Part of Claude Code operator skills: a collection of skills for running a real Claude Code setup.
Agent watchdog
The problem: an agent finishes a session, reports “done,” and you have no independent way to know whether that’s true. Reading the transcript takes as long as doing the work yourself. Trusting the summary means trusting the exact thing you’re trying to verify.
This skill watches another agent’s work like a reviewer with a pager: wait for completion when needed, reconstruct the original request, check the evidence instead of the claim, and close the gap between what was asked and what actually happened.
Choose the mode
Infer the mode from the user’s wording:
- Watch only: monitor a session, PR, branch, CI run, or transcript until it
reaches a terminal state. Do not edit files.
- Audit: read the prompt, transcript, diff, tests, CI, comments, screenshots,
or final claims and return a gap report. Do not edit files.
- Audit and fix: audit first, then make narrow fixes for clear gaps. Avoid
broad rewrites, branch movement, or speculative changes.
- Compare: when given multiple sessions or agents, compare their work against
the same original request and reconcile the important differences.
If authority is unclear, default to audit-only and say what you would fix.
Resolve the target
- Identify every artifact the user supplied: session ID, transcript path,
thread URL, PR, branch, commit, CI run, issue, chat link, or pasted summary.
- Use the host’s native thread/history tools, local transcript files, repo
logs, GitHub tools, or pasted content to resolve the artifact. Prefer the
most direct source over summaries.
- If the artifact is still running and the user asked to watch, poll at a
reasonable interval until it is done, blocked, stale, or clearly waiting on a
human or external system.
- If the artifact cannot be resolved, ask for the missing identifier or path.
Reconstruct the contract
Build a compact contract before judging the work:
- Original user request and any later changes in scope.
- Explicit constraints: branch rules, no-edit requests, deadlines, package
versions, validation expectations, design requirements, or security/privacy
limits.
- Implied acceptance criteria: user-visible behavior, tests, CI, docs, deploys,
screenshots, review replies, or status updates.
- The other agent’s final claims and any “could not do” caveats.
Treat the user’s request as the source of truth, not the other agent’s summary.
Audit the evidence
Inspect evidence, not vibes:
- Read changed files and relevant unchanged files around the touched paths.
- Check git status/diff without reverting unrelated work.
- Compare commands the agent claimed to run with actual output when available.
- Inspect failed or skipped tests, CI logs, browser screenshots, review
comments, deploy output, and error traces.
- For PR/review work, verify unresolved threads and CI state from the source
system when tools are available.
- For UI work, prefer screenshots or browser checks over prose claims.
Classify each issue as:
- Gap: requested behavior is missing or incomplete.
- Bug: the implementation likely fails or regresses behavior.
- Verification miss: the work may be right but the evidence is weak.
- Scope drift: the agent changed something unrelated or skipped a constraint.
- No issue: the concern is already handled, with evidence.
Fix narrowly
When the user authorized repair:
- Fix only gaps with clear evidence.
- Preserve unrelated local changes and do not move branches unless explicitly
asked for that branch operation.
- Use existing repo patterns and targeted tests.
- Re-run the smallest useful validation after each meaningful fix.
- If a fix would require a product decision, credential, destructive action, or
broad rewrite, stop and report the decision instead of guessing.
Report
Lead with the outcome. Keep the report short enough to scan:
Status
- Done, blocked, stale, or still running.
Requested
- What the user asked the watched agent to do.
Observed
- What the watched agent changed, claimed, and verified.
Gaps
- Missing behavior, bugs, weak verification, or scope drift.
Fixes made
- Files changed and validation run. Omit this section for audit-only work.
Remaining risk
- Anything still unverified or waiting on CI/review/deploy/human input.
Name exact files, commands, PRs, or thread IDs when they matter.
1---2name: ww-agent-watchdog3description: Use when asked to watch, babysit, audit, review, compare, or fix another agent’s work from a Codex session ID, Claude Code session/transcript, chat/thread link, PR, branch, log, or pasted run summary. Monitor until the other agent is done or blocked, reconstruct what the user asked, inspect what the agent actually changed and verified, report gaps, and optionally make scoped fixes when the user authorizes repair.4---56> Part of [Claude Code operator skills](https://github.com/lee-fuhr/claude-operator-skills): a collection of skills for running a real Claude Code setup.78# Agent watchdog910**The problem:** an agent finishes a session, reports “done,” and you have no independent way to know whether that’s true. Reading the transcript takes as long as doing the work yourself. Trusting the summary means trusting the exact thing you’re trying to verify.1112This skill watches another agent’s work like a reviewer with a pager: wait for completion when needed, reconstruct the original request, check the evidence instead of the claim, and close the gap between what was asked and what actually happened.1314## Choose the mode1516Infer the mode from the user’s wording:1718- **Watch only:** monitor a session, PR, branch, CI run, or transcript until it19 reaches a terminal state. Do not edit files.20- **Audit:** read the prompt, transcript, diff, tests, CI, comments, screenshots,21 or final claims and return a gap report. Do not edit files.22- **Audit and fix:** audit first, then make narrow fixes for clear gaps. Avoid23 broad rewrites, branch movement, or speculative changes.24- **Compare:** when given multiple sessions or agents, compare their work against25 the same original request and reconcile the important differences.2627If authority is unclear, default to audit-only and say what you would fix.2829## Resolve the target30311. Identify every artifact the user supplied: session ID, transcript path,32 thread URL, PR, branch, commit, CI run, issue, chat link, or pasted summary.332. Use the host’s native thread/history tools, local transcript files, repo34 logs, GitHub tools, or pasted content to resolve the artifact. Prefer the35 most direct source over summaries.363. If the artifact is still running and the user asked to watch, poll at a37 reasonable interval until it is done, blocked, stale, or clearly waiting on a38 human or external system.394. If the artifact cannot be resolved, ask for the missing identifier or path.4041## Reconstruct the contract4243Build a compact contract before judging the work:4445- Original user request and any later changes in scope.46- Explicit constraints: branch rules, no-edit requests, deadlines, package47 versions, validation expectations, design requirements, or security/privacy48 limits.49- Implied acceptance criteria: user-visible behavior, tests, CI, docs, deploys,50 screenshots, review replies, or status updates.51- The other agent’s final claims and any “could not do” caveats.5253Treat the user’s request as the source of truth, not the other agent’s summary.5455## Audit the evidence5657Inspect evidence, not vibes:5859- Read changed files and relevant unchanged files around the touched paths.60- Check git status/diff without reverting unrelated work.61- Compare commands the agent claimed to run with actual output when available.62- Inspect failed or skipped tests, CI logs, browser screenshots, review63 comments, deploy output, and error traces.64- For PR/review work, verify unresolved threads and CI state from the source65 system when tools are available.66- For UI work, prefer screenshots or browser checks over prose claims.6768Classify each issue as:6970- **Gap:** requested behavior is missing or incomplete.71- **Bug:** the implementation likely fails or regresses behavior.72- **Verification miss:** the work may be right but the evidence is weak.73- **Scope drift:** the agent changed something unrelated or skipped a constraint.74- **No issue:** the concern is already handled, with evidence.7576## Fix narrowly7778When the user authorized repair:79801. Fix only gaps with clear evidence.812. Preserve unrelated local changes and do not move branches unless explicitly82 asked for that branch operation.833. Use existing repo patterns and targeted tests.844. Re-run the smallest useful validation after each meaningful fix.855. If a fix would require a product decision, credential, destructive action, or86 broad rewrite, stop and report the decision instead of guessing.8788## Report8990Lead with the outcome. Keep the report short enough to scan:9192```md93Status94- Done, blocked, stale, or still running.9596Requested97- What the user asked the watched agent to do.9899Observed100- What the watched agent changed, claimed, and verified.101102Gaps103- Missing behavior, bugs, weak verification, or scope drift.104105Fixes made106- Files changed and validation run. Omit this section for audit-only work.107108Remaining risk109- Anything still unverified or waiting on CI/review/deploy/human input.110```111112Name exact files, commands, PRs, or thread IDs when they matter.
lee-fuhr