wanshuiyin/meta-optimize

Meta-Optimize: Outer-Loop Harness Optimization for ARIS

Analyze accumulated usage logs and propose optimizations for: $ARGUMENTS

Privilege boundary — this skill is a READ-ONLY PRODUCER

meta-optimize proposes; it does not land. The mutation of the skill corpus is the exclusive job of a separate, human-invoked skill: /meta-apply. This split is structural, not advisory — it is why a missed instruction cannot let this loop apply its own patch (the self-acquittal failure mode):

• No Write/Edit tool. This skill cannot edit a SKILL.md / shared-reference / any corpus file with the frictionless mutators. Its only outputs are the REPORT and staged patch files, written under .aris/meta/ (a scratch area, never the corpus).
• No apply step. There is no in-skill "apply the patch" path (see Step 6). The producer ends by staging approved patches for /meta-apply; a human must then invoke /meta-apply to land them. That human action is the landing gate.
• Bash writes to the corpus are filtered, not impossible — be honest about the layers. What IS fully closed: the accidental / in-flow self-acquittal — this skill has no Write/Edit and no apply step, so an honest run cannot slip into editing the corpus. Defense-in-depth: install the corpus_write_guard PreToolUse hook (like meta_logging.json), which DENIES the common Bash shell-writes (>, tee, sed -i, cp/mv, touch, open(...,'w')) to corpus paths. This is a blacklist, NOT a complete sandbox — a deliberately obscured Bash write (git apply, patch, $var/absolute paths, language file APIs) is not all caught. Full structural prevention requires either removing this skill's Bash or an FS sandbox — over-built for a not-yet-load-bearing producer, so deferred to when the gate carries real auto-modification volume (a brick-3 trigger). The intended backstop against a deliberate write is detection, not prevention — a corpus change with no valid/current provenance stamp (content-hash mismatch) would be catchable in a pre-push integrity check — but that verifier is NOT yet built (provenance.py has content_hash but no integrity-check subcommand, and no pre-push hook runs one). So today the deliberate-write case is neither prevented nor actively detected; track the integrity verifier as a follow-up before this producer goes load-bearing. Its legitimate Bash writes go only to .aris/meta/.

See shared-references/acceptance-gate.md: a loop can DRIVE (propose, review) same-model, but the ACQUITTAL that lands a change must be cross-model (Step 4 jury) and the landing must be a separate human-gated act (/meta-apply).

Context

ARIS is a research harness — a system of skills, bridges, workflows, and artifact contracts that wraps around LLMs to orchestrate research. This skill implements a prototype outer loop that observes how the harness is used and proposes improvements to the harness itself (not to the research artifacts it produces).

Inspired by Meta-Harness (Lee et al., 2026): the key insight is that harness design matters as much as model weights, and harness engineering can be partially automated by logging execution traces and using them to guide improvements.

What This Skill Optimizes (Harness Components)

| Component | Example | Optimizable? | |-----------|---------|:---:| | SKILL.md prompts | Reviewer instructions, quality gates, step descriptions | Yes | | Default parameters | difficulty: medium, MAX_ROUNDS: 4, threshold: 6/10 | Yes | | Convergence rules | When to stop the review loop, retry counts | Yes | | Workflow ordering | Skill chain sequence within a workflow | Yes | | Artifact schemas | What fields go in EXPERIMENT_LOG.md, idea-stage/IDEA_REPORT.md | Cautious | | MCP bridge config | Which reviewer model, routing rules | No (infra) |

Not optimized: The research artifacts themselves (papers, code, experiments). That's what the regular workflows do.

Prerequisites

1. Logging must be active. Copy templates/claude-hooks/meta_logging.json into your project's .claude/settings.json (or merge the hooks section).
2. Sufficient data. At least 5 complete workflow runs logged in .aris/meta/events.jsonl. The skill will check and warn if insufficient.

Workflow

Step 0: Check Data Availability

EVENTS_FILE=".aris/meta/events.jsonl"
if [ ! -f "$EVENTS_FILE" ]; then
    echo "ERROR: No event log found at $EVENTS_FILE"
    echo "Enable logging first: copy templates/claude-hooks/meta_logging.json into .claude/settings.json"
    exit 1
fi

EVENT_COUNT=$(wc -l < "$EVENTS_FILE")
SKILL_INVOCATIONS=$(grep -c '"skill_invoke"' "$EVENTS_FILE" || echo 0)
SESSIONS=$(grep -c '"session_start"' "$EVENTS_FILE" || echo 0)

echo "📊 Event log: $EVENT_COUNT events, $SKILL_INVOCATIONS skill invocations, $SESSIONS sessions"

if [ "$SKILL_INVOCATIONS" -lt 5 ]; then
    echo "⚠️  Insufficient data (<5 skill invocations). Continue using ARIS normally and re-run later."
    exit 0
fi

# Bottleneck succession: what did the LAST cycle say was the limiting stage?
BOTTLENECK_LOG=".aris/meta/bottleneck_log.jsonl"
if [ -f "$BOTTLENECK_LOG" ]; then
    echo "🧭 Prior cycle's bottleneck: $(tail -1 "$BOTTLENECK_LOG")"
fi

If a prior bottleneck entry exists, open the report (Step 5) by stating whether that named bottleneck was resolved (and by which landed patches) and what it has now moved to — bottleneck SUCCESSION, not just existence, is the signal this ledger exists to carry.

Step 1: Analyze Usage Patterns

Read .aris/meta/events.jsonl and compute:

Frequency analysis:

• Which skills are invoked most often?
• Which slash commands do users type most?
• What parameter overrides are most common? (These suggest bad defaults.)

Failure analysis:

• Which tools fail most often? In which skills?
• What error patterns repeat? (OOM, import, compilation, timeout)
• How many auto-debug retries per workflow run?

Convergence analysis (for auto-review-loop):

• Average rounds to reach threshold
• Score trajectory shape (fast improvement? plateau? oscillation?)
• Which review round catches the most critical issues?
• Do users override difficulty mid-run?

Human intervention analysis:

• Where do users interrupt with manual prompts during workflows?
• What manual corrections do users make most? (These indicate skill gaps.)

Model-delta analysis (harness diet):

• Has the session model (session_start events' model field) or the pinned reviewer model changed since a skill's SKILL.md was last touched? (git log -1 --format=%cs -- skills/<skill>/SKILL.md vs the model-bump date.)
• A model bump is a trigger to re-read, not evidence by itself. For each reasoning-scaffolding step or worked example in that SKILL.md, a deletion proposal must cite TARGET-SPECIFIC evidence that the new model no longer needs it: a capability-specific release note, or repeated observed behavior in the event log (e.g. zero failures/interventions in the guarded step since the bump). "The model got newer" alone never justifies a deletion.
• Never deletion candidates, regardless of model: privilege boundaries, acceptance/review gates, corpus- and provenance-integrity rules, output contracts, and safety checks. The diet targets model-compensation scaffolding only — a capability the new model has natively is pure overhead (context weight, drift surface, reading cost). A harness that only ever grows is a harness nobody is re-reading.

Trigger-rate analysis (optional, measured — not from the event log):

• The event log shows which skills were USED, not which were WANTED-but-omitted — the omission failure mode (Claude Code passing over the right skill when the installed list is long) is invisible to it. tools/meta_opt/trigger_eval.py measures it directly: claude -p probes with paraphrased-intent queries run from a neutral cwd (so the realistic long installed corpus is loaded), scored as trigger / confusion(→which skill) / miss.
• Run it when a specific skill is suspected of under- or mis-triggering, or as a before/after check around a description edit: python3 tools/meta_opt/trigger_eval.py --eval-file tools/meta_opt/trigger_evals.sample.json --skills <name> --samples 2
• The confusion matrix is the signal, not just the rate: a query that keeps landing on a sibling skill means the two descriptions overlap on that intent — the fix is disambiguation, not "make the description pushier".
• Measure-only, evidence not verdict. A low trigger rate is an INPUT to a Step-2 proposal (which lands only via /meta-apply), never a self-applied description rewrite. Trigger rate is model-dependent, so compare like with like (record the probe model) and treat it as a proxy — it measures selection under a query set, not the full long-list omission problem.

Present findings as a structured summary table.

Step 1.5: Name the Current Bottleneck

Synthesize the Step-1 analyses into one sentence naming the single most-limiting pipeline stage right now — e.g. "planning", "verification quality", "experiment execution reliability", "writing polish" — with the supporting evidence. The bottleneck always moves: when coding stops being the constraint, planning becomes it; when planning is solved, verification; when verification is automated, taste. This step exists to make the CURRENT constraint visible, so Step 2's ranked table reads as sub-fixes for one named constraint instead of scattered tweaks.

Append the verdict to the append-only ledger .aris/meta/bottleneck_log.jsonl (same never-mutate discipline as .aris/runs/<run_id>.iterations.jsonl):

mkdir -p .aris/meta
# json.dumps, NOT hand-interpolated shell strings: bottleneck/evidence are
# natural language — a stray quote must not break the JSONL (or the shell).
python3 - <<'PY'
import json, datetime
entry = {
    "ts": datetime.datetime.now().astimezone().isoformat(timespec="seconds"),
    "cycle": 3,
    "bottleneck": "verification quality",
    "evidence": "review rounds plateau at 6/10 while tool failures are rare",
    "top_patch_ids": ["P1", "P2"],
}
with open(".aris/meta/bottleneck_log.jsonl", "a", encoding="utf-8") as fh:
    fh.write(json.dumps(entry, ensure_ascii=False) + "\n")
PY

Never edit or delete prior lines — succession history is the point.

Step 2: Identify Optimization Targets

Based on Step 1, rank optimization opportunities by expected impact:

## Optimization Opportunities (ranked)

| # | Target | Signal | Proposed Change | Expected Impact |
|---|--------|--------|-----------------|-----------------|
| 1 | auto-review-loop default threshold | Users override to 7/10 in 60% of runs | Change default from 6/10 to 7/10 | Fewer manual overrides |
| 2 | experiment-bridge retry count | 40% of runs hit max retries on OOM | Add OOM-specific recovery (reduce batch size) | Fewer failed experiments |
| 3 | paper-write de-AI patterns | Users manually fix "delve" in 80% of runs | Add "delve" to default watchword list | Fewer manual edits |
| 4 | experiment-bridge Phase-2 hand-holding steps | Model bump (session_start model changed); scaffold untouched since 2 generations ago; zero tool_failures in the steps it guards | **DELETE steps N–M — the new model does this unprompted** | Smaller harness, less drift surface |

The Proposed-Change column is explicitly allowed to be a deletion — "DELETE step N, new model does this for free" is a first-class optimization, ranked by the same impact logic as additions.

If $ARGUMENTS specifies a target skill, focus analysis on that skill only. If $ARGUMENTS is empty or "all", analyze all skills with sufficient data.

Step 3: Generate Patch Proposals

For each optimization target, generate a concrete diff:

--- a/skills/auto-review-loop/SKILL.md
+++ b/skills/auto-review-loop/SKILL.md
@@ -15,7 +15,7 @@
 ## Constants
 
-- **SCORE_THRESHOLD = 6** — Minimum review score to accept.
+- **SCORE_THRESHOLD = 7** — Minimum review score to accept. (Raised based on usage data: 60% of users overrode to 7+.)

Rules for patch generation:

• One patch per optimization target
• Each patch must include a comment explaining WHY (with data from the log)
• Patches must be minimal — change only what the data supports
• Never change artifact schemas or MCP bridge config in v1
• Never change behavior that would break existing user workflows
• Anti-self-poisoning screen (see shared-references/capture-antipatterns.md): run a proposed patch's rationale through tools/capture_filter.py (resolve via the canonical chain). NEVER propose a change that encodes a negative tool-capability claim ("codex can't…", "gemini is broken") or a one-off / transient failure as a durable rule — those harden into self-cited refusals. Encode the fix / the flag needed / the workaround, not "X can't do Y".

Step 4: Cross-Model Review of Patches (ADVISORY pre-screen)

This review is advisory — it sharpens the Step-5 REPORT so the human can decide what to stage. It is not the landing verdict. The binding cross-model jury runs later, at landing, inside /meta-apply, on the actual staged diff (a producer-relayed verdict would be forgeable). Record this result as advisory_screen only.

Send each patch to GPT-5.6-Sol xhigh for adversarial review:

mcp__codex__codex:
  model: gpt-5.6-sol
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    You are reviewing a proposed optimization to an ARIS SKILL.md file.
    
    ## Original Skill (relevant section)
    [paste original]
    
    ## Proposed Patch
    [paste diff]
    
    ## Evidence from Usage Log
    [paste summary stats]
    
    Review this patch:
    1. Does the evidence support the change?
    2. Could this change hurt other use cases?
    3. Is the change minimal and safe?
    4. Score 1-10: should this be applied?
    
    If score < 7, explain what additional evidence would be needed.

Step 5: Present Results

Output a structured report:

# ARIS Meta-Optimization Report

**Date**: [today]
**Data**: [N] events, [M] skill invocations, [K] sessions
**Target**: [skill name or "all"]

## Current Bottleneck

**[one-phrase name]** — [one-line evidence]. Prior cycle's bottleneck: [name —
resolved by <patch ids> / unresolved / first recorded cycle]. (Ledger:
`.aris/meta/bottleneck_log.jsonl`)

## Proposed Changes

### Change 1: [title]
- **Target**: [skill/file:line]
- **Signal**: [what the data shows]
- **Patch**: [diff]
- **Reviewer Score**: [X/10]
- **Reviewer Notes**: [summary]
- **Status**: ✅ Recommended / ⚠️ Needs more data / ❌ Rejected

### Change 2: ...

## Changes NOT Made (insufficient evidence)
- [pattern observed but too few samples]

## Recommendations
- [ ] Apply Change 1 (reviewer approved)
- [ ] Collect more data for Change 3 (need N more runs)
- [ ] Consider manual review of Change 2

## Next Steps
This skill only **proposes**. To land changes: tell me which to stage, then run
`/meta-apply` (a separate, human-invoked applier that re-checks the cross-model
verdict before mutating anything). meta-optimize never applies.

Step 6: Stage approved patches for /meta-apply (NO in-skill apply)

This skill does not apply anything. After the user has read the Step-5 REPORT and indicated which changes to land, stage them for the privileged applier:

1. For each approved change N, write its unified diff to .aris/meta/pending/<NN>_<skill>.diff and append a row to .aris/meta/pending/manifest.jsonl: {patch: "<NN>_<skill>.diff", target: "<corpus path>", author_model: "<executor>", advisory_screen: "pass|kill", advisory_reason: "<one line>"}. The advisory_screen (your Step-4 codex pre-review) is advisory only — it helps the human read the REPORT. It is NOT the landing verdict and /meta-apply does not trust it: a producer-written verdict would be forgeable. The binding cross-model jury runs at landing, inside /meta-apply, on the actual staged diff.
2. Tell the user: "Staged M patches. Run /meta-apply to judge & land them."

The backup → fresh jury-at-landing → apply → provenance stamp → log all happen inside /meta-apply. meta-optimize never touches the corpus and never produces the acquittal.

Never apply in this skill. Landing is /meta-apply + a fresh jury + a human, always.

Key Rules

• Log-driven, not speculative. Every proposed change must cite specific data from the event log. No "I think this would be better."
• Minimal patches. Change one thing at a time. Don't rewrite entire skills — the one sanctioned large edit is a scaffolding deletion backed by TARGET-SPECIFIC model-delta evidence (a capability-specific release note, or repeated post-bump event-log behavior showing the scaffold is unused — the model name changing is a trigger to look, never sufficient evidence). Privilege boundaries, acceptance gates, corpus/provenance rules, output contracts, and safety checks are never deletion candidates. Deletions go through the same review + approval gates as everything else.
• Reviewer-gated. Every patch goes through cross-model review before recommendation.
• Reversible. Always back up before applying. Always log what changed.
• User-approved. Never auto-apply. Present, explain, let the user decide.
• Honest about uncertainty. If the data is insufficient, say so. Don't optimize on noise.
• Portable. Optimizations should improve the skill for all users, not just one user's style. If a change seems user-specific, flag it.

Event Schema Reference

The log at .aris/meta/events.jsonl contains JSONL records with these shapes:

{"ts":"...","session":"...","event":"skill_invoke","skill":"auto-review-loop","args":"difficulty: hard"}
{"ts":"...","session":"...","event":"PostToolUse","tool":"Bash","input_summary":"pdflatex main.tex"}
{"ts":"...","session":"...","event":"codex_call","tool":"mcp__codex__codex","input_summary":"review..."}
{"ts":"...","session":"...","event":"tool_failure","tool":"Bash","input_summary":"python train.py"}
{"ts":"...","session":"...","event":"slash_command","command":"/auto-review-loop","args":""}
{"ts":"...","session":"...","event":"user_prompt","prompt_preview":"change difficulty to hard"}
{"ts":"...","session":"...","event":"session_start","source":"startup","model":"claude-opus-4-6"}
{"ts":"...","session":"...","event":"session_end"}

Triggering

This skill is NOT part of the standard W1→W1.5→W2→W3→W4 pipeline. It is a maintenance workflow with three trigger mechanisms:

1. Passive logging (always on): Claude Code hooks record events to .aris/meta/events.jsonl automatically during normal usage. Zero user effort.

2. Automatic readiness check (SessionEnd hook): When a Claude Code session ends, check_ready.sh counts skill invocations since the last /meta-optimize run. If ≥5 new invocations have accumulated, it prints a reminder:

   📊 ARIS has logged 8 skill runs since last optimization. Run /meta-optimize to check for improvement opportunities.
   

   It ALSO fires — regardless of invocation count — when the session model has changed since the last optimize (compared against .aris/meta/.last_optimize_model):

   🔁 Model changed since last optimization (claude-opus-4-6 → claude-opus-4-8). Run /meta-optimize — a model bump makes existing scaffolding a deletion candidate (harness diet).
   

   Both are suggestions only — they do not auto-run optimization.

3. Manual trigger: User runs /meta-optimize when they see the reminder or whenever they want.

After each /meta-optimize run, the skill writes the current timestamp to .aris/meta/.last_optimize and the current session model (latest session_start event's model field) to .aris/meta/.last_optimize_model, so the readiness check can detect both new usage and model bumps.

Acknowledgements

Inspired by Meta-Harness (Lee et al., 2026) — end-to-end optimization of model harnesses via filesystem-based experience access and agentic code search.

Output Protocols

Follow these shared protocols for all output files:

• Output Versioning Protocol — write timestamped file first, then copy to fixed name
• Output Manifest Protocol — log every output to MANIFEST.md
• Output Language Protocol — respect the project's language setting

Review Tracing

After each mcp__codex__codex or mcp__codex__codex-reply reviewer call, save the trace following shared-references/review-tracing.md (Policy C — forensic; never silently skip). Use save_trace.sh (resolved per the chain in shared-references/integration-contract.md §2) or write files directly to .aris/traces/<skill>/<date>_run<NN>/. Respect the --- trace: parameter (default: full).