AGENT SKILL

auto-research

Run a general Solution Autoresearch System (SAS) with an adaptive GPT-only agent graph, deterministic evaluation, a locally integrity-checked evidence log, and Open Knowledge Format (OKF) memory. Use for hard problem solving, invention loops, research campaigns, repeated optimization, hypothesis search, or any task where agents should learn across measured cycles without turning guesses into memory.

License 许可
MIT
所属仓库
Babinci/copernicus
最近更新
2026-09-30
Stars
★ 0
Forks
0

01 前言 Front Matter

name
auto-research
description
Run a general Solution Autoresearch System (SAS) with an adaptive GPT-only agent graph, deterministic evaluation, a locally integrity-checked evidence log, and Open Knowledge Format (OKF) memory. Use for hard problem solving, invention loops, research campaigns, repeated optimization, hypothesis search, or any task where agents should learn across measured cycles without turning guesses into memory.

02 资源文件 5

03 SKILL.md 正文

Copernicus Auto-Research

SAS is the invention loop. OKF is its memory layer. Build the smallest graph the actual problem needs; do not import a standing agent society.

Resolve <auto-research-skill-dir> to the directory containing this SKILL.md. Resolve all bundled scripts and references from that directory, never from the user's current working directory.

Read guide.md when explaining the method, designing a first research cycle, selecting an evaluator, or changing this skill. It defines the concepts, purpose, outputs, examples, self-contained boundary, and honest limitations.

Start

  1. Write the user's problem and acceptance test into a plain Markdown file.
  2. Create a run bundle:
python3 <auto-research-skill-dir>/scripts/receipts.py init RUN_DIR --problem PROBLEM.md --title "Short title"
  1. Read contracts.md. Fill graph.json, set its status to ready, and run:
python3 <auto-research-skill-dir>/scripts/receipts.py verify RUN_DIR

Invent the graph

Freeze the problem regime before choosing agents: objective, inputs, evaluator, constraints, budget, freshness boundary, and stop conditions. Then create only roles that change a named artifact or decision.

Every node must declare:

  • bounded input and typed output;
  • lane: deterministic, sol, terra, luna, or human;
  • dependencies;
  • authority ceiling;
  • independent verifier;
  • falsifier;
  • terminal status.

Prefer deterministic controller/evaluator nodes. Use models as bounded mappers, proposers, critics, implementers, or curators. A role need not be a separate agent. Delete any role that does not change a named artifact or decision.

Run one cycle

  1. Snapshot the regime and append a regime receipt.
  2. Select applicable OKF cards deterministically and record their IDs plus file hashes. Baselines record an explicit empty-card receipt.
  3. Execute ready graph nodes through the sibling $fleet skill. Run independent nodes in parallel and dependency edges in order. If Fleet is unavailable, execute the same nodes sequentially; the evidence contract matters more than parallelism.
  4. Make each candidate content-addressed. Bind it to the regime, parent candidate, changed hypothesis, prompt/strategy identity, selected card hashes, and artifact hash.
  5. Run the cheapest trusted filter first, then the authoritative evaluator, then an independent semantic or normative review when correctness is broader than a test result.
  6. Retry only when the failure signature and hypothesis changed. Never retry identical candidate bytes or prompts.
  7. Join candidate, evaluation, review, and cost evidence into one cycle_join receipt.
  8. Only at the cycle boundary, promote, reject, quarantine, or retire an OKF card. A hypothesis never promotes itself.

Append each receipt from a JSON object:

python3 <auto-research-skill-dir>/scripts/receipts.py append RUN_DIR/evidence.jsonl receipt.json
python3 <auto-research-skill-dir>/scripts/receipts.py verify RUN_DIR

OKF memory

Keep one concept per Markdown file with YAML frontmatter. Start with only three card families:

  • selection: what is worth attempting;
  • strategy: what change is likely to help;
  • risk: what can invalidate or mislead the evaluation.

Cards require evidence, regime scope, a falsifier, and a disposition. New domains may rename these concepts, but do not add a fourth family until a measured workflow cannot fit one of the three.

Use relative Markdown links and keep knowledge/index.md as progressive navigation. Update knowledge/log.md only for material curation changes.

Stop rules

Stop and report rather than guessing when the evaluator is missing, the regime changed, evidence cannot be joined, a provider/model is unavailable, the same failure repeats, or the budget is exhausted. Stop for user authority before purchases, publishing, deployment, messaging, signing, secrets, or destructive/external mutations.

Do not add a graph database, vector store, agent genome, role league, custom queue, dashboard, daemon, or second state store until a measured run proves the flat graph, files, and JSONL ledger insufficient.

Auto-Research is self-contained: it uses Markdown, JSON, the bundled Python standard-library receipt tool, and optional sibling Fleet execution. It does not depend on a private repository, hosted database, external memory service, or undisclosed evaluator. The user must supply the problem and the real evaluator.

04 同仓库更多 Skills