evidence remains under existing validated .swarm paths; focused settlement and recovery tests pass.
• 5 (plan durability): touched — phase inference uses the ledger-replayed plan; terminal authorization re-derives scope under the plan lock. Scope-expansion and omitted-scope regressions pass.
• 6 (test_runner safety): not touched — validation used bounded, per-file shell test runs; no broad test_runner call.
• 7 (test writing): touched — bun:test regressions cover empty, omitted, and expanded scope plus legacy gate parity; no new mock.module. No new test file exceeds 500 lines; the existing 522-line gate-status test shrank to 521 under the shrink-only FR-006 ratchet.
• 8 (session state): not touched — no new session/global state.
• 9 (guardrails/retry): touched — exception remains bound to exact scope, generation, and WAL proof; ordinary mutation and repair still require applicable gates.
• 10 (chat/system message): not touched — no chat transform or guidance carrier changed.
• 11 (tool registration): not touched — no metadata, manifest, barrel, or agent map changed.
• 12 (release/cache): touched — the issue release-note fragment is included; package version and cache logic are unchanged.

Test plan

• Frozen issue-tracer acceptance C1–C4: C1/C2 RED→GREEN; C3/C4 GREEN→GREEN, with checkpoint and deferred-work checks recorded in the issue trace.
• Focused exact-worktree validation: 125 tests passed across 11 files, including gate applicability, phase completion, legacy parity, gate status, Turbo bypass, delegation, bridge/summary, and settlement recovery.
• bun run typecheck; bun run lint:ci (exit 0; four non-blocking Biome warnings); targeted Biome; git diff --check; bun run build; Node ESM import.
• Push-protection pattern scan on the follow-up diff: no credential-like matches.
• Fresh CI and independent review for this PR head: pending; this PR remains a draft until checks and review are complete.

Acceptance criteria → evidence

Acceptance criterion	Evidence
AC1: Explicit files_touched: [] with a trusted no-mutation settlement completes after applicable task gates without a fabricated file or Stage A scan.	Frozen C1; tests/unit/tools/empty-scope-completion-2763.test.ts; tests/unit/evidence/gate-evidence-2763.test.ts.
AC2: Status and completion diagnostics agree on the applicable gate set.	Frozen C2; tests/unit/tools/legacy-gate-parity-2763.test.ts compares required, satisfied, and missing arrays, including a satisfied reviewer receipt and missing pre_check.
AC3: Non-empty scope or accepted mutation retains current-generation pre_check.	Frozen C3; existing mutation and scope regressions.
AC4: Malformed inputs, stale or omitted scope, and mutation-bearing tasks remain fail-closed.	Frozen C4; tests/unit/tools/phase-complete-doc-only-evidence.test.ts proves empty scope passes while omitted and expanded scope fail.

Recurrence prevention

• Defect class: completion/status/recovery code independently reconstructing gate applicability from empty or lossy scope instead of trusted, generation-bound evidence.
• The repository-wide census dispositioned semantic hits; no unresolved in-class residual remains.
• The guardrail combines a WAL trust-boundary assertion, shared typed gate derivation, terminal-WAL validation, and regressions for scope expansion and legacy parity.

Risk and rollback

• Risk: medium; this changes durable task evidence and completion decisions, but the exception is limited to authoritative generation-zero empty settlements and remains fail-closed otherwise.
• Rollback: revert the issue commits; the WAL field is optional, preserving legacy parsing.
• Residual behavior: an unrelated context-free no-mutation event after settlement clears eligibility and restores pre_check by design.

Environment notes

• Earlier PR head 01c13fd57d0de2d7babeb877059cb4a701edc2ae had green CI and passed the focused security/adversarial/smoke checks; exact-head checks for this follow-up are pending.
• The canonical isolated integration run had one skill-optimizer-e2e production-store tripwire failure also reproduced on clean origin/main ([Observability PR 05/23] Isolate hive stores and add exact-ID quarantine with rollback #2033). The restricted-home swarm-model F-07 case likewise failed on clean main; its isolated writable-runtime run passed 19/19.
• A supplementary co-run of the existing coder-settlement recovery test once reported a production-store tripwire after nine assertions. Its isolated rerun passed 9/9. The platform store was left untouched, and the cause of that co-run signal is not claimed to be pre-existing.

Waivers

None.

Merge status

Awaiting explicit user approval; not merged. The PR remains draft.

PR head: 2d40c56f640b573aa868f5e25fd379d3c57be940

[github-actions]

Drift check report

Found 2 drift finding(s): 0 error, 0 warning, 2 notice.

required-check-contract (2)

• 🔵 notice scripts/required-check-contract.json: [RULESET_DIVERGENCE] intended-required context "drift" is not yet required by the captured ruleset
• 🔵 notice scripts/required-check-contract.json: [RULESET_DIVERGENCE] intended-required context "drift" is not present for every expected event in captured external workflow evidence

[Copilot]

🔵 Needs a closer look

It changes safety-critical durable task-completion, terminal WAL replay, and gate-derivation trust-boundary semantics whose fail-closed guarantees warrant final human verification.

Pull request overview

This PR fixes issue #2763, where a verification-only task that honestly declares files_touched: [] could never be marked completed because update_task_status required a pre_check gate that pre_check_batch never records for an empty file list, deadlocking the task in in_progress. It also resolves the reported disagreement between check_gate_status and update_task_status over the required-gate set by routing both (plus the durable evidence bridge and phase completion) through one shared deriveApplicableGateSet derivation. The exception is tightly scoped: it is granted only when an authoritative, generation-0 coder settlement proves an empty declared scope and no accepted mutation, and every other path (plan scope alone, null/malformed/non-empty settlement scope, accepted mutation, later generations, stale Stage A, failing secretscan) remains fail-closed.

Changes:

• Add a generation-0 noMutationSettlement marker, bound to the trusted coder-settlement WAL, and a shared deriveApplicableGateSet/isReadOnlyNoMutationEligible derivation reused by completion, status, the durable bridge, and hasPassedAllGates.
• Thread readOnlyNoMutation through the task-terminal transition and terminal WAL schema (with cross-field validation) so a durable, non-forced empty-scope completion replays deterministically.
• Add extensive positive/negative/malformed/replay/idempotence/secretscan/recovery regression tests and a release fragment.

File summaries

File	Description
src/gate-evidence.ts	Adds the noMutationSettlement marker, deriveApplicableGateSet, WAL-bound trust adapter, reducer handling, and stricter hasPassedAllGates.
src/tools/check-gate-status.ts	Routes required/passed/missing gates and status through the shared derivation and readOnlyNoMutation.
src/tools/update-task-status.ts	Uses shared derivation in checkReviewerGate, ignores empty scope for Turbo bypass, and passes resolveTerminal with readOnlyNoMutation.
src/workflow/coder-settlement.ts	Emits declared-scope/transition context and hardens idempotent COMMITTED replay against advisory-gate overwrite.
src/workflow/task-terminal.ts	Propagates readOnlyNoMutation into the terminal event and WAL, including idempotence comparison.
src/workflow/workflow-wal-schema.ts	Adds the optional readOnlyNoMutation field with backward-compat parsing and cross-field validation.
src/evidence/gate-bridge.ts	Derives durable gate status via the shared function instead of treating empty required gates as complete.
scripts/retention-registry.data.ts	Updates task-workflow-evidence citation line numbers for the shifted symbols.
docs/releases/pending/issue-2763-empty-scope-completion.md	Adds the user-visible release fragment.
tests/**	Adds and updates unit/integration coverage for the empty-scope, terminal-replay, secretscan, and receiptless paths.

Review details

• Files reviewed: 20/20 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[zaxbysauce]

Swarm PR Review — #2776 "fix(workflow): complete trusted empty-scope tasks"

Scope: 37ee0ce8b..ea8dba687 (20 files, +1280/-128). Depth tier M (risk triggers: concurrency-state, api-schema-migrations). Profile B (Claude Code, no controller) — 10 parallel explorer lanes (6 base dimensions + 4 mandatory risk-family micro-lanes), 4 independent reviewer lanes, 1 critic challenge, all bound to pr_head_sha=ea8dba68718067cf7fd3b55d5bd01eedabd1ace9.

Verdict: BLOCK

A reviewer-confirmed, critic-upheld CRITICAL finding (evidence-gate bypass) is present and is not mitigated by an unusual/adversarial precondition — it fires on the PR's own advertised happy path.

---

Critical finding

F-1 (CRITICAL, confirmed, upheld by critic) — The new "trusted no-mutation" marker is not proof of no mutation

A plan task with files_touched: [] cannot be dispatched to a coder at all unless a FILE: prompt line or a declare_scope call supplies real, non-empty write authority (resolveCoderScopeSources rejects a wholly empty scope — src/scope/scope-binding.ts:954-957). So every coder-settlement whose declaredFiles is [] necessarily had non-empty write authority (src/hooks/guardrails/file-authority.ts:945-957 grants writes from the resolved scope, not from declaredFiles).

At settlement time, the real git-diff-observed mutation set is filtered through isPathWithinDeclaredScope(file, [], dir) (src/scope/path-identity.ts:85-87), which is scopeEntries.some(...) — vacuously false for an empty array. So observedFiles is always [] and accepted = observedFiles.length > 0 is always false, regardless of what the coder actually wrote to disk.

This PR's new bindTrustedNoMutationSettlement (src/gate-evidence.ts:975-1024) trusts exactly the accepted===false && declaredFiles.length===0 combination as proof of no mutation, and mints a noMutationSettlement marker that lets update_task_status(completed) skip pre_check (build/SAST/secretscan requirement) — specifically on reviewer-required phases, which are the default (src/tools/update-task-status.ts:2090-2107).

Net effect: a coder can be dispatched with files_touched: [] (the plan schema's own default when the field is omitted — src/config/plan-schema.ts:81) plus one FILE: line, write to that file, and the task completes with required_gates: [] — no pre_check, no secretscan, no SAST, no reviewer-required QA — with no forced-completion audit stamp (the new path sets neither qaExempt nor forcedCompletion, converting what used to be an auditable forced-completion into a silent one). The same deriveApplicableGateSet/hasPassedAllGates derivation this PR introduces is also consumed by phase_complete (src/tools/phase-complete.ts:287) and the evidence bridge (src/evidence/gate-bridge.ts:62-68 → preflight/diagnose/evidence-summary services), so the bypass propagates into phase completion too (confirmed: tests/unit/evidence/gate-evidence-2763.test.ts:29-52 asserts hasPassedAllGates === true for zero-gate empty-scope evidence).

This is not an adversarial/unusual-plan construction — it is the only way the PR's own advertised path can ever be reached (a files_touched: [] task with genuinely zero scope throws SCOPE_NOT_DECLARED before any settlement exists at all, so the "pure" issue #2763 scenario described in the bug report is, ironically, still not fixed by this PR).

Falsification probe: dispatch a coder for a files_touched: [] task with one FILE: src/x.ts line, have it write src/x.ts, call update_task_status(completed), and assert the task completes with no pre_check gate and no forcedCompletion/qaExempt stamp.

Suggested fix: record the settlement's declaredFiles from the resolved authority scope (binding.files), not the raw plan files_touched, so declaredFiles: [] genuinely implies zero write authority — or reject bindTrustedNoMutationSettlement whenever rawObservedFiles.length > 0, not just the post-filter observedFiles. Add a regression test where the coder writes a FILE:-granted file and completion is refused.

(Reachability/authority-chain traced independently by two separate agents — an initial reviewer and a final critic — reading src/hooks/delegation-gate.ts, src/scope/scope-binding.ts, src/scope/path-identity.ts, src/hooks/guardrails/file-authority.ts, and src/gate-evidence.ts end to end, with a concrete falsification probe on both passes.)

---

Important non-blockers

F-2 (HIGH, confirmed) — Legacy-migration backfill path is unreachable for its stated purpose (unwired code)

recoverCoderSettlement (src/workflow/coder-settlement.ts:970-986) early-returns for any COMMITTED WAL, so it never reaches the new backfill allowance at :309-318. The three production callers of recoverCoderSettlement therefore cannot use it to backfill a marker onto an already-stuck pre-PR #2763 task. Mitigating context (critic): an ordinary architect re-dispatch (beginCoderSettlement, coder-settlement.ts:402-451) overwrites the stale COMMITTED WAL under a new transitionId and re-enables normal settlement — so the task is not permanently wedged, just not self-healing on deploy alone as the release note implies. Net: the new branch is dead code relative to its own stated intent (project directive #2: "never ship unwired code"). Untested (RP-1, MEDIUM, merge with this finding): the only related test (empty-scope-completion-2763.test.ts:250-282) exercises the marker-present idempotent path, never the marker-absent backfill branch, and that branch also double-increments retryCount/retryHistory when it does run (bounded, capped at 3 — CONC-001/IA-004, LOW-MEDIUM) and emits a duplicate settled audit event without a resumed:true flag (RP-2, LOW).

F-3 (LOW, advisory) — Undocumented check_gate_status reporting-semantics change

check-gate-status.ts's all_passed/status computation now additionally requires workflow.state to be tests_run/complete for all tasks, not just the new empty-scope path (confirmed by the retrofit of check-gate-status-secretscan*.test.ts fixtures). No in-repo programmatic consumer depends on all_passed (it's an agent-facing diagnostic field), so this is advisory, but it's undisclosed in the release fragment and should be called out.

---

Confirmed non-blocking findings (grouped)

• Retention-registry citation drift (LOW): scripts/retention-registry.data.ts's citation repoint for the touched row is incomplete — writerCitations/readerCitations were correctly shifted, but writeLimits.citation, schemaVersion, and readBound.citation still point at pre-diff line numbers that now name unrelated code. Invisible to CI (the registry test only checks non-emptiness).
• "One shared gate derivation" claim is an overstatement (LOW, pre-existing): three hand-rolled gate-derivation blocks in update-task-status.ts (:318-333, :777-816, :1186-1241) and a fourth independent implementation in src/ci/evaluate.ts:524-553 were not migrated to deriveApplicableGateSet. All fail conservatively (no false-green), but they contradict the PR's own "no unresolved in-class residual remains" recurrence-prevention claim.
• Missing test coverage at the phase_complete and crash-recovery boundaries (LOW/MEDIUM): no test exercises a trusted empty-scope task through phase-complete.ts's tool boundary or through recoverCoderSettlement's crash-recovery mint path — both are real production entry points for the new marker.
• repair_idle marker-clearing is correct but untested in combination (LOW): the clearing logic at gate-evidence.ts:915-932 is structurally sound (verified: the marker cannot survive into a mutating generation regardless, due to the generation===0 gate), but no test combines repair_idle with noMutationSettlement.
• Weak fail-closed test assertions (LOW/MEDIUM): empty-scope-completion-2763.test.ts:332-370's negative-path loop only asserts result.success === false for 3 of 4 cases, without checking why — a regression that failed for an unrelated reason would still pass.
• Silently swallowed exception, no logging (LOW): bindTrustedNoMutationSettlement's catch block (gate-evidence.ts:1015-1020) fails closed correctly but logs nothing, making a transient WAL-read failure indistinguishable from a legitimate refusal.
• check_gate_status vs update_task_status reporting/enforcement asymmetry (LOW): check_gate_status omits contradictorySignals/routeGateAllowsTask checks that update_task_status performs — narrower, less severe recurrence of the class of bug this PR claims to fix. No enforcement consumer found.
• New 413-line test file is at 82.6% of the repo's 500-line test-file-cap ratchet (INFO) — currently green, flagged for future headroom.

Pre-existing, not introduced by this PR

• update_task_status's actual completion write path never consults secretscan evidence at all (only the read-only check_gate_status diagnostic tool does) — identical on base and head; not worsened by this PR, but adjacent to F-1's trust boundary and worth its own issue.
• A separate, unrelated null-declaredFiles branch in coder-settlement.ts:1210-1211 also forces observed=[] — untouched by this diff.

Disproved / non-findings

• Explorer claim that "phase completion" isn't wired into the shared derivation at all: disproved — phase-complete.ts:287 calls hasPassedAllGates, which now internally uses the shared derivation (transitively wired, no diff hunk needed there). This nuance matters because it's exactly the path F-1's bypass propagates through.
• Mixed-version WAL-schema passthrough concern (ASM-001): disproved — the new field is explicitly declared in the schema, so it survives normal (non-legacy) parsing; the passthrough comparison to a same-version sibling schema didn't apply.
• Retention-registry's 3-line change: confirmed to be legitimately connected (line-citation bookkeeping from this PR's own insertions), not scope creep.
• The isDuplicateTransition/assertExpectedGeneration replay path: confirmed fail-closed as claimed; the generation guard is enforced upstream before the bypass can matter.

Unverified (flagged, not confirmed or refuted)

• The PR body's "251 passed, 0 failed" broader 22-file affected/recovery matrix claim was not independently reproduced (the exact file list wasn't available to the reviewing lane); the narrower "183 tests" claim also could not be exactly reproduced in isolation (a different subset totaled 126 in one run), though all tests examined passed.
• A narrow rollback-timing edge case (a PREPARED-but-not-yet-COMMITTED WAL with the new marker, rolled back to old code mid-window) is plausible but not reproduced; assessed as fail-closed in the stated direction if it does occur.

---

Validation provenance

• Explorer lanes (10): intent-architecture, correctness-state, tests-falsifiability, security-trust, reliability-performance, compatibility-delivery (base dimensions) + concurrency-state, api-schema-migrations, test-infrastructure, unclassified-risk (mandatory risk-family micro-lanes). ~37 raw candidates generated.
• Independent reviewers (4): every candidate individually re-derived from source (not trusted from the explorer's claim), each with a concrete code citation and falsification probe. Classifications: 1 CRITICAL confirmed, 1 HIGH confirmed (later downgraded to MEDIUM by critic, but spawning a new HIGH finding), several MEDIUM/LOW confirmed, several DISPROVED, several PRE_EXISTING, several UNVERIFIED (disclosed above).
• Critic: independently re-derived the CRITICAL finding's reachability chain from scratch (not trusting the reviewer's citations), strengthened rather than weakened it, corrected an overclaim (secretscan is not bypassed as an overlay — it's just never made a required gate), and surfaced one additional HIGH finding (phase-completion propagation) the reviewers had part-credited to "no consumers" without checking the full diff.
• PR title/body compliance: the PR's own invariant-audit and acceptance-criteria tables are thorough, but two of its own claims are contradicted by code: (1) the "COMMITTED settlement replay preserves advisory gates without incrementing retry state" claim (invariant 9) is false for the marker-absent backfill branch; (2) the empty-scope exception does not actually reach the literal issue update_task_status(completed) permanently blocked for tasks with files_touched: [] — pre_check gate can never be satisfied #2763 scenario (a task with truly empty scope and no FILE:/declare_scope) — that still throws before any settlement exists.
• Mergeability: CI was still running required checks (all previously-completed checks green) at review time; no merge conflicts.

Required actions before merge

1. Fix F-1: bind the trust marker to the actual authority scope, not the raw (possibly-empty) plan declaration; add a test where a coder writes a FILE:-granted file under files_touched: [] and confirm completion is refused.
2. Either wire the backfill path to an actual production entry point (F-2) or remove it and document that stuck-legacy tasks are recovered via ordinary re-dispatch.
3. Add test coverage for the phase_complete boundary and the crash-recovery mint path with a trusted empty-scope task.
4. Fix the retention-registry citation drift and the retry-count double-increment on backfill replay.
5. Amend the release fragment to disclose the check_gate_status semantics change and correct/remove the "proves no mutation was accepted" claim, replacing it with an accurate description of what the marker actually attests.

---

🤖 Generated via /swarm-pr-review (Claude Code, Profile B) — 10 explorer lanes, 4 independent reviewers, 1 critic challenge, all bound to ea8dba68718067cf7fd3b55d5bd01eedabd1ace9.

[zaxbysauce]

🤖 Multi-Stage PR Review

Pipeline: MiniMax-M2.7-highspeed (orientation) (context pack) → MiniMax-M2.7-highspeed (explorer) + MiniMax-M2.7-highspeed (explorer B) (parallel explore, distinct lenses) → GLM-5-turbo (critique) ↔ GLM-5-turbo (critique) (cross-critique) → MiniMax-M2.7-highspeed (fallback arbiter) (arbiter: blind-spot + synthesize)
Commit reviewed: d1183eec4112

---

🔍 PR Intent

Reconstructed from PR text, issue #2763, commits, changed tests, and changed interfaces.

• O-001: Tasks with explicit files_touched: [] and a trusted generation-0 no-mutation coder settlement can complete without pre_check_batch evidence.
• O-002: check_gate_status and update_task_status(completed) report the same applicable gate set for a given task generation.
• O-003: Raw workspace mutations against an empty declared scope → rework_required, not a no-mutation settlement.
• O-004: Non-empty scope, accepted mutation, later repair generations, stale Stage A, and failing secretscan remain fail-closed.
• O-005: The noMutationSettlement marker is cleared on any new dispatch or accepted mutation.
• O-006: WAL replay and terminal recovery preserve the invariant.
• O-007: Existing empty required_gates evidence (no trusted settlement) retains the pre_check obligation.

---

📦 Implementation Summary

The PR adds a noMutationSettlement block to TaskWorkflowMetadata, persisted only via the coder-settlement WAL at dispatch_no_mutation time for generation-0, empty-declared-scope settlements. A new exported deriveApplicableGateSet() function (used by both check_gate_status and update_task_status) reports an empty gate set when the trusted settlement proof is present; otherwise pre_check is included. The marker is cleared by dispatch_attempted and accepted_mutation. hasUnattributedEmptyScopeMutation() flags mutations against an empty declared scope, routing them to rework_required instead. WAL replay, terminal WAL schema, and terminal recovery are hardened. All gate derivation now routes through one shared function.

---

✅ / ⚠️ / ❌ Intended vs Actual

Obligation	Status	Evidence (file:line)
O-001	SUPPORTED	src/gate-evidence.ts:640 — isNoMutationSettlementMetadata; src/gate-evidence.ts:600 — deriveApplicableGateSet returns empty gate set
O-002	SUPPORTED	src/tools/check-gate-status.ts:237 + src/tools/update-task-status.ts:600 — both call deriveApplicableGateSet
O-003	SUPPORTED	src/workflow/coder-settlement.ts:178 — hasUnattributedEmptyScopeMutation; src/workflow/coder-settlement.ts:216 — returns null from scopedObservedFiles
O-004	SUPPORTED	src/gate-evidence.ts:730 — dispatch_attempted clears marker; src/gate-evidence.ts:797 — accepted_mutation clears marker
O-005	SUPPORTED	src/gate-evidence.ts:730 (dispatch_attempted clears), src/gate-evidence.ts:797 (accepted_mutation clears)
O-006	SUPPORTED	src/workflow/workflow-wal-schema.ts:489 — WAL state mismatch check; src/workflow/task-terminal.ts:258 — evidenceMatchesTerminal includes readOnlyNoMutation
O-007	SUPPORTED	src/gate-evidence.ts:649 — !readOnlyNoMutation && !requiredGates.includes('pre_check') → pre_check added; src/evidence/gate-bridge.ts:62 — missingGates updated

---

🚨 Confirmed Findings

[CRITICAL] reduceTaskWorkflowSnapshot task_completed branch accesses context.requiredGates which is never set on the no-mutation path

• Location: src/gate-evidence.ts:909
• Why it matters: When readOnlyNoMutation is true (the fix's primary feature), the reducer throws TypeError: Cannot read properties of undefined (reading 'every') because context.requiredGates is never populated by the dispatch_no_mutation event or bindTrustedNoMutationSettlement. This permanently blocks the exact task the PR intends to unblock.
• Evidence: dispatch_no_mutation events (from src/workflow/coder-settlement.ts:238) construct context as { declaredFiles, settlementTransitionId } — no requiredGates or gates. bindTrustedNoMutationSettlement at :1001 only adds declaredFiles from the WAL and settlementTransitionId. reduceTaskWorkflowSnapshot at :622 defaults context to {}, so context.requiredGates is undefined.
• Fix direction: In the task_completed reducer, replace context.requiredGates.every(...) with derivedGates.requiredGates.every(...) where derivedGates = deriveApplicableGateSet(/* reconstruct from evidence */). Alternatively, construct the applicableGates object in updateEvidenceForTransition before calling reduceTaskWorkflowSnapshot and pass it in.

[MEDIUM] continue in malformed-settlement test case skips the failure assertion

• Location: tests/unit/tools/empty-scope-completion-2763.test.ts:193
• Why it matters: The malformed-settlement case calls expect(...).rejects.toThrow(...) then continues before the expect(result.success, candidate.label).toBe(false) assertion at line 265. If settleNoMutation does NOT throw (e.g., WAL schema accepts scope: 'not-an-array'), the test silently passes with no assertion covering that case.
• Evidence: continue at :193 skips the only assertion in the loop body that checks result.success.
• Fix direction: After the rejects.toThrow assertion, either assert result.success === false directly for that case, or restructure the loop to have a dedicated assertion for each case type.

---

🔬 Unverified but Plausible Risks

• Risk: isNoMutationSettlementMetadata at line 651 checks workflow.state === 'idle' || workflow.state === 'complete'. A task that somehow reaches complete state through another path (e.g., forced completion) and then re-enters dispatch_no_mutation would pass the state check. While repair_idle opens a new generation, the interaction surface between forced completion and the read-only path is complex and deserves a dedicated regression test.
  • What would verify it: A test exercising forced completion followed by a no-mutation dispatch on the same taskId.

---

🧪 Test / Coverage Gaps

• Gap: deriveApplicableGateSet and isNoMutationSettlementMetadata have no unit tests in isolation — coverage is through integration tests. The task_completed reducer path for readOnlyNoMutation is exercised by the integration repro scripts but not by a fast unit test.
  • Evidence: tests/unit/evidence/gate-evidence-2763.test.ts and tests/unit/evidence/gate-bridge-2763.test.ts cover gate derivation and bridge but not the full reducer path including task_completed.

---

📋 Shipped-vs-Claimed Gaps

• Gap: None. All four acceptance criteria have corresponding test coverage (C1/C2/C3/C4).

---

📝 Merge Recommendation

[BLOCK]

Two defects: one is a guaranteed TypeError crash on the primary feature path (CRITICAL), one is a silent test gap (MEDIUM).

Check	Result
No CRITICAL findings	❌ reduceTaskWorkflowSnapshot TypeError on task_completed with readOnlyNoMutation
No unresolved STEALTH_CHANGE	✅
No UNSUPPORTED obligations	✅
Test coverage adequate	⚠️ Missing unit test for task_completed reducer path; malformed-settlement case assertion gap
No hardcoded secrets	✅
All async errors handled	✅
Input validation present	✅
No broken agent role boundaries	✅
Prompt format contracts intact	✅
Lockfile consistent	✅

---

🔁 Validation Provenance

Finding	Source	Outcome
context.requiredGates TypeError	Reviewer confirmed, verified	KEPT (CRITICAL) — traced through bindTrustedNoMutationSettlement and settlementTransitionEvent; context is { declaredFiles, settlementTransitionId }, never requiredGates
accepted: true && settlementFailed: true contradiction	Reviewer confirmed	DROPPED — accepted means "mutation occurred"; settlementFailed means "settlement broke"; both true for out-of-scope mutation is by design per PR docs and test at line 95
String match error.message.includes(...)	Reviewer confirmed	DROPPED — maintenance risk, not a functional defect; other structural guards (WAL assertion, generation check) provide the authorization
continue skips assertion	Reviewer confirmed	KEPT (MEDIUM) — malformed-settlement case skips expect(result.success).toBe(false)
Loop cleanup leak	Reviewer confirmed	DROPPED — resetSwarmState() before cleanup() in each iteration; stale directory reference is recovered on next iteration
result.success not checked by reason	Reviewer confirmed	DROPPED — all 4 cases fail identically (no settlement → blocked by pre_check); error reason is deterministic
Unused settlement variable	Reviewer confirmed	DROPPED — settlement.evidence is used via transitionTaskWorkflowEvidence; beginCoderSettlement → settleCoderDispatch order is tested in other cases
Blind spot: task_completed reducer context access	Blind-spot pass	ADDED as CRITICAL above

---

_{🔒 Reviewed by a 3-model cross-family adversarial debate (architect → dual-lens parallel explorers → cross-critique → arbiter) for high recall with low false-positive noise. Findings are advisory — verify before acting.}

[zaxbysauce]

🤖 Multi-Stage PR Review

Pipeline: MiniMax-M2.7-highspeed (orientation) (context pack) → MiniMax-M2.7-highspeed (explorer) + MiniMax-M2.7-highspeed (explorer B) (parallel explore, distinct lenses) → GLM-5-turbo (critique) ↔ GLM-5-turbo (critique) (cross-critique) → MiniMax-M2.7-highspeed (fallback arbiter) (arbiter: blind-spot + synthesize)
Commit reviewed: 01c13fd57d0d

---

PR Reviewer — opencode-swarm

🔍 PR Intent

Reconstructed from PR text, issue #2763, and diff:

• O-001: Tasks with files_touched: [] and a generation-0 no-mutation coder settlement must be completable without a pre_check gate.
• O-002: check_gate_status and update_task_status(completed) must report the same applicable gate set (converging the historically divergent implementations).
• O-003: Raw workspace mutations against an empty declared scope must be treated as failed mutations, not as no-mutation proof.
• O-004: The no-mutation exception must be fail-closed: plan scope alone, malformed scope, accepted mutation, later generations, stale Stage A, and failing secretscan must all retain pre_check.

---

📦 Implementation Summary

The PR introduces a generation-0 no-mutation settlement path: when a coder dispatch with an empty declared scope produces no observed mutations, a trusted noMutationSettlement marker is persisted in the task evidence. A new deriveApplicableGateSet() function is the single gate-derivation source shared by both check_gate_status and update_task_status, and it reports pre_check as missing unless (a) the marker exists, (b) the workflow state is idle/complete, (c) generation is 0, and (d) the current plan scope is also empty. The hasUnattributedEmptyScopeMutation() guard prevents mutations within an empty declared scope from being silently filtered into a no-mutation proof.

---

✅ / ⚠️ / ❌ Intended vs Actual

Obligation	Status	Evidence
O-001	SUPPORTED	src/gate-evidence.ts:653–688 — deriveApplicableGateSet skips pre_check for trusted no-mutation evidence
O-002	SUPPORTED	src/evidence/gate-bridge.ts:61, src/tools/check-gate-status.ts:250–268 — both call deriveApplicableGateSet
O-003	SUPPORTED	src/workflow/coder-settlement.ts:165–182 — hasUnattributedEmptyScopeMutation forces accepted=true / settlementFailed=true for raw mutations on empty scope
O-004	SUPPORTED	src/gate-evidence.ts:619–660 — isNoMutationSettlementMetadata requires exact generation-0 + empty scope + idle/complete state

---

🚨 Confirmed Findings

None. All three reviewer-confirmed findings were false positives; see §🔁 Validation Provenance.

---

🔬 Unverified but Plausible Risks

None. The empty-scope path is gated by five independent conditions (WAL trust-boundary, generation, settlement metadata, current plan scope, workflow state) and the recovery path also carries the settlementFailed flag.

---

🧪 Test / Coverage Gaps

• Gap: No test covers the case where a no-mutation settlement is recorded, then the plan's files_touched is expanded mid-session before completion, and completion is subsequently attempted.
  • Evidence: The test 'revokes the no-mutation exception when the current plan later gains scope' covers checkReviewerGate but not executeUpdateTaskStatus with the terminal-WAL path.
  • Severity: Low — the readCurrentTaskDeclaredFiles call in update-task-status.ts:604–608 is present; this is a missing regression test rather than a structural gap.

---

📋 Shipped-vs-Claimed Gaps

None.

---

📝 Merge Recommendation

[APPROVE]

The PR correctly implements all four obligations. The fix is precise, fail-closed by construction, and the five-condition gate (WAL trust-boundary + generation + settlement metadata + current plan scope + workflow state) provides defence-in-depth against false-positive read-only completions.

Check	Result
No CRITICAL findings	✅
No unresolved STEALTH_CHANGE	✅
No UNSUPPORTED obligations	✅
Test coverage adequate	✅
No hardcoded secrets	✅
All async errors handled	✅
Input validation present	✅
No broken agent role boundaries	✅
Prompt format contracts intact	✅
Lockfile consistent	✅

---

🔁 Validation Provenance

Confirmed findings CHALLENGED and DROPPED:

Finding	File:Line	Drop reason
Idempotence asymmetry (undefined vs false)	task-terminal.ts:313	Both existingWal.readOnlyNoMutation === true and terminal.readOnlyNoMutation === true evaluate to false when the field is undefined or false. undefined !== true is false, which equals false !== true — idempotence holds. The WAL omit-when-false pattern is also intentional (optional boolean). FALSE POSITIVE.
Idempotence asymmetry (WAL write omit-when-false)	task-terminal.ts:386	Same root cause as above. The omit-when-false pattern is intentional and correct for an optional boolean field; it does not create a mismatch with the idempotence check. FALSE POSITIVE.
Empty declared scope skips scope validation	workflow-wal-schema.ts:277	When declaredFiles === [], the filter condition short-circuits and passes all candidate paths. This is correct by design: (a) scopedObservedFiles returns null for empty scopes in normal flow, so raw paths never reach this validator; (b) the recovery path intentionally preserves raw observed files (for audit) and the hasUnattributedEmptyScopeMutation guard in settleCoderDispatch already forces accepted=true + settlementFailed=true, routing the workflow to rework_required. The validation bypass is harmless here. FALSE POSITIVE.

Blind-spot findings added: None — all obligations are verified by the existing test matrix (104 focused tests + 39 compatibility tests + 117 gate-status assertions + 4 acceptance criteria fixtures).

---

_{🔒 Reviewed by a 3-model cross-family adversarial debate (architect → dual-lens parallel explorers → cross-critique → arbiter) for high recall with low false-positive noise. Findings are advisory — verify before acting.}