Autoresearch

Autonomous optimization loop that runs N iterations against a mechanical metric, keeping improvements and discarding regressions. Every change is git-tracked for safe rollback.

What This Skill Does

Autoresearch takes a measurable goal (test coverage, bundle size, performance score) and iteratively makes atomic code changes to improve it. Each iteration commits before verification, compares against baseline, and keeps or discards the change. A stuck-detection mechanism shifts strategy after consecutive failures.

When to Use

Increasing test coverage from X% to Y%
Reducing bundle size below a target
Improving a performance benchmark score
Any goal expressible as a single numeric metric from a shell command

Core Capabilities

N iterations of autonomous code changes (default: 10)
Git-tracked experiments with automatic rollback on regression
Guard pattern to prevent regressions in other areas (e.g., tsc --noEmit)
Stuck detection: 5+ consecutive failures triggers strategy shift
Results logged to TSV with per-iteration history
Configurable noise tolerance and minimum delta

Arguments

Argument	Required	Description
`Goal`	Yes	What to optimize (natural language)
`Scope`	Yes	Glob patterns for files to modify
`Verify`	Yes	Shell command that outputs a single number
`Guard`	No	Shell command that must pass (exit 0) after each change
`Iterations`	No	Max iterations (default: 10)
`Direction`	No	`higher` or `lower` (default: higher)
`Min-Delta`	No	Minimum improvement to keep a change
`Noise`	No	`low`, `medium`, or `high` — tolerance for metric variance

Example Usage

/ck:autoresearch
Goal: Increase test coverage in src/utils from ~60% to 80%
Scope: src/utils/**/*.ts, tests/utils/**/*.test.ts
Verify: npx jest tests/utils --coverage --coverageReporters=json-summary 2>/dev/null | node -e "const d=require('./coverage-summary.json');console.log(d.total.lines.pct)"
Guard: npx tsc --noEmit && npx jest --passWithNoTests
Iterations: 15
Direction: higher

Workflow

Configure — parse required fields from argument or ask interactively
Baseline — run Verify command and record starting metric
Iterate — for each iteration:
- Make ONE atomic change to files in Scope
- Commit before running Verify
- Compare metric against baseline
- Run Guard if provided
- Keep (update baseline) or discard (git revert)
- Log result to TSV
Stuck detection — 5 consecutive discards triggers strategy shift; 10 consecutive discards forces a full stop
Stop — after N iterations or 10 consecutive discards

Output

Results TSV with columns: iter, timestamp, metric, delta, kept, description

Loop — similar optimization loop with identical protocol
Security — uses autoresearch pattern for iterative security fixes
Test — test suite that Verify commands often wrap