# MCP tool-use eval

- Commit `3d525fab7d-dirty`, gpt-6.1-sol at xhigh effort, user instructions drop, 20 samples per case, template `codex.json.gz`
- T3 context: system text 7609 chars (~1902 tokens), tool descriptions 14905 chars (~3726 tokens), full tool schemas 49159 chars (~12290 tokens) across 71 tools.
- Pass: 100% of 20 samples.
- Cost: the agent is not priced. `vibeproxy` in `~/.codex/config.toml` may use a subscription or bill an API key per token; agent tokens are below.
- Agent tokens: input 157472, output 41535, cacheWrite 0, cacheRead 3656320.
- Shell simulator (gpt-6.1-terra) tokens: input 657108, output 38397, cacheWrite 0, cacheRead 97664 (API-price estimate $0.00).
- Code-mode tools use case mocks and a simulated machine. Native subagent prompts and asynchronous notifications are approximate.

## Cases

| Group | Pass |
| --- | --- |
| child-async-delegate-keeps-working | 100% of 20 |

## By kind

| Group | Pass |
| --- | --- |
| bad | 100% of 20 |

## By pattern

| Group | Pass |
| --- | --- |
| async delegate ends a child's turn with work left | 100% of 20 |