Autoresearch
🦊 阿舟 · Autoresearch
🧪 实验通宵跑，边界先说清。
This package is an Azhou-authored wrapper. The experiment protocol lives in the upstream program.md and is read at runtime from the user's own checkout; this package vendors zero upstream bytes because the upstream repository publishes no license file. See setup and provenance.
Brand protocol
Emit this exact display event once, at the start of every run, with the resolved checkout scope:
🦊 阿舟 · Autoresearch 启动｜mode=<prepare|run|resume|report>｜scope=<checkout>
Use ✅ 验证通过 only after every declared check has run and its output was read back. Use ❌ 验证失败 for a failed environment, checkout, or run check and 🔒 阿舟暂停这一项 when an unattended GPU run or a data download waits for explicit user confirmation. Emoji is display-only; keep JSON keys, schema values, digests, paths, commands, test names, and raw evidence emoji-free. A host without Unicode may remove the leading emoji while preserving the fixed text, ｜ separators, fields, and values.
Workflow
- Resolve the checkout from a user-supplied path only. Verify that
git rev-parse HEAD inside it equals the pinned commit recorded in setup, and refuse to continue on any mismatch. Never scan unrelated directories and never clone into any Git repository.
mode=prepare verifies the environment per setup: uv present, CUDA GPU visible, uv sync clean, data prepared, and one baseline training run possible. Every missing check fails closed; no partial state is reported as ready.mode=run and mode=resume read program.md from the checkout and follow it inside that checkout. Before any unattended sequence, hold with 🔒 阿舟暂停这一项 until the user confirms the GPU hours and disk cost. Results stay in the checkout; this skill never pushes, publishes, or copies results into any repository.mode=report aggregates experiment results that already exist in the checkout. Conversation excerpts, machine paths, and other raw evidence stay out of any committed surface.- End with a receipt containing
schema, status, current_truth, artifacts, verification, holds, next_action, and learning_signal:
## 🦊 阿舟 · Autoresearch receipt
- schema: autoresearch.receipt.v1
- status: pass | fail | hold
- current_truth: <one sentence the checks actually prove>
- artifacts: <paths inside the user checkout>
- verification: <comma-separated check ids>
- holds: <none|fact>
- next_action: <one concrete step>
- learning_signal: <none|one line>
If no pinned checkout is available, stop with status=hold and request one explicit checkout path.
1---2name: autoresearch3description: Wrap Karpathy's autoresearch environment so an agent can run, resume, and report automatic nanochat training experiments inside a user-owned pinned checkout. Requires an NVIDIA GPU and uv; vendors no upstream bytes. Use when the user asks to run or check autoresearch experiments.4---56# Autoresearch78**🦊 阿舟 · Autoresearch**910> 🧪 实验通宵跑，边界先说清。1112This package is an Azhou-authored wrapper. The experiment protocol lives in the upstream `program.md` and is read at runtime from the user's own checkout; this package vendors zero upstream bytes because the upstream repository publishes no license file. See [setup](references/setup.md) and [provenance](references/provenance.md).1314## Brand protocol1516Emit this exact display event once, at the start of every run, with the resolved checkout scope:1718```text19🦊 阿舟 · Autoresearch 启动｜mode=<prepare|run|resume|report>｜scope=<checkout>20```2122Use `✅ 验证通过` only after every declared check has run and its output was read back. Use `❌ 验证失败` for a failed environment, checkout, or run check and `🔒 阿舟暂停这一项` when an unattended GPU run or a data download waits for explicit user confirmation. Emoji is display-only; keep JSON keys, schema values, digests, paths, commands, test names, and raw evidence emoji-free. A host without Unicode may remove the leading emoji while preserving the fixed text, `｜` separators, fields, and values.2324## Workflow25261. Resolve the checkout from a user-supplied path only. Verify that `git rev-parse HEAD` inside it equals the pinned commit recorded in [setup](references/setup.md), and refuse to continue on any mismatch. Never scan unrelated directories and never clone into any Git repository.272. `mode=prepare` verifies the environment per setup: uv present, CUDA GPU visible, `uv sync` clean, data prepared, and one baseline training run possible. Every missing check fails closed; no partial state is reported as ready.283. `mode=run` and `mode=resume` read `program.md` from the checkout and follow it inside that checkout. Before any unattended sequence, hold with `🔒 阿舟暂停这一项` until the user confirms the GPU hours and disk cost. Results stay in the checkout; this skill never pushes, publishes, or copies results into any repository.294. `mode=report` aggregates experiment results that already exist in the checkout. Conversation excerpts, machine paths, and other raw evidence stay out of any committed surface.305. End with a receipt containing `schema`, `status`, `current_truth`, `artifacts`, `verification`, `holds`, `next_action`, and `learning_signal`:3132```text33## 🦊 阿舟 · Autoresearch receipt34- schema: autoresearch.receipt.v135- status: pass | fail | hold36- current_truth: <one sentence the checks actually prove>37- artifacts: <paths inside the user checkout>38- verification: <comma-separated check ids>39- holds: <none|fact>40- next_action: <one concrete step>41- learning_signal: <none|one line>42```4344If no pinned checkout is available, stop with `status=hold` and request one explicit checkout path.
TeFuirnever