refactor: route babysit through structured monitors

[kyleseaman]

Stack: #5171 → #5172 → #5173 → #5174 → #5175 → #5176 → #5177

Position: 7 of 7. Base: #5176.

Routes GitHub babysitting through bounded structured monitors, keeps a finite unsupported-target fallback, and adds executable lifecycle scenarios.

Verification: lifecycle scenarios, monitor regression suites, docs, brand, scrub, and stack-wide gates.

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of 6ce882d5b819008e51e9d7184a8ee12cc3532917 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 6ce882d

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 6ce882d5b819008e51e9d7184a8ee12cc3532917: <one-sentence reason>

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of 6ce882d5b819008e51e9d7184a8ee12cc3532917 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

Moves GitHub-PR watch policy from 600 lines of advisory prose into engine-enforced bounded monitors; the retained legacy fallback stays finite by construction.

Watch

• The legacy monitor_start recipe drops the old skill's stall-tripwire ("3 identical settled cycles → stop and escalate"), so an unsupported-target loop again coasts to its cap indistinguishably from convergence — the exact measured failure (4/4 loops cap-out) the old prose fought. Cost is now capped by the required runtime/cycle bounds, but the outcome ambiguity returns for non-GitHub targets until the engine grows more monitor kinds.

[DESIGN-REVIEWED] 6ce882d

[kyleseaman]

Replaced by #5186 after renaming the branch to remove the prefix.

[github-actions]

First Principles Review (Fable 5) — 🟡 CONCERNS

Premise-level review of 6ce882d5b819008e51e9d7184a8ee12cc3532917 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

All evidence gathered — the contract's counts are done (budget-check call sites, token_usage_known provenance, controller ordering). Emitting the review.

First-Principles-Verdict: CONCERNS

A "refactor" ships an undeclared behavior fix: an exhausted monitor no longer buys one more provider probe or BUSY redispatch — real, derived, but unnamed.

What this change ships

Intent: stop paying a full agent turn per babysit poll by routing GitHub PR watching through the server-side bounded monitor — a FIX to routing/cost.

1. Babysit skill now issues one bounded monitor_watch; 547→78 lines — justified (deletion)
2. Skill loads pointer-on-trigger instead of full-body injection — justified, declared in spec
3. Spent budget now stops the monitor before another probe or BUSY redispatch — undeclared, rides along
4. token_usage_known now visible to the agent via monitor_inspect — justified (skill reads it)
5. monitor_start relabeled legacy fallback for unsupported targets only — justified
6. autonudge_stop relabeled compatibility alias with durable retained outcome — justified
7. monitor_watch reply now states application is pending until turn end — justified
8. Legacy recipe requires finite interval, cycle cap, and runtime — justified
9. New user doc monitoring.md plus two index rows — declared, justified
10. Executable lifecycle scenario suite — declared

Watch

The description says only "Routes GitHub babysitting through bounded structured monitors", but the diff flips test_monitor_controller.py from dispatched.await_count == 2 to == 1: a spent budget previously bought one more redispatched agent turn and one more provider probe. That is a behavior fix under a refactor: type — derived (it enforces decision.py's own "a spent bound must never buy one additional unattended turn" invariant, and the spec hunk documents it), so the only gap is the framing; a human should know this PR changes runtime behavior, not just routing. Grep confirms no duplication: monitor_budget_reason had 4 post-probe call sites in autonudge.py; this adds the only pre-probe/pre-redispatch one, which the post-probe sites cannot reach by construction.

[FIRST-PRINCIPLES-REVIEWED] 6ce882d