babysit

name	babysit
description	Use when the user asks to babysit a PR, run /babysit, keep a PR green, monitor review/CI until ready, or optionally close out a PR through configured merge/capture/retro automation. Can run directly; host goal/persistence support is optional.

Babysit

Babysit the current pull request until configured live evidence says it is green, then perform the configured closeout path when policy allows it.

This is an outer-loop workflow. It does not replace review-response; it repeatedly uses review-response for feedback handling, configured gates for proof before push, and optional host persistence for wait/recheck cycles.

Persistence model

babysit can run directly in any host. Goal/persistence support is optional, not a hard requirement.

• In Claude Code, users may manually wrap babysit in /goal for cross-wait persistence.
• In Pi, bundled /babysit starts a Beislið-owned babysit runtime with loop state and completion tools. It does not depend on pi-goal or invoke slash commands from another extension.
• Without persistence, perform the current audit/feedback/gate step. If the PR is not terminal green or blocked by turn end, stop with evidence and the next recheck/action.

Inputs

Use, in order:

• invocation args such as stop when green, merge, don't merge, skip memento, or skip retro
• .beislid/workflow.md
• current branch PR metadata from configured PR host tooling when available
• configured pr_review_source and pr_review_update
• configured gates, scopes, or gate sets
• configured babysit closeout policy
• optional review_policy AgenticReviewer risk/opt-in policy

Configuration

Optional workflow block:

goal:
  token_budget: 50k
loop:
  use_review_response: true
  run_configured_gates_before_push: true
  wait_interval_seconds: 60
  timeout_minutes: 60
closeout:
  merge:
    mode: ask # off | ask | auto
    method: squash # squash | merge | rebase | repo-default
    delete_branch: true
  memento:
    mode: ask # off | ask | auto
  retro:
    mode: ask # off | ask | auto
    apply_findings: ask # off | ask | auto

Defaults when absent: no goal budget, use review-response, run configured gates before pushes, merge/memento/retro off. Invocation args override config for this run.

Workflow

1. Load project config — read .beislid/workflow.md; hard-fail if missing or wrong version, matching other repo-aware orchestrators.
2. Detect persistence mode — note direct, manually host-goal-wrapped, or Pi Beislið runtime. Do not stop solely because persistence is absent.
3. Find the PR — use configured PR host data or gh pr view when available. If no current PR is found, stop and ask for the PR URL/number.
4. Read live state — collect checks, mergeability/conflict state, review decision, PR-level comments, and inline review threads.
5. Detect feedback and review-policy state — use pr_review_source; never trust a green review-bot status alone. Treat AgenticReviewer provider comments containing Review skipped, Review limit reached, rate limited, or draft detected as not reviewed / deferred review, even if the check is green. If review_policy.agentic_reviewer.mode: opt_in_final_review is configured, classify current PR risk with ready-for-review's path/size rules and require a real AgenticReviewer review when risk > max_auto_closeout_risk. If required review is missing and the configured opt-in label is absent, add it when action policy allows; label is required for automatic opt-in, so if it is missing or label add fails, stop/ask before using description_keyword. Unresolved review comments or requested changes route to review-response when enabled.
6. Feedback handling — if loop.use_review_response is true, invoke review-response; it owns categorization, fixes, replies, commits, pushes, and child-ticket handling. If false, do not fix/reply/commit/push automatically; stop with the feedback summary and ask how to proceed.
7. Gate before push — before any babysit-owned push or merge preparation, run the configured applicable gates. Use existing scope/gate-set selection rules where available. Do not invent hardcoded gates.
8. Wait and recheck — after pushes or pending checks, wait using bounded polling or host monitor facilities. Do not busy-loop. Re-read live PR state after each transition.
9. Green audit — green means required checks successful, mergeable/no conflicts, acceptable review state, no unaddressed actionable feedback, and any risk-required AgenticReviewer review is real. Green status plus deferred-review evidence is not acceptable.
10. Closeout — perform configured merge, memento capture, and retro only when green audit passes and action policy allows. Policy-check closeout side effects: gh.pr.merge or pr.merge as git-remote, memento.capture as workspace-write, retro.run as read plus workspace-write when it may write artifacts, and retro.apply/setup edits as workspace-write. Stop for approval when mode is ask; proceed only when mode is auto and policy allows.
11. Complete persistence loop when present — call completion only after final audit and configured closeout, or the stop-when-green endpoint. In Pi runtime, call update_beislid_babysit({status:"complete", summary:"..."}); if blocked, call update_beislid_babysit({status:"blocked", summary:"..."}).

Safety stops

Stop and ask when any occur:

• no PR can be identified
• red or pending required checks at a merge boundary
• conflicts or unsafe merge resolution
• review feedback requiring architectural, product, scope, or policy judgment
• user-approved pushback is needed
• required credentials or external services are unavailable
• action policy returns ask and no approval has been given
• action policy returns deny
• every closeout step is disabled, in which case stop green and report
• retro findings would require editing workflow/config and apply_findings is not auto

Never force-push or amend published commits. Never merge to bypass a failing/pending required check. Never interpolate review reply bodies into shell commands; use temp JSON payload files.

Closeout semantics

• merge.mode: off — stop once green and report.
• merge.mode: ask — show final audit and merge command/method; wait for approval.
• merge.mode: auto — merge after green audit only if action policy allows. If policy asks, ask; if policy denies, stop.
• memento.mode controls durable knowledge capture after merge/green endpoint.
• retro.mode controls running retro after closeout.
• retro.apply_findings controls whether accepted retro/setup findings may be applied automatically. auto still respects action policy; ambiguous/destructive edits stop.

Output

When stopping, report: PR URL/branch, final state, checks/review/mergeability evidence including deferred-review comments, feedback handled, gates run, closeout side effects, and remaining blockers or human decisions.

Common mistakes

• Treating CI green as enough when review threads are unresolved.
• Treating green AgenticReviewer status as reviewed when comments say skipped, rate-limited, or draft-detected.
• Hardcoding provider labels, gates, or risk thresholds instead of reading workflow config.
• Assuming /goal is required; direct runs are allowed but must stop with next steps when no persistence mechanism is active and the PR is not terminal.
• Posting replies or merging without policy and approval handling.
• Mentioning private provenance in public tracker/PR updates; keep public updates focused on the current repo and PR only.