converge

name	converge
description	Run the Convergence Engine on any prompt file. Iterates up to 100 times, automatically fixing clarity, completeness, efficiency, model fit, and failure resilience until the prompt reaches DEPLOY quality. Auto-triggers on: "/converge", "converge this prompt", "optimize until perfect", "iterate until deploy", "run convergence".
allowed-tools	Bash(python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/convergence.py *) Bash(python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/token-count.py *) Bash(python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/self-eval.py *) Bash(python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/report-gen.py *) Read Write Edit Agent

Convergence Engine

Autonomous prompt optimization. Like gradient descent for prompts — each iteration reduces deviation from perfection.

Usage

The user provides a prompt file path or a prompt folder. Run the full pipeline.

Pipeline

Step 0: Read the Wixie briefing (top-of-context)

Before iteration 1, read the inference-engine's Wixie briefing if present:

BRIEF="${CLAUDE_PLUGIN_ROOT}/../inference-engine/state/briefings/wixie.md"
[ -f "$BRIEF" ] && cat "$BRIEF"

The briefing is machine-generated by the inference-engine and lists elevated patterns — cross-session failures that crossed the Wald SPRT elevation threshold, with calibrated Beta-Binomial posteriors and EMA-decayed weights. Apply any elevated pattern's Counter before iteration 1. The briefing is advisory, never blocking; if missing or stale, proceed honestly and note it in the final report.

Step 1: Locate the prompt

If the user provides:

• A file path → use that file directly
• A folder path → find prompt.* inside it
• A prompt name → look in ${CLAUDE_PLUGIN_ROOT}/../../prompts/<name>/prompt.*
• Nothing → list available prompts from ${CLAUDE_PLUGIN_ROOT}/../../prompts/index.json and ask user to pick

Step 2: Run convergence

python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/convergence.py <prompt-file>

This runs up to 100 iterations:

• Scores the prompt on 5 axes (Clarity, Completeness, Efficiency, Model Fit, Failure Resilience)
• Identifies the weakest axis
• Applies targeted fix (hedge words, missing components, filler, format, fallbacks)
• Re-scores and repeats
• Exits on DEPLOY (overall ≥ 9, all axes ≥ 7) or plateau (3 identical scores)

Step 3: Update artifacts

After convergence:

1. Run token count:

python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/token-count.py <prompt-file> --model <target-model>

2. Run self-eval for final scores:

python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/self-eval.py <prompt-file>

3. Update metadata.json with new scores.

4. Generate report:

python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/report-gen.py <prompt-folder>

5. Update ${CLAUDE_PLUGIN_ROOT}/../../prompts/index.json.

Step 4: Review

Validate the result:

• All files exist and are non-empty
• Metadata scores match self-eval output (tolerance ±1)
• Target model exists in registry
• Format matches model preference

If issues found, fix and re-run convergence (max 3 review cycles).

Step 5: Report

Tell the user:

Convergence complete: X.X → Y.Y in N iterations
Verdict: DEPLOY / BEST EFFORT
[axis scores]

Rules

• Do NOT ask for permission. Run everything autonomously.
• Do NOT modify the prompt's intent or domain content.
• The convergence script handles all text fixes. You handle artifacts and review.
• If convergence.py fails, fall back to manual: read self-eval, apply fixes yourself, re-score, repeat up to 10 times.