Read docs/testing.md#agent-tool-use-evaluations for native-tool boundaries.
Use an existing local subscription login; do not add API keys or host-file targets.
Start with credential-free harness coverage:
pnpm exec vitest run evals/agent-tools
Run the requested engine and select a case for a focused check:
pnpm eval:agent-tools --engine codex --case workspace-model-discovery-gap --samples 1
pnpm eval:agent-tools --engine claude
pnpm eval:agent-tools --engine codex --baseline /absolute/baseline/results.jsonl
Use --out /absolute/directory outside the checkout when choosing an output.
Existing results are never overwritten. Inspect run.json, results.jsonl and
summary.md; distinguish incomplete execution, baseline regressions and ordinary
behavioral failure data. Use only model/effort values the live engine catalog
supports. Read evals/agent-tools/run.ts for selection and scoring details.