Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/midnightdarling/collate/agents-mdgit clone --depth 1 https://github.com/MidnightDarling/collateWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.08385 | $0.08385 |
| Opus 5 | $0.04193 | $0.04193 |
| Sonnet 5 | $0.01677 | $0.01677 |
| Haiku 4.5 | $0.00839 | $0.00839 |
Grade B, and why
collate AGENTS.md scanned grade B with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Reads agent configuration directoriesmediumAgent snooping
.claude/, .codex/, .gemini/ hold keys, settings and other credentials a mod has no legitimate need for.
- `~/.codex/config.toml` 里的 `command_attribution` 必须保持 `false` How it starts
The opening of the file, as written. The whole thing — 560 lines — stays where its author put it; the contents beside it link to each section on GitHub.
AGENTS.md
面向任何运行此 pipeline 的 agent 运行时(Claude Code / Codex CLI / Cursor / Gemini CLI / Hermes agents)。描述完整的自动化工作流、每个 skill 的调用契约、agent 自主决策点、失败处理。人类只负责提供 PDF、接收最终产物与审计报告。
范围说明:本文档覆盖 8 步 pipeline 层。仓库还包含 7 个阅读层 skill(
xray-paper/paper-summary/chunqiu/kaozheng/prometheus/real-thesis/constellatio),它们独立于 pipeline,使用各自skills/<name>/SKILL.md作为调用契约,本文档不重复展开。
总原则
- 端到端自主:从 PDF 到
.docx+.mp.html由 agent 独立完成。不要把中间步骤暴露给人类确认,除非失败降级到人类决策(见「失败处理」)。 - 保留可追溯性:每一步的中间产物全部落盘,不清理。diff-review 的审计报告是交付的一部分,不是调试工具。
- 按 checklist 执行 proofread:结果是机器可读的结构化清单(A/B/C 分级),由 agent 消化后修改
raw.md → final.md。不做超出清单的结构性改写。 - 不替换学术判断:C 类(存疑待考)条目不要替人做决定;在
final.md里保留原文 + 脚注标注,由读者判断。 - 失败要显性:引擎调用失败、依赖缺失、输入不合规 → 立即终止并回传结构化错误,不要静默兜底。
Pipeline 全景
Human ──► PDF
│
▼
┌──────────────────────────────────────────────────┐
│ 1. prep-scan 产 <ws>.ocr/prep/cleaned.pdf │
│ 2. visual-preview 产 previews/visual-prep.html │
│ 3. ocr-run 产 raw.md + meta.json + assets │
│ 4. proofread 产 review/raw.review.md │
│ 5. (agent 按 checklist 写 final.md) │
│ 6. diff-review 产 previews/diff-review.html │
│ 7. to-docx 产 output/<title>_…_final.docx │
│ 8. mp-format 产 output/<title>_…_wechat.* │
└──────────────────────────────────────────────────┘
│
▼
Human ◄── <ws>.ocr/README.md (指向 output/ 与 previews/)
整个 .ocr/ 工作区自描述:根目录 README.md 由 scripts/workspace_readme.py 在每个 skill 结束时刷新,给人类一个清晰的入口。
一条命令模式:
- shell / CI / 本地编排:
python3 scripts/run_full_pipeline.py --pdf <input.pdf> - agent runtime:以
agents/ocr-pipeline-operator.md作为单一入口;它负责调用上面的总编排脚本、在raw.md就位后起historical-proofreadersubagent、再重入总编排脚本完成final.md/diff-review/docx/wechat。
每一步的详细契约见下。
1. setup(首次/环境校验)
何时调用:新环境、依赖缺失、OCR 凭证未配置或怀疑失效。稳定环境下不重复调用。
做什么:按 skills/setup/SKILL.md 检查 Python 依赖、poppler、mineru CLI 可用性;探活 OCR_ENGINE 指向的引擎(本地 mineru 跑一个 --help、云 API 或百度跑一次空探测)。
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 560 lines · 8,385 tokens per session scan B 2cd12b701f4d
collate AGENTS.md is an instructions file published in the GitHub repository MidnightDarling/collate (6 stars, last pushed 3mo ago), licensed Apache-2.0. It adds 8,385 tokens to every session, about $0.0419 per session on Opus 5. A static security scan graded it B with 1 finding (reads agent configuration directories). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
codedb AGENTS.md
AGENTS.md instructions for justrach/codedb, covering codedb agent guidelines, what codedb is (and isn't), review guidelines, pre-merge verification and security-sensitive areas.
honey-for-devs copilot-instructions.md
Instructions for Green-PT/honey-for-devs, covering honey (i shrunk the ai), intensity, lever 1 — minimum code that needs to exist, never cut (lazy ≠ broken) and lever 2 — say less about it.
Puppetmaster AGENTS.md
Instructions for professorpalmer/Puppetmaster, covering agents.md, first: are you a puppetmaster worker?, default subagent runtime: puppetmaster, when to start a puppetmaster swarm (do this by default) and match the verb to the task shape (single feature ≠ swarm).
claude-code-blueprint CLAUDE.md
Instructions for faizkhairi/claude-code-blueprint, covering global claude behavior rules, verify-after-complete (mandatory), what to verify depends on what was built, verification mindset and after verification.
claude-code-blueprint AGENTS.md
Instructions for faizkhairi/claude-code-blueprint, covering agents.md: repository orientation for ai assistants, what this repo is, what this repo is not, repository layout and how to help the user.
agent-native-setup AGENTS.md
Instructions for luca-mastrostefano/agent-native-setup, covering agent-native-setup — agent contract, navigation and command surface.