autoresearch

name

autoresearch

description

Run the knowledge-driven OpenUI research loop that coordinates the training pipeline with curated knowledge. Read and update the repo + personal brains (OKF / Obsidian) and OpenWiki, run the prior-work discovery loop to find related research, drive the autotrain hypothesis loop to generate falsifiable hypotheses, and file new ideas and experiments as Linear issues / milestones / projects. Use to ORCHESTRATE research; to RUN a pipeline phase use autotrain; to RUN an experiment campaign (init/research/validate/execute) use openui-autoresearch; to change harness code use improve-openui-harnesses.

Autoresearch — the OpenUI research loop

Facade for orchestrating research with progressive disclosure: this file routes; each stage's full instructions live in references/ and are read only when that stage runs. This skill coordinates the machinery it does not own — it drives autotrain (pipeline + hypothesis loop), openui-autoresearch (campaign execution), the openwiki CLI, the brains under docs/brains/, and the Linear MCP. It never builds a parallel trainer, evaluator, or knowledge store.

What this loop does

Turns curated knowledge into falsifiable, tracked experiments and folds results back into knowledge:

brains + OpenWiki  ──▶  discovery (prior work)  ──▶  hypotheses  ──▶  Linear issues
      ▲                                                                     │
      └──────────────  update brains from measured results  ◀──────────────┘

Stage routing

Orient once per session with references/loop.md + references/contracts.md — loop.md owns the full ordered sequence. The stages below run in table order; read a reference only when its stage runs.

#	Stage	Reads / writes	Reference
1	Read/update repo + personal brains (OKF / Obsidian)	docs/brains/	references/brains.md
2	Read / refresh codebase knowledge	docs/openwiki/ via the openwiki CLI	references/openwiki.md
3	Find prior work & related research	research-lineage.md, source manifests, campaign bundle	references/discovery.md
4	Drive the auto-training hypothesis loop	hypothesize (owned by openui-autoresearch / autotrain)	references/hypothesis.md
5	File ideas/experiments as Linear issues/milestones/projects	Linear MCP, team SLM	references/linear.md
6	Close the loop: fold results back into brains	docs/design/ → docs/brains/, synthesis-feedback	references/loop.md

Non-negotiable contracts

Digest — full versions in references/contracts.md:

• Knowledge, not results, lives in brains. Measured results go to docs/design/ (documenting-experiment-results); brains link them.
• Grounded & novelty-audited. Every hypothesis links prior work and is checked against recorded dead-ends before it is filed.
• One idea → one issue → one iter-*. A filed experiment maps to a Linear SLM-N issue and, on run, a docs/design/iter-*.md record.
• Inherit the pipeline's laws. Honesty (fixture vs ship), RL fail-closed, no shadow paths, and all approvals carry over from autotrain.
• No secrets / no leakage. Brains and issues never contain credentials, machine-absolute paths, or held-out eval content.