| name | audit-verification |
| model | haiku |
| description | Signal-only audit confronting a finished task against its run evidence: a completion claimed without build/test/typecheck running (unverified- completion-claim), and a diff referencing a file the build would reject (unresolved-reference). Never blocks; runs on demand via /octopus:review- proposals on the cheapest tier. Queued by the zero-LLM verification-check Stop hook.
|
| triggers | {"keywords":["audit verification","verification check","claimed done","unverified","did it run"]} |
Verification Audit
Overview
A formatter, a type checker, and a secret scanner judge syntax, and
audit-grounding judges meaning — but neither asks the simplest question:
was the work actually verified? An agent can end a task asserting "done, tests
pass" without ever running them, and can write code referencing a symbol or file
that does not exist. The compiler would catch the latter — but only if it is run.
audit-verification is that check. It is signal-only and cheap: the
recurring, per-task part is a pure-bash Stop hook (hooks/stop/verification-check.sh)
that queues a proposal only when the work looks unverified. This skill reads the
queued proposal, the diff, and the session, and emits:
unverified-completion-claim — the session asserts done / passing / fixed,
but the hook's run-evidence scan found no build / test / typecheck this session.
Judge the claim against what actually ran; report the gap.unresolved-reference — the hook already detected, deterministically, a
changed file importing a relative path that does not resolve on disk. Confirm
and contextualize it; do not re-derive it.
Cost contract
The only recurring (per-task) component is the bash hook — it never invokes an
LLM. This skill runs only on demand via /octopus:review-proposals, in
batch, and on the cheapest model tier (--model haiku / each assistant's
cheapest): confronting a claim against run evidence is mechanical, not deep
reasoning. Do not spend a frontier model here.
Signal-only
This skill never blocks a commit, task, or merge. The syntactic gate already
blocks at commit; this is the signal for the "claimed done without running" gap.
It reports; the human decides — promote real gaps, archive the rest via
/octopus:review-proposals.
Pairing
guardrails bundle — the syntactic block at commit (formatter / typecheck / secret).audit-grounding — the semantic signal (invented conventions, unsupported facts).audit-verification — the verification signal (this skill). Together they are the local-guardrail triad.
