Background agents complete work but coordinator never reports results #432

Summary

Background-spawned child agents complete their work (files written, commits made) but the coordinator never receives their completion signal. The coordinator goes silent after spawning agents and doesn't present results or trigger follow-up work (Scribe, chaining, etc.).

Environment

• Squad version: 0.8.25
• Platform: GitHub Copilot (VS Code / CLI)
• OS: Windows 11

Observed behavior

1. Coordinator spawns 2-3 agents in parallel via mode: "background"
2. Agents execute successfully — files are written, commits land (verified in git log)
3. Coordinator never comes back to report what the agents did
4. No compact results shown, no Scribe spawned, no follow-up work chained
5. User has to manually ask "What did the team do?" to trigger catch-up

Expected behavior

Per the "After Agent Work" section in squad.agent.md:

1. Coordinator calls read_agent (wait: true, timeout: 300) to collect results
2. If read_agent returns empty, coordinator runs silent success detection — checks filesystem for modified history.md, new decision inbox files, or output files
3. Coordinator presents compact results and spawns Scribe

Reproduction evidence

In one session, the coordinator spawned 3 agents in parallel as background tasks for a demo readiness sprint:

• Ripley — deployment readiness assessment
• Dallas — UI polish across 6 pages
• Lambert — code audit and runtime bug fixes

All three agents completed successfully. Their orchestration logs all show Status: ✅ Complete, and the work landed in a single commit (22 files changed, +1115/-577 lines). The agents clearly did their jobs.

However, the coordinator never came back. The evidence:

1. No compact results presented. The expected post-work output (e.g., 🏗️ Ripley — Deployment ready, no blockers) never appeared. The user discovered results by checking git log.
2. Scribe was never spawned. The Scribe's history.md still contains only its initialization entry ("Agent Scribe initialized and ready for work"). The orchestration logs and session log that do exist were written by the agents themselves, not by a Scribe post-work pass.
3. identity/now.md was never updated. It still reads focus_area: Initial setup from the original session — the coordinator is supposed to update this after work completes, but it never got the chance.
4. Session file has no post-work messages. The session JSON only contains the coordinator's initial greeting and one follow-up — no messages after the background spawns.

This pattern has reproduced on every background fan-out in this project, not just this session.

Notes

The squad.agent.md documents this as a known platform bug ("Silent Success") with an estimated ~7-10% rate, but in my experience the failure rate is 100% — the coordinator has never successfully reported back after background agent spawns. The documented mitigation (filesystem fallback check via read_agent) does not appear to fire at all.

It's unclear whether:

• read_agent is timing out before agents finish
• The coordinator's context is overflowing after fan-out (the "Server Error Retry Loop" bug mentioned in the same comment)
• The mitigation logic is being skipped due to context pressure

Suggestion

Consider adding a more robust completion mechanism — e.g., agents write a structured completion marker file that the coordinator polls for, rather than relying solely on read_agent text responses. Alternatively, surface the read_agent timeout and retry logic more explicitly so it can be tuned.