AGENT SKILL

method-gap-research-status

Read-only family census that recommends one skill, or none, for a target-specific method-gap research census.

License 许可
MIT
最近更新
2026-08-23
Stars
★ 0
Forks
0

01 前言 Front Matter

name
method-gap-research-status
description
Read-only family census that recommends one skill, or none, for a target-specific method-gap research census.
argument-hint
<skill-name-family, for example game-*>
disable-model-invocation
true

02 资源文件 3

03 SKILL.md 正文

Method-Gap Research Status

Run one read-only census across an explicitly named family of skills. Recommend at most one exact target for a later, census-only method-gap commission. This skill is a portfolio selector, not a method-gap auditor: it never concludes that a method gap exists, decides that research is due, authors a brief, or invokes the commission.

Arguments: one skill-name family expressed as a literal name prefix followed by one terminal *, for example game-*. Paths, multiple wildcards, character classes, brace expansion, and a bare * are refused.

Hard boundaries

  • Read only: never modify a target, evidence stream, stored projection, research artifact, report, marker, issue, or repository metadata.
  • Family screen only: never semantically audit every target or build a claim-instrument map. The target-specific commission owns that work.
  • Evidence is not diagnosis: never relabel ordinary friction, a Skill Evolution gate, a conformance defect, calibration pressure, or policy disagreement as a method-gap signal.
  • Safe uncertainty: missing, corrupt, stale, ambiguous, or insufficient evidence never means a skill is healthy or research is not due.
  • No authority escalation: never invoke a method-gap commission, emit a brief-authoring instruction, perform external research, or adopt a prior recommendation.
  • Transient result: do not write a census report. A durable postpone or decline marker belongs only to a later target-specific census with the owner's consent.

Procedure

1. Resolve and inventory the family

From the repository root, run:

cargo run --locked -p skill-evidence -- skills method-gap-research-status \
  "<family>" --root . --now-epoch-milliseconds "<clock-ms>"

Choose one exact census clock in epoch milliseconds before invoking the command; reuse that value for any replay. The compiled command validates the selector, enumerates canonical .claude/skills/ members, hashes their complete packages, validates each matching reports/skill-evidence/ event stream through the shared Rust skill-evidence authority, derives current-hash use and incident facts in memory, discovers target-identifying method-gap lineage candidates, extracts bounded hashed headers and lineage-signal lines, and inspects target-scoped Git status and recent commit subjects when Git is available. It writes nothing.

Fail safely if the compiled command fails, the shared evidence contract is incompatible, or no skill matches. Do not substitute a hand-built family scan for a failed command.

Done when one complete inventory covers every exact family member and identifies every indeterminate store or unavailable Git observation without changing the repository.

2. Reconstruct only decision-relevant lineage

Read Lineage reconstruction completely. Use each candidate's inventory-provided header_excerpt, signal_lines, kind, size, and digest first. Read an exact additional section only when those bounded projections leave an ambiguity that could change the member's selection. Never read embedded evidence payloads or whole reports merely to prove absence. Search by that target's exact name, path, prior names found in those artifacts, and stable report markers only when a named candidate reveals that the inventory is incomplete.

For each family member, establish:

  • whether target-specific research is already in flight or returned but unconsumed;
  • the latest audited target identity and any adopted landing;
  • rejected or deferred recommendations and exact reopening triggers;
  • whether later target changes form a stable, observable interval; and
  • whether current evidence is absent, unreadable, on an old hash, or otherwise incapable of supporting a comparison.

In the census these two count different things: qualifying_uses counts runs, deduplicating the several incidents one run can record, while the outcome counts count records. A target whose runs deviated more than once therefore shows more outcomes than uses, and neither number is wrong. Read exercise off the use count and incident volume off the outcome counts; never subtract one from the other.

Do not read every target package semantically. A current dirty target is not automatically excluded, but it cannot be selected when its method-bearing baseline is changing too rapidly to freeze coherently.

Done when every member has one recoverable lineage state, no irrelevant report body or embedded payload was loaded, and every ambiguity that could change selection is either resolved from an exact bounded section or explicitly indeterminate.

3. Classify candidacy without diagnosing

Read Selection rules completely. Apply its routes, eligibility gates, ordered signals, and comparison rules to the inventory and reconstructed lineage.

Receipts supply interval evidence, not a method-gap verdict. Repeated ordinary friction remains owned by Skill Evolution. A family member becomes a screen next candidate only when the record supports a plausible missing-capability question with material consequence and a target-specific census could decide whether external research is due.

Assign every member exactly one status:

  • screen next;
  • watch;
  • research in flight;
  • insufficient evidence;
  • no current signal; or
  • could not determine.

There may be at most one screen next. When tied evidence cannot support one choice without arbitrary scoring, assign the tied members watch and return no selection.

Done when all members are classified once, ordinary iteration routes remain separate, every positive signal cites concrete current evidence, and at most one exact target survives.

4. Render the transient census and stop

Return:

  1. the family selector and number of canonical members;
  2. Screen next with the exact target, controlling evidence, material consequence, baseline/lineage state, and why a target-specific census can resolve the uncertainty—or None;
  3. the remaining members grouped under their exact statuses with terse reasons;
  4. unavailable or pending evidence;
  5. a literal statement that the census wrote nothing and did not determine that any method gap exists; and
  6. only when one target is selected, this census-only mandate:
Census only for <exact-target-path>. Decide commission now, postpone, or decline; do not author a research brief in this invocation.

Never shorten that mandate to a bare target: a bare target can authorize a brief when the target-specific census recommends commission now.

Done when the output accounts for the whole family, contains zero or one safe handoff, and performs no follow-on action.

Completion terminals

  • Selection: one exact screen next target, one census-only mandate, and no mutation or research authorization.
  • No selection: every family member classified, the controlling absence, tie, insufficiency, in-flight work, or uncertainty stated, and no invocation.
  • Safe failure: the failed contract or instrument named, nothing changed, and no hand-built substitute result.

04 同仓库更多 Skills