agent-watchdog

name	agent-watchdog
description	Audit whether another agent actually did what was asked, not just whether the code is good. Use to watch, audit, compare, or fix another agent's work: a Claude Code or Codex session, a transcript, a PR, a branch, a CI run, or a workflow run.

Agent Watchdog

Watch another agent's work like a reviewer with a pager: reconstruct the request, verify the evidence, and close the gap between what was asked and what actually happened. This audits the agent (did it do the task, did it really verify), not the code itself; pair it with /review or /code-review for the code.

Adapted from BuilderIO/skills (MIT).

Choose the mode

Infer the mode from the user's wording:

• Watch only: monitor a session, PR, branch, CI run, or workflow until it reaches a terminal state. No edits.
• Audit: read the prompt, transcript, diff, tests, CI, and final claims, then return a gap report. No edits.
• Audit and fix: audit first, then make narrow fixes for clear gaps. No broad rewrites, branch moves, or speculative changes.
• Compare: given multiple agents or runs, judge each against the same original request and reconcile the differences.

If authority is unclear, default to audit-only and say what you would fix.

Resolve the target

1. Identify every artifact supplied: session or transcript path, thread URL, PR, branch, commit, CI run, issue, or pasted summary.
2. Resolve it from the most direct source (local transcript, repo logs, gh for PR/CI), not a summary.
3. If it is still running and you were asked to watch, poll at a reasonable interval until it is done, blocked, stale, or waiting on a human.
4. If it cannot be resolved, ask for the missing identifier.

Reconstruct the contract

Build a compact contract before judging the work:

• The original request and any later scope changes. This is the source of truth, NOT the agent's own summary.
• Explicit constraints: branch rules, no-edit requests, versions, validation expectations, security or privacy limits.
• Implied acceptance criteria: user-visible behavior, tests, CI, docs, deploys, review replies.
• The agent's final claims and its "could not do" caveats.

Audit the evidence

Inspect evidence, not vibes:

• Read the changed files and the relevant unchanged files around them.
• Check git status and diff without reverting unrelated work.
• Compare the commands the agent claimed to run against actual output.
• Inspect failed or skipped tests, CI logs, and error traces; for UI work, prefer screenshots or a browser check over prose claims.
• For PR work, verify unresolved threads and CI state from the source system.

Classify each issue:

• Gap: requested behavior missing or incomplete.
• Bug: the implementation likely fails or regresses.
• Verification miss: may be right, but the evidence is weak.
• Scope drift: changed something unrelated or skipped a constraint.
• No issue: already handled, with evidence.

Fix narrowly (only when authorized)

1. Fix only gaps with clear evidence.
2. Preserve unrelated local changes; do not move branches unless explicitly asked.
3. Use existing repo patterns and targeted tests.
4. Re-run the smallest useful validation after each meaningful fix.
5. If a fix needs a product decision, a credential, a destructive action, or a broad rewrite, stop and report it instead of guessing.

Report

Lead with the outcome. Keep it scannable:

Status: done / blocked / stale / running
Requested: what the user asked the agent to do
Observed: what the agent changed, claimed, and verified
Gaps: missing behavior, bugs, weak verification, scope drift
Fixes made: files changed + validation run (omit for audit-only)
Remaining risk: unverified, or waiting on CI/review/deploy/human

Name exact files, commands, PRs, or thread IDs when they matter.

See Also

The post-agent triad, each a different question about a finished run:

• /debrief: recap of what agents accomplished across sessions, with loose ends; this audits one run against its ask
• /sweep: post-op check for artifacts and damage left behind; this checks whether the work got done
• /review: structured review of the code itself, with reasoning capture
• /debugging: when an audited gap turns out to be a real bug, build the loop and isolate it
• RECIPE.md: Agent recipe for parallel decomposition (2 workers)