From autoresearch
Review autoresearch results. Shows what experiments ran, what improved, what failed, and lets you drill into specific experiments.
How this command is triggered — by the user, by Claude, or both
Slash command
/autoresearch:reviewThe summary Claude sees in its command listing — used to decide when to auto-load this command
# Autoresearch Review You are debriefing the human on the results of an autoresearch run. ## Find the initiative First, check what initiatives exist: Each subdirectory is a named initiative (e.g. `sort-optimization`, `api-latency`). - If there's only one, use it automatically. - If there are multiple, list them with their best scores and ask which one to review: - If the human already specified one (e.g. `/autoresearch:review sort-optimization`), use that. ## Present results Read these files from `autoresearch/<name>/`: - `ledger.json` — all directions and their experiment histo...
You are debriefing the human on the results of an autoresearch run.
First, check what initiatives exist:
ls autoresearch/
Each subdirectory is a named initiative (e.g. sort-optimization, api-latency).
Found 3 research initiatives:
sort-optimization — best: 12940933.18 (9 experiments)
api-latency — best: 142.00 (30 experiments)
parser-accuracy — best: 0.91 (15 experiments)
Which one?
/autoresearch:review sort-optimization), use that.Read these files from autoresearch/<name>/:
ledger.json — all directions and their experiment historyroadmap.json — ranked future directionsbest_score.txt — current beststate.json — orchestrator statebranches.jsonl — branch historyPresent a concise summary:
Ask: "Want to inspect a specific experiment? Give me a number, or say 'done'."
If they give a number, show the full log entry: hypothesis, diff, score, branch, worker.
When they're done inspecting, offer:
"The best version of your code is in autoresearch/<name>/best/. Want me to copy it back into your project, replacing the original files?"
If yes:
git status --porcelain. If the target files have uncommitted changes, stop and warn: "Target files have uncommitted changes: {list}. Applying would overwrite them. Commit, stash, or explicitly confirm to proceed."autoresearch/<name>/best/ back to its original location (the editable files listed in program.md). Show what was copied.If no, just tell them where the files are.
autoresearch/ directory exists: "No research initiatives found. Run /autoresearch:design to start one."/reviewReviews staged changes or recent commits across five axes—correctness, readability, architecture, security, performance—producing categorized findings with file:line references and fixes.
/reviewRuns Codex code review on local git state (working tree or vs base branch). Supports --wait/--background, --base <ref>, --scope auto|working-tree|branch.
/reviewInvokes multiple external AI CLIs (Gemini, Claude, Codex, etc.) to independently review a phase plan and produces a structured REVIEWS.md with per-reviewer feedback.
/reviewRuns an enhanced multi-LLM PR review with inline comments, checking availability of local and API-based providers for cross-model escalation.
/reviewReviews HTML file for design anti-patterns, principles violations, and accessibility issues. Generates markdown report with status tables and recommendations.
/reviewReviews specified code scope via four specialists (quality, security, performance, architecture), producing summary, detailed findings, refactoring suggestions, prioritized action plan.
npx claudepluginhub mirceastrugaru/autoresearch