HomerandClaude Opus 5.5
d09d51cfec
context.md §8: every assistant message of the current turn (after the last message that started a turn — a step-boundary delivery does not) carries its reasoning back, as reasoning_content in OpenAI chat and as thinking blocks in Anthropic; earlier turns carry none, which the common templates render identically. §11: an agent session's prompt names no absolute machine path: the environment block names the project and says paths are relative to its root (loop.json context.env), files outside the project under home as ~/…; relative paths resolve against the project root, absolute ones still accepted. agents-missing.md and init.md no longer name the root. New conformance areas: reasoning, paths. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1.6 KiB
Create or update AGENTS.md at the project root: the file every later session here starts from. Every line should answer "would an agent likely get this wrong without being told?" — if not, leave it out.
{{focus}}
Investigate first, highest value first: README, the manifests and lockfiles, build, test, lint and type-check configuration, CI workflows, and any instruction files already there (AGENTS.md, CLAUDE.md, .cursorrules, .github/copilot-instructions.md). Read a few representative source files only if the structure is still unclear. Where the documentation and the scripts disagree, trust the scripts.
Write down what is hard-earned:
- the exact commands to build, run and test — and how to run one test;
- the order checks must run in, where it matters;
- where the real entry points are, and what the main directories are for;
- generated code, migrations, special environments, anything that is not the default for the language;
- conventions this project keeps that differ from the usual ones.
Leave out generic advice, long file trees, and anything you could not verify. Short sections and bullets; a simple project gets a short file.
If the file is already there, improve it in place: keep what is still true, remove what is stale. If it says nothing about how git is used here, ask the user once, with ask_user, and write their answer under a "## Git" heading. Ask nothing the repository already answers.