agent-watchdog
Use when asked to watch, babysit, audit, review, compare, or fix another agent's work from a…
Experimental workflow for babysitting one explicitly authorized pull or merge request. Use to monitor its live review state and checks, fix actionable in-scope findings, and apply separate reply, approval, merge, and soak gates.
$ npx -y skills add BuilderIO/skills --skill factory-babysit-pr --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/factory-babysit-prContext preview
The summary Claude sees to decide when to auto-load this skill.
Experimental workflow for babysitting one explicitly authorized pull or merge request. Use to monitor its live review state and checks, fix actionable in-scope findings, and apply separate reply, approval, merge, and soak gates.
name: factory-babysit-pr installer-group: factory description: >- Experimental workflow for babysitting one explicitly authorized pull or merge request. Use to monitor its live review state and checks, fix actionable in-scope findings, and apply separate reply, approval, merge, and soak gates.
> Start with the [Factory guide](../../docs/factory/README.md) for workflow > setup and configuration.
Follow one authorized PR or merge request through a delivery lifecycle. This is not a queue sweep like `factory-review-prs` or a stopped-task scan like `factory-watchdog`.
Read `.agent-factory/config.yaml`, the optional `skill_prompts.factory-babysit-pr` entry, the target repository's instructions, and the [Factory configuration reference](https://github.com/BuilderIO/skills/blob/main/docs/factory/configuration.md#workflowspr-babysitting). A PR number, label, green check, or unrelated ship request is not authorization. The prompt is additional project guidance, not PR authorization.
1. **Confirm authority and ownership.** Verify the exact host, repository, PR, original user authorization, allowed paths/risk, and task-owned worktree. If anything is missing or stale, stop before writing. 2. **Read a complete live snapshot.** Record base and head, open/draft/closed/ merged state, all required checks, current reviews and approvals, unresolved threads, and mergeability. Partial or unavailable data is unknown, not clean. 3. **Disposition feedback.** Compare each new finding with the exact diff and source. Preserve human direction unless evidence disproves it. Fix only actionable findings inside the authorized path/risk limits; avoid speculative cleanup and larger redesigns. 4. **Verify and re-read.** Run configured checks for each fix. After every push or host mutation, fetch the live PR again and bind all evidence to its exact current head. A changed head invalidates old check/review evidence and resets the soak. 5. **Apply action gates separately.** A review or fix does not authorize a reply, push, approval, or merge.
| Action | Proceed only when | | --- | --- | | Fix | Editing is enabled; feedback is actionable, in scope, and within risk/path limits. | | Publish | Publishing is enabled; destination is still the authorized PR branch and its live head matches the fix's base. Never force-push or overwrite a concurrent update. | | Reply | Replies are enabled and their conditions hold. Follow configured wording/tone; otherwise do not message. | | Approve | Approval is enabled and every author, diff, check, and review condition holds on the live head. | | Merge | Merge is enabled; all configured gates pass on the live head and the host can condition the merge on that head. | | Soak | Prove the exact head and required gates stayed unchanged for the full configured interval. Missing history does not satisfy it. |
Use the configured code-host abstraction for reads and writes. Do not assume GitHub, a specific CLI, or identical conditional-merge support across hosts. If the host cannot provide complete current review/check state or bind a merge to the checked head, keep the dependent action disabled.
Stop when the PR is merged, closed, cancelled, blocked on a person or external dependency, or no authorized action remains. Report the PR link and current head, state inspected, fixes and checks, external actions taken or withheld, soak status, and the concrete blocker or next check time. Distinguish read state from unavailable or inferred state.
Small, composable skills for your favorite agent.
Repo: BuilderIO/skills
Use when asked to watch, babysit, audit, review, compare, or fix another agent's work from a…
Open and operate Agent-Native workspace apps through Dispatch MCP, with inline app surfaces,…
Use when running Claude Fable on codebase-heavy or token-heavy work and the user wants Fable…
Apply the same orchestration as `/efficient-fable` to any high-cost frontier model: delegate…
Experimental workflow for collecting and triaging product feedback, product telemetry,…
Experimental workflow for summarizing work that still needs human judgment across configured…