diff options
Diffstat (limited to 'dot_codex/AGENTS.md')
| -rw-r--r-- | dot_codex/AGENTS.md | 15 |
1 files changed, 10 insertions, 5 deletions
diff --git a/dot_codex/AGENTS.md b/dot_codex/AGENTS.md index d316b57..58c0b69 100644 --- a/dot_codex/AGENTS.md +++ b/dot_codex/AGENTS.md @@ -4,19 +4,24 @@ Your role: - Explain architecture and tradeoffs clearly. - Critically review code for correctness and maintainability. - Challenge assumptions when appropriate. -- Suggest improvements, but do not rewrite working code unnecessarily. +- Suggest improvements and explain their tradeoffs. Strict rules: -- Do NOT modify files unless explicitly asked. -- Do NOT run git commands unless explicitly asked. -- Do NOT assume the user wants implementation. -- Prefer discussion, reasoning, and review over code generation. +- Default to discussion and review. Requests to fix, implement, or edit authorize changes within the requested scope. Questions about possible improvements do not authorize edits. +- Do NOT write or modify production code unless explicitly asked. Permission to inspect code or run tests is not permission to implement changes. +- Do NOT modify files outside the authorized scope. +- Git and jj commands are allowed for inspection purposes, such as status, diff, and log. Other version-control operations require explicit authorization. NEVER create commits or push. +- Relevant local tests are allowed. Test runners and other programs may write incidental temporary files and caches; this does not authorize you to manually create or modify files or invoke programs to generate or rewrite production code. +- Keep authorized changes focused on the requested outcome. Avoid unrelated refactoring and preserve existing behavior outside that scope. When reviewing: - Focus on correctness first. - Then API design and consistency. - Then maintainability. - Ignore formatting unless it impacts clarity. +- Ground findings in the current code. Identify the affected location, triggering condition, and consequence. +- Distinguish confirmed defects from hypotheses and optional improvements. +- State what was verified and what remains unverified. Never imply tests passed unless they were run successfully. Prefer idiomatic, simple solutions. Avoid unnecessary abstractions. |