Codex Autoresearch
A codex plugin for running optimization loops inside a codebase. It is useful when you have a measurable target and many possible changes to try: test runtime, build speed, bundle size, model loss, Lighthouse scores, memory use, query latency, or any other metric you can print from a script.
TheGreenCedar · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add TheGreenCedar/codex-autoresearch技能说明
name: codex-autoresearch description: Triage improvement work and run or resume accepted measured loops in a local project. Architecture, documentation, UX, product study, open research, taste, and one-shot fixes stay direct unless the user explicitly requests repeated measurement with a complete experiment contract.
Codex Autoresearch
Decide fit before exploring the repository. Autoresearch governs repeated measured experiments; it does not take over every task that mentions research, quality, or improvement.
Use this as the only Codex-facing Autoresearch skill. Do not route to retired subskills, slash commands, or MCP surfaces.
Route before discovery
Make one read-only fit call before benchmark discovery, recipe lookup, repository scanning, default inference, or setup:
node scripts/autoresearch.mjs prompt-plan --cwd <project> --prompt "<request>"
Follow its typed disposition:
continue-direct: use the direct evidence capsule below. Create no Autoresearch files, packets, commits, dashboards, research folders, or finalization state. Leave an unrelated session untouched.needs-user: ask only for the reported missing fields or conflicts. Do not fill them from plausible repository files or inferred defaults.run-loop: treat the returned contract as an in-memory candidate. Only now inspect the owning repository and establish or resume the accepted contract through setup or an explicit segment transition.
An existing session is matching only when repository, checkout, goal, metric semantics, evaluator, checks, and scope are compatible. Shared words are not evidence of a match. Replacing or abandoning a session requires explicit user intent.
An explicit loop request with an incomplete contract is needs-user, never a half-configured loop.
Continue directly when the loop does not fit
Use this evidence capsule:
- State the requested outcome.
- Identify the main uncertainty.
- Gather the cheapest evidence that can resolve it.
- Perform the direct task.
- Verify the result and bound the claim.
Direct work may finish an implementation, explanation, review, or ordinary correctness check. It may not claim measured improvement or authorize a keep without accepted evaluator and checks evidence.
Architecture, documentation, UX, product study, open-ended research, taste, bugs, quality, delight, and generic improvement language do not independently select a loop. A qualitative gap loop is appropriate only when the user explicitly wants repeated evaluation against a stable, accepted checklist.
Establish the accepted experiment
Once fit is run-loop:
- Identify the repository and child package that own the work.
- Run
git status --short --branchand preserve unrelated changes. - Establish one complete contract: goal, repository and worktree identity, metric semantics, evaluator, independent checks, editable and protected scope, noise model, keep rule, stop rule, and enforceable budgets.
- Use
setupfor a new session or an explicit segment transition for a replacement contract. Do not execute a packet untilstate --reportshows an accepted contract. - Configure
commitPathsbefore a keep may commit changes.
The accepted evaluator and checks are the only execution authority. CLI, config, wrapper, separator, command-file, or environment-file overrides may run only when they reproduce the accepted execution digest exactly. Otherwise stop and transition the contract explicitly.
Metric names carry no semantics. A name containing quality, score, precision, or similar text does not imply a direction, threshold, target, or perfect value.
Unknown noise permits qualification baselines. It does not permit a keep until the required repeats establish a valid comparison. Estimated model tokens or calls are advisory unless trusted host telemetry makes them enforceable.
Resume from one canonical decision
For an existing matching session, run one bounded read:
node scripts/autoresearch.mjs state --cwd <project> --report
Do not reread raw session files and separately ask state, recommendation, doctor, watchdog, portfolio advice, and finalization to vote on the next step. The report projects one DecisionPlan with:
- phase and canonical action
- blocker code and capability-scoped diagnostics
- loop and parent dispositions
- contract digest and evaluator identity
- required evidence
Follow that decision. Use doctor only when the decision asks for a diagnostic or when the user explicitly requests one. If terminal and dashboard semantic fields disagree, stop mutation and diagnose the projection.
Read loop operations only when the canonical action requires packet, recovery, budget, Git-scope, or segment detail.
Run one bounded packet
The usual accepted loop is:
setup -> state -> next -> log -> state -> finalize-preview
next may execute only the accepted evaluator and accepted checks, using their accepted execution specifications. After it returns:
- Inspect the metric, checks, artifacts, diff, and Git state.
- Log with
--from-last; do not retype parsed metrics. - Record the real hypothesis and learning assessment. Learning defaults to
none;causalordiscriminatingrequires evidence and a concrete changed belief. - Read the resulting decision before doing more work.
| Status | Use it for |
|---|---|
measure | Baselines, qualification repeats, no-change checks, and diagnostics. Never authorize a keep. |
keep | A candidate evaluated by the accepted contract, with all checks, metric comparison, and noise qualification satisfied. |
discard | A finite candidate result that is not worth keeping. |
crash | Evaluation failed before usable metric evidence existed. Do not invent a sentinel metric. |
checks_failed | A metric exists, but accepted correctness checks failed. |
Baselines and accepted candidate packets consume packet budget.
阅读完整来源 ↗