audit

Audit — A Panel of Diverse Models

Announce at start: "Running a dispatch audit with an external panel."

Why

One model reviewing its own work re-applies the biases that produced it. Different models miss different things, so the same audit dispatched to a panel buys perspectives a single reviewer cannot: findings the panel agrees on are strong signals, and findings the panel splits on mark judgment calls worth the user's attention.

The panel is codex and kimi — two external agent CLIs running models unrelated to this session's. Everything about reaching them lives in this plugin's use-external-agents skill: the router that launches them, the brief contract they are dispatched under, and the matrix call that reports whether each one is installed and usable right now.

Protocol

1. Load the dispatch interface. Invoke this plugin's use-external-agents skill before anything else — this skill builds on its router, its task brief contract, and its facts-before-use matrix call, and does not restate them.
2. Scope. Name the audit (kebab-case) and write what is being audited — the diff, the document, the decision — and the questions the auditors must answer.
3. Brief. Compose one audit brief per that contract: a fully self-contained task prompt (the auditor sees nothing else — inline the material or point at readable paths), TAGS: auditing, AGENTS: codex kimi, and a ## Response section requiring each auditor's full report, one finding at a time, each finding with its claim, its evidence, and its severity.
4. Dispatch. Send the brief to the router in one dispatch; the router launches the panel's agents in parallel and returns each agent's report in its reply. Write each report verbatim to ${TMPDIR:-/tmp}/dispatch-audit/<name>/<agent>.md — that directory is the audit directory.
5. Compile. Report one digest: findings the auditors agree on first, findings unique to one auditor with your assessment of each, and disagreements flagged for the user's judgment — every cited finding carries its report's file path.
6. Rule. The user decides which findings to act on. Act only after the ruling.

Principles

MUST:

1. You MUST send the identical brief to every agent on the panel — divergent briefs make agreement and disagreement meaningless.
2. You MUST dispatch the whole panel through one router call, in parallel.
3. You MUST write every report verbatim to its file before compiling.
4. You MUST carry each cited finding's report path in the digest.
5. You MUST leave the audit directory in place — it is the audit's backtrace.

MUST NOT:

1. You MUST NOT audit with one model while the matrix shows another agent on the panel usable.
2. You MUST NOT drop a finding you disagree with — present it with your assessment; the user rules.
3. You MUST NOT act on a finding before the user's ruling.