返回 DeepSeek-Reasonix
use-capability-replay.md
根目录 / docs / eval / use-capability-replay.md
1 # `use_capability` replay evaluation
2
3 This optional paired evaluation compares the cache-stable `use_capability`
4 proxy with a baseline that expands MCP tools into the provider-visible schema.
5 It is a diagnostic benchmark, not a Stable release gate. Live model runs are
6 optional and must use disposable Reasonix homes.
7
8 ## What to measure
9
10 For the same task set, run each task twice:
11
12 1. **Proxy (default):** `use_capability` only. Shared Host plus disk schema
13 cache.
14 2. **Baseline:** a throwaway config that still expands MCP tools into the
15 provider request. Native Tool Search must stay off.
16
17 Record `tools/list` count, first-token latency, and cache-hit tokens. Do not
18 upload prompts, secrets, tool arguments, or workspace paths.
19
20 ## Procedure
21
22 1. Use disposable `REASONIX_HOME` and `REASONIX_CACHE_HOME` directories.
23 2. Pick a representative task set that needs MCP discovery followed by a call.
24 3. Run proxy and baseline for each task with the same model, effort, workspace,
25 skills, agents, and MCP configuration.
26 4. Record content-free pairs matching
27 `internal/eval/replay/testdata/paired_runs.json`.
28 5. Exercise the median helper:
29
30 ```bash
31 go test ./internal/eval/replay/ -run TestMedianReportFivePairedRuns
32 ```
33
34 The repository fixture is synthetic and proves only the median helper. Teams
35 may replace its numbers with live observations for performance analysis, but no
36 paired-run dataset or threshold result is required for Stable publication.
37
38 Native first-party Tool Search stays default-off and is not part of this eval.
39
39 lines MARKDOWN