autoconfig

name	autoconfig
description	24/7 self-improving Claude configuration loop — continuously optimizes runtime config via automated benchmarking
user_invocable	true
commands	["status","run","report","rollback","best","pause","resume"]

AutoConfig — Self-Improving Configuration Loop

Inspired by Karpathy's autoresearch: autonomously mutate config knobs, benchmark via claude -p, keep improvements, discard regressions. Improvements compound — each kept change becomes the new baseline.

Commands

/autoconfig status

Show current state: phase, experiments run, improvements found, rate limit status, daemon state.

Steps:

1. Read ~/.claude/state/autoconfig/program_state.json
2. Query experiment DB for totals: get_experiment_count(), get_total_kept(), get_cumulative_improvement()
3. Check daemon process: pgrep -f experiment_daemon.py
4. Check rate limit state
5. Display summary table

/autoconfig run

Run one experiment cycle manually (foreground).

Steps:

1. Import and call run_one_experiment() from experiment_daemon.py
2. Display the mutation tried, benchmark results, and keep/discard decision
3. If kept, show the improvement percentage

/autoconfig report

Generate a comprehensive report of experiment history.

Steps:

1. Query experiment DB: top improvements, trend analysis, per-phase stats
2. Compute Pareto frontier (best composite at each phase)
3. Show per-knob attribution (which changes contributed most improvement)
4. Show total drift from original config (knobs changed)
5. Show cumulative improvement curve
6. Show convergence status per phase

/autoconfig rollback

Restore the original baseline config.

Steps:

1. Confirm with user: "This will restore the config to the initial baseline. Continue?"
2. Call restore_snapshot("baseline")
3. Verify config is valid (JSON parse, required fields present)
4. Display what changed

/autoconfig best

Apply the best-ever config.

Steps:

1. Check if best snapshot exists
2. Show comparison: current score vs best score
3. Call restore_snapshot("best")
4. Verify config validity

/autoconfig pause

Stop the daemon temporarily.

Steps:

1. Run launchctl unload ~/Library/LaunchAgents/com.chadsimon.autoconfig.plist
2. Verify daemon stopped
3. If mid-experiment, restore checkpoint

/autoconfig resume

Restart the daemon.

Steps:

1. Run launchctl load ~/Library/LaunchAgents/com.chadsimon.autoconfig.plist
2. Verify daemon started
3. Show current phase and experiment count

Config Surface

Target	File	Knobs
Route manifest	~/.claude/state/route_manifest.json	~200 (models, effort, lanes, swarm, dispatch)
Settings	~/.claude/settings.json	~5 (effort level, model defaults)
Agent defs	~/.claude/agents/*.md	~15 (model, effort, behavioral rules per agent)

Safety

• Immutable fields are NEVER touched (permissions, hooks, MCP servers, risk classes)
• All mutations validated against bounds before application
• Checkpoint/rollback on every experiment
• Crash recovery on daemon restart
• See references/safety_spec.md for full guardrail documentation

References

• references/safety_spec.md — Immutable fields, mutation bounds, isolation rules
• references/metric_spec.md — Composite metric formula, noise band, confirmation trials
• program.md — Current research phases and exploration strategy
• scripts/run_benchmark_matrix.py — Files-first repeated benchmark comparison across current, baseline, and best presets