[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6JGFSV5Q1F8K546EDMCYTPV":3,"skill-same:01M6JGFSV5Q1F8K546EDMCYTPV":73},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":39,"linkOnly":40,"markdownContent":41,"markdownBody":42,"frontMatter":43,"refName":52,"commitSha":53,"skillDirectory":54,"forksCount":55,"repositoryHtmlUrl":56,"repositoryHomepageUrl":6,"ownerHtmlUrl":57,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":58,"install":59,"resources":72},"01M6JGFSV5Q1F8K546EDMCYTPV","judge-protocol",null,"HITL protocol for the judge role (Phase 2 of the two-phase judge split) — discuss the Phase 1 draft with the supervisor, ack and execute proposed mutations via CLI, write the final report.","muratovv\u002Fai-hats","muratovv","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F7502801?v=4","01M6JGCBN4KEJX012GPJVSH04T","Fedor Muratov","01M6JGCBSNMN7CW3RGTPPE0MN2",3,"MIT","https:\u002F\u002Fgithub.com\u002Fmuratovv\u002Fai-hats\u002Fblob\u002Fmaster\u002Fpackages\u002Fai-hats-library\u002Fsrc\u002Fai_hats_library\u002Fcore\u002Fskills\u002Fjudge-protocol\u002FSKILL.md",[],[19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38],"judge protocol","protocol","judge","draft","for","pip","split","final report","with","via","ack","cli","hitl","mutations","install","write","supervisor","hats","report","help","2026-11-10 10:12:30",false,"---\nname: judge-protocol\ndescription: >-\n  HITL protocol for the judge role (Phase 2 of the two-phase judge split)\n  — discuss the Phase 1 draft with the supervisor, ack and execute\n  proposed mutations via CLI, write the final report.\nai_hats:\n  requires:\n    cli:\n      - name: ai-hats-rack\n        check: \"rack --help\"\n        hint: \"pip install ai-hats-rack\"\n    mcp: []\nlicense: MIT\n---\n\n# Judge Protocol\n\nHITL protocol for the **judge** role: discuss the Phase 1 draft with\nthe supervisor, ack proposed mutations, execute them via CLI, write\nthe final report. Phase 2 of the two-phase judge split\n(ADR-0007).\n\nRuns via `ai-hats reflect hypothesis` (after Phase 1 `judge-auditor`\nproduces a draft) or `ai-hats execute --role judge` (standalone — see\nEdge Cases).\n\n## When to Use\n\nYou were launched as **judge**. The first user message contains the\nPhase 1 draft body inline (or, in standalone mode, a free-form\nsupervisor prompt). Apply this protocol end-to-end, mutate state ONLY\nthrough CLI (per **rule_backlog_discipline**), and write a report\nbefore the session ends.\n\nYou operate at **L1** (`base-judge` baseline): HITL dialogue with the\nsupervisor, ack'd CLI mutations from a whitelist (see §CLI whitelist),\nno source-file edits without L2 activation, single report artifact at\nexit.\n\n## Step 0 — Inventory the PROP inbox (unconditional)\n\nBefore Step 1, in every shape, whatever the kickoff scoped you to:\n\n```bash\nrack ls --backlog proposal --state open --all\n```\n\n`--all` is load-bearing — the default caps at 30 id-sorted rows.\n\n**Why unconditional.** The inbox is the other half of the reflex loop, and no\nlaunch shape guarantees you see it: Shapes A\u002FB hand you a Phase 1 draft whose\n`## Proposals` section may be thin, and Shape C hands you nothing at all. In\none measured session the kickoff scoped the work to the HYP contour, no one listed the\ninbox, and 136 open proposals went unread — among them PROP-107 and PROP-118,\nfindings the same epic then rediscovered from scratch at full cost. A scope\nthat omits the inbox is a scope you widen, not an instruction to skip this.\n\nCarry three numbers into Step 3 and into the report: total open, how many are\nauto-filed noise (**review-proposal** Step 3b), and the leaders by votes **and**\nby age. Both axes matter: vote counts on cards older than the fix are\nsystematically depressed — `session-reviewer` was emitting `n\u002Fa` for months —\nso ranking on votes alone buries precisely the pre-fix cards.\n\nUnder `reflect hypothesis` the Phase 2 preamble already carries this digest.\nRun the command anyway: it costs one call, and it is the only thing that would\ncatch a digest disagreeing with the catalog.\n\n## Step 1 — Read the first user message\n\nThe first user message is one of three shapes. Detect which and adapt:\n\n### Shape A — Phase 1 draft from `judge-auditor` (default for `reflect hypothesis`)\n\nThe body contains a `## Proposed mutations` section. This is your\nstarting analysis from the headless auditor pass:\n\n- Read the draft cover-to-cover before mutating anything.\n- The `## Proposed mutations` section is a CLI checklist. Each line is\n  a verbatim invocation Phase 1 recommends; you decide which ones to\n  execute after supervisor dialogue.\n- If you disagree with a proposed verdict (e.g. draft says\n  `keep` but recent evidence indicates `extend_window`), say so and\n  ask the supervisor.\n\n### Shape B — Active-HYP \u002F open-PROP inventory (legacy `reflect all`)\n\nThe body has `## Active hypotheses` \u002F `## Open proposals` sections but\nNO `## Proposed mutations` section. This is the legacy single-phase\ninvocation that has not yet migrated to `reflect hypothesis`. Run the\naudit yourself before dialoguing — walk the inventory, draft verdicts\nmentally, then proceed to Step 2 with the supervisor. There is no\nauditor draft to validate; you are doing both phases in one session.\n\n> The legacy `reflect all` pipeline is deprecated but kept\n> for one bake cycle. Don't refuse — supervisor may invoke it\n> intentionally during transition. Apply the same Steps 2 → 3 → 3.5 →\n> 4 flow; just skip the \"validate auditor proposals\" framing.\n\n### Shape C — Free-form supervisor prompt (`ai-hats execute --role judge`)\n\nThe body is an ad-hoc question, not a structured handoff. Glob\n`\u003Cai_hats_dir>\u002Fsessions\u002Fretros\u002Fjudge\u002F*-draft.md`, read the\nlexicographically latest if present, and proceed — or ask the\nsupervisor for context up front.\n\n## Step 2 — Walk hypotheses with supervisor\n\nFor each HYP in the draft's `## Hypotheses` section, surface the\nproposed verdict + recommendation to the supervisor, dialogue if\nneeded, and on ack execute via CLI:\n\n```bash\nrack hyp append-verdict HYP-NNN \\\n  --verdict \u003Cconfirmed|refuted|inconclusive> \\\n  --recommendation \u003Cclose_confirmed|close_refuted|keep|extend_window> \\\n  --evidence \"\u003Creason>\"\n```\n\nWhen a HYP's window closes (`close_confirmed` \u002F `close_refuted` \u002F\n`stalled`):\n\n```bash\nrack transition HYP-NNN \u003Cconfirmed|refuted|stalled>\n```\n\nFollow **review-hypothesis** for verdict-choice rules. The draft\nalready applied them; your job is to validate and adjust based on\ndialogue.\n\n## Step 3 — Walk proposals with supervisor\n\nWalk the inbox from Step 0 — not only the draft's `## Proposals`\nsection, which is a Phase 1 selection and may be thin. For each PROP\nsurface the proposed decision (`accept | reject | defer | duplicate`)\nto the supervisor. On ack for the full batch, run **one** bulk commit:\n\n```bash\nai-hats reflect commit \\\n  --accept PROP-001 --accept PROP-007 \\\n  --reject PROP-003 \\\n  --defer PROP-009 \\\n  --duplicate PROP-012\n```\n\nFor each accepted PROP, spawn the follow-up task as recommended in the\ndraft's `## Proposed mutations` section:\n\n```bash\nrack create \"\u003Ctitle>\" --description \"\u003Cfrom PROP body>\"\nrack transition PROP-NNN --link related_tasks:HATS-NNN\n```\n\nRecord that second command for every PROP whose outcome is a task —\nspawned by an `accept`, or already covered by an existing card on a\n`reject`. The edge is the machine-readable half of the triage; a\nmapping that lives only in this report or a work-log line cannot be\nresolved by anything downstream.\n\nFollow **review-proposal** for decision rules + the cost-citation\nheuristic (cost-cited PROPs get patience; uncited pain claims default\nto `defer` or `reject` after ≥1 sweep cycle), and its Step 3b for the\nauto-filed harness\u002Freviewer cards — those are a single batch decision\nunder a shared criterion, never N individual judgements.\n\n## Step 3.5 — Counter-claims pass (devil's advocate)\n\nThe draft already ran a counter-pass; its `## Counter-claims` section\nrecords each negative observation that was kept \u002F downgraded \u002F\ndropped. Use it as input, but **re-run the pass yourself** on any\nNEW negative observations that emerge from supervisor dialogue\n(things the auditor didn't see).\n\nFor each new draft negative claim, ask:\n\n1. **Count check.** Is the number measured or assumed?\n2. **Variance vs failure.** Is this a failure mode or expected\n   variance for the event class?\n3. **Shipped vs in-flight.** Production contract regression or\n   in-flight dev work?\n4. **Survivor bias.** Are you weighting N problem tickets against zero\n   acknowledgment of M shipped items?\n\nAppend the new claims to the report's `## Counter-claims` section\nalongside the auditor's entries (keep the auditor's entries as-is so\nthe audit trail is visible).\n\n## Step 4 — Write the final judge report\n\nBefore exiting the session write a markdown report to:\n\n```\n\u003Cai_hats_dir>\u002Fsessions\u002Fretros\u002Fjudge\u002F\u003CUTC-ISO-ts>-report.md\n```\n\nUse the `Write` tool (not Bash). Filename example:\n`2026-05-07T14-30-00Z-report.md`. This is the L0 carve-out for this\nrole (per `base-judge` §Contract); everything else goes through CLI.\n\nWrap the body between `BEGIN_JUDGE` \u002F `END_JUDGE` markers so the\npipeline's `extract_marker` step can capture it for audit (the\npipeline's `save_artifact` then writes the same content to the\ncanonical path — your Write tool call is the L0-audit-trail copy\nunder the role's declared write path).\n\nTemplate (section order is load-bearing — Deliverables must precede\nHypotheses; Counter-claims must precede Notes):\n\n```markdown\nBEGIN_JUDGE\n\n# Judge report — \u003CUTC ts>\n\n## Mode\n\nPhase 2 (HITL) — from draft \u003CUTC ts of Phase 1 draft>\n\n## Deliverables since prior report\n\n- \u003CHATS-NNN> — \u003Ctitle> (done, \u003Cdate>)\n- git: \u003Csha> \u003Csubject>\n  (or `(none)`)\n\n## Hypotheses\n\n- HYP-NNN — \u003Cverdict>: \u003Cone-line rationale>\n\n## Proposals\n\ninbox: \u003CN> open · \u003CM> triaged this pass · \u003CK> swept as auto-noise\n\n- PROP-NNN — \u003Cdecision>: \u003Cone-line rationale> (→ HATS-NNN)\n\n## New tasks\n\n- HATS-NNN — \u003Ctitle> (created from PROP-NNN)\n\n## Counter-claims\n\n- \"\u003Cclaim>\" → kept (verified: \u003Ccite>) | downgraded to observation | dropped (\u003Creason>)\n  (or `(none)`)\n\n## Notes\n\n\u003Cdialogue highlights, mid-session changes of mind, items the supervisor\ndeferred. Claims of regress\u002Fpain must cite concrete cost.>\nEND_JUDGE\n```\n\nEmpty sections are fine (use a `(none)` line) — the next sweep needs\nthis file to track history.\n\n## CLI whitelist (L1 per base-judge)\n\nAllowed:\n\n- `rack hyp append-verdict ...`\n- `rack transition \u003CHYP-ID> confirm|refute|stall|revive ...` (HYP status is an FSM edge)\n- `rack transition \u003CPROP-ID> --link related_tasks:\u003CTASK-ID>` (Step 3 triage outcome)\n- `ai-hats reflect commit ...`\n- `rack create ...`\n- `rack ls ...` \u002F `rack context ...` \u002F `ai-hats list ...` (inspection)\n\nForbidden without L2 activation:\n\n- Direct edits of `\u003Cai_hats_dir>\u002Ftracker\u002Fbacklog\u002F**` or\n  `\u003Cai_hats_dir>\u002Ftracker\u002Fhypotheses\u002F**` (use CLI, see\n  `rule_backlog_discipline`).\n- Edits of role \u002F skill \u002F rule \u002F trait source files.\n- Any verb not on the whitelist — escalate via **request-supervisor**.\n\n## Edge Cases\n\n- **Standalone launch (no Phase 1 draft)** — `ai-hats execute --role\n  judge` is supported. Glob the latest `*-draft.md`, or ask the\n  supervisor for ad-hoc context. The protocol from Step 1 onwards is\n  unchanged.\n- **Draft conflicts with recent evidence** — the supervisor or your\n  own re-read may surface evidence the auditor missed. Override the\n  draft's verdict, note the override in `## Notes`, and execute the\n  corrected CLI invocation. The draft is a proposal, not a binding\n  decision.\n- **Empty draft (all `(none)` sections)** — write a report with\n  `(none)` sections and exit. The empty pass is itself a signal.\n- **Mid-session supervisor request out of CLI whitelist** — escalate\n  via **request-supervisor**; do not silently exceed scope.\n- **Counter-claims `(none)` in your report but `## Notes` negative** —\n  anti-pattern; Step 3.5 was skipped on the new observations from\n  dialogue. Return to it before Write.\n- **`## Proposals` with no `inbox:` line** — anti-pattern; Step 0 was\n  skipped, so the section reports what the draft happened to mention\n  rather than what the inbox holds. Run Step 0 and rewrite the section\n  before Write. `0 open` is a fine value; a missing line is not.\n- **Kickoff scopes you away from the inbox** — Step 0 still runs. A\n  narrow kickoff is what produced the miss; report the count\n  even when the pass itself stays inside the scoped contour.\n\n## Scope\n\nThe only allowed direct write under `.agent\u002F` is the judge report at\n`\u003Cai_hats_dir>\u002Fsessions\u002Fretros\u002Fjudge\u002F\u003CUTC-ISO-ts>-report.md` (Step 4).\nAll other side effects — task creation, HYP verdicts\u002Fstatus, PROP\nvotes\u002Fstatus — go through the `rack` CLI (`rack create` \u002F `rack hyp` \u002F\n`rack proposal`). See **rule_backlog_discipline**.\n","# Judge Protocol\n\nHITL protocol for the **judge** role: discuss the Phase 1 draft with\nthe supervisor, ack proposed mutations, execute them via CLI, write\nthe final report. Phase 2 of the two-phase judge split\n(ADR-0007).\n\nRuns via `ai-hats reflect hypothesis` (after Phase 1 `judge-auditor`\nproduces a draft) or `ai-hats execute --role judge` (standalone — see\nEdge Cases).\n\n## When to Use\n\nYou were launched as **judge**. The first user message contains the\nPhase 1 draft body inline (or, in standalone mode, a free-form\nsupervisor prompt). Apply this protocol end-to-end, mutate state ONLY\nthrough CLI (per **rule_backlog_discipline**), and write a report\nbefore the session ends.\n\nYou operate at **L1** (`base-judge` baseline): HITL dialogue with the\nsupervisor, ack'd CLI mutations from a whitelist (see §CLI whitelist),\nno source-file edits without L2 activation, single report artifact at\nexit.\n\n## Step 0 — Inventory the PROP inbox (unconditional)\n\nBefore Step 1, in every shape, whatever the kickoff scoped you to:\n\n```bash\nrack ls --backlog proposal --state open --all\n```\n\n`--all` is load-bearing — the default caps at 30 id-sorted rows.\n\n**Why unconditional.** The inbox is the other half of the reflex loop, and no\nlaunch shape guarantees you see it: Shapes A\u002FB hand you a Phase 1 draft whose\n`## Proposals` section may be thin, and Shape C hands you nothing at all. In\none measured session the kickoff scoped the work to the HYP contour, no one listed the\ninbox, and 136 open proposals went unread — among them PROP-107 and PROP-118,\nfindings the same epic then rediscovered from scratch at full cost. A scope\nthat omits the inbox is a scope you widen, not an instruction to skip this.\n\nCarry three numbers into Step 3 and into the report: total open, how many are\nauto-filed noise (**review-proposal** Step 3b), and the leaders by votes **and**\nby age. Both axes matter: vote counts on cards older than the fix are\nsystematically depressed — `session-reviewer` was emitting `n\u002Fa` for months —\nso ranking on votes alone buries precisely the pre-fix cards.\n\nUnder `reflect hypothesis` the Phase 2 preamble already carries this digest.\nRun the command anyway: it costs one call, and it is the only thing that would\ncatch a digest disagreeing with the catalog.\n\n## Step 1 — Read the first user message\n\nThe first user message is one of three shapes. Detect which and adapt:\n\n### Shape A — Phase 1 draft from `judge-auditor` (default for `reflect hypothesis`)\n\nThe body contains a `## Proposed mutations` section. This is your\nstarting analysis from the headless auditor pass:\n\n- Read the draft cover-to-cover before mutating anything.\n- The `## Proposed mutations` section is a CLI checklist. Each line is\n  a verbatim invocation Phase 1 recommends; you decide which ones to\n  execute after supervisor dialogue.\n- If you disagree with a proposed verdict (e.g. draft says\n  `keep` but recent evidence indicates `extend_window`), say so and\n  ask the supervisor.\n\n### Shape B — Active-HYP \u002F open-PROP inventory (legacy `reflect all`)\n\nThe body has `## Active hypotheses` \u002F `## Open proposals` sections but\nNO `## Proposed mutations` section. This is the legacy single-phase\ninvocation that has not yet migrated to `reflect hypothesis`. Run the\naudit yourself before dialoguing — walk the inventory, draft verdicts\nmentally, then proceed to Step 2 with the supervisor. There is no\nauditor draft to validate; you are doing both phases in one session.\n\n> The legacy `reflect all` pipeline is deprecated but kept\n> for one bake cycle. Don't refuse — supervisor may invoke it\n> intentionally during transition. Apply the same Steps 2 → 3 → 3.5 →\n> 4 flow; just skip the \"validate auditor proposals\" framing.\n\n### Shape C — Free-form supervisor prompt (`ai-hats execute --role judge`)\n\nThe body is an ad-hoc question, not a structured handoff. Glob\n`\u003Cai_hats_dir>\u002Fsessions\u002Fretros\u002Fjudge\u002F*-draft.md`, read the\nlexicographically latest if present, and proceed — or ask the\nsupervisor for context up front.\n\n## Step 2 — Walk hypotheses with supervisor\n\nFor each HYP in the draft's `## Hypotheses` section, surface the\nproposed verdict + recommendation to the supervisor, dialogue if\nneeded, and on ack execute via CLI:\n\n```bash\nrack hyp append-verdict HYP-NNN \\\n  --verdict \u003Cconfirmed|refuted|inconclusive> \\\n  --recommendation \u003Cclose_confirmed|close_refuted|keep|extend_window> \\\n  --evidence \"\u003Creason>\"\n```\n\nWhen a HYP's window closes (`close_confirmed` \u002F `close_refuted` \u002F\n`stalled`):\n\n```bash\nrack transition HYP-NNN \u003Cconfirmed|refuted|stalled>\n```\n\nFollow **review-hypothesis** for verdict-choice rules. The draft\nalready applied them; your job is to validate and adjust based on\ndialogue.\n\n## Step 3 — Walk proposals with supervisor\n\nWalk the inbox from Step 0 — not only the draft's `## Proposals`\nsection, which is a Phase 1 selection and may be thin. For each PROP\nsurface the proposed decision (`accept | reject | defer | duplicate`)\nto the supervisor. On ack for the full batch, run **one** bulk commit:\n\n```bash\nai-hats reflect commit \\\n  --accept PROP-001 --accept PROP-007 \\\n  --reject PROP-003 \\\n  --defer PROP-009 \\\n  --duplicate PROP-012\n```\n\nFor each accepted PROP, spawn the follow-up task as recommended in the\ndraft's `## Proposed mutations` section:\n\n```bash\nrack create \"\u003Ctitle>\" --description \"\u003Cfrom PROP body>\"\nrack transition PROP-NNN --link related_tasks:HATS-NNN\n```\n\nRecord that second command for every PROP whose outcome is a task —\nspawned by an `accept`, or already covered by an existing card on a\n`reject`. The edge is the machine-readable half of the triage; a\nmapping that lives only in this report or a work-log line cannot be\nresolved by anything downstream.\n\nFollow **review-proposal** for decision rules + the cost-citation\nheuristic (cost-cited PROPs get patience; uncited pain claims default\nto `defer` or `reject` after ≥1 sweep cycle), and its Step 3b for the\nauto-filed harness\u002Freviewer cards — those are a single batch decision\nunder a shared criterion, never N individual judgements.\n\n## Step 3.5 — Counter-claims pass (devil's advocate)\n\nThe draft already ran a counter-pass; its `## Counter-claims` section\nrecords each negative observation that was kept \u002F downgraded \u002F\ndropped. Use it as input, but **re-run the pass yourself** on any\nNEW negative observations that emerge from supervisor dialogue\n(things the auditor didn't see).\n\nFor each new draft negative claim, ask:\n\n1. **Count check.** Is the number measured or assumed?\n2. **Variance vs failure.** Is this a failure mode or expected\n   variance for the event class?\n3. **Shipped vs in-flight.** Production contract regression or\n   in-flight dev work?\n4. **Survivor bias.** Are you weighting N problem tickets against zero\n   acknowledgment of M shipped items?\n\nAppend the new claims to the report's `## Counter-claims` section\nalongside the auditor's entries (keep the auditor's entries as-is so\nthe audit trail is visible).\n\n## Step 4 — Write the final judge report\n\nBefore exiting the session write a markdown report to:\n\n```\n\u003Cai_hats_dir>\u002Fsessions\u002Fretros\u002Fjudge\u002F\u003CUTC-ISO-ts>-report.md\n```\n\nUse the `Write` tool (not Bash). Filename example:\n`2026-05-07T14-30-00Z-report.md`. This is the L0 carve-out for this\nrole (per `base-judge` §Contract); everything else goes through CLI.\n\nWrap the body between `BEGIN_JUDGE` \u002F `END_JUDGE` markers so the\npipeline's `extract_marker` step can capture it for audit (the\npipeline's `save_artifact` then writes the same content to the\ncanonical path — your Write tool call is the L0-audit-trail copy\nunder the role's declared write path).\n\nTemplate (section order is load-bearing — Deliverables must precede\nHypotheses; Counter-claims must precede Notes):\n\n```markdown\nBEGIN_JUDGE\n\n# Judge report — \u003CUTC ts>\n\n## Mode\n\nPhase 2 (HITL) — from draft \u003CUTC ts of Phase 1 draft>\n\n## Deliverables since prior report\n\n- \u003CHATS-NNN> — \u003Ctitle> (done, \u003Cdate>)\n- git: \u003Csha> \u003Csubject>\n  (or `(none)`)\n\n## Hypotheses\n\n- HYP-NNN — \u003Cverdict>: \u003Cone-line rationale>\n\n## Proposals\n\ninbox: \u003CN> open · \u003CM> triaged this pass · \u003CK> swept as auto-noise\n\n- PROP-NNN — \u003Cdecision>: \u003Cone-line rationale> (→ HATS-NNN)\n\n## New tasks\n\n- HATS-NNN — \u003Ctitle> (created from PROP-NNN)\n\n## Counter-claims\n\n- \"\u003Cclaim>\" → kept (verified: \u003Ccite>) | downgraded to observation | dropped (\u003Creason>)\n  (or `(none)`)\n\n## Notes\n\n\u003Cdialogue highlights, mid-session changes of mind, items the supervisor\ndeferred. Claims of regress\u002Fpain must cite concrete cost.>\nEND_JUDGE\n```\n\nEmpty sections are fine (use a `(none)` line) — the next sweep needs\nthis file to track history.\n\n## CLI whitelist (L1 per base-judge)\n\nAllowed:\n\n- `rack hyp append-verdict ...`\n- `rack transition \u003CHYP-ID> confirm|refute|stall|revive ...` (HYP status is an FSM edge)\n- `rack transition \u003CPROP-ID> --link related_tasks:\u003CTASK-ID>` (Step 3 triage outcome)\n- `ai-hats reflect commit ...`\n- `rack create ...`\n- `rack ls ...` \u002F `rack context ...` \u002F `ai-hats list ...` (inspection)\n\nForbidden without L2 activation:\n\n- Direct edits of `\u003Cai_hats_dir>\u002Ftracker\u002Fbacklog\u002F**` or\n  `\u003Cai_hats_dir>\u002Ftracker\u002Fhypotheses\u002F**` (use CLI, see\n  `rule_backlog_discipline`).\n- Edits of role \u002F skill \u002F rule \u002F trait source files.\n- Any verb not on the whitelist — escalate via **request-supervisor**.\n\n## Edge Cases\n\n- **Standalone launch (no Phase 1 draft)** — `ai-hats execute --role\n  judge` is supported. Glob the latest `*-draft.md`, or ask the\n  supervisor for ad-hoc context. The protocol from Step 1 onwards is\n  unchanged.\n- **Draft conflicts with recent evidence** — the supervisor or your\n  own re-read may surface evidence the auditor missed. Override the\n  draft's verdict, note the override in `## Notes`, and execute the\n  corrected CLI invocation. The draft is a proposal, not a binding\n  decision.\n- **Empty draft (all `(none)` sections)** — write a report with\n  `(none)` sections and exit. The empty pass is itself a signal.\n- **Mid-session supervisor request out of CLI whitelist** — escalate\n  via **request-supervisor**; do not silently exceed scope.\n- **Counter-claims `(none)` in your report but `## Notes` negative** —\n  anti-pattern; Step 3.5 was skipped on the new observations from\n  dialogue. Return to it before Write.\n- **`## Proposals` with no `inbox:` line** — anti-pattern; Step 0 was\n  skipped, so the section reports what the draft happened to mention\n  rather than what the inbox holds. Run Step 0 and rewrite the section\n  before Write. `0 open` is a fine value; a missing line is not.\n- **Kickoff scopes you away from the inbox** — Step 0 still runs. A\n  narrow kickoff is what produced the miss; report the count\n  even when the pass itself stays inside the scoped contour.\n\n## Scope\n\nThe only allowed direct write under `.agent\u002F` is the judge report at\n`\u003Cai_hats_dir>\u002Fsessions\u002Fretros\u002Fjudge\u002F\u003CUTC-ISO-ts>-report.md` (Step 4).\nAll other side effects — task creation, HYP verdicts\u002Fstatus, PROP\nvotes\u002Fstatus — go through the `rack` CLI (`rack create` \u002F `rack hyp` \u002F\n`rack proposal`). See **rule_backlog_discipline**.\n",{"name":5,"ai_hats":44,"license":15,"description":7},{"requires":45},{"cli":46,"mcp":51},[47],{"hint":48,"name":49,"check":50},"pip install ai-hats-rack","ai-hats-rack","rack --help",[],"master","b4819add93efbaf4d80627118ec34680406b309b","packages\u002Fai-hats-library\u002Fsrc\u002Fai_hats_library\u002Fcore\u002Fskills\u002Fjudge-protocol",1,"https:\u002F\u002Fgithub.com\u002Fmuratovv\u002Fai-hats","https:\u002F\u002Fgithub.com\u002Fmuratovv","2026-11-05 20:59:10",{"source":8,"skillName":5,"commands":60,"prompt":70,"zipAvailable":71},[61,64,67],{"tool":62,"command":63},"npx","npx skills add muratovv\u002Fai-hats --skill judge-protocol",{"tool":65,"command":66},"bunx","bunx skills add muratovv\u002Fai-hats --skill judge-protocol",{"tool":68,"command":69},"pnpm","pnpm dlx skills add muratovv\u002Fai-hats --skill judge-protocol","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fmuratovv\u002Fai-hats\u002Fblob\u002Fmaster\u002Fpackages\u002Fai-hats-library\u002Fsrc\u002Fai_hats_library\u002Fcore\u002Fskills\u002Fjudge-protocol\u002FSKILL.md",true,[],[74,99,123,145,169,182],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":39},"01M6PVFB1SM738R2SRSRSF7CY6","tool-evaluation-protocol","Time-bounded protocol for evaluating new tools, libraries, or methodologies before adoption. Use when evaluating a candidate dependency for the project, researching a methodology (e.g. ML\u002FLLM frameworks, testing approaches), or following up on a hype-driven recommendation.","https:\u002F\u002Fgithub.com\u002Fmuratovv\u002Fai-hats\u002Fblob\u002Fmaster\u002Fpackages\u002Fai-hats-library\u002Fsrc\u002Fai_hats_library\u002Fexperimental\u002Fusage\u002Fskills\u002Ftool-evaluation-protocol\u002FSKILL.md",[],[20,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,23,97,98],"Evaluation","tool","tool evaluation protocol","testing","ml","libraries","adoption","methodology","Tools","use","dependency","up","frameworks","time","project","recommendation","bounded","candidate",{"skillNo":100,"name":101,"nameZh":6,"description":102,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":103,"categories":104,"keywords":105,"publishedAt":39},"01M6PVF7HATBEX1G7N50E2QT57","systematic-debugging","4-phase bug-fix protocol (evidence, pattern, hypothesis, verify). Use when investigating any error, failure, or unexpected behavior, handling bug reports from users or monitoring, or diagnosing test failures with non-obvious causes.","https:\u002F\u002Fgithub.com\u002Fmuratovv\u002Fai-hats\u002Fblob\u002Fmaster\u002Fpackages\u002Fai-hats-library\u002Fsrc\u002Fai_hats_library\u002Fexperimental\u002Fusage\u002Fskills\u002Fsystematic-debugging\u002FSKILL.md",[],[106,107,108,109,20,110,111,112,113,114,90,115,116,117,118,119,120,27,121,122],"systematic debugging","debugging","systematic","test","monitoring","hypothesis","verify","behavior","fix","evidence","bug","reports","pattern","users","error","bug-fix","phase",{"skillNo":124,"name":125,"nameZh":6,"description":126,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":127,"categories":128,"keywords":134,"publishedAt":39},"01M6PVF3MTQN3RPF73AN34856J","skill-template","Canonical template and validation guide for ai-hats skills. Use when creating a new skill, reviewing an existing one for structural compliance, or deciding which pattern fits a new behavior.","https:\u002F\u002Fgithub.com\u002Fmuratovv\u002Fai-hats\u002Fblob\u002Fmaster\u002Fpackages\u002Fai-hats-library\u002Fsrc\u002Fai_hats_library\u002Fexperimental\u002Fusage\u002Fskills\u002Fskill-template\u002FSKILL.md",[129],{"categoryNo":130,"categoryName":131,"taxonomyType":132,"taxonomyTypeDesc":133},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[135,136,137,138,36,113,139,90,140,118,141,142,143,23,144],"template","skill template","validation","compliance","guide","new skill","structural","canonical","which","Existing",{"skillNo":146,"name":147,"nameZh":6,"description":148,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":149,"categories":150,"keywords":152,"publishedAt":39},"01M6PVF06SCV28QARM5MQPANC9","skill-optimization","Audit and refactor library components to eliminate redundancy and staleness. Use during periodic library maintenance, after adding multiple new skills or rules, when a role's token budget is too high, or after retrospectives identify component issues.","https:\u002F\u002Fgithub.com\u002Fmuratovv\u002Fai-hats\u002Fblob\u002Fmaster\u002Fpackages\u002Fai-hats-library\u002Fsrc\u002Fai_hats_library\u002Fexperimental\u002Fusage\u002Fskills\u002Fskill-optimization\u002FSKILL.md",[151],{"categoryNo":130,"categoryName":131,"taxonomyType":132,"taxonomyTypeDesc":133},[153,154,155,156,157,158,159,160,161,162,90,163,164,165,166,167,168],"Optimization","skill optimization","components","component","issues","library","audit","refactor","Budget","Rules","periodic","maintenance","token","token budget","role","high",{"skillNo":170,"name":171,"nameZh":6,"description":172,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":173,"categories":174,"keywords":175,"publishedAt":39},"01M6PVEXM9717016T9MW82HDCT","silent-fallback-gate","PreToolUse hook — refuses an Edit or Write that adds an inert broad except. Use when a Python edit was blocked for a silent fallback.","https:\u002F\u002Fgithub.com\u002Fmuratovv\u002Fai-hats\u002Fblob\u002Fmaster\u002Fpackages\u002Fai-hats-library\u002Fsrc\u002Fai_hats_library\u002Fexperimental\u002Fusage\u002Fskills\u002Fsilent-fallback-gate\u002FSKILL.md",[],[176,177,178,179,34,180,90,181,23],"gate","fallback","silent fallback gate","python","edit","hook",{"skillNo":183,"name":184,"nameZh":6,"description":185,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":186,"categories":187,"keywords":189,"publishedAt":39},"01M6PVESM49C5NT7SB7N3F76VS","retro-to-framework","Convert an observation about agent behavior — from a retrospective or from a single incident — into a framework-level change (hook, rule, skill, skill update). Use after a retrospective identifies problems that are not project-specific, after an incident where an agent misapplied a component it had read, when CLAUDE.md band-aids accumulate (more than 3 per-project rules that could be generic), or when the same problem recurs across multiple projects.","https:\u002F\u002Fgithub.com\u002Fmuratovv\u002Fai-hats\u002Fblob\u002Fmaster\u002Fpackages\u002Fai-hats-library\u002Fsrc\u002Fai_hats_library\u002Fexperimental\u002Fusage\u002Fskills\u002Fretro-to-framework\u002FSKILL.md",[188],{"categoryNo":130,"categoryName":131,"taxonomyType":132,"taxonomyTypeDesc":133},[190,191,192,193,156,194,195,196,162,113,90,197,198,95,199,200,201,202,203,204],"framework","retro","retro to framework","change","retrospective","projects","convert","incident","across","problem","about","more","skill update","single","generic"]