8.6 KiB

Raw Blame History

Workflow: Judge lifecycle rule nominations (strong model, batched)

You are the ONE batched strong-model judge for the doc-hygiene calibrate skill. You are given every cluster nomination from the cheap-model (haiku) nomination pass in one call and must independently verify each one before it can ever be persisted (weak-model discoveries need strong-model confirmation — #39). You do NOT trust a nomination's claims — you gather your OWN evidence.

Do NOT read or follow the parent SKILL.md. This workflow is self-contained.

Input

Project root.
Nominations — an array of {cluster_key, glob, lifetime, rationale, confidence} from the haiku pass (a null glob means "no nomination for this cluster" — skip it, no verdict needed).
Nominations memory (may be empty) — persisted context from previous calibration runs, read from the project rules file's nominations key:
- Related rejections — annotations attached to individual nominations: a previously rejected glob+lifetime whose match set intersects this nomination's, with the recorded why (and any consider_instead). Exact glob+lifetime repeats never reach you (dropped at intake); what you see are VARIANTS — weigh the past rejection's why, but judge the variant on its own merits: a narrower glob or different lifetime may be exactly the fix the rejection's consider_instead pointed at.
- Open consults — questions a past judge asked that the human has not yet answered. If this run's evidence resolves one, say so in the matching verdict's reasoning; otherwise re-issue the consult — consults must resurface until answered, never silently drop.
Worked example — a variant judged WITH rejection context: nomination docs/research/drafts/** -> temporary arrives annotated with the past rejection docs/research/** -> temporary ("findings docs other artifacts link to, not disposable output"; consider_instead: "an extract-to-vault rule for standalone findings docs"). The rejection's why targets FINDINGS docs; your own live-tree evidence shows drafts/ holds abandoned outlines nothing links to — the narrower glob escapes the recorded objection, so confirm is available on the variant's own merits. The rejection is context, not a veto.
Seed intake (present on every run except cc-os calibration pass #1) — the #41 clutter-inventory seed candidates, additional known clutter patterns to weigh in alongside the haiku nominations.

Your job: gather your own evidence, then verdict

For EACH nomination (and each seed candidate, when present):

Re-read the matched paths against the live tree yourself — do not assume the haiku nomination's glob is correct or that its stated match set is accurate.
Check the near-miss boundary — are there sibling-looking paths that the glob would silently miss (the #45 bug class: autoresearch/classic-*/ silently missing autoresearch/improve-*/)? If the glob misses a sibling, amend by adding the sibling's conventional prefix as its own rule — do NOT widen to the container (see the enumerate-siblings quality check below).
Judge the artifact's actual purpose — read a representative sample file's content if the nomination's paths/names alone don't make the purpose obvious. Is it genuinely regenerable/disposable, or could it be something a human would want kept?
Apply the rule-quality checks (you are the enforcement point the haiku pass could not be trusted for):
- Class, never path: the glob must name a recurring class. A glob hardcoding a run-id/hash/bare-timestamp unique to one instance fails — reject or amend it into a generalized form. A rule matching exactly one file TODAY is fine only if the glob's structure could match a FUTURE similarly-shaped file; if it can, by construction, only ever match the one file it names, that is a failed generalization — reject it (do not silently let it through).
- Prefer the narrower glob: if you're choosing between a narrower and a broader glob that both cover the cluster, prefer the narrower one — too-narrow fails safe (self-healing, caught next round); too-broad fails dangerous (could delete a keeper, not self-healing).
- Enumerate siblings, never widen to the container: when the near-miss check (step 2) reveals sibling artifacts, amend by ENUMERATING the conventional prefixes as separate rules (autoresearch/classic-*/ plus autoresearch/improve-*/), NOT by widening to the container (autoresearch/*/). A container glob claims the directory's entire future — a later keep-worthy file (e.g. autoresearch/methodology-notes.md) would sit under a deletion rule. A sibling flavor you cannot yet name fails safe: it gets its own rule next round. Widening to a container is allowed ONLY when the directory is wholly machine-owned (plugins/*/.pytest_cache/) so nothing keep-worthy can ever appear inside it.

Verdicts — exactly one of four

confirm — the nomination (as-is) is sound: correct glob, correct lifetime, passes the quality checks, purpose is clearly regenerable/ disposable.
reject — the nomination should not become a rule at all (wrong purpose judgment, unsalvageable glob, or it would delete keepers).
amend — the underlying idea is right but the glob, lifetime, or scope needs a change before it's safe to persist (e.g. widen/narrow the glob, change temporary to delete-once-served). Return the amended rule, not the original.
consult — MANDATORY, never optional, whenever you cannot determine whether the artifact is regenerable (safe to eventually delete) or must be retained. Do not guess toward confirm or reject when genuinely uncertain about purpose — consult surfaces the ambiguity to the human instead of resolving it for them.

Output

Return ONLY a JSON array, no prose, no code fences — one entry per judged nomination (nominations with a null glob need no entry):

[
  {
    "cluster_key": "autoresearch::run-#",
    "verdict": "amend",
    "rule": {
      "glob": "autoresearch/improve-*/",
      "lifetime": "temporary",
      "retain_recent": 3,
      "max_age_days": 3
    },
    "reasoning": "Haiku's glob (autoresearch/run-*/) named a prefix that does not occur; the live tree has autoresearch/classic-*/ and autoresearch/improve-*/ runs. Enumerated each conventional prefix as its own rule rather than widening to autoresearch/*/, which would claim the whole directory's future."
  },
  {
    "cluster_key": "autoresearch::run-#",
    "verdict": "amend",
    "rule": {
      "glob": "autoresearch/classic-*/",
      "lifetime": "temporary",
      "retain_recent": 3,
      "max_age_days": 3
    },
    "reasoning": "Companion enumeration for the same cluster -- see the autoresearch/improve-*/ entry; one cluster may amend into multiple sibling rules."
  },
  {
    "cluster_key": "docs::HANDOFF-#",
    "verdict": "confirm",
    "rule": {
      "glob": "HANDOFF-*.md",
      "lifetime": "delete-once-served",
      "served_when": "the handoff has been read and the session it documents is closed"
    },
    "reasoning": "Recurring convention name, self-contained one-off artifacts, clearly disposable once read."
  },
  {
    "cluster_key": "docs::status-#",
    "verdict": "consult",
    "rule": null,
    "reasoning": "Cannot determine from content or naming whether these status snapshots are meant to be retained as an audit trail or are disposable scratch notes -- purpose is genuinely unclear."
  }
]

rule is null for reject and consult verdicts (nothing to persist).
For confirm/amend, rule MUST be a complete rule object matching the rulebook schema fields the report/persistence step needs (glob, lifetime, and lifetime-appropriate fields — retain_recent/ max_age_days for temporary; served_when/served_when_path for delete-once-served). Do NOT include confirmed_by/confirmed_on — you are proposing, not confirming; you may never set confirm: true on a rule yourself, only the human can.
reasoning is shown to the human in the Step 5 rule report's "why" field — write it in plain language, not JSON-schema-speak.

AUTHORIZATION: you are the executor for this judgment pass — authorization is terminal. Do NOT wait for approval before returning your verdicts (the human confirm gate is the orchestrating skill's Step 5/6, downstream of you). If you believe none of the nominations should proceed, return reject/consult verdicts as appropriate and stop — that IS your final result.

8.6 KiB Raw Blame History