feat(delegation): follow through when Pylon children stop or finish

[rynfar]

Pylon children could finish, stop, fail, or need user input while their parent stayed idle indefinitely. This adds event-driven parent follow-through for Pylon child threads only; built-in subagents keep upstream behavior.

The server records child lifecycle observations and batches pending updates into an ordinary parent turn at an eligible idle boundary. Admission preserves parent configuration, rejects stale parent/child state, respects stops and blockers, and rechecks delegation settings. Durable per-notification receipts prevent replay across restarts and long histories. Historical completed children are baselined. Three automatic turns after the latest ordinary user message pause follow-through in the existing work log; sending a new message permits continuation. No polling loop, new panel, provider orchestration service, or database migration.

Updated the shipped delegation skill and user guide. Turning delegation off prevents new automatic parent turns without stopping children. Automatic notifications do not approve actions, retry interrupted children, or treat completion as verified work. Busy parents receive updates at their next eligible idle boundary rather than mid-turn steering.

Validation:

• 236 focused tests passed across reactor, pure lifecycle logic, admission decider, real SQLite engine, reactor startup, MCP handlers, and existing work-log presentation.
• Real-engine tests cover atomic parent reservation/receipt, duplicate messages, mixed old/new notices beyond 505 activity rows, settings disabled/missing, and observation ownership. No paid provider calls.
• Reactor tests cover restart recovery, historical baselines, imported history, busy-parent batching, resolved blockers, native-only exclusion, and background completion including Claude status-bearing progress.
• Server, contracts, client-runtime, mobile, and web typechecks passed; scoped lint/format passed.
• Independent adversarial review identified lifecycle, truncated-history deduplication, and settings-admission gaps; each was fixed and regression-tested. Final review found no blockers.

Separate from #597, which fixes stale native-agent status and the misleading sidebar timer. This PR does not alter built-in subagent orchestration.

Model: GPT-6 Astra. Harness: Codex in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
pylon-marketing	Ready	Preview	Sep 17, 2026 6:13pm UTC

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ The exact PR base did not have a successful artifact. Baseline uses the latest successful main measurement shown below.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	14.0 KiB	13.9 KiB	−28 B (−0.2%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.2 KiB	7.2 KiB	+6 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.7 KiB	−34 B (−0.5%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.0 KiB	58.0 KiB	−44 B (−0.1%)	66.4 KiB	✅
Codex	Live turn messages	10	9	−1 (−10.0%)	21	✅
Claude	Total thread wire	14.0 KiB	14.0 KiB	+42 B (+0.3%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.2 KiB	7.2 KiB	+7 B (+0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+35 B (+0.5%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.9 KiB	58.9 KiB	+44 B (+0.1%)	66.4 KiB	✅
Claude	Live turn messages	9	10	+1 (+11.1%)	21	✅

Baseline: 31a5464 · PR result: f9aac53 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 115.6 KiB
• Claude decoded thread snapshot: 116.3 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.