Autoresearch

Overview

autoresearch is a stateful, evaluator-driven iterative improvement loop. Once started against an explicit mission, it keeps iterating through non-passing evaluator results, records each evaluation and decision as durable artifacts, and stops only when an explicit max-runtime ceiling — or another explicit terminal condition — is reached.

This is a Level 4 skill: it owns one mission at a time and does not pause for confirmation between iterations.

When to Use

• You already have a mission and evaluator from /oh-my-claudecode:deep-interview --autoresearch
• You want persistent single-mission improvement with strict evaluation
• You need durable experiment logs under .omc/autoresearch/
• You want a supported path for periodic reruns via Claude Code native cron

When NOT to Use

• You still need to generate the evaluator → use /oh-my-claudecode:deep-interview --autoresearch first
• You need multiple missions orchestrated together — v1 forbids that
• You wanted the deprecated omc autoresearch CLI flow — it is no longer authoritative

Usage

/oh-my-claudecode:autoresearch

Optional arguments:

/oh-my-claudecode:autoresearch --mission-dir <path> --max-runtime <duration>
/oh-my-claudecode:autoresearch --resume <run-id>
/oh-my-claudecode:autoresearch --cron <spec>

Contract

• Single mission only in v1
• Mission setup and evaluator generation belong to deep-interview --autoresearch
• Evaluator output must be structured JSON with a required boolean pass and optional numeric score
• Non-passing iterations do not stop the run
• Stop conditions are explicit and bounded — max-runtime is the primary strict stop hook

State Layout

Canonical persistent storage lives under .omc/autoresearch/<mission-slug>/:

.omc/autoresearch/<mission-slug>/
├── mission.md
├── evaluator.json
└── runs/<run-id>/
    ├── evaluations/
    │   ├── iteration-0001.json
    │   └── iteration-0002.json
    └── decision-log.md

Reuse existing runtime artifacts when available rather than duplicating them.

Required Artifacts

The skill writes, at minimum:

• Mission spec
• Evaluator script or command reference
• Per-iteration evaluation JSON (with required pass boolean)
• Markdown decision log

Workflow

1. Confirm a single mission exists and evaluator setup is already available
2. Activate autoresearch mode state, recording mission slug/dir, evaluator reference, and run-id
3. Enter the loop: run candidate change → evaluate → record iteration JSON + decision-log entry
4. Continue while pass !== true until max-runtime (or another explicit terminal condition) hits
5. Persist final state and write a run summary; surface the latest evaluator JSON and decision-log path

Related Skills

• deep-interview — generates the mission and evaluator
• self-improve — tournament-selection evolutionary loop
• ralph — generic self-referential loop with verifier

Reference