convergence-monitoring

name	convergence-monitoring
user-invocable	true
tags	["autoresearch","convergence","loop-control","wave-executor"]
model	haiku
model-preference	sonnet
model-preference-codex	gpt-5.4-mini
model-preference-cursor	claude-sonnet-4-6
description	Monitor iterative improvement loops for convergence. Three signals — shrinking diff, pass-rate plateau, velocity — drive a Stop/Continue/Investigate decision at each inter-wave checkpoint. Distinct from /evolve (retrospective) and session-reviewer (wave output review): convergence-monitoring answers "are we making progress?" not "was the last wave correct?". Primary consumer: /autoresearch loops and wave-executor inter-wave checkpoints.
attribution	Inspired by cavekit's documented convergence-monitoring concept (MIT, Julius Brussee). Implemented from spec since upstream skill not yet published as of 2026-04-30.

Platform Note: State files use the platform's native directory: .claude/ (Claude Code), .codex/ (Codex CLI), or .cursor/ (Cursor IDE). Shared metrics live in .orchestrator/metrics/. See skills/_shared/platform-tools.md.

Convergence-Monitoring Skill

Platform-native (CC 2.1.105+)

This skill's watcher is registered as a plugin monitor via .claude-plugin/plugin.json's experimental.monitors reference to monitors/monitors.json. Each session that loads this plugin auto-starts the watcher in the background (see scripts/lib/convergence-monitor.mjs). Each NDJSON stdout line from the watcher becomes a <task_notification> event Claude sees mid-session.

For harness < 2.1.105 (no monitor support), the skill's manual probes documented below serve as the fallback path.

When to Invoke

Explicit invocation (/convergence-monitoring): Standalone assessment — reads wave history from .orchestrator/metrics/events.jsonl, computes all three signals, reports verdict + recovery options.

Embedded (loop context): /autoresearch and wave-executor invoke this skill after each Impl-Core or Impl-Polish wave when convergence-monitoring: true is set in Session Config. The skill returns a structured verdict that the caller uses to gate the next wave.

NOT a substitute for quality gates. Quality-gates verify correctness (typecheck, test, lint). Convergence-monitoring verifies progress direction — whether the work is moving toward done or oscillating/stalling.

---

Phase 0: Activation Gate

0.1 Explicit Invocation

If invoked via /convergence-monitoring, skip to Phase 1 unconditionally.

0.2 Embedded Invocation

When called from wave-executor or /autoresearch, check:

1. Read Session Config per skills/_shared/config-reading.md. Store as $CONFIG.
2. Check $CONFIG."convergence-monitoring". If the field is absent or false, return verdict SKIP immediately — the caller MUST NOT block on a skipped monitoring call.
3. Require $WAVE_NUMBER (integer, current wave) and $SESSION_START_REF (git SHA) to be provided by the caller. If either is absent, emit warning and return SKIP.
4. Require at least 2 completed waves of history. If $WAVE_NUMBER < 2, return SKIP — not enough data to assess trend.

---

Phase 1: Signal Collection

Read wave history from .orchestrator/metrics/events.jsonl. Filter to records matching $SESSION_START_REF and event_type in {wave.start, wave.end, quality.incremental}.

Compute the three signals in parallel. Each signal produces:

{ signal: "<name>", value: <float>, trend: "improving" | "plateau" | "regressing", confidence: 0.0-1.0 }

Signal 1 — Shrinking Diff (SD)

Measures: is the diff still shrinking, or has it reached noise level?

Computation:

diff_sizes[] = [git diff --stat $wave_N_start_ref $wave_N_end_ref | tail -1 | grep -oE '[0-9]+ insertion' | awk '{print $1}']
  for each completed wave in this session

• value = ratio of latest wave diff to peak diff (0.0 = unchanged, 1.0 = as large as ever)
• trend:
  • improving — diff_sizes shrinking monotonically for last 2 waves (value < 0.6)
  • plateau — diff_sizes stable within ±15% for 2+ consecutive waves (value 0.0–0.15)
  • regressing — diff_sizes growing (value > 1.0 relative to prior wave)
• confidence = min(1.0, completed_waves / 3) — unreliable below 3 waves

False positive flag: A plateau on diff size with a large diff is not convergence — it may be an oscillation. Cross-check with Signal 3 (velocity) before concluding.

Signal 2 — Pass-Rate Plateau (PR)

Measures: has the test pass-rate stabilized at a level (high or low)?

Computation: Read quality.incremental events from events.jsonl for this session. Extract test.passed / test.total per wave where the field is present.

• value = latest pass-rate (0.0–1.0)
• trend:
  • improving — pass-rate strictly increasing across last 2 waves
  • plateau — pass-rate within ±2% for 2+ consecutive waves
  • regressing — pass-rate decreasing
• confidence = min(1.0, quality_events_count / 2)

False positive flag: A plateau at 1.0 (100% pass) is the desired end-state. A plateau at < 0.9 with no improving trend is a ceiling signal — see Phase 3.

Signal 3 — Velocity (VEL)

Measures: commits and modified-line count per wave — declining velocity is normal as work nears completion; zero velocity for 2+ waves is a stall.

Computation:

commits_per_wave[] = [git rev-list --count $wave_N_start_ref..$wave_N_end_ref]
lines_per_wave[]   = [git diff --shortstat $wave_N_start_ref $wave_N_end_ref | awk '{print $4+$6}']
  for each completed wave in this session

• value = lines changed in latest wave (raw integer, not ratio)
• trend:
  • improving — commits or lines increasing (unusual mid-loop; flag for review)
  • plateau — < 5 lines + 0 commits in latest wave
  • regressing — lines declining (normal trajectory near convergence)
• confidence = min(1.0, completed_waves / 2)

---

Phase 2: Convergence Decision Table

Aggregate signal trends into a single verdict. Apply the table top-to-bottom; use the first matching row.

SD trend	PR trend	VEL trend	Verdict	Rationale
plateau	plateau	plateau	STOP	All three signals stable — loop has converged.
plateau	plateau	regressing	STOP	Velocity decline + stable quality — natural wind-down.
improving	improving	regressing	CONTINUE	Still making progress. Velocity decline is expected.
improving	plateau	regressing	CONTINUE	Diff still shrinking; pass-rate already high.
plateau	improving	regressing	CONTINUE	Pass-rate catching up; diff mostly stable — finishing work.
regressing	any	improving	INVESTIGATE	Diff growing and velocity up — scope expansion or churn.
any	regressing	any	INVESTIGATE	Pass-rate declining — regressions introduced.
plateau	plateau	improving	INVESTIGATE	Stable diff + stable quality but velocity up — odd state.
any	any	plateau	INVESTIGATE	Velocity zero for 2+ waves — stall or early plateau.

Confidence gate: If any signal has confidence < 0.5, downgrade STOP → CONTINUE and INVESTIGATE → CONTINUE (not enough data to act confidently on a Stop or Investigate). Emit a note: low-confidence: <signal-names>.

Verdict Output (structured)

{
  "verdict": "STOP | CONTINUE | INVESTIGATE | SKIP",
  "signals": {
    "shrinking_diff": { "trend": "...", "value": 0.0, "confidence": 0.0 },
    "pass_rate":      { "trend": "...", "value": 0.0, "confidence": 0.0 },
    "velocity":       { "trend": "...", "value": 0,   "confidence": 0.0 }
  },
  "low_confidence": [],
  "notes": []
}

Emit verdict JSON to stdout. Append one JSONL record to .orchestrator/metrics/events.jsonl with event_type: convergence.checkpoint.

---

Phase 3: Non-Convergence Diagnosis

Execute this phase when verdict == INVESTIGATE. Skip if verdict is STOP or CONTINUE.

The core question: ceiling or convergence?

• Convergence — the work is done; signals stabilized at the desired state (pass-rate ~1.0, diff ~0). Stop is the right action but signals were ambiguous.
• Ceiling — progress has halted but the desired state has NOT been reached. Something is blocking further improvement.

3.1 Ceiling vs Convergence Dichotomy

Observable	Classification
pass-rate plateau ≥ 0.95, diff plateau ≈ 0	Convergence
pass-rate plateau < 0.90 with no trend	Ceiling
diff growing (SD regressing) for 2+ waves	Churn / Ceiling
velocity plateau with large outstanding diff	Stall / Ceiling
velocity plateau with near-zero diff	Convergence

3.2 Ceiling Root-Cause Probes

Run these read-only probes to characterize the ceiling. Each probe takes < 5 seconds.

1. Recurring failures: git log --oneline $SESSION_START_REF..HEAD -- '*.test.*' | wc -l — high churn in test files with no new commits to production files suggests a test-fix loop.

2. Oscillating files: identify files that appear in 3+ wave diffs in this session. These are the fragile modules (see architecture.md Module/Seam vocabulary) — their seam is leaking complexity into the loop.

3. Stagnation events: read stagnation_events from the latest session record in sessions.jsonl. A non-empty list confirms the loop is repeating edit-format-friction rather than making functional progress.

4. Scope drift: compare $CONFIG."plan-prd-location" PRD's in-scope items against git diff --name-only $SESSION_START_REF..HEAD. Files outside PRD scope indicate the loop has expanded beyond its brief.

Report the probe results in Phase 4.

---

Phase 4: Recovery Suggestions

Present options based on the Phase 3 diagnosis. Do NOT auto-apply any recovery action — present to the user and await confirmation.

Diagnosis	Recovery Option	When to Use
Convergence	Accept and stop — declare loop done	pass-rate ≥ 0.95, all scope items addressed
Ceiling / Stall	Rollback to last-known-good wave — git reset --soft	Loop is chasing a regression it introduced
Ceiling / Churn	Scope reduction — remove non-PRD files from agent brief	Loop has drifted outside its brief
Stall (fragile)	Agent-count adjustment — reduce agents per wave by 50%	Too many agents editing the same fragile module
Stall (format)	Stagnation break — inject pre-edit grounding prompt	stagnation_events present; loop repeating format errors
Ceiling (test)	Test-fix isolation — spawn dedicated fix wave**	Pass-rate < 0.90 but diff near zero

Rollback caveat: Present the exact command (git reset --soft <wave_N-1_end_ref>) and require explicit user confirmation. Never execute it automatically.

4.1 Recovery Report Format

## Convergence Report — Wave N

Verdict: INVESTIGATE
Diagnosis: Ceiling (pass-rate stuck at 0.82, 3 oscillating files)

Oscillating modules: src/lib/parser.ts, src/lib/tokenizer.ts, tests/parser.test.ts
Stagnation events: 2 (edit-format-friction in src/lib/parser.ts)

Recovery options:
  [A] Scope reduction — remove src/lib/tokenizer.ts from agent brief (outside PRD scope)
  [B] Agent-count adjustment — reduce from 6 to 3 agents next wave (shared fragile module)
  [C] Rollback — git reset --soft <wave_3_end_ref> (destructive, requires confirmation)
  [D] Continue anyway — override monitoring and proceed to next wave

Select option or type 'continue' to override.

---

Anti-Patterns

• Treating STOP as a quality gate. Convergence-monitoring answers "are we done making progress?" not "is the code correct?" Full Gate quality checks are required separately even when verdict is STOP.
• Acting on low-confidence verdicts without the confidence note. Below 3 waves of history the signals are unreliable. The confidence gate in Phase 2 exists for this reason.
• Conflating convergence with correctness. A loop can converge to a broken state (pass-rate plateau at 0.6). Always cross-check the pass-rate value, not just the trend.
• Auto-applying rollback. Phase 4 recovery options are presented to the user, never auto-applied. Rollback is destructive and may discard valid work across parallel sessions (PSA-003).
• Confusing convergence-monitoring with /evolve. evolve is retrospective (what patterns should we learn from finished sessions?). convergence-monitoring is prospective (should the current loop continue?).

---

See Also

• skills/wave-executor/wave-loop.md — intended consumer (inter-wave hook integration pending in a follow-up issue)
• skills/evolve/SKILL.md — retrospective learning; complements but does not replace
• skills/quality-gates/SKILL.md — correctness gate; runs independently of convergence verdict
• convergence-monitoring/SIGNALS.md — signal computation reference + threshold rationale