ShaheerKhawaja/autoloop

autoloop

Autonomous recursive improvement loop for a single target. Runs gap analysis, recursive refinement, evaluation, and convergence checks until the target reaches quality threshold or converges.

Inputs

| Parameter | Values | Default | Description | |-----------|--------|---------|-------------| | target | path or context | cwd | What to operate on |

/autoloop — Autonomous Recursive Improvement

Step 0: Preamble

Before executing, run the shared ProductionOS preamble (templates/PREAMBLE.md).

You are running the /autoloop command. This is an autonomous recursive improvement loop that takes a target and iteratively improves it until convergence.

Input

The user provides:

• Target: A file path, directory, or description of what to improve
• Goal: What "good" looks like (optional -- defaults to "maximize quality score")

Execution Protocol

Step 1: Understand the Target

1. If target is a file path: Read it and assess current state
2. If target is a directory: Scan for key files and assess overall quality
3. If target is a description: Identify what needs to be created or improved

Step 2: Gap Analysis

1. Score current state using the ProductionOS rubric and convergence heuristics already present in this repo
2. Scan ~/repos/ for reference implementations (per CLAUDE.md Auto-Enrichment Protocol)
3. Check ~/.productionos/recursive/reference-corpus/ for similar high-quality outputs
4. Identify specific gaps between current state and goal

Step 3: Initialize Recursion

1. Create session state at ~/.productionos/recursive/recursion-state.json:

   {
     "session_id": "<generated>",
     "target": "<target>",
     "goal": "<goal>",
     "layer": "L17",
     "current_iteration": 0,
     "max_iterations": 10,
     "best_iteration": 0,
     "best_score": 0.0,
     "scores": [],
     "convergence_verdict": "CONTINUE",
     "status": "running"
   }
   

2. Select the appropriate layer:
   • Complex decomposable task -> L16 RecDecomp
   • Quality improvement (default) -> L17 SelfRefine
   • Context too large -> L18 RecSumm
   • Security/factual claims -> L19 RecVerify
   • Plan execution -> L20 PEER

Step 4: Iteration Loop (max 10)

For each iteration:

1. Score: Run confidence scorer on current output
2. Record: Add score to convergence monitor
3. Check Convergence: Run all 5 algorithms from convergence.py:
   • Score delta tracking (stalled if < 0.1 for 2+ iterations)
   • Spectral contraction (converged if cosine > 0.95)
   • Diminishing returns (stalled if DR ratio < 0.15)
   • Oscillation detection (oscillating if sign changes > 60%)
   • EMA velocity (plateau if |EMA delta| < 0.05)
4. If STOP: Return best iteration output
5. If CONTINUE: Apply refinement via the selected layer
6. Quality Gate: Check for monotonic improvement and stop if the loop regresses materially
7. Log: Write metrics to ~/.productionos/recursive/metrics/

Step 5: Completion

1. Return the output from the best-scoring iteration
2. Show convergence trajectory (ASCII visualization)
3. Report: iterations completed, final score, convergence reason
4. Save final state to recursion-state.json

Output Format

AUTOLOOP COMPLETE
Target: <target>
Goal: <goal>
Iterations: <n> / <max>
Best Score: <score> (iteration <i>)
Convergence: <verdict> — <reason>

Trajectory:
  i=0  |***           | 4.20
  i=1  |*********     | 6.50 (+2.30)
  i=2  |***********   | 7.20 (+0.70)
  i=3  |************  | 7.30 (+0.10) <- converged

Applied: <output from best iteration>

Constraints

• Max 10 iterations per autoloop invocation
• Configurable depth per iteration (default: 1)
• Always check token budget before each iteration
• Never modify Phase 1 or Phase 2 RLM scripts
• Log everything to metrics for PromptEvo batch analysis
• Use rlm-recursive-orchestrator agent for depth management when needed

Integration

This command integrates all Phase 1-3 RLM components:

• confidence_scorer.py — scoring each iteration
• quality_gate.py — monotonic improvement enforcement
• convergence.py — 5-algorithm convergence detection
• instinct_scorer.py — weight adjustment from learned patterns
• embedding_corpus.py — reference comparison
• prompt_evolution.py — active prompt selection per layer
• tier2_live_eval.py — evaluation framework
• rlm_classifier.py — budget circuit breaker

Error Handling

| Scenario | Action | |----------|--------| | No target provided | Ask for clarification with examples | | Target not found | Search for alternatives, suggest closest match | | Missing dependencies | Report what is needed and how to install | | Permission denied | Check file permissions, suggest fix | | State file corrupted | Reset to defaults, report what was lost |

Guardrails

1. Do not silently change scope or expand beyond the user request.
2. Prefer concrete outputs and verification over abstract descriptions.
3. Keep scope faithful to the user intent.
4. Preserve existing workflow guardrails and stop conditions.
5. Verify results before concluding.