Harden triage tier against prompt regressions and timeouts

[don-petry]

Summary

Two recent failures in the PR Review Agent workflow each aborted the hourly batch session-fatally — see the investigation report for full analysis.

• Run dev-lead falsely reports "all AI engines are currently rate-limited" when the real cause is stale cancelled check runs #461: triage model returned conversational text asking "Should I triage each PR... Should I design the agent system... Something else?" instead of the required JSON, breaking the contract.
• Run dev-lead: git worktree collision ('branch already used by worktree') breaks review-changes — blocks PR #383 #470: triage process exited 1 on a single PR (likely a timeout or subprocess issue with no diagnostic output captured).

This PR addresses both root causes with minimal, low-risk changes.

Changes

• Bump TRIAGE_TIMEOUT_SEC default 180s → 300s (scripts/engine.sh). The 180s cap was tight for large PRs and offered no headroom; 300s matches ACTION_TIMEOUT_SEC and reduces the chance of a kill being misclassified as a deterministic failure (kills at 124 are retried; exit 1 is not).
• Tighten the triage prompt (prompts/triage.md): explicitly forbid clarifying questions, plan proposals, and prose. Add a fallback rule so even on incomplete context the model must still emit JSON (with escalate=true and a signal explaining what was missing) rather than respond conversationally.

Test plan

• Hourly cron run completes without aborting on the same PR set
• Verify TRIAGE_TIMEOUT_SEC=300 shows in the workflow env block
• Spot-check a triage log to confirm prompt update was picked up
• Watch the next 24h of runs for any non-JSON failures

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR hardens the tier-1 triage path in the PR review workflow to reduce session-fatal failures from model regressions (non-JSON output) and overly tight timeouts.

Changes:

• Increase the default triage timeout from 180s to 300s.
• Tighten the triage prompt to forbid clarifying questions/plans/prose and to require JSON output even when context is incomplete.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
scripts/engine.sh	Raises default TRIAGE_TIMEOUT_SEC to reduce triage subprocess timeouts.
prompts/triage.md	Strengthens prompt constraints to reduce non-JSON / conversational triage outputs.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The new fallback instruction says to emit JSON with "escalate": true and "a signal" explaining what's missing. Since the required schema below uses a signals array (and scripts parse .signals | join(...)), this wording risks the model outputting a signal field or omitting signals, which would break the contract. Suggest rephrasing to explicitly say to populate the signals array (and still include required risk and summary fields).

Suggested change

	with `"escalate": true` and a signal explaining what was missing — never
	respond with prose.
	with `"escalate": true`, populate the `signals` array with what was missing,
	and still include the required `risk` and `summary` fields — never respond
	with prose.