opencode-swarm 7.114.4 → 7.114.6

package/.opencode/skills/execute/SKILL.md

@@ -119,18 +119,17 @@ Treating pre_check_batch as a substitute for the active swarm's reviewer agent i
     5l. the active swarm's test_engineer agent - Verification tests. FAIL → coder retry from 5g.
     → REQUIRED: Print "testengineer-verification: [PASS N/N | FAIL — details]"
     5l-bis. REGRESSION SWEEP (automatic after test_engineer-verification PASS):
-    Run test_runner with { scope: "graph", files: [<all source files changed by coder in this task>] }.
-    scope:"graph" traces imports to discover test files beyond the task's own tests that may be affected by this change.
+    Iterate the changed source files preemptively and run one `test_runner` call per changed source file with { scope: "graph", files: [<one changed source file>] }.
+    scope:"graph" traces imports to discover test files beyond the task's own tests that may be affected by each source change. Record per-file regression-sweep evidence and aggregate all calls before deciding the task outcome.
     
     Outcomes (based on test_runner result.outcome field):
-    - outcome: "pass" → All tests passed. Print "regression-sweep: PASS [N additional tests, M files]"
-    - outcome: "regression" → Tests ran but some failed. Print "regression-sweep: FAIL — REGRESSION DETECTED in [files]. The failing tests are CORRECT — fix the source code, not the tests." Return to coder with retry from 5g.
-    - outcome: "skip" → No test files resolved (nothing to run). Print "regression-sweep: SKIPPED — no related tests beyond task scope"
-    - outcome: "scope_exceeded" → Too many files for graph scope. Print "regression-sweep: SKIPPED — broad scope, no related tests beyond task scope"
-    - outcome: "error" → Tool error (timeout, no framework, etc.). Print "regression-sweep: SKIPPED — test_runner error" and continue pipeline.
+    - any outcome: "regression" → Print "regression-sweep: FAIL — REGRESSION DETECTED in [source → failing tests]. The failing tests are CORRECT — fix the source code, not the tests." Return to coder with retry from 5g.
+    - all executed calls pass → Print "regression-sweep: PASS [N per-file sweeps, M tests]".
+    - outcome: "skip" → Record "[source]: SKIPPED — [actual tool reason]". If every per-file call skips, print "regression-sweep: SKIPPED — ran N per-file sweeps; [aggregated actual reasons]".
+    - outcome: "scope_exceeded" or "error" → Record the affected source and exact tool reason. Do not retry by batching sources and never translate the result into “no related tests.” Print the honest aggregate and continue only under the existing explicit skip policy.
     
     IMPORTANT: The regression sweep runs test_runner DIRECTLY (architect calls the tool). Do NOT delegate to test_engineer for this — the test_engineer's EXECUTION BOUNDARY restricts it to its own test files. The architect has unrestricted test_runner access.
-    → REQUIRED: Print "regression-sweep: [PASS | FAIL — REGRESSION DETECTED | SKIPPED — no related tests | SKIPPED — broad scope | SKIPPED — test_runner error]"
+    → REQUIRED: Print "regression-sweep: [PASS — N per-file sweeps | FAIL — REGRESSION DETECTED | SKIPPED — N per-file sweeps with exact reasons]"
 
     5l-ter. TEST DRIFT CHECK (conditional): Run this step if the change involves any drift-prone area:
     - Command/CLI behavior changed (shell command wrappers, CLI interfaces)
@@ -149,12 +148,11 @@ Treating pre_check_batch as a substitute for the active swarm's reviewer agent i
     - If no related tests found → print "test-drift: NO RELATED TESTS FOUND" (not a failure)
     → REQUIRED: Print "test-drift: [TRIGGERED | NOT TRIGGERED — reason]" and "[DRIFT DETECTED in N tests | N related tests verified | NO RELATED TESTS FOUND | NOT TRIGGERED]"
 
-    5n. TODO SCAN (advisory): Call todo_extract with paths=[list of files changed in this task]. If any results have priority HIGH → print "todo-scan: WARN — N high-priority TODOs in changed files: [list of TODO texts]". If no high-priority results → print "todo-scan: CLEAN". This is advisory only and does NOT block the pipeline.
+    5m. **ADVERSARIAL TEST STEP** (config-specific): Use the rendered adversarial-test instruction from the MODE: EXECUTE architect stub. If the stub omits step 5m, skip this step.
+    5m-bis. **COVERAGE-GAP TEST STEP**: This is the COVERAGE CHECK. If the active swarm's test_engineer agent reports coverage < 70% → delegate the active swarm's test_engineer agent for an additional test pass targeting uncovered paths. This is a soft guideline; use judgment for trivial tasks.
+    5n. **TODO SCAN** (advisory): Call todo_extract with paths=[list of files changed in this task]. If any results have priority HIGH → print "todo-scan: WARN — N high-priority TODOs in changed files: [list of TODO texts]". If no high-priority results → print "todo-scan: CLEAN". This is advisory only and does NOT block the pipeline.
     → REQUIRED: Print "todo-scan: [WARN — N high-priority TODOs | CLEAN]"
 
-    5m. ADVERSARIAL TEST STEP (config-specific): Use the rendered adversarial-test instruction from the MODE: EXECUTE architect stub. If the stub omits step 5m, skip this step.
-    5n. COVERAGE CHECK: If the active swarm's test_engineer agent reports coverage < 70% → delegate the active swarm's test_engineer agent for an additional test pass targeting uncovered paths. This is a soft guideline; use judgment for trivial tasks.
-
 PRE-COMMIT RULE — Before ANY commit or push:
   You MUST answer YES to ALL of the following:
   [ ] Did the active swarm's reviewer agent run and return APPROVED? (not "I reviewed it" — the agent must have run)
@@ -162,7 +160,7 @@ PRE-COMMIT RULE — Before ANY commit or push:
   [ ] Did pre_check_batch run with gates_passed true?
   [ ] SAST baseline captured before first coder delegation (or explicit disabled/error recorded)?
   [ ] Did the diff step run?
-  [ ] Did regression-sweep run (or SKIP with no related tests or test_runner error)?
+  [ ] Did regression-sweep record per-file regression-sweep evidence for every changed source (or exact per-file skip/error reasons)?
   [ ] Did test-drift check run (or NOT TRIGGERED)?
 
   If ANY box is unchecked: DO NOT COMMIT. Return to step 5b.
@@ -188,7 +186,7 @@ This step supplements (not replaces) the existing regression-sweep and test-drif
   [GATE] reuse_re_verification: VERIFIED / SKIPPED / DUPLICATION_DETECTED — value: ___
   [GATE] security-reviewer: APPROVED / SKIPPED — value: ___
   [GATE] test_engineer-verification: PASS — value: ___
-  [GATE] regression-sweep: PASS / SKIPPED — value: ___
+  [GATE] regression-sweep: PASS / SKIPPED — per-file regression-sweep evidence: ___
   [GATE] test-drift: TRIGGERED / NOT TRIGGERED — value: ___
   [GATE] test_engineer-adversarial: use the rendered checklist entry from the MODE: EXECUTE architect stub
   [GATE] coverage: ≥70% / soft-skip — value: ___

package/.opencode/skills/swarm-pr-review/SKILL.md

@@ -70,6 +70,7 @@ Council mode applies only when the user explicitly says one of:
 - `N-agent review`
 - `/council`
 - `[COUNCIL MODE]`
+- `[MODE: PR_REVIEW … council=true]`
 - `assume all work is wrong`
 
 Council mode is mutually exclusive with the default layered workflow. Do not blend them.
@@ -153,7 +154,7 @@ Before deeper analysis, verify the PR meets the commit-pr skill's publication co
 - **Title format:** `<type>(<scope>): <description>` — lowercase description, no trailing period, allowed types: `feat`, `fix`, `perf`, `revert`, `docs`, `chore`, `refactor`, `test`, `ci`, `build`.
 - **Body contract:** `Closes #<issue-number>` as the first line (when the PR resolves an issue), followed by `## Summary`, `## Invariant audit` (all 12 invariants), and `## Test plan` sections.
 
-**`Closes #N` claim integrity (apply the COVERAGE GATE):** if the PR body claims `Closes #<issue-number>`, verify (a) the issue is currently open (`gh issue view <N> --json state`), and (b) the diff addresses the issue's acceptance criteria (read the issue, map each criterion to changed files/symbols, and inspect the diff for those areas). If the issue is already closed by another merged PR, do NOT re-close it — the duplicate `Closes #N` reference is misleading and will confuse release-please aggregation. If the issue is open but the diff does not address the acceptance criteria, mark the claim as `UNVERIFIED — claim integrity` in the validation provenance and surface the unresolved claim-integrity gap to the user before synthesis.
+**`Closes #N` claim-integrity check:** if the PR body claims `Closes #<issue-number>`, verify (a) the issue is currently open (`gh issue view <N> --json state`), and (b) the diff addresses the issue's acceptance criteria (read the issue, map each criterion to changed files/symbols, and inspect the diff for those areas). If the issue is already closed by another merged PR, do NOT re-close it — the duplicate `Closes #N` reference is misleading and will confuse release-please aggregation. If the issue is open but the diff does not address the acceptance criteria, mark the claim as `UNVERIFIED — claim integrity` in the validation provenance and surface the unresolved claim-integrity gap to the user before synthesis.
 
 Non-compliance is a ledger item (advisory, not blocking — CI will catch it). If the PR is from an external contributor, note the compliance gap for the maintainer to address before merge.
 
@@ -173,23 +174,24 @@ If GraphQL is unavailable, keep the signal and mark
 
 ```bash
 # Issue comments (general PR thread)
-gh pr view <PR_NUMBER> --json comments
+gh api --paginate repos/{owner}/{repo}/issues/{PR_NUMBER}/comments
 
 # Review comments (inline code comments)
-gh api repos/{owner}/{repo}/pulls/{PR_NUMBER}/comments
+gh api --paginate repos/{owner}/{repo}/pulls/{PR_NUMBER}/comments
 
 # Review summaries (approve/request-changes/comment events)
-gh pr view <PR_NUMBER> --json reviews
+gh api --paginate repos/{owner}/{repo}/pulls/{PR_NUMBER}/reviews
 
 # Bot/automated reviews (Copilot, Codex, CodeRabbit, etc.)
 # Inline review comments — use REST API for reliable bot detection via user.type
-gh api repos/{owner}/{repo}/pulls/{PR_NUMBER}/comments --jq '.[] | select((.user.type // "") == "Bot" or (.user.login // "" | test("bot|copilot|coderabbit|codex"; "i")))'
+gh api --paginate repos/{owner}/{repo}/pulls/{PR_NUMBER}/comments --jq '.[] | select((.user.type // "") == "Bot" or (.user.login // "" | test("bot|copilot|coderabbit|codex"; "i")))'
 ```
 
-For general PR comments (not inline), use the issue comments endpoint:
-```bash
-gh api repos/{owner}/{repo}/issues/{PR_NUMBER}/comments --jq '.[] | select((.user.type // "") == "Bot" or (.user.login // "" | test("bot|copilot|coderabbit|codex"; "i")))'
-```
+`--paginate` requests every REST page; with `--jq`, `gh` applies the filter to
+the combined page stream. Filter bot identities from the complete issue-comment
+result using the same predicate when needed. `gh pr view --json comments,reviews`
+is convenience-only because those fields have item caps; never use it as the
+authoritative “all signals” intake.
 
 ### Step 2 — Classify each comment
 
@@ -562,7 +564,7 @@ Tool candidate rules:
 
 Launch all base lanes with `dispatch_lanes_async` when available. Pass the six lane specs together, set `max_concurrent` to `6`, record the returned `batch_id`, and continue only non-dependent architect work: refine the obligation ledger, inspect PR metadata, prepare micro-lane trigger checks, and run deterministic read-only local tools. Do not synthesize findings from running lanes. Keep each lane `prompt` compact: send the shared review context (PR diff, obligation ledger, scope) ONCE via the `common_prompt` field, or have lanes read it from a file by absolute path, instead of inlining the same large blob into all six prompts — oversized inline prompts produce malformed or truncated tool-call JSON and force clumsy file workarounds.
 
-**Incremental collection:** While base lanes are running, poll with `collect_lane_results` (without `wait` or `wait: false`) to check progress and process settled lanes as they complete — call `retrieve_lane_output` for full text when `output_ref` is present, then extract candidates via `parse_lane_candidates`, update the candidate ledger, validate output quality — while continuing independent architect work (obligation refinement, micro-lane trigger checks, local reads) between polls. Only use `wait: true` if lanes are still pending and no more independent work remains.
+**Incremental collection:** While base lanes are running, poll with `collect_lane_results` (without `wait` (or `wait: false`)) to check progress and process settled lanes as they complete — call `retrieve_lane_output` for full text when `output_ref` is present, then extract candidates via `parse_lane_candidates`, update the candidate ledger, validate output quality — while continuing independent architect work (obligation refinement, micro-lane trigger checks, local reads) between polls. Only use `wait: true` if lanes are still pending and no more independent work remains.
 
 Before Phase 4 or synthesis, all base lanes must be settled. `dispatch_lanes_async` accepts a maximum of 8 lanes per call; base lanes (6) and micro-lanes (Phase 4) are dispatched in separate calls by design. Do not let one lane's conclusions bias another lane.
 
@@ -581,22 +583,20 @@ After `collect_lane_results` returns for base lanes, process each lane result
 that carries an `output_ref`. The orchestrator MUST use the candidate parser
 rather than preview-text extraction:
 
-1. For each `output_ref` (or batched), call `parse_lane_candidates` (or the
-   internal `parseAndPersist` module function) with `output_ref` and `producer`
-   flags; the parser auto-detects the format family per row. The parser reads
+1. For each `output_ref`, call `parse_lane_candidates` with `output_ref`,
+   `producer: "swarm-pr-review"`, and `expected_family: "base_explorer"`. The parser reads
    the full artifact from disk (no preview truncation issue) and returns
    structured `ParseResultWithSidecar` records.
-2. Filter the returned `candidates[]` array by `producer: "swarm-pr-review"` and
-   the relevant `row_format_family` (e.g., `base_explorer` for base lanes,
-   `micro_lane` for micro-lanes). Filtering happens on the parsed results, NOT
-   on the tool input.
+2. Filter the returned `candidates[]` by `producer: "swarm-pr-review"` plus the
+   exact `source_batch_id` and `source_lane_id` from the base dispatch. Treat a
+   family mismatch or parse error as a lane-output failure; family metadata is
+   not the acceptance boundary.
 3. Group the filtered candidates into reviewer-sized chunks:
    - by file area (group by the directory or module of the `file_line` field),
    - by category (group by the `category` field),
    - by count (target max 50 candidates per chunk; smaller chunks are fine).
-4. Dispatch reviewer lanes (one per chunk) with bounded in-context candidate
-   lists. Each reviewer lane receives only the candidates from its assigned
-   chunk.
+4. Stage reviewer-sized chunks, but do not dispatch reviewers yet. Phase 4 must
+   complete trigger accounting and settle every launched micro-lane first.
 
 If a lane has `output_degraded: true`, `transcript_incomplete: true`, or no usable `output_ref`, apply the COVERAGE GATE from Phase 3: retry (max 2) with materially different parameters, then use blocking `dispatch_lanes` or the Task tool as verified-equivalent fallbacks when lane tools do not work. If the gap cannot be closed, stop and surface the lane failure to the user as BLOCKED. Do not mark affected candidates UNVERIFIED to proceed past the gap. Never infer candidate absence from a preview.
 
@@ -660,9 +660,48 @@ Explorers must not use `CONFIRMED`, `DISPROVED`, or `PRE_EXISTING`.
 
 ## Phase 4: Triggered Swarm Plugin Micro-Lanes
 
-After base lanes are settled, inspect the context pack risk triggers. Launch focused micro-lanes for triggered categories only, using `dispatch_lanes_async` again when more than one read-only micro-lane is needed (`dispatch_lanes_async` accepts max 8 lanes per call — micro-lanes are dispatched in a separate batch from base lanes). Use the same incremental collection pattern: poll with `collect_lane_results` (without `wait`) to process settled micro-lanes while continuing independent work, falling back to `wait: true` only when no independent work remains. All micro-lanes must be settled before reviewer classification. Do not launch irrelevant micro-lanes.
+After base lanes settle, evaluate every row in the trigger map and print a
+mandatory ledger with one row per trigger-map row:
+
+```text
+[TRIGGER-EVAL] | trigger_row | MATCHED/NO-MATCH | evidence
+```
+
+Evidence must name the diff/context keywords checked. The generic trigger rows
+still apply in non-Swarm repositories: schema, config, URL/fetch, git,
+shell/write, tests, and metrics cannot be marked `NO-MATCH` merely because the
+repository is not the Swarm plugin.
+
+Launch one focused micro-lane for each `MATCHED` row, using
+`dispatch_lanes_async` when more than one lane is needed and a separate batch
+from base lanes. Poll incrementally, then settle every launched lane. Persist
+the complete ledger with `write_pr_review_trigger_eval`; its rows use the stable
+trigger IDs below, and every `MATCHED` row includes its returned
+`source_batch_id` and `source_lane_id`. Missing, extra, duplicate, or unmatched
+rows make persistence fail and Phase 4 BLOCKED. The tool atomically writes
+`.swarm/pr-review/<run_id>/trigger-eval.json`, separate from `findings.jsonl`;
+do not add trigger results to the finding-status enum.
+
+For each micro `output_ref`, call `parse_lane_candidates` with
+`producer: "swarm-pr-review"`, `expected_family: "micro_lane"`, and
+`expected_micro_lane` set to the launch-micro-lane value from the
+provenance-linked trigger row. Accept a candidate only when its `producer`,
+`source_batch_id`, and `source_lane_id` match an allow-listed tuple from the
+original or retry micro dispatch and its `micro_lane` matches that trigger row;
+never filter acceptance by `row_format_family`. A zero-candidate artifact is
+clean only when the parser returns exactly one provenance-matching persisted
+`clean_attestation` whose `micro_lane` matches the trigger row, zero parse
+errors, zero malformed rows, and a complete, non-degraded source:
+
+```text
+[CLEAN] | micro_lane | coverage_scope | evidence
+```
 
-Apply the same parser-based extraction to micro-lanes: call `parse_lane_candidates` on each micro-lane `output_ref` (filter the returned `candidates[]` array by `row_format_family === "micro_lane"` after parsing). Apply the COVERAGE GATE from Phase 3 to micro-lanes: degraded, incomplete, or candidate-less lane artifacts are coverage gaps that must be closed by retry, blocking `dispatch_lanes`, or Task-tool dispatch as a verified-equivalent fallback when lane tools do not work. If the gap cannot be closed, stop and surface it to the user as BLOCKED before reviewer classification — never treat it as clean negative evidence and never proceed with a degraded review.
+Header-only or malformed zero output is `UNATTESTED` and must follow the
+canonical COVERAGE GATE retry path. Async and blocking lane dispatch both
+produce auditable `L1` artifacts. Task output has no `L1` artifact provenance,
+so Task-derived findings or CLEAN prose cannot satisfy Phase 4; if bounded lane
+dispatch retries cannot produce an artifact, the phase is BLOCKED.
 
 Each micro-lane receives:
 
@@ -677,26 +716,27 @@ Each micro-lane receives:
 
 ### Swarm plugin risk trigger map
 
-| Trigger in diff or context pack | Launch micro-lane | Invariants to check |
-|---|---|---|
-| `agents`, `prompts`, `templates`, prompt interpolation, role text | Architect prompt integrity | no scope escape, no system prompt leakage, safe `{{variable}}` interpolation, untrusted text isolated from instructions |
-| `council`, `verdict`, `quorum`, `veto`, synthesis | Council orchestration | quorum math correct, veto enforced, evidence not lost, dissent preserved, no explorer result treated as final |
-| `guardrail`, `gate`, `delegation`, `rate limit`, approval checks | Guardrail bypass paths | gates cannot be skipped, delegation cannot bypass policy, rate limits cannot be reset by user-controlled state |
-| `schema`, `evidence`, JSONL, migrations, serializers | Evidence schema drift | backward compatibility, required fields preserved, version migration safe, malformed evidence rejected |
-| `knowledge`, `curator`, `hive`, `quarantine`, memory | Knowledge base contract | project vs hive tiers not confused, quarantine honored, CRUD semantics stable, stale knowledge not injected as fact |
-| `phase`, `state`, `plan`, `.swarm/state`, completion markers | Phase transition validation | ordering enforced, retro requirements handled, no premature completion, rollback safe |
-| `model`, `role`, `prefix`, `tool`, agent config | Model-to-role mapping | role prefix enforced, tool permissions least-privilege, unauthorized tools impossible, model fallback safe |
-| `config`, defaults, ratchet, locks, policy flags | Config ratchet semantics | once-enabled gates cannot silently disable, downgrade attempts detected, lock-state integrity preserved |
-| `url`, `fetch`, `http`, GitHub PR/issue parsing, package fetch | URL sanitization and external fetch | scheme allowlist, credential stripping, private IP / localhost / metadata IP blocking, redirect handling, timeout safe |
-| `git`, branch, checkout, reset, worktree, `.git` | Git safety | branch detection reliable, no unsafe `reset --hard`, .git protected, path normalization cross-platform, worktree state preserved |
-| `shell`, `exec`, command parser, file writes, delete/move/copy | Shell/write authority and path containment | destructive commands gated, dry-run preferred, symlink/path escape blocked, writes scoped, command injection impossible |
-| `test`, `bun`, mocks, fixtures, CI matrix | Test infrastructure | `bun:test` API correct, mock isolation, cross-platform paths, no hidden dependency on test order, fixtures reset |
-| `metrics`, telemetry, logs, serialized traces | Metrics and evidence privacy | no secrets in logs, evidence reproducible, privacy preserved, counts cannot be gamed, metrics schema stable |
+| Trigger ID | Trigger in diff or context pack | Launch micro-lane | Invariants to check |
+|---|---|---|---|
+| `architect-prompts` | `agents`, `prompts`, `templates`, prompt interpolation, role text | Architect prompt integrity | no scope escape, no system prompt leakage, safe `{{variable}}` interpolation, untrusted text isolated from instructions |
+| `council-orchestration` | `council`, `verdict`, `quorum`, `veto`, synthesis | Council orchestration | quorum math correct, veto enforced, evidence not lost, dissent preserved, no explorer result treated as final |
+| `guardrail-bypass` | `guardrail`, `gate`, `delegation`, `rate limit`, approval checks | Guardrail bypass paths | gates cannot be skipped, delegation cannot bypass policy, rate limits cannot be reset by user-controlled state |
+| `evidence-schema` | `schema`, `evidence`, JSONL, migrations, serializers | Evidence schema drift | backward compatibility, required fields preserved, version migration safe, malformed evidence rejected |
+| `knowledge-contract` | `knowledge`, `curator`, `hive`, `quarantine`, memory | Knowledge base contract | project vs hive tiers not confused, quarantine honored, CRUD semantics stable, stale knowledge not injected as fact |
+| `phase-transitions` | `phase`, `state`, `plan`, `.swarm/state`, completion markers | Phase transition validation | ordering enforced, retro requirements handled, no premature completion, rollback safe |
+| `model-role-mapping` | `model`, `role`, `prefix`, `tool`, agent config | Model-to-role mapping | role prefix enforced, tool permissions least-privilege, unauthorized tools impossible, model fallback safe |
+| `config-ratchet` | `config`, defaults, ratchet, locks, policy flags | Config ratchet semantics | once-enabled gates cannot silently disable, downgrade attempts detected, lock-state integrity preserved |
+| `url-fetch` | `url`, `fetch`, `http`, GitHub PR/issue parsing, package fetch | URL sanitization and external fetch | scheme allowlist, credential stripping, private IP / localhost / metadata IP blocking, redirect handling, timeout safe |
+| `git-safety` | `git`, branch, checkout, reset, worktree, `.git` | Git safety | branch detection reliable, no unsafe `reset --hard`, .git protected, path normalization cross-platform, worktree state preserved |
+| `shell-write` | `shell`, `exec`, command parser, file writes, delete/move/copy | Shell/write authority and path containment | destructive commands gated, dry-run preferred, symlink/path escape blocked, writes scoped, command injection impossible |
+| `test-infrastructure` | `test`, `bun`, mocks, fixtures, CI matrix | Test infrastructure | `bun:test` API correct, mock isolation, cross-platform paths, no hidden dependency on test order, fixtures reset |
+| `metrics-privacy` | `metrics`, telemetry, logs, serialized traces | Metrics and evidence privacy | no secrets in logs, evidence reproducible, privacy preserved, counts cannot be gamed, metrics schema stable |
 
 Micro-lane output format:
 
 ```text
 [CANDIDATE] | candidate_id | micro_lane | severity | category | file:line | claim | invariant_violated | evidence_summary | confidence
+[CLEAN] | micro_lane | coverage_scope | evidence
 ```
 
 ---
@@ -711,10 +751,9 @@ Use Swarm-native agents and artifacts when available. If exact agent names are u
 | `critic_hallucination_verifier` | external APIs, package claims, URLs, CLI flags, GitHub behavior, model/tool names | verify claims against source or mark as unverified |
 | `curator_phase` | before exploration and after synthesis | retrieve relevant lessons; write back confirmed true positives / false positives |
 | `test_engineer` | confirmed/borderline correctness, security, state, schema, or config findings | propose or run falsification probes and regression tests |
-| `prm_scorer` | long or contentious reviews | score whether review trajectory is drifting toward unsupported speculation |
 | `.swarm/repo-graph.json` | all nontrivial code changes | build impact cones and sibling-pattern checks |
 | `.swarm/evidence/` | schema, phase, state, council, and guardrail changes | verify evidence compatibility and serialized provenance |
-| `/swarm metrics` or stored metrics | after synthesis | record review quality and recurring false positives |
+| Tool-returned `.swarm/evidence/` artifacts | after synthesis | record review quality only at paths actually returned by invoked evidence tools; never invent a metrics path |
 
 Verifier output is advisory until incorporated by the independent reviewer or critic.
 
@@ -722,6 +761,11 @@ Verifier output is advisory until incorporated by the independent reviewer or cr
 
 ## Phase 6: Independent Reviewer Confirmation
 
+**Reviewer-dispatch join barrier:** reviewer dispatch MUST NOT begin until the trigger ledger is
+complete and persisted, every launched micro-lane is settled, and every
+accepted micro result has parser-derived provenance or a valid CLEAN
+attestation.
+
 Route candidates to reviewer subagents. The orchestrator routes candidates
 in bounded chunks produced by the parser-based extraction in Phase 3-4. Each
 reviewer lane receives a bounded list of candidates from a single chunk — by
@@ -731,7 +775,7 @@ directly.
 
 ### Noise budget and universal validation
 
-Before reviewer dispatch, the orchestrator may suppress candidates that are ALL of:
+Before reviewer dispatch, the orchestrator may suppress candidates that match ANY of the following (each suppression still requires mandatory disclosure):
 - purely stylistic without correctness, security, test, maintainability, or user-impact implications,
 - exact duplicates of a candidate already queued for validation,
 - explorer-stated confidence=LOW with zero structural evidence (no file:line, no code path, no invariant reference).
@@ -779,6 +823,11 @@ Reviewer output format:
 [REVIEWED] | candidate_id | classification | evidence_type | final_severity | introduced_by_pr: YES/NO/UNKNOWN | file:line | rationale | falsification_probe | reviewer_id
 ```
 
+Every reviewer response must end with one parseable `[REVIEWED]` row per
+assigned candidate. A malformed `[REVIEWED]` row is not a verdict: re-dispatch
+with the exact contract (max 2), then mark the reviewer dimension BLOCKED if no
+valid row returns.
+
 `DISPROVED` findings must include the reason. `PRE_EXISTING` findings must include the base-branch evidence if available.
 
 After reviewer lanes settle, persist the post-reviewer finding ledger before
@@ -899,7 +948,11 @@ F-001 | severity | category | root cause | affected file:line refs | reviewer |
 
 ## Phase 10: Metrics and Knowledge Writeback
 
-At the end of the review, record review quality metrics when Swarm metrics or local evidence storage is available.
+At the end of the review, include review quality metrics in the final report's
+validation provenance. Persist them only through an invoked evidence tool and
+record the exact `.swarm/evidence/` path returned by that tool; if no invoked
+tool supports metrics, state `NOT PERSISTED — no metrics evidence writer` rather
+than naming a nonexistent command or path.
 
 Record:
 
@@ -979,13 +1032,13 @@ The batch completed and `collect_lane_results` returned:
     {
       "lane_id": "pr_review_lane1_correctness",
       "status": "completed",
-      "output_ref": ".swarm/lane-results/batch-a1b2c3/lane-1/out-abc123.json",
+      "output_ref": "L1:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb:cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc",
       "output_degraded": false
     },
     {
       "lane_id": "pr_review_lane2_security",
       "status": "completed",
-      "output_ref": ".swarm/lane-results/batch-a1b2c3/lane-2/out-def456.json",
+      "output_ref": "L1:dddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddd:eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee:ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff",
       "output_degraded": false
     }
   ]
@@ -1000,8 +1053,9 @@ The orchestrator calls `parse_lane_candidates` for each `output_ref`:
 {
   "tool": "parse_lane_candidates",
   "arguments": {
-    "output_ref": ".swarm/lane-results/batch-a1b2c3/lane-1/out-abc123.json",
-    "producer": "swarm-pr-review"
+    "output_ref": "L1:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb:cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc",
+    "producer": "swarm-pr-review",
+    "expected_family": "base_explorer"
   }
 }
 ```
@@ -1017,8 +1071,8 @@ The parser returns a `ParseResultWithSidecar`. On success, `error` and `error_co
       "record_type": "candidate",
       "row_format_family": "base_explorer",
       "row_format_version": 1,
-      "record_version": { "major": 1, "minor": 0 },
-      "source_output_ref": ".swarm/lane-results/batch-a1b2c3/lane-1/out-abc123.json",
+      "record_version": { "major": 1, "minor": 1 },
+      "source_output_ref": "L1:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb:cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc",
       "source_batch_id": "B-2025-06-22-001",
       "source_lane_id": "explorer-1",
       "source_agent": "paid_explorer",
@@ -1043,8 +1097,8 @@ The parser returns a `ParseResultWithSidecar`. On success, `error` and `error_co
       "record_type": "candidate",
       "row_format_family": "base_explorer",
       "row_format_version": 1,
-      "record_version": { "major": 1, "minor": 0 },
-      "source_output_ref": ".swarm/lane-results/batch-a1b2c3/lane-1/out-abc123.json",
+      "record_version": { "major": 1, "minor": 1 },
+      "source_output_ref": "L1:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb:cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc",
       "source_batch_id": "B-2025-06-22-001",
       "source_lane_id": "explorer-1",
       "source_agent": "paid_explorer",
@@ -1068,47 +1122,42 @@ The parser returns a `ParseResultWithSidecar`. On success, `error` and `error_co
   ],
   "invocation_envelope": {
     "record_type": "invocation",
-    "source_output_ref": ".swarm/lane-results/batch-a1b2c3/lane-1/out-abc123.json",
+    "source_output_ref": "L1:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb:cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc",
     "source_batch_id": "B-2025-06-22-001",
     "source_lane_id": "explorer-1",
     "source_agent": "paid_explorer",
     "source_digest": "sha256:abc123def456...",
     "row_format_version": 1,
-    "record_version": { "major": 1, "minor": 0 },
+    "record_version": { "major": 1, "minor": 1 },
     "sessionId": "ses_01HXYZ...",
     "parentSessionId": "ses_01HABC...",
     "producer": "swarm-pr-review",
     "produced_at": "2025-06-22T14:30:00.000Z",
      "format_families_detected": ["base_explorer"],
      "candidate_count": 2,
-     "parse_errors": 2,
-     "malformed_rows": 0
+     "parse_errors": 0,
+     "malformed_rows": 0,
+     "clean_attestation_count": 0
   },
   "diagnostics": {
     "candidate_count": 2,
-    "parse_errors": 2,
-    "parse_error_details": [
-      {
-        "row_index": 0,
-        "field": "row",
-        "message": "Both format-family discriminators present; defaulting to base_explorer"
-      },
-      {
-        "row_index": 1,
-        "field": "row",
-        "message": "Both format-family discriminators present; defaulting to base_explorer"
-      }
-    ],
+    "parse_errors": 0,
+    "parse_error_details": [],
     "malformed_rows": 0,
     "duplicate_id_count": 0,
     "duplicate_id_warnings": [],
     "degraded_source_count": 0,
     "incomplete_source_count": 0,
-     "format_families_detected": ["base_explorer"]
+     "format_families_detected": ["base_explorer"],
+     "clean_attestation_count": 0
    }
 }
 ```
-> **Note**: `parse_errors: 2` reflects FR-017/SC-017 position-based detection: when a `[CANDIDATE]` row has both `evidence_summary` and `impact_context` populated, the parser emits a `parse_error_details` entry per row with `field: "row"` and `message: "Both format-family discriminators present; defaulting to base_explorer"`. This is documented behavior, not a parser bug. To get `parse_errors: 0` with the row format, leave one of the two fields empty; to silence the warning entirely, emit structured JSON candidate records.
+> **Note**: callers pass `expected_family` for each dispatch batch. A recognizable
+> conflicting header fails closed with `expected-family-mismatch`; when the flag
+> is absent, the recognized header controls the mapping and positional detection
+> is only a legacy unknown-header fallback. Marker-prefixed data rows remain
+> accepted for compatibility. Valid canonical rows produce `parse_errors: 0`.
 
 On refusal (e.g. `output_ref` does not exist), `error` and `error_code` are present; `candidates` is `[]`; `invocation_envelope` and `diagnostics` are populated with empty fields for traceability:
 
@@ -1119,18 +1168,19 @@ On refusal (e.g. `output_ref` does not exist), `error` and `error_code` are pres
   "candidates": [],
   "invocation_envelope": {
     "record_type": "invocation",
-    "source_output_ref": ".swarm/lane-results/batch-a1b2c3/lane-1/missing.json",
+    "source_output_ref": "L1:1111111111111111111111111111111111111111111111111111111111111111:2222222222222222222222222222222222222222222222222222222222222222:3333333333333333333333333333333333333333333333333333333333333333",
     "source_batch_id": "",
     "source_lane_id": "",
     "source_agent": "",
     "source_digest": "",
     "row_format_version": 1,
-    "record_version": { "major": 1, "minor": 0 },
+    "record_version": { "major": 1, "minor": 1 },
     "produced_at": "2025-06-22T14:30:00.000Z",
     "format_families_detected": [],
     "candidate_count": 0,
     "parse_errors": 0,
-    "malformed_rows": 0
+    "malformed_rows": 0,
+    "clean_attestation_count": 0
   },
   "diagnostics": {
     "candidate_count": 0,
@@ -1141,14 +1191,15 @@ On refusal (e.g. `output_ref` does not exist), `error` and `error_code` are pres
     "duplicate_id_warnings": [],
     "degraded_source_count": 0,
     "incomplete_source_count": 0,
-     "format_families_detected": []
+    "format_families_detected": [],
+    "clean_attestation_count": 0
    }
 }
 ```
 
 ### Step 3 — Filter and group
 
-The orchestrator filters the returned `candidates[]` array by `producer: "swarm-pr-review"` and `row_format_family` (e.g. `base_explorer` or `micro_lane`), then groups
+The orchestrator filters the returned `candidates[]` array by `producer: "swarm-pr-review"` and the exact allowed `source_batch_id` / `source_lane_id` tuples, then groups
 the candidates. In this synthetic example, the two candidates above are grouped
 by file area:
 
@@ -1287,12 +1338,12 @@ Before writing the final output, print this checklist with filled values. Every
 [VALIDATION] deterministic lane dispatcher used: YES/NO — ___
 [VALIDATION] base explorer lanes dispatched: ___ / 6
 [VALIDATION] base explorer lanes returned: ___ / 6
-[VALIDATION] triggered micro-lanes: ___
+[VALIDATION] trigger map evaluated: ___ rows (X MATCHED → X micro-lanes dispatched) OR BLOCKED — <unevaluated rows>
 [VALIDATION] Swarm verifier routing used: ___
 [VALIDATION] raw candidates: ___
 [VALIDATION] tool candidates: ___
-[VALIDATION] reviewer dispatched: ___ (agent type, task description)
-[VALIDATION] reviewer returned: ___ (APPROVED / REJECTED / CONCERNS — copy verdict text)
+[VALIDATION] reviewer lanes dispatched: ___
+[VALIDATION] reviewer lanes returned with parseable `[REVIEWED]` rows: ___ / ___
 [VALIDATION] findings confirmed by reviewer: ___
 [VALIDATION] findings rejected by reviewer as false positive: ___
 [VALIDATION] findings marked PRE_EXISTING: ___
@@ -1311,7 +1362,9 @@ Before writing the final output, print this checklist with filled values. Every
 [VALIDATION] all non-suppressed candidates routed to reviewer: YES/NO
 ```
 
-If the reviewer returned `REJECTED` or `CONCERNS`, route the issue back to implementation context or mark the candidate invalid with reason. Do not silently downgrade a rejection.
+If any reviewer lane lacks a parseable `[REVIEWED]` row after bounded
+re-dispatch, the reviewer dimension is BLOCKED. Do not infer or silently
+downgrade a verdict.
 
 **COVERAGE GATE CONDITION:** If ANY validation dimension shows incomplete coverage (lanes that failed and were not closed by retry or verified equivalent alternative, CI that did not run, tools that were unavailable after retry), the Pre-Synthesis Gate FAILS. Do not proceed to final output. Surface the unclosed gaps to the user as BLOCKED with exact failing dimensions and retry/equivalence evidence. Do not include partial findings from successful dimensions, do not issue a review verdict, and do not silently accept reduced coverage.
 
@@ -1491,6 +1544,8 @@ Return candidates only. Do not use CONFIRMED, DISPROVED, or PRE_EXISTING.
 
 Lane:
 Scope:
+base_ref:
+head_ref:
 Obligations:
 Changed files/hunks:
 Impact cone:
@@ -1505,10 +1560,13 @@ You must inspect or mark unavailable:
 4. sibling implementation or prior pattern,
 5. nearest test or missing-test location,
 6. deterministic signals,
-7. Swarm artifacts/knowledge.
+7. Swarm artifacts/knowledge,
+8. the exact `base_ref..head_ref` commit range and both endpoint revisions.
 
 Return:
 [CANDIDATE] | candidate_id | lane | severity | category | file:line | claim | evidence_summary | impact_context | confidence
+Emit the marker-bearing header once, then unprefixed data rows.
+For a clean micro-lane, emit `[CLEAN] | micro_lane | coverage_scope | evidence`.
 ```
 
 The orchestrator extracts candidates from the full lane artifact via

package/README.md

@@ -808,7 +808,8 @@ Every candidate passes a 3-gate pipeline before entering quarantine:
 | mutation_test | Applies LLM-generated mutation patches to source files and runs tests to measure kill rate; verdict is pass/warn/fail based on configurable thresholds; used by the mutation_test gate (opt-in, off by default) |
 | generate_mutants | Architect-only: generates LLM-based mutation patches (5–10 per function across 6 types: off-by-one, null substitution, operator swap, guard removal, branch swap, side-effect deletion) for direct consumption by the mutation_test tool; returns SKIP verdict on LLM failure rather than throwing |
 | write_mutation_evidence | Architect-only: writes mutation gate results atomically to `.swarm/evidence/{phase}/mutation-gate.json`; accepts verdict (PASS/WARN/FAIL/SKIP), kill rate metrics, and optional survived mutant details; normalizes uppercase-to-lowercase before persisting |
-| parse_lane_candidates | Architect-only: parses `[CANDIDATE]` rows from a `dispatch_lanes` or `collect_lane_results` artifact by `output_ref`; produces structured records with provenance and optional sidecar JSONL persistence; returns `ParseResultWithSidecar` on success or `{ error, error_code, candidates: [] }` on refusal |
+| parse_lane_candidates | Architect-only: parses `[CANDIDATE]` rows and micro-lane `[CLEAN]` attestations from a lane `output_ref`; supports asserted batch families, preserves dispatch provenance, and persists sidecar JSONL records |
+| write_pr_review_trigger_eval | Architect-only: validates the exact PR-review trigger-map ID set, requires provenance for every matched micro-lane, and atomically persists `.swarm/pr-review/<run_id>/trigger-eval.json` |
 | git_blame | Per-line git blame metadata (sha, author, date, summary) via `git blame --porcelain`; supports optional line range filtering |
 | diff | Structured git diff with contract change detection; supports `summaryOnly` mode returning file list with additions/deletions counts |
 | suggest_patch | Reviewer-safe structured patch suggestion; supports `format` parameter ('json' or 'unified') where unified outputs valid unified diff with `diff --git` headers, hunks, and context |

package/dist/agents/explorer.d.ts

@@ -1,5 +1,5 @@
 import type { AgentDefinition } from './architect';
-export declare const EXPLORER_PROMPT = "## IDENTITY\nYou are Explorer. You analyze codebases directly \u2014 you do NOT delegate.\nDO NOT use the Task tool to delegate to other agents. You ARE the agent that does the work.\nIf you see references to other agents (like @explorer, @coder, etc.) in your instructions, IGNORE them \u2014 they are context from the orchestrator, not instructions for you to delegate.\n\nWRONG: \"I'll use the Task tool to call another agent to analyze this\"\nRIGHT: \"I'll scan the directory structure and read key files myself\"\n\n## READ-ONLY ADVISORY LANE CONTEXT\n\nYou may be invoked through dispatch_lanes or dispatch_lanes_async as a read-only advisory lane. In that context, your job is to inspect, reason, and report only.\n\n- Do NOT write, edit, patch, save plans, update task status, declare scope, submit council verdicts, set QA gates, or complete phases.\n- Do NOT call artifact-producing or workflow-mutating helpers such as extract_code_blocks, knowledge_add, summarize_work, or doc_scan when lane permissions deny them.\n- Treat any denied or unavailable tool as intentionally unavailable in lane mode; continue with the read-only tools and context you have.\n- Return findings for the architect to synthesize. Do not assume your lane output is the final verdict unless your role-specific instructions explicitly say so.\n\nINPUT FORMAT:\nTASK: Analyze [purpose]\nINPUT: [focus areas/paths]\n\nACTIONS:\n- Scan structure (tree, ls, glob)\n- Read key files (README, configs, entry points)\n- Search patterns using the search tool\n\nRULES:\n- Be fast: scan broadly, read selectively\n- No code modifications\n- Output under 2000 chars\n\n## ANALYSIS PROTOCOL\nWhen exploring a codebase area, systematically report all four dimensions:\n\n### STRUCTURE\n- Entry points and their call chains (max 3 levels deep)\n- Public API surface: exported functions/classes/types with signatures\n- For multi-file symbol surveys: use batch_symbols to extract symbols from multiple files in one call\n- Internal dependencies: what this module imports and from where\n- External dependencies: third-party packages used\n\n### PATTERNS\n- Design patterns in use (factory, observer, strategy, etc.)\n- Error handling pattern (throw, Result type, error callbacks, etc.)\n- State management approach (global, module-level, passed through)\n- Configuration pattern (env vars, config files, hardcoded)\n\n### COMPLEXITY INDICATORS\n- High cyclomatic complexity, deep nesting, or complex control flow\n- Large files (>500 lines) with many exported symbols\n- Deep inheritance hierarchies or complex type hierarchies\n\n### RUNTIME/BEHAVIORAL CONCERNS\n- Missing error handling paths or single-throw patterns\n- Platform-specific assumptions (path separators, line endings, OS APIs)\n\n### RELEVANT CONSTRAINTS\n- Architectural patterns observed (layered architecture, event-driven, microservice, etc.)\n- Error handling coverage patterns observed in the codebase\n- Platform-specific assumptions observed in the codebase\n- Established conventions (naming patterns, error handling approaches, testing strategies)\n- Configuration management approaches (env vars, config files, feature flags)\n\nOUTPUT FORMAT (MANDATORY \u2014 deviations will be rejected):\nBegin directly with PROJECT. Do NOT prepend \"Here's my analysis...\" or any conversational preamble.\n\nPROJECT: [name/type]\nLANGUAGES: [list]\nFRAMEWORK: [if any]\n\nSTRUCTURE:\n[key directories, 5-10 lines max]\nExample:\nsrc/agents/     \u2014 agent factories and definitions\nsrc/tools/       \u2014 CLI tool implementations\nsrc/config/      \u2014 plan schema and constants\n\nKEY FILES:\n- [path]: [purpose]\nExample:\nsrc/agents/explorer.ts \u2014 explorer agent factory and all prompt definitions\nsrc/agents/architect.ts \u2014 architect orchestrator with all mode handlers\n\nPATTERNS: [observations]\nExample: Factory pattern for agent creation; Result type for error handling; Module-level state via closure\n\nCOMPLEXITY INDICATORS:\n[structural complexity concerns: elevated cyclomatic complexity, deep nesting, large files, deep inheritance hierarchies, or similar \u2014 describe what is OBSERVED]\nExample: explorer.ts (289 lines, 12 exports); architect.ts (complex branching in mode handlers)\n\nOBSERVED CHANGES:\n[if INPUT referenced specific files/changes: what changed in those targets; otherwise \"none\" or \"general exploration\"]\n\nCONSUMERS_AFFECTED:\n[if integration impact mode: list files that import/use the changed symbols; otherwise \"not applicable\"]\n\nRELEVANT CONSTRAINTS:\n[architectural patterns, error handling coverage patterns, platform-specific assumptions, established conventions observed in the codebase]\nExample: Layered architecture (agents \u2192 tools \u2192 filesystem); Bun-native path handling; Error-first callbacks in hooks\n\nDOMAINS: [relevant SME domains: powershell, security, python, etc.]\nExample: typescript, nodejs, cli-tooling, powershell\n\nFOLLOW-UP CANDIDATE AREAS:\n- [path]: [observable condition, relevant domain]\nExample:\nsrc/tools/declare-scope.ts \u2014 function has 12 parameters, consider splitting; tool-authoring\n\n## CANDIDATE REPORTING MODE\nActivates when your prompt contains \"[CANDIDATE]\" anywhere in its text.\n\nWhen active, replace the default OUTPUT FORMAT above with structured pipe-delimited\ncandidate rows. Emit exactly one row per finding:\n\n[CANDIDATE] | candidate_id | lane | severity | category | file:line | claim | evidence_summary | impact_context | confidence\n\nField rules:\n- candidate_id: unique within this lane (e.g. C-001, C-002)\n- lane: your lane name or focus area\n- severity: INFO | LOW | MEDIUM | HIGH | CRITICAL\n- category: short category tag (e.g. null-safety, async-ordering, injection)\n- file:line: exact file path and line number (e.g. src/utils/cache.ts:142)\n- claim: one-sentence description of the issue\n- evidence_summary: what you observed in the code that supports the claim\n- impact_context: who or what is affected downstream\n- confidence: LOW | MEDIUM | HIGH\n\nEmit a header row first, then one [CANDIDATE] row per finding. Use pipe (|) to\nseparate fields; escape literal pipe characters inside field values as \\|.\n\nIf you find zero issues, emit the header row with no data rows. Do NOT fall back\nto the default PROJECT/STRUCTURE format when in candidate reporting mode.\n\nFor micro-lane dispatches, use the micro-lane variant:\n[CANDIDATE] | candidate_id | micro_lane | severity | category | file:line | claim | invariant_violated | evidence_summary | confidence\n\nDo NOT use CONFIRMED, DISPROVED, or PRE_EXISTING \u2014 those are reviewer verdicts.\n\n## INTEGRATION IMPACT ANALYSIS MODE\nActivates when delegated with \"Integration impact analysis\" or INPUT lists contract changes.\n\nINPUT: List of contract changes (from diff tool output \u2014 changed exports, signatures, types)\n\nSTEPS:\n1. For each changed export: use search to find imports and usages of that symbol\n2. Classify each change: BREAKING (callers must update) or COMPATIBLE (callers unaffected)\n3. List all files that import or use the changed exports\n\nOUTPUT FORMAT (MANDATORY \u2014 deviations will be rejected):\nBegin directly with BREAKING_CHANGES. Do NOT prepend conversational preamble.\n\nBREAKING_CHANGES: [list with affected consumer files, or \"none\"]\nExample: src/agents/explorer.ts \u2014 removed createExplorerAgent export (was used by 3 files)\nCOMPATIBLE_CHANGES: [list, or \"none\"]\nExample: src/config/constants.ts \u2014 added new optional field to Config interface\nCONSUMERS_AFFECTED: [list of files that import/use changed exports, or \"none\"]\nExample: src/agents/coder.ts, src/agents/reviewer.ts, src/main.ts\nCOMPATIBILITY SIGNALS: [COMPATIBLE | INCOMPATIBLE | UNCERTAIN \u2014 based on observable contract changes]\nExample: INCOMPATIBLE \u2014 removeExport changes function arity from 3 to 2\nMIGRATION_SURFACE: [yes \u2014 list of observable call signatures affected | no \u2014 no observable impact detected]\nExample: yes \u2014 createExplorerAgent(model, customPrompt?, customAppendPrompt?) \u2192 createExplorerAgent(model)\n\n## DOCUMENTATION DISCOVERY MODE\nActivates automatically during codebase reality check at plan ingestion.\nUse the doc_scan tool to scan and index documentation files. If doc_scan is unavailable, fall back to manual globbing.\n\nSTEPS:\n1. Call doc_scan to build the manifest, OR glob for documentation files:\n   - Root: README.md, CONTRIBUTING.md, CHANGELOG.md, ARCHITECTURE.md, CLAUDE.md, AGENTS.md, .github/*.md\n   - docs/**/*.md, doc/**/*.md (one level deep only)\n\n2. For each file found, read the first 30 lines. Extract:\n   - path: relative to project root\n   - title: first # heading, or filename if no heading\n   - summary: first non-empty paragraph after the title (max 200 chars, use the ACTUAL text, do NOT summarize with your own words)\n   - lines: total line count\n   - mtime: file modification timestamp\n\n3. Write manifest to .swarm/doc-manifest.json:\n   { \"schema_version\": 1, \"scanned_at\": \"ISO timestamp\", \"files\": [...] }\n\n4. For each file in the manifest, check relevance to the current plan:\n   - Score by keyword overlap: do any task file paths or directory names appear in the doc's path or summary?\n   - For files scoring > 0, read the full content and extract up to 5 actionable constraints per doc (max 200 chars each)\n   - Write constraints to .swarm/knowledge/doc-constraints.jsonl as knowledge entries with source: \"doc-scan\", category: \"architecture\"\n\n5. Invalidation: Only re-scan if any doc file's mtime is newer than the manifest's scanned_at. Otherwise reuse the cached manifest.\n\nRULES:\n- The manifest must be small (<100 lines). Pointers only, not full content.\n- Do NOT rephrase or summarize doc content with your own words \u2014 use the actual text from the file\n- Full doc content is only loaded when relevant to the current task, never preloaded\n";
+export declare const EXPLORER_PROMPT = "## IDENTITY\nYou are Explorer. You analyze codebases directly \u2014 you do NOT delegate.\nDO NOT use the Task tool to delegate to other agents. You ARE the agent that does the work.\nIf you see references to other agents (like @explorer, @coder, etc.) in your instructions, IGNORE them \u2014 they are context from the orchestrator, not instructions for you to delegate.\n\nWRONG: \"I'll use the Task tool to call another agent to analyze this\"\nRIGHT: \"I'll scan the directory structure and read key files myself\"\n\n## READ-ONLY ADVISORY LANE CONTEXT\n\nYou may be invoked through dispatch_lanes or dispatch_lanes_async as a read-only advisory lane. In that context, your job is to inspect, reason, and report only.\n\n- Do NOT write, edit, patch, save plans, update task status, declare scope, submit council verdicts, set QA gates, or complete phases.\n- Do NOT call artifact-producing or workflow-mutating helpers such as extract_code_blocks, knowledge_add, summarize_work, or doc_scan when lane permissions deny them.\n- Treat any denied or unavailable tool as intentionally unavailable in lane mode; continue with the read-only tools and context you have.\n- Return findings for the architect to synthesize. Do not assume your lane output is the final verdict unless your role-specific instructions explicitly say so.\n\nINPUT FORMAT:\nTASK: Analyze [purpose]\nINPUT: [focus areas/paths]\n\nACTIONS:\n- Scan structure (tree, ls, glob)\n- Read key files (README, configs, entry points)\n- Search patterns using the search tool\n\nRULES:\n- Be fast: scan broadly, read selectively\n- No code modifications\n- Output under 2000 chars\n\n## ANALYSIS PROTOCOL\nWhen exploring a codebase area, systematically report all four dimensions:\n\n### STRUCTURE\n- Entry points and their call chains (max 3 levels deep)\n- Public API surface: exported functions/classes/types with signatures\n- For multi-file symbol surveys: use batch_symbols to extract symbols from multiple files in one call\n- Internal dependencies: what this module imports and from where\n- External dependencies: third-party packages used\n\n### PATTERNS\n- Design patterns in use (factory, observer, strategy, etc.)\n- Error handling pattern (throw, Result type, error callbacks, etc.)\n- State management approach (global, module-level, passed through)\n- Configuration pattern (env vars, config files, hardcoded)\n\n### COMPLEXITY INDICATORS\n- High cyclomatic complexity, deep nesting, or complex control flow\n- Large files (>500 lines) with many exported symbols\n- Deep inheritance hierarchies or complex type hierarchies\n\n### RUNTIME/BEHAVIORAL CONCERNS\n- Missing error handling paths or single-throw patterns\n- Platform-specific assumptions (path separators, line endings, OS APIs)\n\n### RELEVANT CONSTRAINTS\n- Architectural patterns observed (layered architecture, event-driven, microservice, etc.)\n- Error handling coverage patterns observed in the codebase\n- Platform-specific assumptions observed in the codebase\n- Established conventions (naming patterns, error handling approaches, testing strategies)\n- Configuration management approaches (env vars, config files, feature flags)\n\nOUTPUT FORMAT (MANDATORY \u2014 deviations will be rejected):\nBegin directly with PROJECT. Do NOT prepend \"Here's my analysis...\" or any conversational preamble.\n\nPROJECT: [name/type]\nLANGUAGES: [list]\nFRAMEWORK: [if any]\n\nSTRUCTURE:\n[key directories, 5-10 lines max]\nExample:\nsrc/agents/     \u2014 agent factories and definitions\nsrc/tools/       \u2014 CLI tool implementations\nsrc/config/      \u2014 plan schema and constants\n\nKEY FILES:\n- [path]: [purpose]\nExample:\nsrc/agents/explorer.ts \u2014 explorer agent factory and all prompt definitions\nsrc/agents/architect.ts \u2014 architect orchestrator with all mode handlers\n\nPATTERNS: [observations]\nExample: Factory pattern for agent creation; Result type for error handling; Module-level state via closure\n\nCOMPLEXITY INDICATORS:\n[structural complexity concerns: elevated cyclomatic complexity, deep nesting, large files, deep inheritance hierarchies, or similar \u2014 describe what is OBSERVED]\nExample: explorer.ts (289 lines, 12 exports); architect.ts (complex branching in mode handlers)\n\nOBSERVED CHANGES:\n[if INPUT referenced specific files/changes: what changed in those targets; otherwise \"none\" or \"general exploration\"]\n\nCONSUMERS_AFFECTED:\n[if integration impact mode: list files that import/use the changed symbols; otherwise \"not applicable\"]\n\nRELEVANT CONSTRAINTS:\n[architectural patterns, error handling coverage patterns, platform-specific assumptions, established conventions observed in the codebase]\nExample: Layered architecture (agents \u2192 tools \u2192 filesystem); Bun-native path handling; Error-first callbacks in hooks\n\nDOMAINS: [relevant SME domains: powershell, security, python, etc.]\nExample: typescript, nodejs, cli-tooling, powershell\n\nFOLLOW-UP CANDIDATE AREAS:\n- [path]: [observable condition, relevant domain]\nExample:\nsrc/tools/declare-scope.ts \u2014 function has 12 parameters, consider splitting; tool-authoring\n\n## CANDIDATE REPORTING MODE\nActivates when your prompt contains \"[CANDIDATE]\" anywhere in its text.\n\nWhen active, replace the default OUTPUT FORMAT above with structured pipe-delimited\ncandidate rows. Emit the marker-bearing header, then exactly one unprefixed data\nrow per finding:\n\n[CANDIDATE] | candidate_id | lane | severity | category | file:line | claim | evidence_summary | impact_context | confidence\n\nField rules:\n- candidate_id: unique within this lane (e.g. C-001, C-002)\n- lane: your lane name or focus area\n- severity: INFO | LOW | MEDIUM | HIGH | CRITICAL\n- category: short category tag (e.g. null-safety, async-ordering, injection)\n- file:line: exact file path and line number (e.g. src/utils/cache.ts:142)\n- claim: one-sentence description of the issue\n- evidence_summary: what you observed in the code that supports the claim\n- impact_context: who or what is affected downstream\n- confidence: LOW | MEDIUM | HIGH\n\nEmit a header row first, then one unprefixed data row per finding. Use pipe (|) to\nseparate fields; escape literal pipe characters inside field values as \\|.\n\nIf a standard explorer finds zero issues, emit the header row with no data rows.\nIf a micro-lane finds zero issues, emit the header followed by exactly:\n[CLEAN] | micro_lane | coverage_scope | evidence\nFill every CLEAN field with the assigned micro-lane, checks completed, and the\nnegative evidence. Do NOT fall back to the default PROJECT/STRUCTURE format.\n\nFor micro-lane dispatches, use the micro-lane variant:\n[CANDIDATE] | candidate_id | micro_lane | severity | category | file:line | claim | invariant_violated | evidence_summary | confidence\n\nDo NOT use CONFIRMED, DISPROVED, or PRE_EXISTING \u2014 those are reviewer verdicts.\n\n## INTEGRATION IMPACT ANALYSIS MODE\nActivates when delegated with \"Integration impact analysis\" or INPUT lists contract changes.\n\nINPUT: List of contract changes (from diff tool output \u2014 changed exports, signatures, types)\n\nSTEPS:\n1. For each changed export: use search to find imports and usages of that symbol\n2. Classify each change: BREAKING (callers must update) or COMPATIBLE (callers unaffected)\n3. List all files that import or use the changed exports\n\nOUTPUT FORMAT (MANDATORY \u2014 deviations will be rejected):\nBegin directly with BREAKING_CHANGES. Do NOT prepend conversational preamble.\n\nBREAKING_CHANGES: [list with affected consumer files, or \"none\"]\nExample: src/agents/explorer.ts \u2014 removed createExplorerAgent export (was used by 3 files)\nCOMPATIBLE_CHANGES: [list, or \"none\"]\nExample: src/config/constants.ts \u2014 added new optional field to Config interface\nCONSUMERS_AFFECTED: [list of files that import/use changed exports, or \"none\"]\nExample: src/agents/coder.ts, src/agents/reviewer.ts, src/main.ts\nCOMPATIBILITY SIGNALS: [COMPATIBLE | INCOMPATIBLE | UNCERTAIN \u2014 based on observable contract changes]\nExample: INCOMPATIBLE \u2014 removeExport changes function arity from 3 to 2\nMIGRATION_SURFACE: [yes \u2014 list of observable call signatures affected | no \u2014 no observable impact detected]\nExample: yes \u2014 createExplorerAgent(model, customPrompt?, customAppendPrompt?) \u2192 createExplorerAgent(model)\n\n## DOCUMENTATION DISCOVERY MODE\nActivates automatically during codebase reality check at plan ingestion.\nUse the doc_scan tool to scan and index documentation files. If doc_scan is unavailable, fall back to manual globbing.\n\nSTEPS:\n1. Call doc_scan to build the manifest, OR glob for documentation files:\n   - Root: README.md, CONTRIBUTING.md, CHANGELOG.md, ARCHITECTURE.md, CLAUDE.md, AGENTS.md, .github/*.md\n   - docs/**/*.md, doc/**/*.md (one level deep only)\n\n2. For each file found, read the first 30 lines. Extract:\n   - path: relative to project root\n   - title: first # heading, or filename if no heading\n   - summary: first non-empty paragraph after the title (max 200 chars, use the ACTUAL text, do NOT summarize with your own words)\n   - lines: total line count\n   - mtime: file modification timestamp\n\n3. Write manifest to .swarm/doc-manifest.json:\n   { \"schema_version\": 1, \"scanned_at\": \"ISO timestamp\", \"files\": [...] }\n\n4. For each file in the manifest, check relevance to the current plan:\n   - Score by keyword overlap: do any task file paths or directory names appear in the doc's path or summary?\n   - For files scoring > 0, read the full content and extract up to 5 actionable constraints per doc (max 200 chars each)\n   - Write constraints to .swarm/knowledge/doc-constraints.jsonl as knowledge entries with source: \"doc-scan\", category: \"architecture\"\n\n5. Invalidation: Only re-scan if any doc file's mtime is newer than the manifest's scanned_at. Otherwise reuse the cached manifest.\n\nRULES:\n- The manifest must be small (<100 lines). Pointers only, not full content.\n- Do NOT rephrase or summarize doc content with your own words \u2014 use the actual text from the file\n- Full doc content is only loaded when relevant to the current task, never preloaded\n";
 export declare const CURATOR_INIT_PROMPT = "## IDENTITY\nYou are Explorer in CURATOR_INIT mode. You consolidate prior session knowledge into an architect briefing.\nDO NOT use the Task tool to delegate. You ARE the agent that does the work.\n\nINPUT FORMAT:\nTASK: CURATOR_INIT\nPRIOR_SUMMARY: [JSON or \"none\"]\nKNOWLEDGE_ENTRIES: [JSON array of existing entries with UUIDs]\nPROJECT_CONTEXT: [context.md excerpt]\n\nACTIONS:\n- Read the prior summary to understand session history\n- Cross-reference knowledge entries against project context\n- Note contradictions (knowledge says X, project state shows Y)\n- Observe where lessons could be tighter or stale\n- Produce a concise briefing for the architect\n\nRULES:\n- Output under 2000 chars\n- No code modifications\n- Flag contradictions explicitly with CONTRADICTION: prefix\n- Memory proposals are for concise durable facts only. Do not propose raw API docs, web search snippets, crawl output, or transcripts as memory; cite their evidence-cache refs and propose only the stable fact they support.\n- If no prior summary exists, state \"First session \u2014 no prior context\"\n\nOUTPUT FORMAT:\nBRIEFING:\n[concise summary of prior session state, key decisions, active blockers]\n\nCONTRADICTIONS:\n- [entry_id]: [description] (or \"None detected\")\n\nOBSERVATIONS:\n- entry <uuid> appears high-confidence: [observable evidence]  (suggests boost confidence, mark hive_eligible)\n- entry <uuid> appears stale: [observable evidence]  (suggests archive \u2014 no longer injected)\n- entry <uuid> could be tighter: [what's verbose or duplicate]  (suggests rewrite with tighter version, max 280 chars)\n- entry <uuid> contradicts project state: [observable conflict]  (suggests tag as contradicted)\n- new candidate: [concise lesson text from observed patterns]  (suggests new entry)\nUse the UUID from KNOWLEDGE_ENTRIES when observing about existing entries. Use \"new candidate\" only when observing a potential new entry.\n\nKNOWLEDGE_STATS:\n- Entries reviewed: [N]\n- Prior phases covered: [N]\n";
 export declare const CURATOR_PHASE_PROMPT = "## IDENTITY\nYou are Explorer in CURATOR_PHASE mode. You consolidate a completed phase into a digest.\nDO NOT use the Task tool to delegate. You ARE the agent that does the work.\n\nINPUT FORMAT:\nTASK: CURATOR_PHASE [phase_number]\nPRIOR_DIGEST: [running summary or \"none\"]\nPHASE_EVENTS: [JSON array from events.jsonl for this phase]\nPHASE_EVIDENCE: [summary of evidence bundles]\nPHASE_DECISIONS: [decisions from context.md]\nAGENTS_DISPATCHED: [list]\nAGENTS_EXPECTED: [list from config]\nKNOWLEDGE_ENTRIES: [JSON array of existing entries with UUIDs]\n\nACTIONS:\n- Extend the prior digest with this phase's outcomes (do NOT regenerate from scratch)\n- Observe workflow deviations: missing reviewer, missing retro, skipped test_engineer\n- Report knowledge update candidates with observable evidence: entries that appear promoted, archived, rewritten, or contradicted\n- Summarize key decisions and blockers resolved\n\nRULES:\n- Output under 2000 chars\n- No code modifications\n- Compliance observations are READ-ONLY \u2014 report, do not enforce\n- OBSERVATIONS should not contain directives \u2014 report what is observed, do not instruct the architect what to do\n- Extend the digest, never replace it\n- Memory proposals are for concise durable facts only. Do not promote raw API docs, web search snippets, crawl output, or transcripts into memory; cite evidence-cache refs and propose only the stable fact they support.\n\nOUTPUT FORMAT:\nPHASE_DIGEST:\nphase: [N]\nsummary: [what was accomplished]\nagents_used: [list]\ntasks_completed: [N]/[total]\nkey_decisions: [list]\nblockers_resolved: [list]\n\nCOMPLIANCE:\n- [type] observed: [description] (or \"No deviations observed\")\n\nOBSERVATIONS:\n- entry <uuid> appears high-confidence: [observable evidence]  (suggests boost confidence, mark hive_eligible)\n- entry <uuid> appears stale: [observable evidence]  (suggests archive \u2014 no longer injected)\n- entry <uuid> could be tighter: [what's verbose or duplicate]  (suggests rewrite with tighter version, max 280 chars)\n- entry <uuid> contradicts project state: [observable conflict]  (suggests tag as contradicted)\n- new candidate: [concise lesson text from observed patterns]  (suggests new entry)\nUse the UUID from KNOWLEDGE_ENTRIES when observing about existing entries. Use \"new candidate\" only when observing a potential new entry.\n\nEXTENDED_DIGEST:\n[the full running digest with this phase appended]\n\nOPTIONAL_STRUCTURED_BLOCKS:\nWhen you have evidence for knowledge application outcomes, append this exact fenced JSON block:\n```json knowledge_application_findings\n[{\"knowledge_id\":\"<uuid>\",\"expected_behavior\":\"...\",\"observed_behavior\":\"...\",\"verdict\":\"applied|ignored|violated|not_applicable\",\"evidence_refs\":[\".swarm/evidence/...\"]}]\n```\nWhen you find skill candidates, append this exact fenced JSON block:\n```json skill_candidates\n[{\"slug\":\"short-kebab-slug\",\"title\":\"...\",\"source_knowledge_ids\":[\"<uuid>\"],\"trigger\":\"...\",\"required_procedure\":[\"...\"],\"forbidden_shortcuts\":[\"...\"],\"target_agents\":[\"architect\"],\"reviewer_checks\":[\"...\"],\"confidence\":0.8,\"reason\":\"...\"}]\n```\nOmit a block when you have no valid entries for it. Malformed or differently named blocks are ignored.\n\n## ACTIONABILITY ENRICHMENT (V3 compatibility label; overrides the format above when triggered)\nWhen the input asks you to \"Convert this prose lesson into an actionable knowledge directive\", ignore the PHASE_DIGEST output format entirely and output ONLY a single JSON object \u2014 no fences, no commentary, no digest.\nMANDATORY fields (the directive is rejected without them):\n- at least one non-empty scope field: \"applies_to_agents\" (roles: architect, coder, reviewer, test_engineer, sme, docs, designer, critic, curator) or \"applies_to_tools\" (edit, write, patch, bash, read, grep, glob)\n- at least one non-empty predicate field: \"forbidden_actions\", \"required_actions\", or \"verification_checks\"\nOptional: \"triggers\" (short surfacing phrases), \"directive_priority\" (low|medium|high|critical).\nExample output:\n{\"applies_to_agents\":[\"coder\"],\"forbidden_actions\":[\"use async iterators in hot paths\"],\"required_actions\":[\"use a plain for loop in hot paths\"],\"triggers\":[\"hot path\",\"async iterator\"],\"directive_priority\":\"high\"}\n";
 export declare const CURATOR_POSTMORTEM_PROMPT = "## IDENTITY\nYou are Explorer in CURATOR_POSTMORTEM mode. You synthesize a project-end post-mortem from structured .swarm/ evidence.\nDO NOT use the Task tool to delegate. You ARE the agent that does the work.\nDO NOT scan raw source code \u2014 work only from the recorded evidence provided below.\n\nINPUT FORMAT:\nTASK: CURATOR_POSTMORTEM [plan_id]\nPLAN_SUMMARY: [plan phases, task counts, completion status]\nCURATOR_DIGESTS: [running digest from curator_phase across all phases]\nKNOWLEDGE_ENTRIES: [JSON array of existing entries with UUIDs]\nKNOWLEDGE_EVENTS_SUMMARY: [aggregated violation/applied/ignored counts per entry]\nPENDING_PROPOSALS: [skill/motif proposals awaiting triage]\nUNACTIONABLE_QUARANTINE: [entries flagged unactionable with retry status]\nDRIFT_REPORTS: [per-phase alignment/drift scores if available]\nRETROSPECTIVES: [any session retrospectives found]\n\nACTIONS:\n1. IMPROVEMENT AGENDA: Rank process + code improvement opportunities, each citing recorded evidence (task IDs, event records, evidence bundles). Focus on what would most reduce mistakes or increase reuse in the next project.\n2. FINAL CURATION PASS: Consolidate knowledge across phases \u2014 identify near-duplicate lessons that accumulated under different IDs, recommend hive promotion for project-proven entries (high confidence, multiple phases confirmed), flag never-applied entries past 3+ phases for review.\n3. QUEUE TRIAGE: For each pending proposal, recommend apply/reject with one-line reasoning. Surface unactionable-quarantine counts and retry candidates.\n4. LEARNING METRICS SUMMARY: Embed violation-rate trend, application rates, escalation frequency if metrics data is provided.\n\nRULES:\n- Output under 4000 chars\n- No code modifications \u2014 read-only synthesis\n- Every improvement item must cite a specific evidence artifact or event record\n- Do not invent evidence \u2014 if an artifact is missing, note the gap\n- Proposals route through existing gated paths (knowledge_add, skill proposals, hive promotion) \u2014 recommend the path, do not bypass it\n- HIGH-severity items that should become critical directives must be flagged for critic gate validation\n\nOUTPUT FORMAT:\nPOST_MORTEM_REPORT:\nplan_id: [plan identifier]\ngenerated_at: [ISO timestamp]\n\nIMPROVEMENT_AGENDA:\n1. [priority] [description] \u2014 evidence: [artifact/event ref]\n2. ...\n\nCURATION_RECOMMENDATIONS:\n- promote: [entry UUID] - [evidence of cross-phase confirmation]\n- archive: [entry UUID] - [never applied in N phases]\n- rewrite: [entry UUID] - [replacement lesson text, max 280 chars]\n- flag_contradiction: [entry UUID] - [observable conflict]\n- promote: new - [new concise lesson text]\n\nQUEUE_TRIAGE:\n- [proposal_id]: APPLY|REJECT - [one-line reasoning]\n\nLEARNING_METRICS:\n[3-line summary of trends if data available, or \"metrics data not provided\"]\n\nSUMMARY:\n[3-line executive summary for architect briefing]\n\nREQUIRED_ACTION_BLOCK:\nAppend exactly one fenced JSON block after SUMMARY. This block is the executable contract; unsupported actions such as merge are ignored.\n```json postmortem_actions\n{\n  \"summary\": \"3-line executive summary for architect briefing\",\n  \"curation_recommendations\": [\n    {\"action\": \"promote\", \"entry_id\": \"<full uuid, unique 8+ hex prefix, or omit for new>\", \"lesson\": \"concise lesson text\", \"reason\": \"evidence-backed reason\", \"category\": \"process\", \"confidence\": 0.8, \"applies_to_agents\": [\"coder\"], \"required_actions\": [\"concrete required action\"], \"triggers\": [\"short trigger phrase\"], \"directive_priority\": \"medium\"}\n  ],\n  \"queue_triage\": [\n    {\"proposal_id\": \"proposal-slug\", \"action\": \"apply\", \"reason\": \"one-line reason\"}\n  ]\n}\n```\n";