CLAUDE.md

Read AGENTS.md first — single source of truth. Everything below is Claude-Code-specific.

Slash commands (.claude/commands/)

/test-unit — make unit-test
/test-all — make test (long; benchmarks + notebooks too)
/format — make format + uv run mypy .
/typecheck — uv run mypy .
/build-docs — sources .env, runs uv run build-docs
/verify-model <model_name> — guided single-model verify_models.py run
/add-model-support <hf_repo> — checklist-driven new-adapter workflow
/task-complete — end-of-task gate: clean new comments, /format, make test-pr (unit + docstring + acceptance + integration), loop until clean. Skips notebook + benchmark tiers.

.claude/settings.json is the checked-in scaffold (hooks, env); ships with no permission allowlist.
First-time permission setup: every command prompts on first use. Either (1) run /fewer-permission-prompts to populate .claude/settings.local.json from your transcripts, or (2) approve with "Don't ask again." Both write to the gitignored local file.
.claude/settings.local.json, .claude/agents/, .claude/worktrees/ are gitignored — per-user.

AGENTS.md §10 — hard rules; load-bearing.
AGENTS.md §2 — the single model system (the Hooked* model classes were removed in 4.0).
tests/QUARANTINES.md — check before debugging any failing test. A skip with a matching reason is not your bug; a failure not on that list is real.
debugging_numerical_divergence.md — Bridge-vs-HF logit drift bisection.
compatibility_mode.md — bridge.enable_compatibility_mode() contract; read before adding tests that use it.
devtools/adapter_builder/ — autonomous adapter builder (contributor tooling; agent-teams mode needs Max, solo mode runs on any tier); manual path is /add-model-support.

good first issue · help wanted
verification-request — models awaiting verification; pick one that fits your machine, run /verify-model <model_id>.
Backfilling per-adapter unit tests in tests/unit/model_bridge/supported_architectures/ is high-leverage; copy a sibling adapter test.