fix(harness): raise idle watchdog default to 600s so long compaction survives

[TomeHirata]

Related issue

N/A

Summary

• The per-turn idle watchdog fails a turn that emits no non-heartbeat events for its window (default 240s). It resets on every real progress event, so it's meant to catch a wedged turn — not a slow-but-healthy one.
• Context compaction's summarizing LLM call (runtime/compaction.py → summarize_history) runs as a single long await that emits nothing until it returns. On a near-full context that call can take longer than 240s, so the idle watchdog trips it.
• This wedges the session in a loop reported in Slack: Prompt is too long → auto-compaction → turn exceeded the 240s harness idle watchdog (run_turn emitted no events for 240s; likely a wedged LLM or tool call) → retry re-triggers the same slow compaction → repeat.
• Fix: raise the idle watchdog default from 240s to 600s in omnigent/runtime/harnesses/_scaffold.py, giving a healthy long compaction room to finish. The HARNESS_TURN_TIMEOUT_S env knob and the absolute per-turn ceiling (HARNESS_TURN_ABSOLUTE_TIMEOUT_S, 3600s) are unchanged.

Test Plan

• pytest tests/runtime/harnesses/test_scaffold.py -k "watchdog or idle or wedged or busy or absolute or heartbeat" — 6 passed. These drive the watchdog via HARNESS_TURN_TIMEOUT_S overrides, so they're independent of the default and confirm idle/absolute/heartbeat behavior still holds.
• No test asserts the 240 default; the only other 240 literals in the tree are unrelated (MCP startup settle, pi-native park timeout).

Demo

N/A

Type of change

• Bug fix
• Feature
• UI / frontend change
• Refactor / chore
• Docs
• Test / CI
• Breaking change

Test coverage

• Unit tests added / updated
• Integration tests added / updated
• E2E tests added / updated
• Manual verification completed
• Existing tests cover this change
• Not applicable

Coverage notes

Behavior is unchanged except for the default window length. Existing scaffold watchdog tests exercise idle-timeout, absolute-cap, and heartbeat-ignoring paths via explicit HARNESS_TURN_TIMEOUT_S overrides, so they cover the mechanism regardless of the default. I ran that subset locally (6 passed). No new test is added because the change is a single default constant with no new branch.

Changelog

Long context compaction no longer fails with a 240s idle-watchdog timeout on near-full sessions

---

This pull request and its description were written by Isaac.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[omnigent-ci]

Polly AI Review

Summary

Low-risk, single-constant change: raises the idle-watchdog default from 240s to 600s in _scaffold.py, plus an explanatory comment. The reasoning is sound and matches the code — every non-heartbeat ctx.emit resets the idle deadline, so the watchdog is meant to catch a wedged turn, and a long single await (context compaction's summarize_history) genuinely emits nothing until it returns, so it can trip a too-tight window. Raising the default gives healthy compaction room without touching the absolute per-turn ceiling.

Blocking issues

None. Verified:

• _TURN_IDLE_TIMEOUT_S is the idle watchdog; the new default 600 remains well under the unchanged absolute cap _TURN_ABSOLUTE_TIMEOUT_S = 3600, so the backstop still bounds a runaway-but-active turn.
• The HARNESS_TURN_TIMEOUT_S env knob and <= 0 disable path are untouched — ops can still override.
• No test asserts the 240 literal; the scaffold watchdog tests drive the mechanism via explicit HARNESS_TURN_TIMEOUT_S overrides ("2", "10") and a direct _TURN_IDLE_TIMEOUT_S monkeypatch (0.2), so they remain valid regardless of the default.

Security vulnerabilities

None. No lockfile, dependency, extras, or trust-boundary changes.

Non-blocking notes

• The stated failure mode is specifically the in-process LLM compaction path (_summarize_history_uncached → llm_client.responses.create), which has no inner timeout and is what the raised window protects. The runner-delegated path (_summarize_via_runner_uncached) already caps its HTTP call at timeout=120.0, so it would fail its own way before ever reaching even the old 240s watchdog — worth being aware of, but not something this PR needs to address.
• 600s is a reasonable heuristic, but it's a fixed guess at "longest healthy compaction." If near-full compactions routinely approach that, a follow-up could consider emitting a periodic non-heartbeat progress event from the long compaction await (the actual root cause) rather than continuing to widen the window. Not required here — the default bump is the pragmatic fix and the env knob remains for tuning.

Overall: correct, minimal, and safe to merge. The change does exactly what it claims and the existing tests still cover the watchdog mechanism.

---

_{Automated review by Polly · workflow run}

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[github-actions]

🏷️ Doc impact: no-doc-update

Only changes an internal ops env-var default (HARNESS_TURN_TIMEOUT_S 240s→600s) for the turn idle watchdog, which is not a documented user-facing surface or default.

_{Auto-classified on merge. Set the label manually before merging to override. · run}