graphrag-stocktake

name

graphrag-stocktake

description

vault の定期棚卸し(ライフサイクル + 配線負債のクリーニング)。①state無しレガシーや決着済みなのに active のままの Investigation を state:closed に、②「あとで」のまま停滞した Goal (planned/active) を続行/achieved/abandoned に、③未ガード Constraint (enforcement debt) と参照腐敗マーカーの返済を、④偽陽性の多い authority echo 指紋 (echo_stats) の引退を、機械検出(stocktake / constraint-check / xref-check --root)+ 裏取り裁定で整える。「棚卸しして」「vault を掃除/クリーニングして」「Investigation を整理して」「active が溜まってる」「enforcement debt を返済して」で発火。resume の stocktake_hint / open_goals や ask の enforcement_debt が出た時にも。削除はしない(閉じる/宣言するだけ)。スラッシュ: /graphrag-knowledge:graphrag-stocktake

Vault stocktake (lifecycle + wiring-debt periodic cleaning)

A periodic sweep that tidies the vault's lifecycle state and wiring debt, via machine detection + corroborated adjudication. For the read/write foundation and CLI details, follow the parent skill graphrag-knowledge and $REF/ (= ${CLAUDE_PLUGIN_ROOT}/references). $CLI = node --experimental-strip-types ${CLAUDE_PLUGIN_ROOT}/graphrag/cli.ts.

Scope (what this skill does and does not do)

• The primary channel for cleanup is the always-on triggers (pre-commit nudge + delta-check headlines, checkpoint, the closed judgment at write-back when implementation reaches a milestone). This skill is the later-stage periodic sweep that picks up their leftovers/misses, not a replacement for the always-on triggers.
• Detection is delegated to the deterministic verbs (stocktake, constraint-check, xref-check --root); this skill handles adjudication only. Do not make the LLM read the whole vault (per §Anti-patterns in the parent skill).
• Fire on an explicit instruction like "stocktake", or when you see the stocktake_hint / open_goals returned by brief --mode resume, or the enforcement_debt riding on ask.
• Why a dedicated event exists at all: field data shows debt is not repaid "in passing" — enforcement_debt rode along on every ask for weeks and nobody repaid it, while the same kind of prompt at a dedicated moment produced action. Repayment needs its own moment; this skill is that moment.

Procedure

1. Lifecycle sweep: run $CLI stocktake [--days <N=14>] [--vault <dir>] (read-only, deterministic). Suspects carry type (Investigation | Goal | Constraint) and signals: stateless | stale-active | no-generated-at | progress-claim | stale-planned-goal | stale-active-goal | settled-premise (+ abandoned-premise). If suspects are zero, say so and move to step 3.
2. Adjudicate each suspect with corroboration (the core of this skill):
   • Investigations — corroboration against code, tests, git evidence, and chat evidence beats the summary's self-report (unimplemented / in-progress / incomplete). If tests or an implementation actually exist / it shipped, the investigation is settled = closed even if the summary says "unimplemented". Remaining code work is owned by the Decision/Risk side — an Investigation closes on "the inquiry being settled". A genuinely ongoing focus stays active (assign state:"active" explicitly to stateless ones that still move). If you cannot corroborate, do not close — report it.
   • Goals (deferred work) — three honest outcomes, corroborate before choosing: the work happened (git/code evidence) → achieved; the premise died or nobody will ever do it → abandoned (say so — a dead promise kept "planned" is a lie that resurfaces forever); still genuinely wanted → keep, expressed mechanically: include an op:update on the Goal in the same plan (even a no-op update refreshes generated_at to now), which silences re-prompting for the threshold window (14d planned / 90d active). A keep that doesn't touch generated_at will be re-flagged next session — the update IS the machine-readable form of "I looked at this and it stands." Also check the wiring: if the debt has a findable home (including the extension seed file for new-capability reservations) and documented_by is missing, add it now; if the Goal is genuinely placeless, record that judgment instead — it rides the resume/stocktake lanes only, and that is a legitimate state, not a gap.
   • Constraints with settled-premise (debt shadows) — the "until X is done, Y cannot be trusted" warning whose premise Goals have all settled. Two different endings, split by the abandoned-premise signal: (a) premises — corroborate that Y actually got fixed (the Goal claims it; verify against code/tests), then delete the shadow (301 successor if a durable invariant replaced it); if Y is somehow still broken, reopen honestly (new Goal + rewire the shadow's premise to it). (b) any premise — the breakage became permanent, so do NOT just delete the still-true warning: convert the shadow into a premise-less permanent Constraint (keep the constrains wiring, drop has_premise) or downgrade to a Risk, whichever is honest. 2b. — when the stocktake output carries (authority-echo firing history from the commit lane), adjudicate the hottest aliases: a fingerprint that fires on every legitimate use is crying wolf and trains readers to ignore echoes. Corroborate before dropping (check the logged paths — one hot alias can be a REAL serial re-implementation): if the firings are legitimate uses of shared/library vocabulary, retire the alias via on the owning node's ; if they are genuine re-implementations, the echo is doing its job — leave it and surface the finding instead.

Off-limits (never do)

• Do NOT delete. Lineage preservation is the way — terminal states are discounted 0.6× in ranking, so they naturally sink in search results (not excluded). abandoned is a state, not a deletion.
• Do NOT step into whether knowledge nodes (Decision/Risk/OK, etc.) are "still true" — that is the domain of drift-reconciliation ($REF/drift-reconciliation.md). This skill handles lifecycle state and wiring debt only.
• Do NOT force repayment. Enforcement debt stays advisory (the migration-rail decision): this skill is the offered moment, not a gate. Declining with a reason is a legitimate outcome — record it via enforcement:"none" + enforcement_reason so it stays visibly unenforceable instead of silently unguarded.