Skip to content
Octopus Octopus Octopus

audit-verification

/octopus:audit-verification is the check that asks the simplest question about a finished task: was it actually verified? It surfaces two specific failure modes — a completion claimed without any build, test, or typecheck having run, and a diff that references a file or symbol that does not exist on disk.

What it solves

A formatter and a type checker judge syntax. audit-grounding judges whether the agent invented conventions or unsupported facts. But neither asks whether the agent actually ran the build before declaring success. An agent can close a task with “done, all tests pass” without ever executing a test runner. It can write code importing a module that was never created. The compiler would catch the second problem — but only if it is run. audit-verification is that catch, applied consistently, at the cheapest possible cost.

How it works

The recurring, per-task component is a pure-bash Stop hook (hooks/stop/verification-check.sh) that runs after every task completes. It performs a deterministic, zero-LLM scan of the session and the diff:

  • unverified-completion-claim — the session contains a completion assertion (“done”, “tests pass”, “fixed”) but no evidence of a build, test, or typecheck command having been executed. When this pattern is detected, the hook queues a verification proposal.
  • unresolved-reference — the diff imports a relative path that does not resolve on disk. The hook detects this deterministically.

When the hook queues a proposal, /octopus:audit-verification (or /octopus:review-proposals) reads it and contextualizes the finding. This skill runs on the cheapest model tier — confronting a claim against run evidence is mechanical, not deep reasoning.

The skill is signal-only. It never blocks a commit, task, or merge. It reports; the human decides. Blocking is the guardrails bundle’s job.

Usage & parameters

/octopus:audit-verification

This command takes no flags. It is typically reached through /octopus:review-proposals when a queued verification proposal is present, or run directly after a task to check the current session.

Source: commands/audit-verification.md