review-bench/davidortinau_sentencestudio/agent-progress-diagnostic

Context

Squad agents like Wash (publish), Kaylee (feature work), Jayne (E2E) routinely make 50–150 tool calls per task. From the Coordinator's seat, a long-quiet agent at ~80+ tool calls looks identical to a hung agent. The reflex is to assume it's stuck and either kill it (stop_bash) or escalate to Captain. Both are usually wrong.

Apply this skill when: an async background agent has been running long enough that you're tempted to ask "is it hung?" Before any kill action, run the rubric.

Patterns

Tool-call envelope per role (rough budgets)

These are observational — refine as more data lands.

Diagnostic rubric (run in order)

1. Check elapsed seconds and tool-call count. Within envelope? → almost certainly fine, just wait. Past envelope? → continue.

1. Send a status `write_agent` ping. One line, low-cost:

`` "status check — what step are you on?" ``

• If reply arrives within 30 seconds → agent is alive and processing. Wait.
• If reply is queued behind a long turn → still alive. Wait for the in-flight turn to finish.
• If no reply after 2+ minutes → continue to step 3.

1. Check filesystem for files the agent should be writing. Most Squad agents land an inbox decision file (/.squad/decisions/inbox/<agent>-<task>.md) or a screenshot near completion. Use ls -la on the expected output dir:

``bash ls -la /Users/davidortinau/work/SentenceStudio/.squad/decisions/inbox/ `` A file with a recent timestamp (mtime within last 60s) means the agent IS writing — give it more time.

1. Check the agent's history/log file. Most Squad agents append to .squad/agents/<name>/history.md or write orchestration log entries. Recent timestamp → alive.

1. Only consider stop_bash if all four signals say stuck. Specifically: tool-call count past envelope AND ping unanswered AND no recent file writes AND no log activity. Even then, prefer write_agent with a "wrap up and report what you have" message over a hard kill — graceful shutdown preserves work in progress.

Heuristics

• A long-running agent at 80+ tool calls is making progress 95% of the time. The only reason it looks hung is that you can't see the in-flight turn.
• The "is it hung?" question is usually a Coordinator anxiety, not an agent problem. When in doubt, send a status ping — it's cheap and definitive.
• Never kill an agent that's writing a decision file. Inbox file writes can be the last 5–10 tool calls of an otherwise complete task. Killing during that window destroys the work product.
• If Captain asks "is it hung?", run the rubric out loud — narrate each check. This documents the diagnosis and trains the rubric.

Examples

2026-05-07 — Wash Publish #9 (the founding case)

• Symptom: Wash at ~83 tool calls, ~12 minutes elapsed, no recent stdout. Captain: "is it hung?"
• Rubric applied:

1. Tool-call count: 83. Wash envelope is 60–120. Within range. → Likely fine.
2. Status write_agent: "status check?" → Wash replied within seconds with "writing decision file."
3. Filesystem check: .squad/decisions/inbox/ had a fresh wash-publish-9-*.md mtime within 30s. → Confirmed alive.
4. No kill needed. Wash completed successfully ~2 minutes later.

• Outcome: Coordinator avoided a false-positive kill. This case is why the skill exists.

Anti-Patterns

• ❌ Reaching for `stop_bash` because "it's been a while." Set a stopwatch, run the rubric, only kill on hard evidence.
• ❌ Escalating to Captain with "I think Wash is hung" before pinging the agent yourself. Captain will rightly ask "did you check?"
• ❌ Comparing tool-call count against the WRONG envelope. Wash's normal range is not Scribe's normal range. Use the table above.
• ❌ Killing during inbox-file write window. If .squad/decisions/inbox/ has a fresh mtime, the agent is in its closing moves. Wait.

References

• Founding case: Wash Publish #9, 2026-05-07. See .squad/agents/wash/history.md (most recent entry) and .squad/decisions.md (Publish #9 section).
• Tools: write_agent, read_agent, list_agents, stop_bash. See Copilot CLI tool reference.
• Companion: .squad/orchestration.md — broader Coordinator runbook.
• Confidence note: Currently low (n=1). Upgrade to medium after 3+ observations validate the rubric across different agent roles.