The repository guide for every coding agent is AGENTS.md; it is imported here so
Claude Code loads it in full. Edit conventions there, not in this file. This file
holds only what is specific to Claude Code.
@AGENTS.md
Pick the subagent model for the task; never run everything on the largest
model. Pass model explicitly on every Agent call:
haiku: search, file discovery, mechanical work (renames, checking docs against code, collecting test output).sonnet: routine implementation, functional and end-to-end testing, regression runs, convention surveys.opus: design, security and isolation reasoning, adversarial review, where a wrong answer is expensive.
The Codex CLI (@openai/codex) is installed and authenticated on some machines,
not on cloud sessions. Probe once per session with
command -v codex && codex login status; if it is missing or prints
Not logged in, use Claude subagents only, and do not install it, npx it or
retry. When it is authenticated, treat it as a first-class subagent and run it in
parallel with Claude subagents, each on its own slice, with the same brief
discipline and the same per-slice verification:
- Read-only slices (verification, surveys, review):
codex exec -s read-only "<brief>", andcodex exec -s read-only review --base main(or--uncommitted) for a diff review. A second model's independent read is the main value here, so route review and fact-check slices to Codex first. - Write slices (an isolated implementation or test-fixing task): give Codex
its own checkout, never the one this session is editing:
git worktree add ../<repo>-codex-<slice> <branch>, thencodex exec -s workspace-write -C ../<repo>-codex-<slice> "<brief>". Review its diff and run the repo checks before merging it into the branch. - Never pass
danger-full-access,--approve-for-meor--dangerously-bypass-approvals-and-sandboxhere.