メニュー
Replay the hash chain over an existing audit JSONL; returns the chain head on success, the first integrity failure reason on failure.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Append one hash-chained audit event to okrs/<id>/audit/events/<run-id>.jsonl. Returns the new chain head.
Query the mesh's threat library for entries matching a concern (e.g. "identity", "data-licensing").
Academic paper search via arXiv. Returns ranked paper results per query envelope.
Aggregated architecture context for a scope — CALM nodes + ADRs + fitness functions + quality attributes. PURE data; the Architect persona is in the parent agent's prompt.
Aggregated quality context for a scope — quality attribute definitions + performance SLOs + reliability targets per BAR.
Aggregated security context for a scope — STRIDE threats + OWASP refs + NIST controls + prior threat-model docs. PURE data; the Security persona is in the parent agent's prompt.
| name | audit-verify-chain |
| description | Replay the hash chain over an existing audit JSONL; returns the chain head on success, the first integrity failure reason on failure. |
| version | 0.1.0 |
| purity | pure-data |
| runtime | research-runner |
| command | npx @maintainabilityai/research-runner@~0.1.64 skill-audit-verify-chain |
CI defense against forged audit logs. Called by prd-agent.yml and market-research-agent.yml after the agent's JSONL is committed; verdict fails + chain-integrity-failed label applied on ok: false.
This skill is NOT called by agents. Agents use audit-emit-event (which produces valid chains); audit-verify-chain is the post-hoc check that catches any agent that ignored the rule and wrote events by hand.
| Field | Type | Required | Description |
|---|---|---|---|
okrId | string | yes | OKR whose audit log to verify. |
runId | string | yes | The run's JSONL filename (without .jsonl). |
| Field | Type | Description |
|---|---|---|
ok | boolean | True iff the chain verifies end-to-end. |
chainHead | string | Final event_hash (= chain root for the run). Present only when ok: true. |
eventCount | integer | Number of events verified. Present only when ok: true. |
reason | string | First failure cause. Present only when ok: false. Examples: forged-hash-line-3: recorded=abc123… recomputed=def456…, prev-hash-mismatch-line-7, event-id-mismatch-line-2. |
Same chain rules implemented by the audit-verify-chain handler in packages/research-runner/src/runner/skills.ts:
prev_event_hash MUST be null.prev_event_hash MUST equal the preceding event's event_hash.event_hash MUST equal sha256(canonicalStringify({...event, event_hash: ''})) — i.e. the canonical key-sorted JSON of the event with its own hash field blanked.event_id MUST be monotonic from 1 (matches the on-disk line number).echo '{"okrId":"OKR-2026Q2-IMDB-001-celeb-api","runId":"HOW-2026-05-21-728gdt"}' \
| npx @maintainabilityai/research-runner@~0.1.64 skill-audit-verify-chain
The skill returns {ok: false, reason} for any verification failure; the calling workflow treats ok: false as blocking and labels the PR chain-integrity-failed. There is no "partial success" — even one tampered event invalidates the chain.