fix(pilot): debounce process resource alarms

[Monsooooon]

Background

Process resource alarms currently fire on single CPU or memory threshold samples, which creates noisy alerts from transient CPU spikes and repeated memory readings near 512MB. CPU and memory also shared the same process alarm type, allowing aggregation windows to overwrite one resource signal with the other.

Core Changes

• Split process CPU and memory alarms into independent alarm types to avoid aggregation overwrite.
• Normalize CPU usage by CPU count and require consecutive samples before alerting.
• Add memory soft / hard / critical levels, with soft debounce and alert cooldown.
• Add focused unit coverage for debounce behavior, severity selection, independent aggregation, and cooldown.

Modified Files

File	Notes
src/metrics/alarm-manager.ts	Adds independent process CPU and memory alarm types.
src/metrics/metrics-writer.ts	Adds CPU normalization/debounce and memory levels/debounce/cooldown.
tests/unit/metrics/metrics-writer.test.ts	Covers resource alarm debounce, severity, aggregation, and cooldown behavior.

Verification

• npx vitest run tests/unit/metrics/metrics-writer.test.ts — 21/21 PASS
• npx vitest run tests/unit/metrics/ — 62/62 PASS
• npm run typecheck — PASS
• Real Codex CLI E2E trace validation — PASS, 0 ERROR
  • Target CLI: /opt/homebrew/bin/codex, codex-cli 0.142.5
  • Probe coverage: 4 STEP spans, 3 TOOL spans
  • Trace validation: Summary: 1 traces, 13 spans, Pass: 33 Warn: 138 Error: 0 Skipped: 0
  • Trace ID: 2981a5fc3b47cbe517ba4f9039ee90d6

[ralf0131]

Summary

Splits the generic PROCESS_RESOURCE_ALARM into dedicated PROCESS_CPU_ALARM and PROCESS_MEMORY_ALARM types, with CPU normalization by core count (os.cpus()), tiered memory thresholds (soft 512MB / hard 1024MB / critical 2048MB), consecutive-sample requirements (3 samples) to suppress transient spikes, and a 1-hour cooldown to prevent alarm flapping. Comprehensive test coverage for all threshold tiers, normalization, suppression, and cooldown behavior. All CI green (Node 18/20/22). Clean, well-structured change.

Findings

No issues found — the implementation is correct, properly typed, and thoroughly tested.

---

*Automated review by *

[ralf0131]

Summary

Re-review after commit 011f1785. The new commit correctly addresses @Snssn's feedback — CPU alarms now compare raw process CPU percent (from calcCpuPercent()) against the 80% threshold, without normalizing by host core count. The 3-consecutive-sample debounce and tiered memory thresholds (soft/hard/critical) with independent per-type cooldowns are well-implemented and comprehensively tested.

Findings

• [Info] src/metrics/metrics-writer.ts — The PR body still states "Normalize CPU usage by CPU count", but the implementation (after the fix) uses raw process CPU percent. Consider updating the body to match the actual behavior.

Verdict

LGTM. The fix is correct, code is clean, and test coverage (6 cases) is thorough. CLA is signed.

---

Automated review by github-manager-bot