[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6C46RSKVPJZADD3E9S6NRA4":3,"skill-same:01M6C46RSKVPJZADD3E9S6NRA4":83},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":33,"linkOnly":34,"markdownContent":35,"markdownBody":36,"frontMatter":37,"refName":38,"commitSha":39,"skillDirectory":40,"forksCount":41,"repositoryHtmlUrl":42,"repositoryHomepageUrl":6,"ownerHtmlUrl":43,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":44,"install":45,"resources":58},"01M6C46RSKVPJZADD3E9S6NRA4","review-judgment",null,"Reusable single-shot review judgment methodology for scoring prosecution and defense ledgers, verifying evidence, and emitting judge output, plus the parent-owned post-judge steps every review lane runs against that output — the close-out record's amendment rule (its single home on every lane) and the Post-Judge Disposition Gate. Use when ruling on review findings after prosecution and defense are available, and when a lane needs the close-out amendment trigger. DO NOT USE FOR: GitHub review intake routing, response-location policy for review responses, or fix execution ownership (keep those in Code-Review-Response.agent.md).","Grimblaz\u002Fagent-orchestra","Grimblaz","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F74755849?v=4","01M6C430M9XZDHGHPWQVTQ3KMA","01M6C430RJY3MMX0QAZ3X8DR3Y",3,"MIT","https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Freview-judgment\u002FSKILL.md",[],[18,19,20,21,22,23,24,25,26,27,28,29,30,31,32],"review","review judgment","github","policy","routing","code","code-review","intake","lane","methodology","defense","execution","fix","use","evidence","2026-11-10 10:12:30",false,"---\nname: review-judgment\ndescription: \"Reusable single-shot review judgment methodology for scoring prosecution and defense ledgers, verifying evidence, and emitting judge output, plus the parent-owned post-judge steps every review lane runs against that output — the close-out record's amendment rule (its single home on every lane) and the Post-Judge Disposition Gate. Use when ruling on review findings after prosecution and defense are available, and when a lane needs the close-out amendment trigger. DO NOT USE FOR: GitHub review intake routing, response-location policy for review responses, or fix execution ownership (keep those in Code-Review-Response.agent.md).\"\n---\n\n\u003C!-- platform-assumptions: markdown skill guidance for VS Code custom agents in Agent Orchestra; assumes the calling agent owns intake routing, categorization policy, and handoff to implementation. -->\n\u003C!-- markdownlint-disable-file MD041 MD003 -->\n\n# Review Judgment\n\nReusable judgment method for a single referee pass over prosecution and defense.\n\n## When to Use\n\n- When prosecution findings and defense responses are both available\n- When a judge must independently verify claims before ruling\n- When a scored review summary and machine-readable ruling block are required\n- When external review comments have already been converted into a prosecution ledger\n\n## Purpose\n\nMake one final, evidence-backed ruling per finding. The goal is not to split the difference between prosecutor and defense, but to decide whether the proposed change would improve the code and to record that decision in a format the pipeline can consume.\n\nProsecution ledgers are coverage-first by contract: prosecution reports every finding with a statable failure mode, including low-confidence and low-severity ones. The judge is the `filter of record` — expect a wide, uneven ledger as normal input, not a sign that prosecution failed. Filtering happens here, at judgment, not upstream.\n\n## Single-Shot Judgment Workflow\n\n1. Read the prosecution finding, including severity, points, citation, and failure mode.\n2. Read the defense response and note whether it disproves, concedes, or cannot disprove the claim.\n3. Verify the evidence independently.\n4. Rule once: prosecution sustained or defense sustained.\n5. Emit score, confidence, and structured output.\n\nNo rebuttal rounds. Uncertain items still need a ruling.\n\n## Improvement Test\n\nReminder: because prosecution is coverage-first, this test will routinely see low-confidence and low-severity items — that is expected input, not noise to wave through. Rule on each the same way, evidence-first.\n\nAsk this first for every item:\n\n1. Will acting on this finding improve the code?\n\nOutcomes:\n\n- Yes -> accept the improvement\n- No -> reject it\n- Unclear even after verification -> reject it for now\n\nUncertainty is not a deferral bucket. If improvement cannot be shown with evidence, do not accept it.\n\n## Independent Verification Expectations\n\nBefore sustaining a finding:\n\n- Read the cited code, config, test, or document directly\n- Confirm the claimed defect actually exists\n- State what was verified, not just what the prosecutor said\n\nWhen the cited evidence does not support the claim, sustain the defense and explain the mismatch clearly.\n\n**POST-FIX-SCOPED**: when this verification pass follows post-fix targeted prosecution with mutation-tested verification, judgment coverage extends to every branch the fix commit modified, per the canonical post-fix scope constraint in `skills\u002Fvalidation-methodology\u002Freferences\u002Freview-reconciliation.md`. This note does not apply to a non-post-fix (main-review) judgment pass.\n\n## Scoring Model\n\nSeverity maps to points as follows:\n\n- `critical` or `high` -> 10 points\n- `medium` -> 5 points\n- `low` -> 1 point\n\nJudges may override the prosecution severity when verification shows the impact is lower or higher than claimed.\n\nConfidence guidance:\n\n- `high` -> direct structural proof, test output, or explicit code evidence\n- `medium` -> evidence leans one way but is not fully conclusive\n- `low` -> honest uncertainty after reasonable verification\n\n## Score Summary Output\n\nEmit a score table after ruling all findings.\n\n```markdown\n### Adversarial Review Score Summary\n\n| Finding     | Pass | Prosecution (severity, pts) | Defense verdict | Ruling                   | Confidence | Points    |\n| ----------- | ---- | --------------------------- | --------------- | ------------------------ | ---------- | --------- |\n| F1: {title} | {N}  | {severity} ({pts} pts)      | conceded        | ✅ Sustained             | high       | P+{pts}   |\n| F2: {title} | {N}  | {severity} ({pts} pts)      | disproved       | ❌ Defense sustained     | medium     | D+{pts}   |\n| F3: {title} | {N}  | {severity} ({pts} pts)      | disproved       | ✅ Prosecution sustained | high       | D-{2×pts} |\n\n**Totals**\n\n- Prosecutor: {sum of sustained prosecution points} pts ({N} findings sustained)\n- Defense: {net points after rejected-disproof penalties} pts\n- Judge rulings: {total} ({N} pending user scoring)\n```\n\nUse `—` in the Pass column when the prosecution mode does not carry a pass number.\n\n## Structured Judge Output\n\n### Sentinel emission (issue #441, D-new-4)\n\n**Immediately after the judge ruling is finalized** and before any pipeline-metrics persistence,\npersist an idempotent sentinel PR comment for the PR being reviewed:\n\n```text\n\u003C!-- review-judge-produced-{PR} -->\n```\n\nThis sentinel is separate from the judge-rulings comment. `skills\u002Fsession-memory-contract\u002Fscripts\u002Fpersist-marker.ps1` is the ONLY documented write path for this marker (family `review-judge-produced`, `post-new`, `sentinel-empty` validator adapter — pins an empty, marker-only payload) — never a hand-composed `gh pr comment` call. **What that rule buys is a single audited writer — not protection from `updated_at` advancement.** This family is `post-new`, so the write POSTs a fresh comment and advances nothing; the hazard arrives when a write replaces a **whole body** already carrying someone else's marker. See `skills\u002Fsession-memory-contract\u002Freferences\u002Fhandoff-markers.md` § What the write-path rule buys for the write-shape split. The script's own post-new idempotency comparison (compare the candidate against the latest marker match; normalized-identical → no-op) supersedes the separate `gh pr view {PR} --json comments` presence check this section previously directed: pass a body file whose only content is the marker line, and the script either posts a new sentinel or reports a `no-op` action against the already-present one.\n\n**Ordering rule**: sentinel comment → judge-rulings comment. The sentinel must be written first so that the warn-only hook can detect \"review completed but credit not yet written\" during the window between sentinel emission and PR-body update.\n\n**SMC-16 governance**: the sentinel marker `\u003C!-- review-judge-produced-{PR} -->` is governed by SMC-16. See `Documents\u002FDesign\u002Fsession-memory-contract.md`. Both Copilot and Claude judge runs write this sentinel; Code-Conductor reads it via the warn-only hook synthesis path.\n\n### Judge-rulings comment\n\nAfter the sentinel, emit the `judge-rulings` block in the same PR comment as the Markdown score summary:\n\n```yaml\n\u003C!-- judge-rulings\n- id: F1\n  judge_ruling: sustained\n  judge_confidence: high\n  points_awarded: P+10\n- id: F2\n  judge_ruling: defense-sustained\n  judge_confidence: medium\n  points_awarded: D+5\n-->\n```\n\nKeep the Markdown score summary and the `judge-rulings` block together in the same response payload. On GitHub, keep them in the same PR comment rather than splitting them across separate comments. **This comment does not include `\u003C!-- code-review-complete-{PR} -->`** — that marker is retired as of issue #441 Step 11; Code-Conductor reads `credits[]` from the `\u003C!-- pipeline-metrics -->` PR body block directly. **This comment does not include pipeline-metrics body emission** — that is owned by Code-Conductor's `## Pipeline Metrics` emitter at PR creation time.\n\n### Phase-containment emission\n\nIn the same PR comment as the `\u003C!-- judge-rulings ... -->` block, emit one `\u003C!-- phase-containment-{PR} -->` block per sustained finding (`judge_ruling: sustained`). The two block families are different shapes and are not interchangeable: `judge-rulings` (above) stays **bare** — one unclosed `\u003C!-- judge-rulings ... -->` comment — while `phase-containment` is always **paired** — a self-closed `\u003C!-- phase-containment-{PR} -->` open tag followed by plain-text YAML fields and a separate `\u003C!-- \u002Fphase-containment-{PR} -->` close tag, because the close tag powers `Get-PhaseContainmentBlock`'s pair-matching malformation detection (issue #772 D6). **Scope boundary**: this section and its Observer variant below document the block *shape* only — routing a given finding's emission to the shared `Add-JudgeRulingsBlock`\u002F`Get-PhaseContainmentBlock` helper machinery is deferred to AC6, not covered here.\n\n- `finding_key`: `code-review:{stable_finding_key}`\n- `introduced_phase`: set by explicit agent judgment — no default; reason which phase originated this defect\n- `catchable_phase`: set by explicit agent judgment — no default; reason which phase was the earliest this defect could have been caught\n- `caught_stage: code-review`\n- `escape_distance`: recomputed as `3 - ordinal(catchable_phase)` (code-review projection = 3; phase ordinals: experience=0, design=1, plan=2, implementation=3)\n- `severity`, `systemic_fix_type`, `category`: carry forward from the finding\n- `apparatus_meta: false` unless a stated criterion justifies `true`; when `apparatus_meta: true`, the entry is audited\n- `appended_at`: stamp the current UTC instant in the strict `yyyy-MM-ddTHH:mm:ssZ` form (863 M1 fix) — this block is hand-authored directly into the PR comment (no script primitive writes it on this surface), so the judge\u002Fagent authoring the block is responsible for stamping this field itself\n\nA fully literal canonical example, for a sustained code-review finding on PR 879:\n\n```markdown\n\u003C!-- phase-containment-879 -->\nfinding_key: code-review:gh-1234\nintroduced_phase: implementation\ncatchable_phase: implementation\ncaught_stage: code-review\nescape_distance: 0\nseverity: high\nsystemic_fix_type: instruction\ncategory: security\napparatus_meta: false\nappended_at: 2026-07-18T22:20:00Z\n\u003C!-- \u002Fphase-containment-879 -->\n```\n\n#### Observer emission variant (post-review-observer)\n\n`reviewer_source` values referenced below are resolved per § `reviewer_source` Lookup Order (later in this document) — read that section first if the resolution mechanism itself is in question; this rule only consumes the already-resolved value.\n\nWhen a sustained finding's `reviewer_source` resolves to a real external identity (not the reserved `local` sentinel) and its `internal_match.match_status` is `novel`, emit the observer variant of the phase-containment block instead of the standard code-review block above — one block per finding, never both:\n\n- `finding_key`: `post-review-observer:{stable_finding_key}` — this exact prefix is load-bearing (M26): `Get-EmissionGap` attributes a block to a surface by checking the `finding_key` prefix, so a block emitted with the wrong prefix is invisible to, or miscounted by, the reconciliation sweep even though it would pass schema validation.\n- `caught_stage: post-review-observer`\n- `escape_distance`: recomputed as `4 - ordinal(catchable_phase)` (post-review-observer projection = 4; same phase ordinals as above)\n- `introduced_phase`, `catchable_phase`, `severity`, `systemic_fix_type`, `category`, `apparatus_meta`: same setter rule and carry-forward as the code-review block above\n\nA fully literal canonical example, for a sustained post-review-observer finding on PR 879 — same paired shape as the code-review block above, distinguished by the `post-review-observer:` `finding_key` prefix, `caught_stage`, and its own `escape_distance` projection:\n\n```markdown\n\u003C!-- phase-containment-879 -->\nfinding_key: post-review-observer:gh-5678\nintroduced_phase: implementation\ncatchable_phase: implementation\ncaught_stage: post-review-observer\nescape_distance: 1\nseverity: medium\nsystemic_fix_type: skill\ncategory: architecture\napparatus_meta: false\nappended_at: 2026-07-18T22:20:00Z\n\u003C!-- \u002Fphase-containment-879 -->\n```\n\n**Novel-gating is a trinary, not a two-way rule (M25)**:\n\n- `reviewer_source` is the reserved `local` sentinel → standard `code-review` block, unchanged.\n- `reviewer_source` is a resolved real external identity AND `internal_match.match_status: novel` → observer block, which REPLACES the standard block. One defect, one block, never both.\n- `internal_match.match_status` is `duplicate` or `ambiguous`, OR `reviewer_source` is the `unresolved` lookup-failure sentinel → NEITHER block is emitted.\n\nWriting the unqualified two-way version of this rule — \"any resolved external identity gets an observer block\" — is wrong: it would emit an observer block for a `duplicate`-matched finding, double-counting one defect as both a catch-side overlap and an escape-side miss.\n\n**Dispatch is exact-equality only (M40)**: test `reviewer_source -eq 'local'`, never `-like`, `-match`, or other containment-style matching. A real GitHub login literally named `local` is normalized to `ext-local` by the intake process specifically so it can be told apart from the reserved `local` sentinel (see `skills\u002Fcode-review-intake\u002FSKILL.md` § GitHub Review Mode, around line 21, for the normalization rule) — a containment-style dispatch would misclassify `ext-local` as pipeline-native and silently delete a real escape.\n\n**Setter rule**: `catchable_phase` and `introduced_phase` must each be set by explicit agent judgment with no default — the agent must reason about which phase was the earliest in which this specific defect was catchable. Validate each block against `skills\u002Fcalibration-pipeline\u002Fschemas\u002Fphase-containment.schema.json`.\n\n**Emission check (hub maintainers only)**: after posting the `judge-rulings` PR comment with its phase-containment blocks, run `pwsh .\u002F.github\u002Fscripts\u002Fphase-containment-emission-check.ps1 -Pr {N}` and treat its output as advisory — warn-only, never blocking. The repo-relative script path does not resolve from a consumer repo's CWD, so this nudge applies only when working in the Agent Orchestra hub repo itself; see the script header for the full contract.\n\n**Detective-sample extension**: the Code-Critic detective-sample (from `agents\u002FCode-Critic.agent.md:110`) is extended to sample `apparatus_meta: true` entries and `catchable_phase == caught_stage` entries for plausibility review.\n\nField values:\n\n- `judge_ruling`: `sustained` or `defense-sustained`\n- `judge_confidence`: `high`, `medium`, or `low`\n- `points_awarded`: `P+{pts}`, `D+{pts}`, or `D-{2×pts}`\n\n## General Judgment Workflow\n\n### Default Path\n\n- Verify the finding\n- Rule on it\n- Categorize it according to the calling agent's policy\n- Stop after emitting the judgment output\n\n### Evidence-First Rejection\n\nWhen rejecting, cite the reason explicitly:\n\n- The code contradicts the finding\n- Tests or types already guarantee the claimed invariant\n- The documented design makes the proposed change harmful\n- The reviewer's cited evidence is factually wrong\n\n### Escalation Boundary\n\nJudgment does not implement fixes. It produces the ruling and the evidence package needed for the owning orchestrator to route accepted work.\n\n## Related Guidance\n\n- Load `adversarial-review` for prosecution and defense methodology\n- Load `code-review-intake` when GitHub review retrieval and ledger construction are the main problem\n- Read [references\u002Fmultiline-capture-audit.md](references\u002Fmultiline-capture-audit.md) before adding or changing any command-capture site in this skill, and before trusting an `ac_cross_check` recorded before plugin `3.9.1` (issue #977)\n\n## Gotchas\n\n| Trigger                                   | Gotcha                                                   | Fix                                                          |\n| ----------------------------------------- | -------------------------------------------------------- | ------------------------------------------------------------ |\n| The judge repeats the prosecutor verbatim | The ruling becomes a rubber stamp instead of independent | Read the cited artifact directly and state what was verified |\n\n| Trigger                             | Gotcha                                                         | Fix                                                                |\n| ----------------------------------- | -------------------------------------------------------------- | ------------------------------------------------------------------ |\n| A ruling omits the `judge-rulings` block | Downstream consumers can miss completion or fail to route cleanly | Emit the `judge-rulings` block immediately after the score summary in the same payload; ensure the `\u003C!-- review-judge-produced-{PR} -->` sentinel was written as a separate PR comment before |\n\n## Close-Out Record Amendment\n\n**Single home, every lane.** This section is the one place the close-out record's amendment rule is stated. Every lane that runs a judge reaches it by naming it from that lane's own entry document:\n\n| Lane | Entry document that names this section |\n| --- | --- |\n| `\u002Forchestra:review` | `commands\u002Forchestra-review.md` |\n| `\u002Forchestra:review-lite` | `commands\u002Forchestra-review-lite.md` |\n| `\u002Forchestra:review-judge` | `commands\u002Forchestra-review-judge.md` |\n| `\u002Freview-github` | `commands\u002Freview-github.md` |\n| `\u002Forchestrate`, bare `review` via `\u002Fcode-conductor` | `commands\u002Forchestrate.md` and `commands\u002Fcode-conductor.md`, which mandate reading `agents\u002FCode-Conductor.agent.md` § Review Reconciliation Loop |\n| `\u002Fgoal-run` | `commands\u002Fgoal-run.md`, whose body `agents\u002FGoal-Run.agent.md` § Stage 3 dispatches the `standard` adapter — panel, defense, **and judge** |\n\n**Out of population, stated rather than omitted.** AC8's standard is that an excluded path carries its reason, so each is named here rather than left to silence:\n\n- **`\u002Forchestra:review-prosecute` and `\u002Forchestra:review-defend`** — both stop before judge by their own contract, so no pass on them can sustain a finding.\n- **`\u002Fspine-run` and `\u002Forchestra:spine`** — Spine-Runner **runs no judge**. Its frame port `review` is dispatched `skill-only` (`agents\u002FSpine-Runner.agent.md` § Dispatch Table), which *inspects the PR body's `pipeline-metrics` block to verify a review row another lane emitted*; it never dispatches `code-review-response`. Its only adversarial-review mention is the Senior Engineer adversarial-pattern guard, which **halt-returns** rather than dispatching. A verifier of someone else's judge output sustains no findings of its own, so it triggers no amendment — the lane that produced those findings already did. Note the near-miss: `#552 D11`, quoted below to ground this step's executor, names \"Code-Conductor **or spine-runner**\" as orchestrating executors. That is about who may perform non-adversarial post-judgment work when a judge has run, not a claim that spine-runner runs one.\n- **The `design-challenge` and `proxy-github` adapters** — both declare no judge stage in their contracts.\n\nA path that reaches a judge and is missing from the table above is a gap in this section, not a lane that owes nothing — add it rather than reasoning about whether it \"really\" applies.\n\nEvery other surface that mentions the amendment cites this one rather than restating it: two statements of one trigger are two things to keep in sync, and the drift between them is invisible until one is already wrong.\n\n**Executor: the owning parent workflow — never the judge subagent.** The judge's scope stops at its own emission — `agents\u002FCode-Review-Response.agent.md`: *\"Scope boundary: Code-Review-Response owns the sentinel and the `judge-rulings` comment only\"* — and non-adversarial post-judgment work belongs to *\"the orchestrating executor (Code-Conductor or spine-runner), **not** Code-Critic or Code-Review-Response\"* per adversarial-independence `#552 D11` (`skills\u002Fvalidation-methodology\u002Freferences\u002Freview-reconciliation.md` § Response Commit & Push). Do not place this step inside the judge's verification, scoring, or emission steps.\n\n**Why this file.** It already houses the post-judge steps the *parent* owns and each lane's entry document names by section — § Post-Judge Disposition Gate below is one, and is explicitly *\"owned by the calling workflow … not by this skill's judgment pass\"*. The close-out record is an issue-level account of the run that predates and outlives the review, not a review response, so this section is not the *\"response-location policy\"* this skill's frontmatter scopes away: it says nothing about where review responses are posted.\n\n### When it fires\n\n**Once per judge pass, on that pass's emission** — as soon as the parent holds that pass's judge output, and **before** the Post-Judge Disposition Gate below. Not after dispositions, and not after fixes.\n\n\"That pass's judge output\" is whatever form the pass emitted, not a fixed artifact list. On a pull-request target it is the `\u003C!-- review-judge-produced-{PR} -->` sentinel, the `judge-rulings` block, and its phase-containment blocks. On a **non-PR review target** none of those three exists — all are PR-comment-scoped (§ Sentinel emission, § Judge-rulings comment, § Phase-containment emission), and `agents\u002Fcode-review-response.md` § Persistence routes such a run to emit the score summary and `judge-rulings` payload directly in chat instead. There the trigger is that chat payload. Do not read the PR-shaped list as a precondition: a run that waits for artifacts its own lane never produces skips forever, which is the failure this section exists to remove.\n\nThe moment is the judge's emission because that is where this step's input is produced: the record's item 1 is *\"one line per sustained finding — where it was introduced, where it was catchable, where it was caught\"* (`skills\u002Fpost-pr-review\u002FSKILL.md` § 9), which is the phase-containment blocks' own grain. Nothing downstream adds to it — a disposition rules `incorporate | dismiss | escalate`, and a fix lands or does not; neither changes where a defect was introduced, was catchable, or was caught. Firing here therefore does not make the record describe a state that never became true. It also means the amendment text must not describe fix status or disposition status; if it does, the wording is wrong for the moment, not the moment wrong for the wording.\n\n**Every judge pass, not only the first.** A run can reach a judge more than once — the main pass and the post-fix pass (`skills\u002Fvalidation-methodology\u002Freferences\u002Freview-reconciliation.md` § Post-Fix Targeted Prosecution Pass), and a standalone `\u002Forchestra:review-judge` re-run is a pass of its own. The post-fix pass by construction carries the latest findings a run has, which is exactly the population a late amendment exists for.\n\n### Resolving the issue\n\nThis step is the only issue-keyed thing in an otherwise PR-keyed or target-keyed pipeline, so resolve the issue first, in this order:\n\n1. **When the review target is a pull request** — `gh pr view {PR} --json closingIssuesReferences`. This is the PR's own authoritative statement of what it closes. **Zero results**: fall through to route 2. **One result**: that is the issue. **More than one**: evaluate each separately rather than picking one.\n2. **Otherwise, the active issue id the parent already holds.** Every lane's dispatcher carries one when the run has one — the local-review commands pass *\"active issue id if available\"* as pre-dispatch context, and Code-Conductor, Goal-Run and Spine-Runner each run against an issue. This is the ordinary route on a run with **no** pull request, and it is why this step is not a documented no-op on lanes that review a non-PR target.\n3. **Neither available** — a degraded input, not the normal path. Report `no-issue-resolved` per § Advisory, and reported.\n\n**When both routes resolve and disagree, the pull request wins, and say so.** The parent can hold a different issue than the one the PR closes — a designed parent while the PR closes one of its chunks is the ordinary case, and a chunk's findings appended to the parent's record is a public misattribution on someone else's comment. Amend the issue route 1 returned, and name the held issue in the outcome so the divergence is visible rather than silently resolved.\n\n### Whether it applies\n\nPer resolved issue, **applies only** when both hold:\n\n- the issue carries a lawful **open-for-work affirmation record** — the lookup is `skills\u002Fpost-pr-review\u002FSKILL.md` § 9 \"How to check\". Use it rather than a substitute: `gh issue view --json comments` carries no `updated_at` and so cannot tell a lawful record from one edited after creation, failing in the permissive direction; **and**\n- a close-out record already exists on that issue, recognised by its first line (§ 9 \"Action\").\n\n**Does not apply otherwise**, and this step never manufactures a first record. It re-keys *when an existing record is amended*; it does not change which issues owe one, nor the two moments at which one is written — those stay at `skills\u002Fplan-authoring\u002FSKILL.md` § The close-out obligation on an affirmation-record issue and at § 9.\n\n**`not-applicable` on a first-PR run is expected, not a coverage gap.** On a run that opens the issue's first pull request, the review completes before the pull request is created (`agents\u002FCode-Conductor.agent.md` § Review Completion Gate), and the record is written after that review at the pre-PR moment stated in `skills\u002Fplan-authoring\u002FSKILL.md` § The close-out obligation on an affirmation-record issue — this section does not restate that moment. So at judge time no record exists yet, and the record written moments later already accounts for this pass's findings. Nothing is owed and nothing is missed. This step's live population is a judge pass on an issue whose record **already** exists: a review of an already-open pull request, a second pull request, a post-fix pass after the record was written, or a resumed run. Read a first-PR `not-applicable` as the rule working, and do not \"fix\" it by making the step manufacture a record.\n\n### What it does\n\nWhen this pass sustained findings the existing record does not account for, **amend that record in place** — do not post a second one, which would read as two close-outs with no way to tell which is current — and say the amendment is one, per § 9.\n\n**The write route, because this section owns it and the obvious shortcuts mis-target.** The record has no marker family and no `persist-marker.ps1` write path, deliberately (§ 9), so the transport is a plain comment PATCH keyed by the comment's own id:\n\n1. The lawfulness lookup above already paginates the issue's comments through `gh api repos\u002F{owner}\u002F{repo}\u002Fissues\u002F{ID}\u002Fcomments --paginate`. Take the **`id`** of the comment whose first line is the record's, from that same read.\n2. **Re-read that comment's live body immediately before writing** — `gh api repos\u002F{owner}\u002F{repo}\u002Fissues\u002Fcomments\u002F{comment_id} --jq .body`. The record accumulates appended lines, and a body composed from an earlier read silently drops anything appended in between (`skills\u002Fsafe-operations\u002Freferences\u002Fgit-and-gh-traps.md`).\n3. Append the amendment line to **that** live body and write it back: `gh api --method PATCH repos\u002F{owner}\u002F{repo}\u002Fissues\u002Fcomments\u002F{comment_id} -F body=@{file}`.\n\n**Two shortcuts are forbidden here, both documented mis-write traps** (same reference): `gh issue comment --edit-last` targets the author's most recent comment on the issue, which is whatever the run posted last rather than the record; and `Find-OrUpsertComment`'s `-Body` **replaces the existing body verbatim on the PATCH path** — it never re-reads and merges — so pointing it at a record that accumulates appended lines destroys them. That second hazard is about the *write*, not the *selection*: #1031 narrowed the selector to line-1-exact, and the close-out record is not marker-keyed at all, so it was never a candidate for that upserter. A mis-targeted or body-replacing write here damages a public comment whose first writer is often a person, which is why the id comes from the same read that recognised the record and never from a guess.\n\n**Bounded by what the record already carries.** Before amending, read it and name only the sustained findings it does not already account for — whether in its original items or in any amendment line already on it, whichever run appended that line. When this pass adds none, report `no-new-findings` and append nothing. Several near-identical dated lines on one record is the duplicate account this bound exists to prevent, not evidence that the step fired.\n\n**What \"already account for\" is matched on.** The record's item-1 lines are prose, and the judge's findings carry a `stable_finding_key` (§ Stable Finding Key), so the two do not share a key. Match on the **finding's own identity as the record states it** — the file-and-symptom pair a record line names, not string equality with a finding title. When a prior line plainly describes this pass's finding, it is accounted for; when you cannot tell, say so in the outcome rather than appending a line that may be a duplicate. Recording an uncertainty is cheap; a second near-identical dated line is the harm.\n\n### Advisory, and reported\n\n**A skipped or failed amendment never halts the loop** — nor the run, on a lane that has no loop. Emit the loud literal and carry it into this lane's accountability channel:\n\n- `⚠️ close-out record amendment skipped — {reason}`\n\n**The accountability channel, named per lane.** This step is advisory, so its report is the only trace it leaves; a lane with no named channel lets a skip disappear into a report that satisfies every other requirement. **The channel is the run's own terminal report** — the block a lane already returns or posts when it finishes — and every lane has one:\n\n| Lane | Channel |\n| --- | --- |\n| `\u002Freview-github` (GitHub intake) | Response Summary item 5 (`skills\u002Fvalidation-methodology\u002Freferences\u002Freview-reconciliation.md` § Response Commit & Push) |\n| `\u002Forchestrate`, bare `review` via `\u002Fcode-conductor` — **existing-PR path** | the same Response Summary item 5 |\n| `\u002Forchestrate` — **new-PR path** (Conductor Step 4 `gh pr create`) | the **Close-out record amendment** row of the PR-body pipeline-metrics block. That path fires no `persist-changes` and assembles no Response Summary — the two push sites are mutually exclusive — so item 5 does not exist there and pointing at it would name a slot the run never produces. |\n| `\u002Forchestra:review`, `-lite`, `-judge` (local review) | a `Close-out record amendment:` line in the **judgment payload the command returns**, immediately after the `judge-rulings` block and beside the disposition-gate outcome. It is the parent's own line, appended to the payload — it does not alter the judge output the command returns unchanged. |\n| `\u002Fgoal-run` | the Stage 3 report; on a halt, the typed halt report's own body |\n\n**Report one outcome per resolved issue, per judge pass** — not one per pass. Route 1's \"more than one: evaluate each separately\" means a single pass can resolve to several issues (a PR closing more than one), and each gets its own applies-only check and its own outcome; a pass that touched three issues reports three outcome lines, one per issue id, even when they differ — `amended` on one and `not-applicable` on another is the ordinary shape, not a conflict to resolve into a single line. State how many passes fired and, within each pass, how many issues it resolved. The outcomes partition per issue — exactly one applies to each:\n\n- `amended` — naming what it added, and naming the held issue when it differed from the PR's.\n- `no-new-findings` — the record already accounts for every finding this pass sustained.\n- `not-applicable` — an issue **was** resolved, and it carries no affirmation record or no existing record. Name which.\n- `no-issue-resolved` — route 3: neither an issue nor a pull request was available. Carries the loud literal below.\n- `skipped` — the step was reached but could not complete (a failed read, a failed write, a lookup error). Carries the loud literal below with its reason.\n\n**Not a gate.** Nothing refuses a judgment, a fix, a pull request, a merge, or a close because this step did not run, and nothing re-checks it once the run ends. A blocking check here would be the detector `Documents\u002FDesign\u002Fopen-for-work.md` § The close-out habit deliberately declined, wearing another name.\n\n## Post-Judge Disposition Gate\n\n### Purpose\n\nAfter the judge emits its rulings (sentinel + judge-rulings comment), the owning parent agent runs the review-disposition engagement gate over the judge-sustained findings. This prevents cognitive surrender at the PR code-review verdict point: the engineer must consciously disposition each sustained finding rather than having the agent assume outcomes.\n\nThis gate is owned by the calling workflow (e.g. `\u002Forchestra:review-judge`, `\u002Forchestra:review`), not by this skill's judgment pass. The judgment pass ends with the judge-rulings comment. Disposition begins immediately after.\n\n### When to Run\n\nThis section splits two previously-conflated rules — marker emission and the maintainer-interaction gate — so the always-emit behavior below does not read as contradicting the sustained-only scoping that follows it.\n\n**Marker emission** (persisting `\u003C!-- review-dispositions-{PR} -->` and `\u003C!-- engagement-record-review-{PR} -->`) fires on **every** judge pass, including a zero-sustained pass:\n\n- After the `\u003C!-- review-judge-produced-{PR} -->` sentinel is confirmed written\n- Even when no finding was judge-sustained this pass — per the M9 coverage semantics above (§ Scope, around line 277: \"coverage means measurement, not presence\"), a zero-sustained pass still emits both markers, with `entries: []` on the dispositions marker\n\n**Disposition gate** (the maintainer-interaction step, classification into routine vs load-bearing) fires only when there are findings to disposition:\n\n- Over the judge-sustained findings from the `\u003C!-- judge-rulings ... -->` block\n- Only for sustained findings (`judge_ruling: sustained`); defense-sustained findings (`judge_ruling: defense-sustained`) are skipped silently (not disposition-gated)\n- When there are zero judge-sustained findings, no finding enters classification and the gate never fires this pass — but marker emission (above) still runs\n\n### Classification\n\nFor each judge-sustained finding, run the solution-authoring classification gate (`skills\u002Fsolution-authoring\u002FSKILL.md § Rule: Classification gate`) to determine whether it is **load-bearing** or **routine**.\n\nA finding is **load-bearing** iff all three legs pass:\n\n1. **Reversibility** — acting on this finding would change a published or durable artifact (source file, test, config, doc, or skill).\n2. **Non-inheritance** — no specific inherited artifact (prior design decision, existing AC, approved plan statement, locked methodology rule) already settles the required action. The agent MUST attempt to cite one before declaring non-inheritable.\n3. **Audit-plausibility** — the agent can write a substantive `disposition_rationale` sentence.\n\nFailing any leg collapses the finding to **routine**. The tier rule from `solution-authoring § Applying the gate to adversarial-review dispositions` applies: load-bearing adversarial-review dispositions use the **escalation tier** decision brief; routine findings are recorded silently.\n\n### Stable Finding Key\n\nBefore the gate fires, compute a `stable_finding_key` for each finding. This key must survive re-reviews of the same PR. Use the first available:\n\n1. **GitHub comment ID** — if the finding originated from a GitHub review comment, use its comment ID as the key. A comment that yields a single finding keeps the tier-1 key `gh-{comment_id}` — unchanged, and backward compatible with dispositions persisted before this discriminator existed. When one comment yields multiple distinct findings (the code-review-intake skill's split rule — see `skills\u002Fcode-review-intake\u002FSKILL.md`), each finding's key becomes `gh-{comment_id}-{normalized-title-hash[:8]}`, reusing tier 2's normalization convention below to disambiguate findings sharing a comment ID.\n2. **file:line:hash** — normalize the finding title (lowercase, remove punctuation, replace spaces with hyphens), then form `{relative_file}:{line}:{normalized-title-hash[:8]}` where hash is the first 8 hex chars of SHA-256 over the UTF-8 normalized title.\n3. **finding_id fallback** — if neither is available, use the sequential `finding_id` (e.g. `F1`) with a `warn:` prefix to signal instability: `warn:F1`.\n\nThe `stable_finding_key` is what the resume-read mechanism uses to detect prior dispositions across re-reviews of the same PR (see § Stable-Key Resume below).\n\n### `reviewer_source` Lookup Order (GitHub-Sourced Findings Only)\n\nThis lookup order applies only to findings that originated from an external GitHub review comment. Pipeline-native findings (local prosecution\u002Fdefense\u002Fjudge findings) never enter this lookup procedure — the disposition-recording sites below write `reviewer_source: local` directly for those findings, without evaluating any of the three tiers.\n\nBoth disposition-recording sites below (§ Routine Findings and § Load-Bearing Findings) write a `reviewer_source` field. For GitHub-sourced findings, resolve its value in this order:\n\n1. **In-context intake ledger** — when the session is continuous with the GitHub-intake pass that built the finding, read the `reviewer_source` value the intake skill (`skills\u002Fcode-review-intake\u002FSKILL.md`) already recorded for this finding.\n2. **`gh api` re-derivation** — otherwise (standalone or resumed judge pass), extract the `comment_id` embedded in the finding's `stable_finding_key` and re-derive the reviewer identity via `gh api`, trying the PR-review-comment, issue-comment, and review endpoints in turn. Parse rule: the `comment_id` is the numeric run immediately after the `gh-` prefix, up to the next `-` or end of string; this covers both the single-finding `gh-{comment_id}` key shape and the multi-finding `gh-{comment_id}-{hash}` key shape identically. Normalize the re-derived login per the canonical `reviewer_source` normalization rule in `skills\u002Fcode-review-intake\u002FSKILL.md` § GitHub Review Mode (lowercase, `.login`-only, trailing `[bot]` strip, quoted-scalar output, reserved-value\u002F`ext-`-prefix collision escape) — this site does not restate that mechanism.\n3. **`unresolved` sentinel** — on any failure (comment deleted, API error, no comment ID embedded in the key to re-derive from), write the sentinel `unresolved`. Never write `local` on lookup failure — `local` is reserved exclusively for pipeline-native (non-GitHub-sourced) findings, which are written directly per the guard above and never fall through to this sentinel.\n\n### `internal_match` Writer Rule (GitHub-Sourced Findings Only, DD2)\n\n`internal_match.match_status` and the PR-level `external_sources_reconciled` field are **written by the judge**, not merely consumed — § Observer emission variant (above) only ever reads the already-resolved value; this section is the writer contract those readers depend on.\n\nAt reconciliation time — with **both** the internal prosecution ledger (this PR's own pipeline-native findings) and the external review's findings in context simultaneously, per design decision DD2 — the judge sets `internal_match.match_status` on **every** external-source disposition entry (any entry whose `reviewer_source` resolved to a value other than `local`):\n\n- **`duplicate`** — the external finding describes the same defect as an already-caught internal finding. Also write `matched_finding_key` with that internal finding's `stable_finding_key`.\n- **`novel`** — the external finding has no matching internal finding; the pipeline missed it. This is what the observer emission variant later turns into a `post-review-observer` escape block.\n- **`ambiguous`** — the judge cannot confidently resolve a match either way after verification. Excluded from `n₂`\u002F`m`\u002Fcoverage math downstream (Seam Specification) — an honest \"could not resolve\" must never be forced into `duplicate` or `novel`.\n\n`reviewer_source: local` entries (pipeline-native findings) never carry `internal_match` — the field only applies to external-source entries. Field order matters: `internal_match` is written **before** `disposition_rationale` (M42) so a `disposition_rationale` block-scalar's free text can never be mistaken for the real field by a downstream line-regex parser.\n\n**Scope (post-fix batch 4, issue #854 s6; corrected G-CR8, PR #859 GitHub-review post-fix): only when this pass ran in GitHub Review Mode.** The PR-level `external_sources_reconciled` field records that an external review was actually reconciled against — it is meaningless, and actively misleading, on a pass that never had one. Emit it once in the same posted marker **only when this pass is GitHub Review Mode** — determined from the session context\u002Fcommand that invoked this pass (`skills\u002Fcode-review-intake\u002FSKILL.md` § GitHub Review Mode (Proxy Prosecution Pipeline) — the `\u002Freview-github` proxy-prosecution path that ingests GitHub-sourced findings), never from entry presence. Within a GitHub Review Mode pass, emit it **even when empty** (`external_sources_reconciled: []`): a genuinely zero-finding external review is a legal, required coverage record (M9 — \"coverage means measurement, not presence\"), and omitting the field on a GitHub Review Mode pass is indistinguishable downstream from \"this PR was never measured at all.\" A zero-finding external review pass has zero entries by definition, so a detection test keyed on entry presence (e.g. \"at least one entry this pass whose `reviewer_source` resolved to a value other than `local`\") can never fire on exactly the case M9 requires the field to be emitted for — GitHub Review Mode is a property of the SESSION, not a derived property of what got written. On a purely internal-only pass — plain `\u002Forchestra:review`, no GitHub-sourced findings ingested this pass — **omit the field entirely**. Do not emit `external_sources_reconciled: []` on an internal-only pass: that would falsely claim an external review was reconciled and never occurred, manufacturing a measured-looking zero that was never actually measured (the exact false-clean coverage vector this issue exists to eliminate).\n\n### Routine Findings — Silent Recording\n\nFor routine findings, the agent records the disposition silently in the `review-dispositions-{PR}` accumulator without asking the maintainer. Use `schema_version: 4` (current emission format); write `severity`, `stage`, and `reviewer_source` (the reviewer identity or class that produced the finding — use `local` for pipeline-native prosecution\u002Fdefense\u002Fjudge findings; external identities are resolved per § `reviewer_source` Lookup Order above) for all entries, `internal_match` for external-source entries (§ `internal_match` Writer Rule above), and include `ac_cross_check` for any `dismiss` or `defer` entry with severity ≥ medium. Remember the PR-level `external_sources_reconciled` field once per posted marker (§ `internal_match` Writer Rule above; § Persistence — Ordering below shows the full marker shape):\n\n> **Pre-condition**: for any `dismiss` entry with severity ≥ medium, run the AC cross-check (see § AC Cross-Check — Blocking Pre-Condition) before writing this entry.\n\n```yaml\n- stable_finding_key: \"src\u002Ffoo.ts:42:null-check-missing-a1b2c3d4\"\n  finding_id: F2\n  pass: 1\n  disposition: incorporate   # or dismiss — agent's judgment based on judge ruling\n  classification: routine\n  severity: medium           # v3: required field\n  stage: code-review         # v3: required field\n  reviewer_source: local     # v3: writer always emits this field (writer obligation); the reader\u002Fconsumer treats an absent field as local for backward compat with pre-v3 entries (reader-semantics fallback) — local is reserved for pipeline-native findings; external identities resolve per § reviewer_source Lookup Order\n  disposition_rationale: \"Trivial null-guard already required by the existing type contract; no maintainer choice required.\"\n  artifact_citation: \"src\u002Ftypes\u002Findex.ts:18 (NonNullable\u003CT> constraint)\"\n```\n\nThe agent chooses `incorporate` or `dismiss` based on the judge ruling direction. Routine findings do not fire the gate; they feed directly into the accumulator.\n\n### Load-Bearing Findings — the disposition question\n\nFor load-bearing findings, render an **escalation-tier decision brief** (three required elements, all present before the option list):\n\n1. **Concrete element → current state**: what the current code\u002Fartifact actually does, with evidence (file:line or artifact citation from the finding).\n2. **Decision setup → the conflict**: why this finding's proposed change conflicts with or extends beyond the current state.\n3. **Conditional misconception → customer failure mode**: the concrete failure the engineer would cause by taking the wrong path.\n\nThen ask, with exactly these options:\n\n- `Incorporate — apply the fix` (Recommended if prosecution was sustained on strong evidence)\n- `Dismiss — the finding does not warrant a change`\n- `Escalate — this needs a design decision or a separate issue`\n- `Decline engagement — proceed without classification`\n\nCapture the engineer's choice verbatim.\n\n> **Pre-condition**: if the engineer chooses `dismiss` with severity ≥ medium, run the AC cross-check (see § AC Cross-Check — Blocking Pre-Condition) before writing this entry.\n\nRecord as (v4 format — include `severity`, `stage`, `reviewer_source`, `internal_match` for external-source entries (§ `internal_match` Writer Rule above), and `ac_cross_check` for dismiss\u002Fdefer entries with severity ≥ medium):\n\n```yaml\n- stable_finding_key: \"src\u002Fauth\u002Fsession.ts:88:token-expiry-not-checked-b7c1a2f3\"\n  finding_id: F1\n  pass: 2\n  disposition: incorporate   # or dismiss or escalate per engineer choice\n  classification: load-bearing\n  severity: high             # v3: required field\n  stage: code-review         # v3: required field\n  reviewer_source: local     # v3: writer always emits this field (writer obligation); the reader\u002Fconsumer treats an absent field as local for backward compat with pre-v3 entries (reader-semantics fallback) — local is reserved for pipeline-native findings; external identities resolve per § reviewer_source Lookup Order\n  disposition_rationale: \"Engineer chose incorporate: the expiry check was confirmed missing and the fix is bounded to one function.\"\n```\n\nThis example's `reviewer_source: local` means no `internal_match` is written (pipeline-native). For a load-bearing finding whose `reviewer_source` resolved to an external identity, add `internal_match: { match_status: ... }` before `disposition_rationale`, same shape as the routine external-source example above.\n\n### Escalate Semantics\n\nWhen an engineer selects `Escalate`:\n\n1. Record `disposition: escalate` in the entry.\n2. Emit a concise escalation note inline (not a question): `Finding {finding_id} escalated — recommend filing a follow-up issue for: {finding title}. Proceeding without implementing this finding.`\n3. Do NOT implement the finding in the current PR. The current work continues; the escalated finding is noted for tracking.\n4. Record the `escalate` outcome faithfully in `review-dispositions-{PR}` and in the L0 gate-decision token.\n\n### AC Cross-Check — Blocking Pre-Condition\n\nBefore writing any entry with `disposition: dismiss` or `disposition: defer` **and** `severity` ≥ medium — whether the entry originates from a code-review finding (`stage: code-review`) or a CE Gate defect deferral (`stage: ce`) — the agent MUST complete an AC cross-check:\n\n1. Call `Get-AcRefsFromIssue -IssueNumber {parent_issue}` to extract AC file-path references (`{ac_refs}`, ARM 1) **and** `Get-AcTermsFromIssue -IssueNumber {parent_issue}` to extract behavioral AC terms (`{ac_terms}`, ARM 2). **Both**, not one: the two arms are independently sufficient — a file-path match alone reaches `force-accept` without any term match — so calling only the term helper silently disables half the gate. An earlier revision of this step named only the term helper while step 2 went on to pass an `{ac_refs}` that nothing had produced (issue #977).\n2. Call `Get-StructuralVerdict -Finding {finding} -PrFileSet {pr_files} -AcRefs {ac_refs} -RepoRoot {repo_root} -AcTerms {ac_terms}` to obtain the `ac_cross_check` object.\n3. Write the returned `ac_cross_check` object into the disposition entry.\n\n**This pre-condition is blocking.** The gate MUST NOT commit a `dismiss` or `defer` entry with severity ≥ medium that has a null or absent `ac_cross_check`. If either helper returns empty (no AC section, or no parseable tokens), the cross-check still runs — pass `@()` for that arm; when **both** are empty the verdict's `ac_cross_check.source` will be `no-ac-section` and `routed` will be `defer`.\n\n**Low-severity exemption.** Entries with severity `low` are exempt from this pre-condition (the validator also exempts them). Record them without `ac_cross_check`.\n\n> **Any machine-computed `ac_cross_check` recorded before plugin `3.9.1` is untrustworthy — every field, not just `source: no-ac-section`.** Until issue #977 landed, both AC helpers captured the issue body as a per-line array and split it as if it were one string, so they read the body's **second line** rather than the acceptance-criteria section.\n>\n> Do not read that as \"the gate was merely blind.\" It read the **wrong input**, which fails in both directions:\n>\n> - When line 2 carried no backticked token — the usual case, since line 2 of a conventionally formatted issue is blank — both arrays came back empty, `source` resolved to `no-ac-section`, and the finding took the `defer` arm regardless of what the acceptance criteria actually said.\n> - When line 2 *did* carry a backticked token, that token was returned as though it were an acceptance-criteria reference. `source` was then `issue`, not `no-ac-section` (the verdict function only reports `no-ac-section` when **both** arrays are empty), and a fabricated file-arm match could reach `force-accept` — which overrides a structural deferral. This arm was **reachable but never observed** in the 60-issue sweep; it is why the distrust above is scoped to *every* pre-`3.9.1` machine-computed cross-check rather than to one `source` value. Scoping it to `no-ac-section` would have excluded exactly the records that carried override authority.\n>\n> **Anchor on the release, not on a date.** The fix shipped in `3.9.1`; several pull requests merged earlier the same day carry pre-fix cross-checks, so \"before 2026-08-02\" would wrongly vouch for them.\n>\n> Historical dispositions were deliberately **annotated rather than recomputed** (issue #977, maintainer decision) — a recompute reads issue bodies as amended since, which is a new claim, not a reconstruction.\n>\n> **`no-ac-section` remains imprecise even post-fix**, and this is unchanged by #977: the verdict function emits it whenever both arrays are empty, which includes an issue that *has* a populated acceptance-criteria section whose tokens are all stop-listed or unbackticked. It means \"nothing extractable\", not \"no section\".\n>\n> Both helpers now join the captured lines before splitting; do not remove that join, and see [references\u002Fmultiline-capture-audit.md](references\u002Fmultiline-capture-audit.md) before adding any new command-capture site.\n\n**`Add-FollowUpIssue` guard.** When the cross-check routes to `defer` and the agent calls `Add-FollowUpIssue` to file a follow-up issue, it MUST pass the `ac_cross_check` outcome as part of the issue body. Include a fenced YAML block in the body:\n\n```yaml\nac_cross_check:\n  file_arm: {bool}\n  term_arm: {bool}\n  result: {matched-high|matched-ambiguous|no-match}\n  source: {issue|pr-body|no-ac-section}\n  routed: defer\n```\n\nThis ensures the follow-up issue carries AC-provenance for the deferral decision, which is the AC4 contract.\n\n### `DEFERRED-SIGNIFICANT (structural)` and `disposition: defer` name one outcome\n\nThis skill owns the deferral mechanism, and until issue #998 its prose carried only one of the outcome's two names. A reader who arrived holding the consumer-facing label — the spelling every downstream surface uses — found no occurrence of it here and no way to learn that the mechanism described above is what produces it. The traversal dead-ended inside the skill that owns the thing.\n\n`scripts\u002FTest-DeferralCriteria.ps1` — this skill's own script — returns `verdict = 'DEFERRED-SIGNIFICANT (structural)'` when a finding meets the structural deferral criteria. **That verdict and the `disposition: defer` value recorded in the accumulator are one outcome under two names**, not two stages, two tiers, or two severities. The verdict is what the script emits; `disposition: defer` is how that same outcome is written down in the durable record. A finding carrying either carries both, and the `ac_cross_check` obligations above attach to it under whichever name you met it by.\n\nConsumers see the script's spelling. `agents\u002FCode-Review-Response.agent.md` § Vocabulary note already maps its own `SIGNIFICANT` tier onto `DEFERRED-SIGNIFICANT (structural)`, and the token travels the tree in that form. Nothing is renamed here — the seam was that this skill never said which of its own outputs the label referred to, and this section is that statement.\n\n### Legitimate Partial-AC Defer — Loud Guard\n\nWhen the AC cross-check returns `routed: defer` (because `result: no-match` or `source: no-ac-section`), the agent has confirmed that this finding genuinely lacks plan AC coverage. **Silently recording a `defer` entry at this point is the exact anti-pattern this feature was built to prevent.**\n\nThe agent MUST follow this sequence instead:\n\n1. **Emit a loud inline note** (not a question — this is a guard, not a decision the maintainer makes): `⚠️ Finding {finding_id} deferred without AC coverage (ac_cross_check.result: {result}) — mandatory proposal required.`\n\n2. **Enter a mandatory proposal** into the `§2e Filing Approval Gate` batch (`skills\u002Fsafe-operations\u002FSKILL.md` § 2e), pre-checked and recommended-approve, annotated `AC-uncovered defer`. The canonical title uses `ConvertTo-CanonicalFollowupTitle`. The proposal body MUST include the finding title, the judge ruling, and the `ac_cross_check` YAML block — this payload travels into the durable drop record if the maintainer drops the proposal. This is mandatory regardless of the finding's classification tier — routine findings that lack AC coverage still enter the gate as a proposal.\n\n3. **Record in the accumulator** with `disposition: defer`, the `ac_cross_check` object, and `disposition_rationale` that cites the `no-match`\u002F`no-ac-section` outcome and references the gate proposal (and the resulting issue URL once the maintainer approves it and it is filed).\n\nThe loud guard does not apply when `routed: force-accept` (high-confidence AC match) or `routed: disposition-gate` (ambiguous match fires the disposition question normally). It applies only to the `routed: defer` arm.\n\n**Low-severity exemption applies here too**: findings with severity `low` are exempt from this guard (the low-severity exemption from the blocking pre-condition applies throughout this section).\n\n### L0 Gate-Decision Tokens\n\nFor each finding that was gate-classified (routine or load-bearing), emit an L0 gate-decision token per `skills\u002Fsolution-authoring\u002Fschemas\u002Fgate-decision-token.schema.json`:\n\n```yaml\ndecision_id: \"{stable_finding_key}\"\nphase: review\noutcome: asked          # 'asked' for load-bearing, 'gate-fails' for routine\nclassification: load-bearing   # or routine\nwindow_position: review-disposition\ntimestamp: \"{ISO-8601 UTC}\"\npull_request_number: {PR}   # NOT issue_number\nskip_reason: \"routine-finding\"   # omit when outcome=asked\n```\n\nWrite each token to the authoritative L0 location per `skills\u002Fsolution-authoring\u002FSKILL.md` § L0 Gate Token: `\u002Fmemories\u002Fsession\u002Fgate-events-{session_key}.jsonl` (fallback: `.copilot-tracking\u002Fgate-events.jsonl`) as a single JSON line per the existing L0 emission pattern.\n\n### Stable-Key Resume\n\nAt the start of the disposition pass, call `Read-EngagementRecords -Phase review -PullRequestNumber {PR}` (from `.github\u002Fscripts\u002Flib\u002Fframe-engagement-record-core.ps1`). If a prior `engagement-record-review-{PR}` exists, extract its `load_bearing_decisions[].decision_id` values. These are `stable_finding_key` values of previously-gate-fired findings.\n\nFor each finding in the current judge-sustained set, check whether its `stable_finding_key` appears in the prior record:\n\n- **Match found** → `same-decision-resume` skip: reuse the prior `engineer_choice`, log `Reusing prior {stable_finding_key}: {engineer_choice}`, do not ask again.\n- **No match** → run the gate normally.\n\nThis enables re-review of a PR without re-asking for findings already dispositioned.\n\n### Persistence — Ordering\n\nWrite in this order (atomic marker first, engagement-record second). `skills\u002Fsession-memory-contract\u002Fscripts\u002Fpersist-marker.ps1` is the ONLY documented write path for the `review-dispositions-{PR}` marker (family `review-dispositions`, `post-new`) — never a hand-composed `gh pr comment` call. **What that rule buys is a single audited writer — not protection from `updated_at` advancement.** This family is `post-new`, so the write POSTs a fresh comment and advances nothing; the hazard arrives when a write replaces a **whole body** already carrying someone else's marker. See `skills\u002Fsession-memory-contract\u002Freferences\u002Fhandoff-markers.md` § What the write-path rule buys for the write-shape split. **`engagement-record-review-{PR}` is a known v1 registry gap and stays hand-authored for now**: the `engagement-record` family registry row declares `TargetSurface: issue` for every phase, but the `review` phase is PR-keyed — a `review`-phase write through the script would be refused by the surface preflight (`persist-marker-core.ps1`'s documented \"Known v1 scope gap\"). Do not attempt to route `engagement-record-review-{PR}` through the script until that gap is closed; hand-compose it per the existing convention below.\n\n1. **`\u003C!-- review-dispositions-{PR} -->`** — Persist via `persist-marker.ps1` (family `review-dispositions`). Payload: `schema_version: 4`, `passes_run: [...]`, `entries: [...]` (all findings, routine and load-bearing, one entry per finding), plus the PR-level `external_sources_reconciled` field. v2 added per-entry `severity`, `ac_cross_check`, and `stage` fields; v3 added per-entry `reviewer_source` (the reviewer identity or class — `local` for pipeline-native findings; external identities resolve per § `reviewer_source` Lookup Order above); v4 adds per-entry `internal_match` and the PR-level `external_sources_reconciled` field (see § `internal_match` Writer Rule below). This is the atomic per-finding record.\n\n   ````text\n   \u003C!-- review-dispositions-{PR} -->\n\n   ```yaml\n   schema_version: 4\n   passes_run: [1, 2, 3, 4, 5]\n   entries:\n     - stable_finding_key: \"...\"\n       pass: 1\n       disposition: incorporate\n       classification: routine\n       severity: medium\n       stage: code-review\n       reviewer_source: local   # pipeline-native finding: no internal_match (see below)\n       disposition_rationale: \"...\"\n       ac_cross_check:\n         file_arm: false\n         term_arm: true\n         result: matched-high\n         ac_ref: \"- the AC line that matched\"\n         source: issue\n         routed: force-accept\n     - stable_finding_key: \"gh-123456789\"\n       pass: 1\n       disposition: incorporate\n       classification: routine\n       severity: medium\n       stage: code-review\n       reviewer_source: jdoe    # external identity, resolved per § reviewer_source Lookup Order\n       internal_match:          # written BEFORE disposition_rationale (M42 field order)\n         match_status: novel    # duplicate | novel | ambiguous — see § internal_match Writer Rule\n       disposition_rationale: \"...\"\n   external_sources_reconciled: [\"gh-123456789\"]   # PR-level; emitted ONLY because this pass ran in GitHub Review Mode (a property of the SESSION\u002Fcommand, e.g. \u002Freview-github — never inferred from entry presence). [] is a legal, required zero-finding coverage record WITHIN a GitHub Review Mode pass, never an omission; on a purely internal-only pass (no external-source entries this pass) OMIT this field entirely instead.\n   ```\n   ````\n\n   > **v3 per-entry requirements** (carried over unchanged from v2): For entries with `disposition: dismiss` or `disposition: defer` and `severity` ≥ medium, `ac_cross_check` is required. The `ac_cross_check` object records which arms ran (`file_arm`, `term_arm`), the result tier (`matched-high | matched-ambiguous | no-match`), the matched AC reference if any, the source, and the routing outcome. Legacy `schema_version: 1` entries are exempt from this check. `artifact_citation` covers non-AC inherited artifacts; `ac_cross_check.ac_ref` is the AC-specific channel.\n   >\n   > **v4 requirements** (additive on top of v3): every external-source entry (`reviewer_source` not `local`) carries `internal_match.match_status`, written per § `internal_match` Writer Rule below. The PR-level `external_sources_reconciled` field is emitted once per posted marker **only on a GitHub Review Mode pass** (§ `internal_match` Writer Rule above) — including as `external_sources_reconciled: []` when that external review found nothing to reconcile. An absent field means this pass never attempted external reconciliation (the normal, expected state for a plain internal-only `\u002Forchestra:review` pass); an empty array means a GitHub Review Mode pass measured and found zero (M9). Never emit the field on an internal-only pass, even as `[]` — that would misrepresent an unmeasured pass as a measured one. `reviewer_source` stays required on every entry, same as v3.\n   >\n   > **`stage` field values**: The `stage` field records which pipeline stage produced this entry: `code-review` for the post-judge disposition gate, `ce` for CE Gate defect deferral. Both stages use the same `ac_cross_check` pre-condition at severity ≥ medium.\n   >\n   > **In-session schema audit is now pre-write and blocking (893-D3 amendment, s9)**: `persist-marker.ps1`'s `review-dispositions` family validator adapter invokes `.github\u002Fscripts\u002Flib\u002Freview-dispositions-validator-core.ps1` **in-process, via the call operator (`&`), with mandatory try\u002Fcatch containment** — never as a `pwsh -File` subprocess (the original 893-D3 subprocess design is superseded; a subprocess cannot bind that script's `-InMemoryMarkers` array parameter without hitting the recorded #866 flattening trap, since `-InMemoryMarkers` is a top-level array parameter). Any finding the validator returns is converted into a hard pre-write refusal on this write path: v4 schema violations (e.g., missing `ac_cross_check` on dismiss\u002Fdefer entries at severity ≥ medium, or a missing `reviewer_source`) are refused with the offending field named, before any network write. This is a property of the `persist-marker.ps1` write path only — the standalone validator script's own read-side warn-only contract (SMC-23 detection-at-review) is unchanged for any other caller.\n\n2. **`\u003C!-- engagement-record-review-{PR} -->`** — Post as a separate PR comment (not the same comment as review-dispositions), via a hand-composed `gh pr comment` call — **not** `persist-marker.ps1` (see the known v1 registry gap noted above this list). Payload follows `skills\u002Fengagement-record-emission\u002FSKILL.md` shape at `schema_version: 4`, `phase: review`. Load-bearing findings that fired the disposition question appear in `load_bearing_decisions[]` with their `engineer_choice` and `audit_rationale`. Routine findings do not appear in the engagement-record.\n\n   The engagement-record carries the `same-decision-resume` identity; the review-dispositions carries the per-finding outcome record. Never merge the two into a single comment.\n\n### SMC References\n\n- **SMC-19** — `finding_dispositions` is design-only (issue-keyed on `design-phase-complete`). This section's `review-dispositions` is a distinct, PR-keyed path.\n- **SMC-23** — governs `review-dispositions-{PR}` + `engagement-record-review-{PR}` survival, write path (PR comments), and cross-tool fungibility.\n- **SMC-20** — extended to include `review` phase; governs `engagement-record-review-{PR}` cross-session resume semantics.\n\n### Gotcha: Gate fires post-judge, not inside judge\n\nThe classification gate fires in the **owning parent workflow** after receiving the judge output. The judge body (this skill's judgment pass) ends at the judge-rulings comment. Do not add gate logic inside the judge's verification or scoring steps — doing so would conflate two distinct phases and make re-running the judge independent of dispositions impossible.\n","\u003C!-- platform-assumptions: markdown skill guidance for VS Code custom agents in Agent Orchestra; assumes the calling agent owns intake routing, categorization policy, and handoff to implementation. -->\n\u003C!-- markdownlint-disable-file MD041 MD003 -->\n\n# Review Judgment\n\nReusable judgment method for a single referee pass over prosecution and defense.\n\n## When to Use\n\n- When prosecution findings and defense responses are both available\n- When a judge must independently verify claims before ruling\n- When a scored review summary and machine-readable ruling block are required\n- When external review comments have already been converted into a prosecution ledger\n\n## Purpose\n\nMake one final, evidence-backed ruling per finding. The goal is not to split the difference between prosecutor and defense, but to decide whether the proposed change would improve the code and to record that decision in a format the pipeline can consume.\n\nProsecution ledgers are coverage-first by contract: prosecution reports every finding with a statable failure mode, including low-confidence and low-severity ones. The judge is the `filter of record` — expect a wide, uneven ledger as normal input, not a sign that prosecution failed. Filtering happens here, at judgment, not upstream.\n\n## Single-Shot Judgment Workflow\n\n1. Read the prosecution finding, including severity, points, citation, and failure mode.\n2. Read the defense response and note whether it disproves, concedes, or cannot disprove the claim.\n3. Verify the evidence independently.\n4. Rule once: prosecution sustained or defense sustained.\n5. Emit score, confidence, and structured output.\n\nNo rebuttal rounds. Uncertain items still need a ruling.\n\n## Improvement Test\n\nReminder: because prosecution is coverage-first, this test will routinely see low-confidence and low-severity items — that is expected input, not noise to wave through. Rule on each the same way, evidence-first.\n\nAsk this first for every item:\n\n1. Will acting on this finding improve the code?\n\nOutcomes:\n\n- Yes -> accept the improvement\n- No -> reject it\n- Unclear even after verification -> reject it for now\n\nUncertainty is not a deferral bucket. If improvement cannot be shown with evidence, do not accept it.\n\n## Independent Verification Expectations\n\nBefore sustaining a finding:\n\n- Read the cited code, config, test, or document directly\n- Confirm the claimed defect actually exists\n- State what was verified, not just what the prosecutor said\n\nWhen the cited evidence does not support the claim, sustain the defense and explain the mismatch clearly.\n\n**POST-FIX-SCOPED**: when this verification pass follows post-fix targeted prosecution with mutation-tested verification, judgment coverage extends to every branch the fix commit modified, per the canonical post-fix scope constraint in `skills\u002Fvalidation-methodology\u002Freferences\u002Freview-reconciliation.md`. This note does not apply to a non-post-fix (main-review) judgment pass.\n\n## Scoring Model\n\nSeverity maps to points as follows:\n\n- `critical` or `high` -> 10 points\n- `medium` -> 5 points\n- `low` -> 1 point\n\nJudges may override the prosecution severity when verification shows the impact is lower or higher than claimed.\n\nConfidence guidance:\n\n- `high` -> direct structural proof, test output, or explicit code evidence\n- `medium` -> evidence leans one way but is not fully conclusive\n- `low` -> honest uncertainty after reasonable verification\n\n## Score Summary Output\n\nEmit a score table after ruling all findings.\n\n```markdown\n### Adversarial Review Score Summary\n\n| Finding     | Pass | Prosecution (severity, pts) | Defense verdict | Ruling                   | Confidence | Points    |\n| ----------- | ---- | --------------------------- | --------------- | ------------------------ | ---------- | --------- |\n| F1: {title} | {N}  | {severity} ({pts} pts)      | conceded        | ✅ Sustained             | high       | P+{pts}   |\n| F2: {title} | {N}  | {severity} ({pts} pts)      | disproved       | ❌ Defense sustained     | medium     | D+{pts}   |\n| F3: {title} | {N}  | {severity} ({pts} pts)      | disproved       | ✅ Prosecution sustained | high       | D-{2×pts} |\n\n**Totals**\n\n- Prosecutor: {sum of sustained prosecution points} pts ({N} findings sustained)\n- Defense: {net points after rejected-disproof penalties} pts\n- Judge rulings: {total} ({N} pending user scoring)\n```\n\nUse `—` in the Pass column when the prosecution mode does not carry a pass number.\n\n## Structured Judge Output\n\n### Sentinel emission (issue #441, D-new-4)\n\n**Immediately after the judge ruling is finalized** and before any pipeline-metrics persistence,\npersist an idempotent sentinel PR comment for the PR being reviewed:\n\n```text\n\u003C!-- review-judge-produced-{PR} -->\n```\n\nThis sentinel is separate from the judge-rulings comment. `skills\u002Fsession-memory-contract\u002Fscripts\u002Fpersist-marker.ps1` is the ONLY documented write path for this marker (family `review-judge-produced`, `post-new`, `sentinel-empty` validator adapter — pins an empty, marker-only payload) — never a hand-composed `gh pr comment` call. **What that rule buys is a single audited writer — not protection from `updated_at` advancement.** This family is `post-new`, so the write POSTs a fresh comment and advances nothing; the hazard arrives when a write replaces a **whole body** already carrying someone else's marker. See `skills\u002Fsession-memory-contract\u002Freferences\u002Fhandoff-markers.md` § What the write-path rule buys for the write-shape split. The script's own post-new idempotency comparison (compare the candidate against the latest marker match; normalized-identical → no-op) supersedes the separate `gh pr view {PR} --json comments` presence check this section previously directed: pass a body file whose only content is the marker line, and the script either posts a new sentinel or reports a `no-op` action against the already-present one.\n\n**Ordering rule**: sentinel comment → judge-rulings comment. The sentinel must be written first so that the warn-only hook can detect \"review completed but credit not yet written\" during the window between sentinel emission and PR-body update.\n\n**SMC-16 governance**: the sentinel marker `\u003C!-- review-judge-produced-{PR} -->` is governed by SMC-16. See `Documents\u002FDesign\u002Fsession-memory-contract.md`. Both Copilot and Claude judge runs write this sentinel; Code-Conductor reads it via the warn-only hook synthesis path.\n\n### Judge-rulings comment\n\nAfter the sentinel, emit the `judge-rulings` block in the same PR comment as the Markdown score summary:\n\n```yaml\n\u003C!-- judge-rulings\n- id: F1\n  judge_ruling: sustained\n  judge_confidence: high\n  points_awarded: P+10\n- id: F2\n  judge_ruling: defense-sustained\n  judge_confidence: medium\n  points_awarded: D+5\n-->\n```\n\nKeep the Markdown score summary and the `judge-rulings` block together in the same response payload. On GitHub, keep them in the same PR comment rather than splitting them across separate comments. **This comment does not include `\u003C!-- code-review-complete-{PR} -->`** — that marker is retired as of issue #441 Step 11; Code-Conductor reads `credits[]` from the `\u003C!-- pipeline-metrics -->` PR body block directly. **This comment does not include pipeline-metrics body emission** — that is owned by Code-Conductor's `## Pipeline Metrics` emitter at PR creation time.\n\n### Phase-containment emission\n\nIn the same PR comment as the `\u003C!-- judge-rulings ... -->` block, emit one `\u003C!-- phase-containment-{PR} -->` block per sustained finding (`judge_ruling: sustained`). The two block families are different shapes and are not interchangeable: `judge-rulings` (above) stays **bare** — one unclosed `\u003C!-- judge-rulings ... -->` comment — while `phase-containment` is always **paired** — a self-closed `\u003C!-- phase-containment-{PR} -->` open tag followed by plain-text YAML fields and a separate `\u003C!-- \u002Fphase-containment-{PR} -->` close tag, because the close tag powers `Get-PhaseContainmentBlock`'s pair-matching malformation detection (issue #772 D6). **Scope boundary**: this section and its Observer variant below document the block *shape* only — routing a given finding's emission to the shared `Add-JudgeRulingsBlock`\u002F`Get-PhaseContainmentBlock` helper machinery is deferred to AC6, not covered here.\n\n- `finding_key`: `code-review:{stable_finding_key}`\n- `introduced_phase`: set by explicit agent judgment — no default; reason which phase originated this defect\n- `catchable_phase`: set by explicit agent judgment — no default; reason which phase was the earliest this defect could have been caught\n- `caught_stage: code-review`\n- `escape_distance`: recomputed as `3 - ordinal(catchable_phase)` (code-review projection = 3; phase ordinals: experience=0, design=1, plan=2, implementation=3)\n- `severity`, `systemic_fix_type`, `category`: carry forward from the finding\n- `apparatus_meta: false` unless a stated criterion justifies `true`; when `apparatus_meta: true`, the entry is audited\n- `appended_at`: stamp the current UTC instant in the strict `yyyy-MM-ddTHH:mm:ssZ` form (863 M1 fix) — this block is hand-authored directly into the PR comment (no script primitive writes it on this surface), so the judge\u002Fagent authoring the block is responsible for stamping this field itself\n\nA fully literal canonical example, for a sustained code-review finding on PR 879:\n\n```markdown\n\u003C!-- phase-containment-879 -->\nfinding_key: code-review:gh-1234\nintroduced_phase: implementation\ncatchable_phase: implementation\ncaught_stage: code-review\nescape_distance: 0\nseverity: high\nsystemic_fix_type: instruction\ncategory: security\napparatus_meta: false\nappended_at: 2026-07-18T22:20:00Z\n\u003C!-- \u002Fphase-containment-879 -->\n```\n\n#### Observer emission variant (post-review-observer)\n\n`reviewer_source` values referenced below are resolved per § `reviewer_source` Lookup Order (later in this document) — read that section first if the resolution mechanism itself is in question; this rule only consumes the already-resolved value.\n\nWhen a sustained finding's `reviewer_source` resolves to a real external identity (not the reserved `local` sentinel) and its `internal_match.match_status` is `novel`, emit the observer variant of the phase-containment block instead of the standard code-review block above — one block per finding, never both:\n\n- `finding_key`: `post-review-observer:{stable_finding_key}` — this exact prefix is load-bearing (M26): `Get-EmissionGap` attributes a block to a surface by checking the `finding_key` prefix, so a block emitted with the wrong prefix is invisible to, or miscounted by, the reconciliation sweep even though it would pass schema validation.\n- `caught_stage: post-review-observer`\n- `escape_distance`: recomputed as `4 - ordinal(catchable_phase)` (post-review-observer projection = 4; same phase ordinals as above)\n- `introduced_phase`, `catchable_phase`, `severity`, `systemic_fix_type`, `category`, `apparatus_meta`: same setter rule and carry-forward as the code-review block above\n\nA fully literal canonical example, for a sustained post-review-observer finding on PR 879 — same paired shape as the code-review block above, distinguished by the `post-review-observer:` `finding_key` prefix, `caught_stage`, and its own `escape_distance` projection:\n\n```markdown\n\u003C!-- phase-containment-879 -->\nfinding_key: post-review-observer:gh-5678\nintroduced_phase: implementation\ncatchable_phase: implementation\ncaught_stage: post-review-observer\nescape_distance: 1\nseverity: medium\nsystemic_fix_type: skill\ncategory: architecture\napparatus_meta: false\nappended_at: 2026-07-18T22:20:00Z\n\u003C!-- \u002Fphase-containment-879 -->\n```\n\n**Novel-gating is a trinary, not a two-way rule (M25)**:\n\n- `reviewer_source` is the reserved `local` sentinel → standard `code-review` block, unchanged.\n- `reviewer_source` is a resolved real external identity AND `internal_match.match_status: novel` → observer block, which REPLACES the standard block. One defect, one block, never both.\n- `internal_match.match_status` is `duplicate` or `ambiguous`, OR `reviewer_source` is the `unresolved` lookup-failure sentinel → NEITHER block is emitted.\n\nWriting the unqualified two-way version of this rule — \"any resolved external identity gets an observer block\" — is wrong: it would emit an observer block for a `duplicate`-matched finding, double-counting one defect as both a catch-side overlap and an escape-side miss.\n\n**Dispatch is exact-equality only (M40)**: test `reviewer_source -eq 'local'`, never `-like`, `-match`, or other containment-style matching. A real GitHub login literally named `local` is normalized to `ext-local` by the intake process specifically so it can be told apart from the reserved `local` sentinel (see `skills\u002Fcode-review-intake\u002FSKILL.md` § GitHub Review Mode, around line 21, for the normalization rule) — a containment-style dispatch would misclassify `ext-local` as pipeline-native and silently delete a real escape.\n\n**Setter rule**: `catchable_phase` and `introduced_phase` must each be set by explicit agent judgment with no default — the agent must reason about which phase was the earliest in which this specific defect was catchable. Validate each block against `skills\u002Fcalibration-pipeline\u002Fschemas\u002Fphase-containment.schema.json`.\n\n**Emission check (hub maintainers only)**: after posting the `judge-rulings` PR comment with its phase-containment blocks, run `pwsh .\u002F.github\u002Fscripts\u002Fphase-containment-emission-check.ps1 -Pr {N}` and treat its output as advisory — warn-only, never blocking. The repo-relative script path does not resolve from a consumer repo's CWD, so this nudge applies only when working in the Agent Orchestra hub repo itself; see the script header for the full contract.\n\n**Detective-sample extension**: the Code-Critic detective-sample (from `agents\u002FCode-Critic.agent.md:110`) is extended to sample `apparatus_meta: true` entries and `catchable_phase == caught_stage` entries for plausibility review.\n\nField values:\n\n- `judge_ruling`: `sustained` or `defense-sustained`\n- `judge_confidence`: `high`, `medium`, or `low`\n- `points_awarded`: `P+{pts}`, `D+{pts}`, or `D-{2×pts}`\n\n## General Judgment Workflow\n\n### Default Path\n\n- Verify the finding\n- Rule on it\n- Categorize it according to the calling agent's policy\n- Stop after emitting the judgment output\n\n### Evidence-First Rejection\n\nWhen rejecting, cite the reason explicitly:\n\n- The code contradicts the finding\n- Tests or types already guarantee the claimed invariant\n- The documented design makes the proposed change harmful\n- The reviewer's cited evidence is factually wrong\n\n### Escalation Boundary\n\nJudgment does not implement fixes. It produces the ruling and the evidence package needed for the owning orchestrator to route accepted work.\n\n## Related Guidance\n\n- Load `adversarial-review` for prosecution and defense methodology\n- Load `code-review-intake` when GitHub review retrieval and ledger construction are the main problem\n- Read [references\u002Fmultiline-capture-audit.md](references\u002Fmultiline-capture-audit.md) before adding or changing any command-capture site in this skill, and before trusting an `ac_cross_check` recorded before plugin `3.9.1` (issue #977)\n\n## Gotchas\n\n| Trigger                                   | Gotcha                                                   | Fix                                                          |\n| ----------------------------------------- | -------------------------------------------------------- | ------------------------------------------------------------ |\n| The judge repeats the prosecutor verbatim | The ruling becomes a rubber stamp instead of independent | Read the cited artifact directly and state what was verified |\n\n| Trigger                             | Gotcha                                                         | Fix                                                                |\n| ----------------------------------- | -------------------------------------------------------------- | ------------------------------------------------------------------ |\n| A ruling omits the `judge-rulings` block | Downstream consumers can miss completion or fail to route cleanly | Emit the `judge-rulings` block immediately after the score summary in the same payload; ensure the `\u003C!-- review-judge-produced-{PR} -->` sentinel was written as a separate PR comment before |\n\n## Close-Out Record Amendment\n\n**Single home, every lane.** This section is the one place the close-out record's amendment rule is stated. Every lane that runs a judge reaches it by naming it from that lane's own entry document:\n\n| Lane | Entry document that names this section |\n| --- | --- |\n| `\u002Forchestra:review` | `commands\u002Forchestra-review.md` |\n| `\u002Forchestra:review-lite` | `commands\u002Forchestra-review-lite.md` |\n| `\u002Forchestra:review-judge` | `commands\u002Forchestra-review-judge.md` |\n| `\u002Freview-github` | `commands\u002Freview-github.md` |\n| `\u002Forchestrate`, bare `review` via `\u002Fcode-conductor` | `commands\u002Forchestrate.md` and `commands\u002Fcode-conductor.md`, which mandate reading `agents\u002FCode-Conductor.agent.md` § Review Reconciliation Loop |\n| `\u002Fgoal-run` | `commands\u002Fgoal-run.md`, whose body `agents\u002FGoal-Run.agent.md` § Stage 3 dispatches the `standard` adapter — panel, defense, **and judge** |\n\n**Out of population, stated rather than omitted.** AC8's standard is that an excluded path carries its reason, so each is named here rather than left to silence:\n\n- **`\u002Forchestra:review-prosecute` and `\u002Forchestra:review-defend`** — both stop before judge by their own contract, so no pass on them can sustain a finding.\n- **`\u002Fspine-run` and `\u002Forchestra:spine`** — Spine-Runner **runs no judge**. Its frame port `review` is dispatched `skill-only` (`agents\u002FSpine-Runner.agent.md` § Dispatch Table), which *inspects the PR body's `pipeline-metrics` block to verify a review row another lane emitted*; it never dispatches `code-review-response`. Its only adversarial-review mention is the Senior Engineer adversarial-pattern guard, which **halt-returns** rather than dispatching. A verifier of someone else's judge output sustains no findings of its own, so it triggers no amendment — the lane that produced those findings already did. Note the near-miss: `#552 D11`, quoted below to ground this step's executor, names \"Code-Conductor **or spine-runner**\" as orchestrating executors. That is about who may perform non-adversarial post-judgment work when a judge has run, not a claim that spine-runner runs one.\n- **The `design-challenge` and `proxy-github` adapters** — both declare no judge stage in their contracts.\n\nA path that reaches a judge and is missing from the table above is a gap in this section, not a lane that owes nothing — add it rather than reasoning about whether it \"really\" applies.\n\nEvery other surface that mentions the amendment cites this one rather than restating it: two statements of one trigger are two things to keep in sync, and the drift between them is invisible until one is already wrong.\n\n**Executor: the owning parent workflow — never the judge subagent.** The judge's scope stops at its own emission — `agents\u002FCode-Review-Response.agent.md`: *\"Scope boundary: Code-Review-Response owns the sentinel and the `judge-rulings` comment only\"* — and non-adversarial post-judgment work belongs to *\"the orchestrating executor (Code-Conductor or spine-runner), **not** Code-Critic or Code-Review-Response\"* per adversarial-independence `#552 D11` (`skills\u002Fvalidation-methodology\u002Freferences\u002Freview-reconciliation.md` § Response Commit & Push). Do not place this step inside the judge's verification, scoring, or emission steps.\n\n**Why this file.** It already houses the post-judge steps the *parent* owns and each lane's entry document names by section — § Post-Judge Disposition Gate below is one, and is explicitly *\"owned by the calling workflow … not by this skill's judgment pass\"*. The close-out record is an issue-level account of the run that predates and outlives the review, not a review response, so this section is not the *\"response-location policy\"* this skill's frontmatter scopes away: it says nothing about where review responses are posted.\n\n### When it fires\n\n**Once per judge pass, on that pass's emission** — as soon as the parent holds that pass's judge output, and **before** the Post-Judge Disposition Gate below. Not after dispositions, and not after fixes.\n\n\"That pass's judge output\" is whatever form the pass emitted, not a fixed artifact list. On a pull-request target it is the `\u003C!-- review-judge-produced-{PR} -->` sentinel, the `judge-rulings` block, and its phase-containment blocks. On a **non-PR review target** none of those three exists — all are PR-comment-scoped (§ Sentinel emission, § Judge-rulings comment, § Phase-containment emission), and `agents\u002Fcode-review-response.md` § Persistence routes such a run to emit the score summary and `judge-rulings` payload directly in chat instead. There the trigger is that chat payload. Do not read the PR-shaped list as a precondition: a run that waits for artifacts its own lane never produces skips forever, which is the failure this section exists to remove.\n\nThe moment is the judge's emission because that is where this step's input is produced: the record's item 1 is *\"one line per sustained finding — where it was introduced, where it was catchable, where it was caught\"* (`skills\u002Fpost-pr-review\u002FSKILL.md` § 9), which is the phase-containment blocks' own grain. Nothing downstream adds to it — a disposition rules `incorporate | dismiss | escalate`, and a fix lands or does not; neither changes where a defect was introduced, was catchable, or was caught. Firing here therefore does not make the record describe a state that never became true. It also means the amendment text must not describe fix status or disposition status; if it does, the wording is wrong for the moment, not the moment wrong for the wording.\n\n**Every judge pass, not only the first.** A run can reach a judge more than once — the main pass and the post-fix pass (`skills\u002Fvalidation-methodology\u002Freferences\u002Freview-reconciliation.md` § Post-Fix Targeted Prosecution Pass), and a standalone `\u002Forchestra:review-judge` re-run is a pass of its own. The post-fix pass by construction carries the latest findings a run has, which is exactly the population a late amendment exists for.\n\n### Resolving the issue\n\nThis step is the only issue-keyed thing in an otherwise PR-keyed or target-keyed pipeline, so resolve the issue first, in this order:\n\n1. **When the review target is a pull request** — `gh pr view {PR} --json closingIssuesReferences`. This is the PR's own authoritative statement of what it closes. **Zero results**: fall through to route 2. **One result**: that is the issue. **More than one**: evaluate each separately rather than picking one.\n2. **Otherwise, the active issue id the parent already holds.** Every lane's dispatcher carries one when the run has one — the local-review commands pass *\"active issue id if available\"* as pre-dispatch context, and Code-Conductor, Goal-Run and Spine-Runner each run against an issue. This is the ordinary route on a run with **no** pull request, and it is why this step is not a documented no-op on lanes that review a non-PR target.\n3. **Neither available** — a degraded input, not the normal path. Report `no-issue-resolved` per § Advisory, and reported.\n\n**When both routes resolve and disagree, the pull request wins, and say so.** The parent can hold a different issue than the one the PR closes — a designed parent while the PR closes one of its chunks is the ordinary case, and a chunk's findings appended to the parent's record is a public misattribution on someone else's comment. Amend the issue route 1 returned, and name the held issue in the outcome so the divergence is visible rather than silently resolved.\n\n### Whether it applies\n\nPer resolved issue, **applies only** when both hold:\n\n- the issue carries a lawful **open-for-work affirmation record** — the lookup is `skills\u002Fpost-pr-review\u002FSKILL.md` § 9 \"How to check\". Use it rather than a substitute: `gh issue view --json comments` carries no `updated_at` and so cannot tell a lawful record from one edited after creation, failing in the permissive direction; **and**\n- a close-out record already exists on that issue, recognised by its first line (§ 9 \"Action\").\n\n**Does not apply otherwise**, and this step never manufactures a first record. It re-keys *when an existing record is amended*; it does not change which issues owe one, nor the two moments at which one is written — those stay at `skills\u002Fplan-authoring\u002FSKILL.md` § The close-out obligation on an affirmation-record issue and at § 9.\n\n**`not-applicable` on a first-PR run is expected, not a coverage gap.** On a run that opens the issue's first pull request, the review completes before the pull request is created (`agents\u002FCode-Conductor.agent.md` § Review Completion Gate), and the record is written after that review at the pre-PR moment stated in `skills\u002Fplan-authoring\u002FSKILL.md` § The close-out obligation on an affirmation-record issue — this section does not restate that moment. So at judge time no record exists yet, and the record written moments later already accounts for this pass's findings. Nothing is owed and nothing is missed. This step's live population is a judge pass on an issue whose record **already** exists: a review of an already-open pull request, a second pull request, a post-fix pass after the record was written, or a resumed run. Read a first-PR `not-applicable` as the rule working, and do not \"fix\" it by making the step manufacture a record.\n\n### What it does\n\nWhen this pass sustained findings the existing record does not account for, **amend that record in place** — do not post a second one, which would read as two close-outs with no way to tell which is current — and say the amendment is one, per § 9.\n\n**The write route, because this section owns it and the obvious shortcuts mis-target.** The record has no marker family and no `persist-marker.ps1` write path, deliberately (§ 9), so the transport is a plain comment PATCH keyed by the comment's own id:\n\n1. The lawfulness lookup above already paginates the issue's comments through `gh api repos\u002F{owner}\u002F{repo}\u002Fissues\u002F{ID}\u002Fcomments --paginate`. Take the **`id`** of the comment whose first line is the record's, from that same read.\n2. **Re-read that comment's live body immediately before writing** — `gh api repos\u002F{owner}\u002F{repo}\u002Fissues\u002Fcomments\u002F{comment_id} --jq .body`. The record accumulates appended lines, and a body composed from an earlier read silently drops anything appended in between (`skills\u002Fsafe-operations\u002Freferences\u002Fgit-and-gh-traps.md`).\n3. Append the amendment line to **that** live body and write it back: `gh api --method PATCH repos\u002F{owner}\u002F{repo}\u002Fissues\u002Fcomments\u002F{comment_id} -F body=@{file}`.\n\n**Two shortcuts are forbidden here, both documented mis-write traps** (same reference): `gh issue comment --edit-last` targets the author's most recent comment on the issue, which is whatever the run posted last rather than the record; and `Find-OrUpsertComment`'s `-Body` **replaces the existing body verbatim on the PATCH path** — it never re-reads and merges — so pointing it at a record that accumulates appended lines destroys them. That second hazard is about the *write*, not the *selection*: #1031 narrowed the selector to line-1-exact, and the close-out record is not marker-keyed at all, so it was never a candidate for that upserter. A mis-targeted or body-replacing write here damages a public comment whose first writer is often a person, which is why the id comes from the same read that recognised the record and never from a guess.\n\n**Bounded by what the record already carries.** Before amending, read it and name only the sustained findings it does not already account for — whether in its original items or in any amendment line already on it, whichever run appended that line. When this pass adds none, report `no-new-findings` and append nothing. Several near-identical dated lines on one record is the duplicate account this bound exists to prevent, not evidence that the step fired.\n\n**What \"already account for\" is matched on.** The record's item-1 lines are prose, and the judge's findings carry a `stable_finding_key` (§ Stable Finding Key), so the two do not share a key. Match on the **finding's own identity as the record states it** — the file-and-symptom pair a record line names, not string equality with a finding title. When a prior line plainly describes this pass's finding, it is accounted for; when you cannot tell, say so in the outcome rather than appending a line that may be a duplicate. Recording an uncertainty is cheap; a second near-identical dated line is the harm.\n\n### Advisory, and reported\n\n**A skipped or failed amendment never halts the loop** — nor the run, on a lane that has no loop. Emit the loud literal and carry it into this lane's accountability channel:\n\n- `⚠️ close-out record amendment skipped — {reason}`\n\n**The accountability channel, named per lane.** This step is advisory, so its report is the only trace it leaves; a lane with no named channel lets a skip disappear into a report that satisfies every other requirement. **The channel is the run's own terminal report** — the block a lane already returns or posts when it finishes — and every lane has one:\n\n| Lane | Channel |\n| --- | --- |\n| `\u002Freview-github` (GitHub intake) | Response Summary item 5 (`skills\u002Fvalidation-methodology\u002Freferences\u002Freview-reconciliation.md` § Response Commit & Push) |\n| `\u002Forchestrate`, bare `review` via `\u002Fcode-conductor` — **existing-PR path** | the same Response Summary item 5 |\n| `\u002Forchestrate` — **new-PR path** (Conductor Step 4 `gh pr create`) | the **Close-out record amendment** row of the PR-body pipeline-metrics block. That path fires no `persist-changes` and assembles no Response Summary — the two push sites are mutually exclusive — so item 5 does not exist there and pointing at it would name a slot the run never produces. |\n| `\u002Forchestra:review`, `-lite`, `-judge` (local review) | a `Close-out record amendment:` line in the **judgment payload the command returns**, immediately after the `judge-rulings` block and beside the disposition-gate outcome. It is the parent's own line, appended to the payload — it does not alter the judge output the command returns unchanged. |\n| `\u002Fgoal-run` | the Stage 3 report; on a halt, the typed halt report's own body |\n\n**Report one outcome per resolved issue, per judge pass** — not one per pass. Route 1's \"more than one: evaluate each separately\" means a single pass can resolve to several issues (a PR closing more than one), and each gets its own applies-only check and its own outcome; a pass that touched three issues reports three outcome lines, one per issue id, even when they differ — `amended` on one and `not-applicable` on another is the ordinary shape, not a conflict to resolve into a single line. State how many passes fired and, within each pass, how many issues it resolved. The outcomes partition per issue — exactly one applies to each:\n\n- `amended` — naming what it added, and naming the held issue when it differed from the PR's.\n- `no-new-findings` — the record already accounts for every finding this pass sustained.\n- `not-applicable` — an issue **was** resolved, and it carries no affirmation record or no existing record. Name which.\n- `no-issue-resolved` — route 3: neither an issue nor a pull request was available. Carries the loud literal below.\n- `skipped` — the step was reached but could not complete (a failed read, a failed write, a lookup error). Carries the loud literal below with its reason.\n\n**Not a gate.** Nothing refuses a judgment, a fix, a pull request, a merge, or a close because this step did not run, and nothing re-checks it once the run ends. A blocking check here would be the detector `Documents\u002FDesign\u002Fopen-for-work.md` § The close-out habit deliberately declined, wearing another name.\n\n## Post-Judge Disposition Gate\n\n### Purpose\n\nAfter the judge emits its rulings (sentinel + judge-rulings comment), the owning parent agent runs the review-disposition engagement gate over the judge-sustained findings. This prevents cognitive surrender at the PR code-review verdict point: the engineer must consciously disposition each sustained finding rather than having the agent assume outcomes.\n\nThis gate is owned by the calling workflow (e.g. `\u002Forchestra:review-judge`, `\u002Forchestra:review`), not by this skill's judgment pass. The judgment pass ends with the judge-rulings comment. Disposition begins immediately after.\n\n### When to Run\n\nThis section splits two previously-conflated rules — marker emission and the maintainer-interaction gate — so the always-emit behavior below does not read as contradicting the sustained-only scoping that follows it.\n\n**Marker emission** (persisting `\u003C!-- review-dispositions-{PR} -->` and `\u003C!-- engagement-record-review-{PR} -->`) fires on **every** judge pass, including a zero-sustained pass:\n\n- After the `\u003C!-- review-judge-produced-{PR} -->` sentinel is confirmed written\n- Even when no finding was judge-sustained this pass — per the M9 coverage semantics above (§ Scope, around line 277: \"coverage means measurement, not presence\"), a zero-sustained pass still emits both markers, with `entries: []` on the dispositions marker\n\n**Disposition gate** (the maintainer-interaction step, classification into routine vs load-bearing) fires only when there are findings to disposition:\n\n- Over the judge-sustained findings from the `\u003C!-- judge-rulings ... -->` block\n- Only for sustained findings (`judge_ruling: sustained`); defense-sustained findings (`judge_ruling: defense-sustained`) are skipped silently (not disposition-gated)\n- When there are zero judge-sustained findings, no finding enters classification and the gate never fires this pass — but marker emission (above) still runs\n\n### Classification\n\nFor each judge-sustained finding, run the solution-authoring classification gate (`skills\u002Fsolution-authoring\u002FSKILL.md § Rule: Classification gate`) to determine whether it is **load-bearing** or **routine**.\n\nA finding is **load-bearing** iff all three legs pass:\n\n1. **Reversibility** — acting on this finding would change a published or durable artifact (source file, test, config, doc, or skill).\n2. **Non-inheritance** — no specific inherited artifact (prior design decision, existing AC, approved plan statement, locked methodology rule) already settles the required action. The agent MUST attempt to cite one before declaring non-inheritable.\n3. **Audit-plausibility** — the agent can write a substantive `disposition_rationale` sentence.\n\nFailing any leg collapses the finding to **routine**. The tier rule from `solution-authoring § Applying the gate to adversarial-review dispositions` applies: load-bearing adversarial-review dispositions use the **escalation tier** decision brief; routine findings are recorded silently.\n\n### Stable Finding Key\n\nBefore the gate fires, compute a `stable_finding_key` for each finding. This key must survive re-reviews of the same PR. Use the first available:\n\n1. **GitHub comment ID** — if the finding originated from a GitHub review comment, use its comment ID as the key. A comment that yields a single finding keeps the tier-1 key `gh-{comment_id}` — unchanged, and backward compatible with dispositions persisted before this discriminator existed. When one comment yields multiple distinct findings (the code-review-intake skill's split rule — see `skills\u002Fcode-review-intake\u002FSKILL.md`), each finding's key becomes `gh-{comment_id}-{normalized-title-hash[:8]}`, reusing tier 2's normalization convention below to disambiguate findings sharing a comment ID.\n2. **file:line:hash** — normalize the finding title (lowercase, remove punctuation, replace spaces with hyphens), then form `{relative_file}:{line}:{normalized-title-hash[:8]}` where hash is the first 8 hex chars of SHA-256 over the UTF-8 normalized title.\n3. **finding_id fallback** — if neither is available, use the sequential `finding_id` (e.g. `F1`) with a `warn:` prefix to signal instability: `warn:F1`.\n\nThe `stable_finding_key` is what the resume-read mechanism uses to detect prior dispositions across re-reviews of the same PR (see § Stable-Key Resume below).\n\n### `reviewer_source` Lookup Order (GitHub-Sourced Findings Only)\n\nThis lookup order applies only to findings that originated from an external GitHub review comment. Pipeline-native findings (local prosecution\u002Fdefense\u002Fjudge findings) never enter this lookup procedure — the disposition-recording sites below write `reviewer_source: local` directly for those findings, without evaluating any of the three tiers.\n\nBoth disposition-recording sites below (§ Routine Findings and § Load-Bearing Findings) write a `reviewer_source` field. For GitHub-sourced findings, resolve its value in this order:\n\n1. **In-context intake ledger** — when the session is continuous with the GitHub-intake pass that built the finding, read the `reviewer_source` value the intake skill (`skills\u002Fcode-review-intake\u002FSKILL.md`) already recorded for this finding.\n2. **`gh api` re-derivation** — otherwise (standalone or resumed judge pass), extract the `comment_id` embedded in the finding's `stable_finding_key` and re-derive the reviewer identity via `gh api`, trying the PR-review-comment, issue-comment, and review endpoints in turn. Parse rule: the `comment_id` is the numeric run immediately after the `gh-` prefix, up to the next `-` or end of string; this covers both the single-finding `gh-{comment_id}` key shape and the multi-finding `gh-{comment_id}-{hash}` key shape identically. Normalize the re-derived login per the canonical `reviewer_source` normalization rule in `skills\u002Fcode-review-intake\u002FSKILL.md` § GitHub Review Mode (lowercase, `.login`-only, trailing `[bot]` strip, quoted-scalar output, reserved-value\u002F`ext-`-prefix collision escape) — this site does not restate that mechanism.\n3. **`unresolved` sentinel** — on any failure (comment deleted, API error, no comment ID embedded in the key to re-derive from), write the sentinel `unresolved`. Never write `local` on lookup failure — `local` is reserved exclusively for pipeline-native (non-GitHub-sourced) findings, which are written directly per the guard above and never fall through to this sentinel.\n\n### `internal_match` Writer Rule (GitHub-Sourced Findings Only, DD2)\n\n`internal_match.match_status` and the PR-level `external_sources_reconciled` field are **written by the judge**, not merely consumed — § Observer emission variant (above) only ever reads the already-resolved value; this section is the writer contract those readers depend on.\n\nAt reconciliation time — with **both** the internal prosecution ledger (this PR's own pipeline-native findings) and the external review's findings in context simultaneously, per design decision DD2 — the judge sets `internal_match.match_status` on **every** external-source disposition entry (any entry whose `reviewer_source` resolved to a value other than `local`):\n\n- **`duplicate`** — the external finding describes the same defect as an already-caught internal finding. Also write `matched_finding_key` with that internal finding's `stable_finding_key`.\n- **`novel`** — the external finding has no matching internal finding; the pipeline missed it. This is what the observer emission variant later turns into a `post-review-observer` escape block.\n- **`ambiguous`** — the judge cannot confidently resolve a match either way after verification. Excluded from `n₂`\u002F`m`\u002Fcoverage math downstream (Seam Specification) — an honest \"could not resolve\" must never be forced into `duplicate` or `novel`.\n\n`reviewer_source: local` entries (pipeline-native findings) never carry `internal_match` — the field only applies to external-source entries. Field order matters: `internal_match` is written **before** `disposition_rationale` (M42) so a `disposition_rationale` block-scalar's free text can never be mistaken for the real field by a downstream line-regex parser.\n\n**Scope (post-fix batch 4, issue #854 s6; corrected G-CR8, PR #859 GitHub-review post-fix): only when this pass ran in GitHub Review Mode.** The PR-level `external_sources_reconciled` field records that an external review was actually reconciled against — it is meaningless, and actively misleading, on a pass that never had one. Emit it once in the same posted marker **only when this pass is GitHub Review Mode** — determined from the session context\u002Fcommand that invoked this pass (`skills\u002Fcode-review-intake\u002FSKILL.md` § GitHub Review Mode (Proxy Prosecution Pipeline) — the `\u002Freview-github` proxy-prosecution path that ingests GitHub-sourced findings), never from entry presence. Within a GitHub Review Mode pass, emit it **even when empty** (`external_sources_reconciled: []`): a genuinely zero-finding external review is a legal, required coverage record (M9 — \"coverage means measurement, not presence\"), and omitting the field on a GitHub Review Mode pass is indistinguishable downstream from \"this PR was never measured at all.\" A zero-finding external review pass has zero entries by definition, so a detection test keyed on entry presence (e.g. \"at least one entry this pass whose `reviewer_source` resolved to a value other than `local`\") can never fire on exactly the case M9 requires the field to be emitted for — GitHub Review Mode is a property of the SESSION, not a derived property of what got written. On a purely internal-only pass — plain `\u002Forchestra:review`, no GitHub-sourced findings ingested this pass — **omit the field entirely**. Do not emit `external_sources_reconciled: []` on an internal-only pass: that would falsely claim an external review was reconciled and never occurred, manufacturing a measured-looking zero that was never actually measured (the exact false-clean coverage vector this issue exists to eliminate).\n\n### Routine Findings — Silent Recording\n\nFor routine findings, the agent records the disposition silently in the `review-dispositions-{PR}` accumulator without asking the maintainer. Use `schema_version: 4` (current emission format); write `severity`, `stage`, and `reviewer_source` (the reviewer identity or class that produced the finding — use `local` for pipeline-native prosecution\u002Fdefense\u002Fjudge findings; external identities are resolved per § `reviewer_source` Lookup Order above) for all entries, `internal_match` for external-source entries (§ `internal_match` Writer Rule above), and include `ac_cross_check` for any `dismiss` or `defer` entry with severity ≥ medium. Remember the PR-level `external_sources_reconciled` field once per posted marker (§ `internal_match` Writer Rule above; § Persistence — Ordering below shows the full marker shape):\n\n> **Pre-condition**: for any `dismiss` entry with severity ≥ medium, run the AC cross-check (see § AC Cross-Check — Blocking Pre-Condition) before writing this entry.\n\n```yaml\n- stable_finding_key: \"src\u002Ffoo.ts:42:null-check-missing-a1b2c3d4\"\n  finding_id: F2\n  pass: 1\n  disposition: incorporate   # or dismiss — agent's judgment based on judge ruling\n  classification: routine\n  severity: medium           # v3: required field\n  stage: code-review         # v3: required field\n  reviewer_source: local     # v3: writer always emits this field (writer obligation); the reader\u002Fconsumer treats an absent field as local for backward compat with pre-v3 entries (reader-semantics fallback) — local is reserved for pipeline-native findings; external identities resolve per § reviewer_source Lookup Order\n  disposition_rationale: \"Trivial null-guard already required by the existing type contract; no maintainer choice required.\"\n  artifact_citation: \"src\u002Ftypes\u002Findex.ts:18 (NonNullable\u003CT> constraint)\"\n```\n\nThe agent chooses `incorporate` or `dismiss` based on the judge ruling direction. Routine findings do not fire the gate; they feed directly into the accumulator.\n\n### Load-Bearing Findings — the disposition question\n\nFor load-bearing findings, render an **escalation-tier decision brief** (three required elements, all present before the option list):\n\n1. **Concrete element → current state**: what the current code\u002Fartifact actually does, with evidence (file:line or artifact citation from the finding).\n2. **Decision setup → the conflict**: why this finding's proposed change conflicts with or extends beyond the current state.\n3. **Conditional misconception → customer failure mode**: the concrete failure the engineer would cause by taking the wrong path.\n\nThen ask, with exactly these options:\n\n- `Incorporate — apply the fix` (Recommended if prosecution was sustained on strong evidence)\n- `Dismiss — the finding does not warrant a change`\n- `Escalate — this needs a design decision or a separate issue`\n- `Decline engagement — proceed without classification`\n\nCapture the engineer's choice verbatim.\n\n> **Pre-condition**: if the engineer chooses `dismiss` with severity ≥ medium, run the AC cross-check (see § AC Cross-Check — Blocking Pre-Condition) before writing this entry.\n\nRecord as (v4 format — include `severity`, `stage`, `reviewer_source`, `internal_match` for external-source entries (§ `internal_match` Writer Rule above), and `ac_cross_check` for dismiss\u002Fdefer entries with severity ≥ medium):\n\n```yaml\n- stable_finding_key: \"src\u002Fauth\u002Fsession.ts:88:token-expiry-not-checked-b7c1a2f3\"\n  finding_id: F1\n  pass: 2\n  disposition: incorporate   # or dismiss or escalate per engineer choice\n  classification: load-bearing\n  severity: high             # v3: required field\n  stage: code-review         # v3: required field\n  reviewer_source: local     # v3: writer always emits this field (writer obligation); the reader\u002Fconsumer treats an absent field as local for backward compat with pre-v3 entries (reader-semantics fallback) — local is reserved for pipeline-native findings; external identities resolve per § reviewer_source Lookup Order\n  disposition_rationale: \"Engineer chose incorporate: the expiry check was confirmed missing and the fix is bounded to one function.\"\n```\n\nThis example's `reviewer_source: local` means no `internal_match` is written (pipeline-native). For a load-bearing finding whose `reviewer_source` resolved to an external identity, add `internal_match: { match_status: ... }` before `disposition_rationale`, same shape as the routine external-source example above.\n\n### Escalate Semantics\n\nWhen an engineer selects `Escalate`:\n\n1. Record `disposition: escalate` in the entry.\n2. Emit a concise escalation note inline (not a question): `Finding {finding_id} escalated — recommend filing a follow-up issue for: {finding title}. Proceeding without implementing this finding.`\n3. Do NOT implement the finding in the current PR. The current work continues; the escalated finding is noted for tracking.\n4. Record the `escalate` outcome faithfully in `review-dispositions-{PR}` and in the L0 gate-decision token.\n\n### AC Cross-Check — Blocking Pre-Condition\n\nBefore writing any entry with `disposition: dismiss` or `disposition: defer` **and** `severity` ≥ medium — whether the entry originates from a code-review finding (`stage: code-review`) or a CE Gate defect deferral (`stage: ce`) — the agent MUST complete an AC cross-check:\n\n1. Call `Get-AcRefsFromIssue -IssueNumber {parent_issue}` to extract AC file-path references (`{ac_refs}`, ARM 1) **and** `Get-AcTermsFromIssue -IssueNumber {parent_issue}` to extract behavioral AC terms (`{ac_terms}`, ARM 2). **Both**, not one: the two arms are independently sufficient — a file-path match alone reaches `force-accept` without any term match — so calling only the term helper silently disables half the gate. An earlier revision of this step named only the term helper while step 2 went on to pass an `{ac_refs}` that nothing had produced (issue #977).\n2. Call `Get-StructuralVerdict -Finding {finding} -PrFileSet {pr_files} -AcRefs {ac_refs} -RepoRoot {repo_root} -AcTerms {ac_terms}` to obtain the `ac_cross_check` object.\n3. Write the returned `ac_cross_check` object into the disposition entry.\n\n**This pre-condition is blocking.** The gate MUST NOT commit a `dismiss` or `defer` entry with severity ≥ medium that has a null or absent `ac_cross_check`. If either helper returns empty (no AC section, or no parseable tokens), the cross-check still runs — pass `@()` for that arm; when **both** are empty the verdict's `ac_cross_check.source` will be `no-ac-section` and `routed` will be `defer`.\n\n**Low-severity exemption.** Entries with severity `low` are exempt from this pre-condition (the validator also exempts them). Record them without `ac_cross_check`.\n\n> **Any machine-computed `ac_cross_check` recorded before plugin `3.9.1` is untrustworthy — every field, not just `source: no-ac-section`.** Until issue #977 landed, both AC helpers captured the issue body as a per-line array and split it as if it were one string, so they read the body's **second line** rather than the acceptance-criteria section.\n>\n> Do not read that as \"the gate was merely blind.\" It read the **wrong input**, which fails in both directions:\n>\n> - When line 2 carried no backticked token — the usual case, since line 2 of a conventionally formatted issue is blank — both arrays came back empty, `source` resolved to `no-ac-section`, and the finding took the `defer` arm regardless of what the acceptance criteria actually said.\n> - When line 2 *did* carry a backticked token, that token was returned as though it were an acceptance-criteria reference. `source` was then `issue`, not `no-ac-section` (the verdict function only reports `no-ac-section` when **both** arrays are empty), and a fabricated file-arm match could reach `force-accept` — which overrides a structural deferral. This arm was **reachable but never observed** in the 60-issue sweep; it is why the distrust above is scoped to *every* pre-`3.9.1` machine-computed cross-check rather than to one `source` value. Scoping it to `no-ac-section` would have excluded exactly the records that carried override authority.\n>\n> **Anchor on the release, not on a date.** The fix shipped in `3.9.1`; several pull requests merged earlier the same day carry pre-fix cross-checks, so \"before 2026-08-02\" would wrongly vouch for them.\n>\n> Historical dispositions were deliberately **annotated rather than recomputed** (issue #977, maintainer decision) — a recompute reads issue bodies as amended since, which is a new claim, not a reconstruction.\n>\n> **`no-ac-section` remains imprecise even post-fix**, and this is unchanged by #977: the verdict function emits it whenever both arrays are empty, which includes an issue that *has* a populated acceptance-criteria section whose tokens are all stop-listed or unbackticked. It means \"nothing extractable\", not \"no section\".\n>\n> Both helpers now join the captured lines before splitting; do not remove that join, and see [references\u002Fmultiline-capture-audit.md](references\u002Fmultiline-capture-audit.md) before adding any new command-capture site.\n\n**`Add-FollowUpIssue` guard.** When the cross-check routes to `defer` and the agent calls `Add-FollowUpIssue` to file a follow-up issue, it MUST pass the `ac_cross_check` outcome as part of the issue body. Include a fenced YAML block in the body:\n\n```yaml\nac_cross_check:\n  file_arm: {bool}\n  term_arm: {bool}\n  result: {matched-high|matched-ambiguous|no-match}\n  source: {issue|pr-body|no-ac-section}\n  routed: defer\n```\n\nThis ensures the follow-up issue carries AC-provenance for the deferral decision, which is the AC4 contract.\n\n### `DEFERRED-SIGNIFICANT (structural)` and `disposition: defer` name one outcome\n\nThis skill owns the deferral mechanism, and until issue #998 its prose carried only one of the outcome's two names. A reader who arrived holding the consumer-facing label — the spelling every downstream surface uses — found no occurrence of it here and no way to learn that the mechanism described above is what produces it. The traversal dead-ended inside the skill that owns the thing.\n\n`scripts\u002FTest-DeferralCriteria.ps1` — this skill's own script — returns `verdict = 'DEFERRED-SIGNIFICANT (structural)'` when a finding meets the structural deferral criteria. **That verdict and the `disposition: defer` value recorded in the accumulator are one outcome under two names**, not two stages, two tiers, or two severities. The verdict is what the script emits; `disposition: defer` is how that same outcome is written down in the durable record. A finding carrying either carries both, and the `ac_cross_check` obligations above attach to it under whichever name you met it by.\n\nConsumers see the script's spelling. `agents\u002FCode-Review-Response.agent.md` § Vocabulary note already maps its own `SIGNIFICANT` tier onto `DEFERRED-SIGNIFICANT (structural)`, and the token travels the tree in that form. Nothing is renamed here — the seam was that this skill never said which of its own outputs the label referred to, and this section is that statement.\n\n### Legitimate Partial-AC Defer — Loud Guard\n\nWhen the AC cross-check returns `routed: defer` (because `result: no-match` or `source: no-ac-section`), the agent has confirmed that this finding genuinely lacks plan AC coverage. **Silently recording a `defer` entry at this point is the exact anti-pattern this feature was built to prevent.**\n\nThe agent MUST follow this sequence instead:\n\n1. **Emit a loud inline note** (not a question — this is a guard, not a decision the maintainer makes): `⚠️ Finding {finding_id} deferred without AC coverage (ac_cross_check.result: {result}) — mandatory proposal required.`\n\n2. **Enter a mandatory proposal** into the `§2e Filing Approval Gate` batch (`skills\u002Fsafe-operations\u002FSKILL.md` § 2e), pre-checked and recommended-approve, annotated `AC-uncovered defer`. The canonical title uses `ConvertTo-CanonicalFollowupTitle`. The proposal body MUST include the finding title, the judge ruling, and the `ac_cross_check` YAML block — this payload travels into the durable drop record if the maintainer drops the proposal. This is mandatory regardless of the finding's classification tier — routine findings that lack AC coverage still enter the gate as a proposal.\n\n3. **Record in the accumulator** with `disposition: defer`, the `ac_cross_check` object, and `disposition_rationale` that cites the `no-match`\u002F`no-ac-section` outcome and references the gate proposal (and the resulting issue URL once the maintainer approves it and it is filed).\n\nThe loud guard does not apply when `routed: force-accept` (high-confidence AC match) or `routed: disposition-gate` (ambiguous match fires the disposition question normally). It applies only to the `routed: defer` arm.\n\n**Low-severity exemption applies here too**: findings with severity `low` are exempt from this guard (the low-severity exemption from the blocking pre-condition applies throughout this section).\n\n### L0 Gate-Decision Tokens\n\nFor each finding that was gate-classified (routine or load-bearing), emit an L0 gate-decision token per `skills\u002Fsolution-authoring\u002Fschemas\u002Fgate-decision-token.schema.json`:\n\n```yaml\ndecision_id: \"{stable_finding_key}\"\nphase: review\noutcome: asked          # 'asked' for load-bearing, 'gate-fails' for routine\nclassification: load-bearing   # or routine\nwindow_position: review-disposition\ntimestamp: \"{ISO-8601 UTC}\"\npull_request_number: {PR}   # NOT issue_number\nskip_reason: \"routine-finding\"   # omit when outcome=asked\n```\n\nWrite each token to the authoritative L0 location per `skills\u002Fsolution-authoring\u002FSKILL.md` § L0 Gate Token: `\u002Fmemories\u002Fsession\u002Fgate-events-{session_key}.jsonl` (fallback: `.copilot-tracking\u002Fgate-events.jsonl`) as a single JSON line per the existing L0 emission pattern.\n\n### Stable-Key Resume\n\nAt the start of the disposition pass, call `Read-EngagementRecords -Phase review -PullRequestNumber {PR}` (from `.github\u002Fscripts\u002Flib\u002Fframe-engagement-record-core.ps1`). If a prior `engagement-record-review-{PR}` exists, extract its `load_bearing_decisions[].decision_id` values. These are `stable_finding_key` values of previously-gate-fired findings.\n\nFor each finding in the current judge-sustained set, check whether its `stable_finding_key` appears in the prior record:\n\n- **Match found** → `same-decision-resume` skip: reuse the prior `engineer_choice`, log `Reusing prior {stable_finding_key}: {engineer_choice}`, do not ask again.\n- **No match** → run the gate normally.\n\nThis enables re-review of a PR without re-asking for findings already dispositioned.\n\n### Persistence — Ordering\n\nWrite in this order (atomic marker first, engagement-record second). `skills\u002Fsession-memory-contract\u002Fscripts\u002Fpersist-marker.ps1` is the ONLY documented write path for the `review-dispositions-{PR}` marker (family `review-dispositions`, `post-new`) — never a hand-composed `gh pr comment` call. **What that rule buys is a single audited writer — not protection from `updated_at` advancement.** This family is `post-new`, so the write POSTs a fresh comment and advances nothing; the hazard arrives when a write replaces a **whole body** already carrying someone else's marker. See `skills\u002Fsession-memory-contract\u002Freferences\u002Fhandoff-markers.md` § What the write-path rule buys for the write-shape split. **`engagement-record-review-{PR}` is a known v1 registry gap and stays hand-authored for now**: the `engagement-record` family registry row declares `TargetSurface: issue` for every phase, but the `review` phase is PR-keyed — a `review`-phase write through the script would be refused by the surface preflight (`persist-marker-core.ps1`'s documented \"Known v1 scope gap\"). Do not attempt to route `engagement-record-review-{PR}` through the script until that gap is closed; hand-compose it per the existing convention below.\n\n1. **`\u003C!-- review-dispositions-{PR} -->`** — Persist via `persist-marker.ps1` (family `review-dispositions`). Payload: `schema_version: 4`, `passes_run: [...]`, `entries: [...]` (all findings, routine and load-bearing, one entry per finding), plus the PR-level `external_sources_reconciled` field. v2 added per-entry `severity`, `ac_cross_check`, and `stage` fields; v3 added per-entry `reviewer_source` (the reviewer identity or class — `local` for pipeline-native findings; external identities resolve per § `reviewer_source` Lookup Order above); v4 adds per-entry `internal_match` and the PR-level `external_sources_reconciled` field (see § `internal_match` Writer Rule below). This is the atomic per-finding record.\n\n   ````text\n   \u003C!-- review-dispositions-{PR} -->\n\n   ```yaml\n   schema_version: 4\n   passes_run: [1, 2, 3, 4, 5]\n   entries:\n     - stable_finding_key: \"...\"\n       pass: 1\n       disposition: incorporate\n       classification: routine\n       severity: medium\n       stage: code-review\n       reviewer_source: local   # pipeline-native finding: no internal_match (see below)\n       disposition_rationale: \"...\"\n       ac_cross_check:\n         file_arm: false\n         term_arm: true\n         result: matched-high\n         ac_ref: \"- the AC line that matched\"\n         source: issue\n         routed: force-accept\n     - stable_finding_key: \"gh-123456789\"\n       pass: 1\n       disposition: incorporate\n       classification: routine\n       severity: medium\n       stage: code-review\n       reviewer_source: jdoe    # external identity, resolved per § reviewer_source Lookup Order\n       internal_match:          # written BEFORE disposition_rationale (M42 field order)\n         match_status: novel    # duplicate | novel | ambiguous — see § internal_match Writer Rule\n       disposition_rationale: \"...\"\n   external_sources_reconciled: [\"gh-123456789\"]   # PR-level; emitted ONLY because this pass ran in GitHub Review Mode (a property of the SESSION\u002Fcommand, e.g. \u002Freview-github — never inferred from entry presence). [] is a legal, required zero-finding coverage record WITHIN a GitHub Review Mode pass, never an omission; on a purely internal-only pass (no external-source entries this pass) OMIT this field entirely instead.\n   ```\n   ````\n\n   > **v3 per-entry requirements** (carried over unchanged from v2): For entries with `disposition: dismiss` or `disposition: defer` and `severity` ≥ medium, `ac_cross_check` is required. The `ac_cross_check` object records which arms ran (`file_arm`, `term_arm`), the result tier (`matched-high | matched-ambiguous | no-match`), the matched AC reference if any, the source, and the routing outcome. Legacy `schema_version: 1` entries are exempt from this check. `artifact_citation` covers non-AC inherited artifacts; `ac_cross_check.ac_ref` is the AC-specific channel.\n   >\n   > **v4 requirements** (additive on top of v3): every external-source entry (`reviewer_source` not `local`) carries `internal_match.match_status`, written per § `internal_match` Writer Rule below. The PR-level `external_sources_reconciled` field is emitted once per posted marker **only on a GitHub Review Mode pass** (§ `internal_match` Writer Rule above) — including as `external_sources_reconciled: []` when that external review found nothing to reconcile. An absent field means this pass never attempted external reconciliation (the normal, expected state for a plain internal-only `\u002Forchestra:review` pass); an empty array means a GitHub Review Mode pass measured and found zero (M9). Never emit the field on an internal-only pass, even as `[]` — that would misrepresent an unmeasured pass as a measured one. `reviewer_source` stays required on every entry, same as v3.\n   >\n   > **`stage` field values**: The `stage` field records which pipeline stage produced this entry: `code-review` for the post-judge disposition gate, `ce` for CE Gate defect deferral. Both stages use the same `ac_cross_check` pre-condition at severity ≥ medium.\n   >\n   > **In-session schema audit is now pre-write and blocking (893-D3 amendment, s9)**: `persist-marker.ps1`'s `review-dispositions` family validator adapter invokes `.github\u002Fscripts\u002Flib\u002Freview-dispositions-validator-core.ps1` **in-process, via the call operator (`&`), with mandatory try\u002Fcatch containment** — never as a `pwsh -File` subprocess (the original 893-D3 subprocess design is superseded; a subprocess cannot bind that script's `-InMemoryMarkers` array parameter without hitting the recorded #866 flattening trap, since `-InMemoryMarkers` is a top-level array parameter). Any finding the validator returns is converted into a hard pre-write refusal on this write path: v4 schema violations (e.g., missing `ac_cross_check` on dismiss\u002Fdefer entries at severity ≥ medium, or a missing `reviewer_source`) are refused with the offending field named, before any network write. This is a property of the `persist-marker.ps1` write path only — the standalone validator script's own read-side warn-only contract (SMC-23 detection-at-review) is unchanged for any other caller.\n\n2. **`\u003C!-- engagement-record-review-{PR} -->`** — Post as a separate PR comment (not the same comment as review-dispositions), via a hand-composed `gh pr comment` call — **not** `persist-marker.ps1` (see the known v1 registry gap noted above this list). Payload follows `skills\u002Fengagement-record-emission\u002FSKILL.md` shape at `schema_version: 4`, `phase: review`. Load-bearing findings that fired the disposition question appear in `load_bearing_decisions[]` with their `engineer_choice` and `audit_rationale`. Routine findings do not appear in the engagement-record.\n\n   The engagement-record carries the `same-decision-resume` identity; the review-dispositions carries the per-finding outcome record. Never merge the two into a single comment.\n\n### SMC References\n\n- **SMC-19** — `finding_dispositions` is design-only (issue-keyed on `design-phase-complete`). This section's `review-dispositions` is a distinct, PR-keyed path.\n- **SMC-23** — governs `review-dispositions-{PR}` + `engagement-record-review-{PR}` survival, write path (PR comments), and cross-tool fungibility.\n- **SMC-20** — extended to include `review` phase; governs `engagement-record-review-{PR}` cross-session resume semantics.\n\n### Gotcha: Gate fires post-judge, not inside judge\n\nThe classification gate fires in the **owning parent workflow** after receiving the judge output. The judge body (this skill's judgment pass) ends at the judge-rulings comment. Do not add gate logic inside the judge's verification or scoring steps — doing so would conflate two distinct phases and make re-running the judge independent of dispositions impossible.\n",{"name":5,"description":7},"main","d5486aacc32c9ca1b07173cfd530f8fe75c29e58","skills\u002Freview-judgment",0,"https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra","https:\u002F\u002Fgithub.com\u002FGrimblaz","2026-08-16 01:12:46",{"source":8,"skillName":5,"commands":46,"prompt":56,"zipAvailable":57},[47,50,53],{"tool":48,"command":49},"npx","npx skills add Grimblaz\u002Fagent-orchestra --skill review-judgment",{"tool":51,"command":52},"bunx","bunx skills add Grimblaz\u002Fagent-orchestra --skill review-judgment",{"tool":54,"command":55},"pnpm","pnpm dlx skills add Grimblaz\u002Fagent-orchestra --skill review-judgment","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Freview-judgment\u002FSKILL.md",true,[59,65,69,75,79],{"path":60,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":63,"sourceHtmlUrl":64},"platforms\u002Fclaude.md","REFERENCE","参考文档",1474,"https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Freview-judgment\u002Fplatforms\u002Fclaude.md",{"path":66,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":67,"sourceHtmlUrl":68},"references\u002Fmultiline-capture-audit.md",12820,"https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Freview-judgment\u002Freferences\u002Fmultiline-capture-audit.md",{"path":70,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":73,"sourceHtmlUrl":74},"scripts\u002FGet-AcRefsFromIssue.ps1","SCRIPT","脚本",5345,"https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Freview-judgment\u002Fscripts\u002FGet-AcRefsFromIssue.ps1",{"path":76,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":77,"sourceHtmlUrl":78},"scripts\u002FGet-AcTermsFromIssue.ps1",9766,"https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Freview-judgment\u002Fscripts\u002FGet-AcTermsFromIssue.ps1",{"path":80,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":81,"sourceHtmlUrl":82},"scripts\u002FTest-DeferralCriteria.ps1",15931,"https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Freview-judgment\u002Fscripts\u002FTest-DeferralCriteria.ps1",[84,117,137,152,179,200],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":98,"publishedAt":33},"01M6C48YMJVFHM7DVGG1WWW5RJ","webapp-testing","Playwright end-to-end testing guidance for web apps, with practical patterns and setup steps. Use when creating or improving browser-based E2E coverage, test stability, and CI execution. DO NOT USE FOR: React component tests (use ui-testing), canvas game interaction (use browser-canvas-testing), TDD workflow (use test-driven-development), debugging failures (use systematic-debugging), or randomized property tests (use property-based-testing).","https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Fwebapp-testing\u002FSKILL.md",[90,95],{"categoryNo":91,"categoryName":92,"taxonomyType":93,"taxonomyTypeDesc":94},"01JBSEEDCAT000000000000008","测试与质量","SCENARIO","使用场景",{"categoryNo":96,"categoryName":97,"taxonomyType":93,"taxonomyTypeDesc":94},"01JBSEEDCAT000000000000009","前端开发",[99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,29,114,115,116,31],"testing","webapp testing","webapp","canvas","tdd","workflow","playwright","test","web","react","component","debugging","development","browser","ci","e2e","game","coverage",{"skillNo":118,"name":119,"nameZh":6,"description":120,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":121,"categories":122,"keywords":123,"publishedAt":33},"01M6C48TNNP2KCSQXV8YK7SK49","verification-before-completion","Evidence-based verification checklist before marking work complete. Use before PRs, releases, marking tickets done, or any \"I'm finished\" declaration; also when naming a baseline commit for a differential suite claim, when designing a mutation campaign or red-run evidence, when making an absence claim, a universal claim, or a one-time demonstration offered as a standing guarantee, when attributing a measurement or claiming a sweep or verification loop is complete, and when stating what a check or a CI suite will do about a change. DO NOT USE FOR: post-merge cleanup or archival (use post-pr-review) or processing GitHub review comments (use code-review-intake).","https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Fverification-before-completion\u002FSKILL.md",[],[124,125,18,126,127,128,129,130,131,23,24,25,113,132,133,134,135,31,32,136],"verification","verification before completion","change","commit","merge","pr","cleanup","run","loop","verification loop","work","check","campaign",{"skillNo":138,"name":139,"nameZh":6,"description":140,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":141,"categories":142,"keywords":143,"publishedAt":33},"01M6C48QZWNTP7JGKD32QZAKKX","validation-methodology","Reusable validation and review methodology for staged validation, failure triage, and prosecution-depth setup. Use when running validation ladders, triaging failures, or executing adversarial review passes. DO NOT USE FOR: CE Gate orchestration, specialist dispatch ownership, or step execution flow (keep those in Code-Conductor).","https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Fvalidation-methodology\u002FSKILL.md",[],[144,27,145,18,146,147,148,149,23,150,29,31,151],"validation","validation methodology","triage","orchestration","setup","adversarial review","flow","dispatch",{"skillNo":153,"name":154,"nameZh":6,"description":155,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":156,"categories":157,"keywords":166,"publishedAt":33},"01M6C48MCAHG5Z79BJ369CRYRZ","upstream-onboarding","Shared opening-phase protocol for upstream agents (Experience-Owner, Solution-Designer, Issue-Planner) and Code-Conductor when invoked on an existing GitHub issue. Renders a scaled context brief and runs a standards check on inherited work at each phase boundary. Use when a user-invocable agent receives a request referencing an existing GitHub issue and must load solution-authoring and upstream-onboarding at the opening phase. DO NOT USE FOR: subagent dispatches (which already operate within an assessed session context); post-merge review (use post-pr-review); research or non-tree subagents.","https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Fupstream-onboarding\u002FSKILL.md",[158,161],{"categoryNo":159,"categoryName":160,"taxonomyType":93,"taxonomyTypeDesc":94},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":162,"categoryName":163,"taxonomyType":164,"taxonomyTypeDesc":165},"01JBSEEDCAT000000000000038","人力资源","PROFESSION","适用职业",[167,168,18,20,169,128,170,129,23,171,172,173,174,175,176,134,135,177,178,31],"onboarding","upstream onboarding","research","standards check","protocol","brief","boundary","designer","subagent","planner","user","session",{"skillNo":180,"name":181,"nameZh":6,"description":182,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":183,"categories":184,"keywords":193,"publishedAt":33},"01M6C48GPVE120TGPMD1CS365H","ui-testing","Resilient React component testing strategies focusing on user behavior. Use when writing or reviewing component-level React tests, fixing flaky component tests, or establishing React testing patterns. DO NOT USE FOR: Playwright E2E tests (use webapp-testing), canvas game interaction (use browser-canvas-testing), TDD workflow (use test-driven-development), debugging failures (use systematic-debugging), UI component design (use frontend-design), or randomized property tests (use property-based-testing).","https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Fui-testing\u002FSKILL.md",[185,186,187,190],{"categoryNo":91,"categoryName":92,"taxonomyType":93,"taxonomyTypeDesc":94},{"categoryNo":96,"categoryName":97,"taxonomyType":93,"taxonomyTypeDesc":94},{"categoryNo":188,"categoryName":189,"taxonomyType":93,"taxonomyTypeDesc":94},"01JBSEEDCAT000000000000015","视觉设计",{"categoryNo":191,"categoryName":192,"taxonomyType":164,"taxonomyTypeDesc":165},"01JBSEEDCAT000000000000032","设计师",[99,194,195,196,197,198,102,104,105,106,108,109,110,111,112,199,114,115,31,101],"ui","ui testing","frontend","design","writing","testing patterns",{"skillNo":201,"name":202,"nameZh":6,"description":203,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":204,"categories":205,"keywords":208,"publishedAt":33},"01M6C48DRJ9XBDR2FW0W37DY53","ui-iteration","Reusable screenshot-driven UI polish workflow for iterative visual refinement. Use when improving layout, hierarchy, spacing, feedback, or Tailwind-based presentation through repeated review-and-adjust passes. DO NOT USE FOR: functional bug fixing, accessibility-only audits, or major redesign strategy work (use Code-Smith or frontend-design).","https:\u002F\u002Fgithub.com\u002FGrimblaz\u002Fagent-orchestra\u002Fblob\u002Fmain\u002Fskills\u002Fui-iteration\u002FSKILL.md",[206,207],{"categoryNo":96,"categoryName":97,"taxonomyType":93,"taxonomyTypeDesc":94},{"categoryNo":191,"categoryName":192,"taxonomyType":164,"taxonomyTypeDesc":165},[194,209,196,197,18,104,210,23,211,212,213,214,215,134,216,217,31,218],"ui iteration","accessibility","screenshot","Tailwind","polish","ui polish","feedback","redesign","strategy","only"]