agents.md

Rules for coding agents (Claude Code, Codex and others), kept in two parts.

Part 1: general files (repository root)

These are repository-independent and reusable as they are:

• GENERAL.md: roles and spawning, delegation, decision rights and escalation, briefs and reports, tests and review, handover between sessions, safe git and process habits, and the general skills.
• USER.md: the contract between the user and the agents (the user’s part and the options).
• skills//SKILL.md: the general skills (procedures an agent loads when a task matches).

Part 2: llmservingsim/

The rule set of one concrete repository (LLMServingSim), as a worked example of how the general files are adopted: its AGENTS.md entry file, the role files under agents/, the agent definitions and hooks under .claude/, and the generator and publishing scripts.

Adopting the general files in another repository

1. Copy GENERAL.md and USER.md into the repository, and the skills into .claude/skills/.
2. Write the repository’s own AGENTS.md, with a term map in the style of section 0.4 of llmservingsim/AGENTS.md that maps the general terms (handoff file, decisions list, task trail, …) to the repository’s own files.
3. Mark the sections that are not tied to the repository with `` on the line after their heading, and write them in general terms.
4. Run a generator like llmservingsim/agents/gen_general.py to rebuild GENERAL.md from the marked sections, and a publisher like llmservingsim/agents/publish_agentsmd.py to sync this layout.

Provenance

The root files are generated from LLMServingSim and are edited there, never in this repository: change the marked section in the source repository, regenerate, and run python3 agents/publish_agentsmd.py.

Source commit of the last sync: 6734dd35