expermimental build

[ychhabria]

Summary

• tune Codex file-search defaults from a realistic quality/performance eval loop
• raise the default result budget to 100 across the shared file-search crate and the TUI/app-server entrypoints
• set the shared file-search default worker count to 8

Evaluation

• built a realistic search scenario set centered on Codex code navigation and MCP-heavy use cases
• corpus roots: codex, openai, and docs
• exact-target set: 44 search cases across MCP/tool lookup, approvals/sandbox prompts, protocol/schema lookup, OAuth/MCP client files, connector catalog/docs, and large-repo application files
• broad probes: mcp, approval, sandbox, oauth, connector, github, and slack
• tried engine-ranking changes and rerank formulas in parallel; they either regressed quality or failed to reproduce end-to-end, so they are not included in this PR
• ran a ten-way config sweep: l50_t2, l50_t4, l50_t8, l75_t2, l75_t4, l75_t8, l100_t2, l100_t4, l100_t8, l150_t4

Among the top-quality tier, l100_t8 was the fastest setting:

• MRR = 0.4445
• hit@1 = 0.3864
• hit@5 = 0.5000
• hit@10 = 0.5682
• mean latency = 3642.54 ms
• median latency = 745.91 ms

Other tied top-quality configs were slower:

• l75_t4 mean 3999.77 ms
• l100_t4 mean 4014.52 ms
• l50_t8 mean 4414.51 ms
• l100_t2 mean 4688.22 ms

Changes kept

• codex-file-search default limit = 100
• codex-file-search default threads = 8
• app-server fuzzy search MATCH_LIMIT = 100
• app-server fuzzy search MAX_THREADS = 8
• TUI and TUI app-server FILE_SEARCH_LIMIT = 100

Verification

• git diff --check
• cargo test -p codex-file-search (15 passed; 0 failed)
• GitHub CI started successfully; early code-related checks are green and the remaining matrix is pending

Notes

• this is performance-only work; no UX/search-presentation changes are included
• raw eval artifacts were saved locally under ~/repos/docs/codex

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}