ai-lab

Close the current (or given) ai-lab session — the archivist files it into the knowledge base (incl. null results and campaign lessons), the digest and research map are regenerated, the lab is checked for consistency, the handoff note and cycle progress are updated, and the session is committed to the lab's git history.

Multi-agent brainstorm in an ai-lab workspace — several ideator subagents with different lenses generate ideas independently and in parallel; the ideas are merged, checked for prior-art collisions, ranked by pairwise comparison, refined, optionally piloted, and the shortlist is brought to 교수님 to decide. Use when the lab needs new research ideas or hypotheses on a topic.

Autonomous (e.g. overnight) exploration in an ai-lab workspace — 교수님 approves a bounded search contract once; then the explorer proposes one change per trial while lab.py runs every trial on a fixed time budget and decides keep/discard from the metric; the campaign ends with a fresh-seed confirmation, a summary, a review, and a decision request. Use for searches over one metric (hyperparameters, architecture tweaks, training tricks).

Run a lab meeting in an ai-lab workspace with 교수님 (the user) — cycle planning, progress check, results review, journal club, retrospective, a 1:1 with one researcher, or a general discussion. Researchers argue from their roles (the critic answers every proposal), 교수님 decides class-C matters, and minutes, decisions and action items are recorded.

Cheap overview of an ai-lab workspace — long-term goal, current cycle, handoff note, decisions waiting for 교수님, running campaigns, consistency warnings, open actions — and the suggested next step.

Create a new, separate ai-lab research workspace (its own folder/git repo) from the plugin template.

Draft a Korean paper or technical report from the lab's findings in an ai-lab workspace — claims only from reviewed findings, numbers only from result files, citations only from verified paper entries — then audit it, have it reviewed like a conference submission, and revise once.

Package accepted research results as markdown-only documents that 교수님 can carry into the company (only .md files can be brought in) — a self-contained report, plus optionally the code packed into a restorable .md.

Execute open lab action items in an ai-lab workspace by delegating each to the right researcher subagent (scout, engineer, critic, writer) with a compact task brief, enforcing the decision rules (pilots autonomous, full runs and findings need 교수님), then log results. Use after a lab meeting or brainstorm produced action items.