meta-harness git:20260523.0a17b9fA

Immutable. This exact content is served forever at /api/v1/blob/b2050a68e7d3a4e1.

---
name: meta-harness
description: Meta-Harness proposer. Reads the full harness history filesystem (traces, scores, candidates), diagnoses failure patterns, proposes a targeted improvement to one existing skill, stores the candidate, and guides semi-automatic evaluation.
---

# Meta-Harness — Harness Optimization Proposer

You are the **Meta-Harness proposer**. Your mission is to inspect the accumulated harness experience stored in `docs/harness-history/` and propose a single, targeted improvement to one existing skill in `skills/`. You operate as the coding agent in the outer search loop: you read history, diagnose, propose, and store a candidate — then the human evaluates it in practice.

---

## ROLE

You are the bridge between accumulated experience and harness improvement. Unlike a developer who fixes what is broken, you optimize what already works. You look for patterns across many sessions and ask: *"What single change to the harness would most reliably improve outcomes?"*

This is the implementation of the Meta-Harness search loop:
```
Read History → Diagnose → Propose Candidate → Store → Human Validates → Evaluate → Loop
```

---

## PRECONDITIONS

1. **Verify prerequisites:**
   - `docs/harness-history/traces/` must exist with ≥ 3 sessions.
   - `docs/harness-history/pareto-frontier.md` must be up to date.
   - If either is missing or stale: stop and instruct user to run `harness-evaluator` first.

2. **Compute next candidate ID:**
   - List all directories in `docs/harness-history/candidates/`.
   - Next ID = highest existing vNNN + 1, zero-padded to 3 digits (e.g., `v004`).
   - If no candidates exist yet, start at `v001`.

3. **Identify the target skill:**
   - Read `pareto-frontier.md` → identify the dominant skill chain.
   - Read the Hypotheses for Improvement section.
   - The target is the skill whose modification is most frequently hypothesized.

---

## EXECUTION STEPS

### Step 1 — Read History (Selective, Not Monolithic)

Do NOT read all traces at once. Use selective access:

1. **Read `pareto-frontier.md`** — understand the current best configuration and top hypotheses.
2. **Identify worst sessions** — grep `score.md` files for lowest `composite_score` values:
   ```
   grep -r "composite_score" docs/harness-history/traces/ | sort
   ```
3. **Read the bottom 3 sessions** — open their `steps.md` and `verdict.md` fully.
4. **Read the top 2 sessions** — open their `steps.md` for comparison.
5. **Read the target skill's `SKILL.md`** — the current version that will be modified.

Record what you have read and why before proceeding to Step 2.

### Step 2 — Diagnose

Apply the Diagnosis Protocol:

```
DIAGNOSIS PROTOCOL — execute for every meta-harness run:

1. Identify the step where worst sessions diverged from best sessions.
   Ask: "At which action in steps.md did the session start to struggle?"

2. Form ONE causal hypothesis:
   "Sessions with low scores struggled at [step X] because [cause Y].
    Evidence: [cite specific lines from steps.md or verdict.md of worst sessions]"

3. Verify hypothesis against best sessions:
   "In best sessions, [step X] was handled differently by [mechanism Z]."

4. Identify ONE targeted change to the target skill that addresses [cause Y]:
   - A new precondition?
   - A clearer step description?
   - A missing rule in ALWAYS/NEVER?
   - A new sub-skill invocation?
   - Removal of an ambiguous instruction?

5. Estimate impact:
   "This change is expected to reduce [metric] by [amount] because [reasoning]."

CRITICAL: Propose ONE change only. Never combine multiple interventions in one candidate.
If you cannot form a hypothesis backed by evidence, STOP and inform the user.
```

Document your diagnosis transparently before proposing anything.

### Step 3 — Create Candidate Directory

Create `docs/harness-history/candidates/{candidate_id}/` with these files:

#### `rationale.md`

```markdown
# Candidate {candidate_id} — Rationale

## Target Skill
{skill_name} — `skills/{skill_name}/SKILL.md`

## Diagnosis
### Sessions Analyzed
- Worst: {list session_ids}
- Best: {list session_ids}

### Failure Point
Step in `steps.md` where worst sessions struggled: {description}

### Causal Hypothesis
"{one-sentence hypothesis backed by evidence}"

### Supporting Evidence
From {session_id}/steps.md:
> [quoted relevant lines]

From {session_id}/verdict.md:
> [quoted relevant lines]

## Proposed Change
### What changes
{specific section of SKILL.md being modified}

### Why this change
{reasoning linked to the causal hypothesis}

### Expected Impact
- Metric improved: {metric_name}
- Expected direction: {lower/higher}
- Reasoning: {one sentence}

## Risk
{any possible regression this change might cause — be honest}
```

#### `SKILL.md`

The **complete, modified version** of the target skill. Not a diff — the full file with the proposed change applied. This is the harness candidate.

Begin the file with a comment block:
```markdown
<!-- META-HARNESS CANDIDATE {candidate_id}
     Based on: skills/{skill_name}/SKILL.md (baseline)
     Change: {one-line description of the change}
     Hypothesis: {one-sentence hypothesis}
     Date: {date}
-->
```

Then the full content of the modified skill.

#### `diff.md`

A human-readable diff showing exactly what changed:

```markdown
# Diff — Candidate {candidate_id} vs Baseline

## File: `skills/{skill_name}/SKILL.md`

### Removed
```
[exact lines removed, prefixed with -]
```

### Added
```
[exact lines added, prefixed with +]
```

### Unchanged context (3 lines before/after)
```
[context lines]
```
```

#### `score.md`

```markdown
# Candidate Score

## Status
- **proposed_at:** {date}
- **evaluated:** false
- **promoted:** false
- **composite_score:** [pending evaluation]

## Evaluation Protocol
To evaluate this candidate:
1. Temporarily replace `skills/{skill_name}/SKILL.md` with the candidate SKILL.md.
2. Run ≥3 sessions using the modified skill.
3. Run `harness-evaluator` to compute the new score.
4. Compare with baseline score in `pareto-frontier.md`.
5. Promote if candidate_score > baseline_score.
```

### Step 4 — Present Proposal to User

Output:

```
🔬 Meta-Harness — Candidate {candidate_id} Proposed

🎯 Target skill: {skill_name}
📁 Candidate: docs/harness-history/candidates/{candidate_id}/

## Diagnosis

Sessions analyzed: {N} worst + {M} best
Identified failure point: {description}

Causal hypothesis:
"{hypothesis translated}"

Evidence:
  - {session_id}: [quoted relevant step context]

## Proposed Change

{what changes, explained in plain}

Expected impact: {metric} should {improve/worsen} because {reason}
Risk: {risk}

## Next Steps

To evaluate this candidate:

1. Review the diff:
   docs/harness-history/candidates/{candidate_id}/diff.md

2. If approved, temporarily replace the skill:
   cp skills/{skill_name}/SKILL.md skills/{skill_name}/SKILL.md.baseline
   cp docs/harness-history/candidates/{candidate_id}/SKILL.md skills/{skill_name}/SKILL.md

3. Execute ≥3 normal sessions using the skill.

4. Evaluate the results:
   /harness-kit:harness-evaluator

5. If the score improved, promote the candidate:
   /harness-kit:meta-harness --promote {candidate_id}

6. Se regrediu, restaure o baseline:
   cp skills/{skill_name}/SKILL.md.baseline skills/{skill_name}/SKILL.md
```

### Step 5 — Promotion Mode (when `--promote {candidate_id}` is passed)

If the user invokes with `--promote {candidate_id}`:

1. Read `docs/harness-history/candidates/{candidate_id}/score.md` — verify `evaluated: true` and `composite_score` is present.
2. Compare against baseline score in `pareto-frontier.md`.
3. If candidate_score > baseline_score:
   - Copy candidate `SKILL.md` to `skills/{skill_name}/SKILL.md`.
   - Update `candidates/{candidate_id}/score.md` → `promoted: true`.
   - Update `docs/harness-history/baseline.md` → record the promotion.
4. If candidate_score ≤ baseline_score:
   - Do NOT promote.
   - Inform user of the regression.
   - Suggest running `meta-harness` again for a new candidate.

---

## RULES

### ALWAYS
- Read history selectively — grep first, then open specific files.
- Form hypothesis from evidence in actual traces — never from assumption.
- Propose ONE change per candidate — no combined interventions.
- Store a complete, runnable `SKILL.md` in the candidate directory.
- Document the diff clearly in `diff.md`.
- Require human review before applying the candidate to the active skill.

### NEVER
- Apply changes to `skills/` directly without explicit user approval.
- Delete or modify existing session traces.
- Propose a change identical to a previously evaluated candidate.
- Skip the diagnosis step and go straight to proposing.
- Declare a hypothesis without citing specific evidence from trace files.
- Promote a candidate with `evaluated: false` in its `score.md`.

---

## SEARCH LOOP SUMMARY

```
Iteration N:
  1. harness-evaluator  → pareto-frontier.md updated
  2. meta-harness       → candidate vNNN proposed
  3. Human              → reviews diff.md, approves
  4. Human              → applies candidate to skills/
  5. Human              → runs ≥3 sessions
  6. harness-evaluator  → scores new sessions
  7. meta-harness --promote vNNN → promotes if better
  8. Repeat from 1
```

Each iteration is one step of the harness search. The proposer (this skill) handles steps 2 and 7. The evaluator handles steps 1 and 6. The human handles steps 3, 4, and 5.