一键导入
auto-paper
Internal Harness instruction source for auto-paper. Route through visible Harness aliases or hook contracts instead of invoking directly.
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。
菜单
Internal Harness instruction source for auto-paper. Route through visible Harness aliases or hook contracts instead of invoking directly.
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
Internal Harness instruction source for auto-iterate-goal. Route through visible Harness aliases or hook contracts instead of invoking directly.
Internal Harness instruction source for evaluate. Route through visible Harness aliases or hook contracts instead of invoking directly.
Internal Harness instruction source for iterate. Route through visible Harness aliases or hook contracts instead of invoking directly.
Internal Harness instruction source for orchestrator. Route through visible Harness aliases or hook contracts instead of invoking directly.
Visible Harness write entry. Use for manuscript writing, citation-supported paper work, final documentation, README hardening, and GitHub Pages preparation.
Visible Harness write entry. Use for manuscript writing, citation-supported paper work, final documentation, README hardening, and GitHub Pages preparation.
| name | auto-paper |
| description | Internal Harness instruction source for auto-paper. Route through visible Harness aliases or hook contracts instead of invoking directly. |
Use this as the auto-paper orchestrator, not a polishing tool. It routes phases, checks artifacts, and keeps paper state separate from auto-iterate state.
docs/30_evidence/Experiment_Evidence_Index.{json,md} and artifact paths
named there. Direct iteration_log.json reads are weak intent signals; paper
claims require reports, configs, logs, metrics, or run artifacts.docs/context/experiments.md when present for Research Wiki context,
stable observations, run requests, assurance gaps, and open questions.
Legacy docs/45_discoveries/Research_Wiki.md is a fallback before
dynamic-context-v2 migration; neither file is approval.iteration_log.json or .auto_iterate/**; auto-paper state
uses auto_paper_log.json, .auto_paper/, and
auto_paper_output/<paper_id>/.Before the loop, establish paper_id, workflow, venue/target, draft path, TeX
roots, bibliography paths, figures, materials, references, artifact dir,
optional experiment evidence index, output language, citation target count,
figure/table requirement status, compile command, human gate policy, and
forbidden directions in auto_paper_output/<paper_id>/config.yaml or an intake
artifact. If missing, route to $auto-paper-intake or return USER_GATE.
After Grill establishes an Automation Policy, paper phases auto-proceed within
that policy. Use USER_GATE only for missing operator intent, explicit approval
tools, or irreversible external submit.
Use the controller for resumable routing when the operator requests an automatic paper chain:
tooling/auto_paper/scripts/auto_paper_ctl.sh start \
--paper-id <paper_id> \
--artifact-dir auto_paper_output/<paper_id>
The controller checks artifacts and writes .auto_paper/; it does not invent
manuscript content.
Run in order unless an audit routes back: $auto-paper-research,
$auto-paper-argument, $auto-paper-citation, $auto-paper-layout,
$auto-paper-patch, $auto-paper-harden.
Optional branches: $auto-paper-response, $auto-paper-data,
$auto-paper-figure. Enter $auto-paper-figure when the operator asks for
figure work, TeX or source files mention needed figures/tables, or harden finds
figure/caption risk.
Durable decisions live under auto_paper_output/<paper_id>/.
config.yaml, source_index.md, tex_inventory.json,
intake_report.md, experiment_source_map.md, optional
figure_requirement_scan.mdrun_request_register.{json,md} when missing experiments are neededUse references/artifact-contract.md for identifiers and stale-artifact policy.
Do not patch around missing upstream artifacts. Every audit finding names
root_cause, owning_phase, required_artifact, fix_action, and
downstream_risk.
Decisions: NEXT_UNIT, REWORK_RESEARCH, REWORK_ARGUMENT,
REWORK_CITATION, REWORK_LAYOUT, REWORK_PATCH, RUN_REQUEST,
USER_GATE, COMPLETE, ABORT.
RUN_REQUEST means writing found missing experiment evidence for $run. Record
the blocking claim, needed evidence, minimum artifacts, suggested prompt, and
acceptance check in run_request_register.{json,md}.
When a draft introduces, narrows, or removes a result claim, record Claim Delta
Evidence in the phase artifact or Gate ledger. Unsupported claims become
RUN_REQUEST, not a default approval pause.
Load only active-phase references from references/: loop, auto-iterate
boundary, artifact contract, writing rationale, motivation thread, citation
support, and LaTeX source control.
Use deterministic helpers when available:
.agents/skills/auto-paper/scripts/figure_requirement_scan.py../../../.agents/references/research-supervision-patterns.md../../../.agents/references/research-supervision/README.mdFinal responses report artifact manifest, modified files, guard/compile results, remaining reviewer risks, and next human gates.