菜单
Run Codex-native AutoResearch loops with versioned scorecards, durable session files, METRIC packets, ASI logging, and keep/discard/crash/checks_failed decisions.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | autoresearch |
| description | Run Codex-native AutoResearch loops with versioned scorecards, durable session files, METRIC packets, ASI logging, and keep/discard/crash/checks_failed decisions. |
Use this skill when a research or improvement loop needs to test candidate changes against a measurable target before deciding whether to keep them. Codex main owns the decision. External models may advise only when the content is GREEN or sanitized YELLOW.
Use the unified loop:
Target -> Onboard -> Setup -> Doctor -> Packet -> Log -> Continue or Finalize
Start by selecting a scorecard from config/scorecards. Record the id and version before candidate work begins. Scorecards, eval scripts, and fixtures are read-only during eval runs.
For real repo loops, create durable state in the target repo:
autoresearch.md for goal, metric, scope, constraints, and stop conditions.autoresearch.jsonl for append-only config and packet decisions.autoresearch.ideas.md for deferred hypotheses and next-action notes.autoresearch.last-run.json for the latest packet evidence.autoresearch.research/<slug>/ only for optional quality-gap research rounds.Benchmarks must print the primary metric as METRIC name=value. Secondary METRIC lines may explain tradeoffs but do not drive keep/discard. Missing, null, crashed, clipped, or ineligible metrics are unknown; never report them as zero or as wins.
Every packet decision must log ASI: hypothesis, evidence, next_action_hint, and rollback_reason for discard, crash, or checks_failed. Optional ASI fields such as lane, family, and risk are useful when a loop spans many attempts.
Keep a candidate only when the primary metric is finite, hard gates pass, the latest packet is fresh, protected eval inputs are unchanged when applicable, scoped commit paths are clear, and the score or metric delta meets the suite threshold. Otherwise log discard, crash, or checks_failed and preserve evidence.
Persist PASS results to MiVault as GREEN knowledge. RED content must never be saved, logged, or externalized.
Use the local Ralph adapter first. From this repo:
python3 scripts/autoresearch/doctor.py --cwd <target-repo>
python3 scripts/autoresearch/setup.py --cwd <target-repo> --goal "<goal>" --metric <name> --direction lower|higher --benchmark-command "<command>"
python3 scripts/autoresearch/next.py --cwd <target-repo>
python3 scripts/autoresearch/log.py --cwd <target-repo> --from-last --status keep|discard|crash|checks_failed
python3 scripts/autoresearch/state.py --cwd <target-repo> --compact
If the upstream codex-autoresearch MCP or CLI is installed, use its read-only planning/state tools as an optional backend for setup guidance. Mutation still goes through Codex approval, Ralph CLI logging, scorecards, and gates.
For medium or high-complexity hypotheses, GLM-5.1 may be used through ralph_coding_models.zai_coding_deep as an optional counterpart. MiniMax-M2.7-highspeed may be used through ralph_coding_models.minimax_agentic_fast for compact summaries, log review, diff review, or test ideas.
Do not use Z.ai or MiniMax as direct model_provider backends. Do not route RED data to external MCPs.
Use scripts/evals/autoresearch_dry_run.py with tests/evals/fixtures/autoresearch_toy_speed for the deterministic toy validation.
Improve, audit, simplify, rewrite, or migrate prompts, tool descriptions, agent instructions, and prompt stacks for GPT-5.6 Sol or the GPT-5.6 family. Use for outcome-first prompt design, autonomy boundaries, tool routing, PTC, grounding, verbosity, reasoning effort, and prompt evals.
Apply a deep, design-minded engineering workflow for complex work that needs careful planning, iteration, and simplification.
Use when preparing or running a Claude CLI agentic engineering review through claude -p for repository audits, architecture analysis, system design, security review, large refactors, specs, RFCs, or evidence-grounded long-form engineering analysis.
Use when preparing or running a ZCode GLM-5.2 agentic builder workflow through zcode --prompt for fast implementation, iterative code generation, focused fixes, and validation on an existing repository.
Apply adversarial opposite-analysis to plans, specs, architecture, code changes, and claims. Use when the user asks for adversarial review, opposing analysis, contrarian review, red-team reasoning, or Z.ai and MiniMax cross-checks through the Ralph MCP router.
Review and adjudicate Bugbot, Cursor, Seer, and similar automated PR feedback with local evidence before accepting, fixing, or dismissing findings.