Semantic reranking of tool_search results

What it does

After the literal match, reorders tool_search's candidates with one Jev choice, with a "none fits" answer. Only tool descriptions and the search words leave the machine, never the user's files.

Source

Outside beads: one of four proposals for using Jev judgments in ganja, carried into the map by its destination (#33, #34, #35).

File and line references in this issue are to main as of 75a7a92.

Against the three criteria

• Direction (README): The README's only word on tools is "permission-gated tools" (README.md:6); it says nothing about MCP or Jev.
• Today's terminal users: A user with many deferred tools gets the right tool found by meaning rather than by spelling. Matching is literal today (crates/ganja-tool/src/deferral.rs:439, crates/ganja-tool/src/deferral.rs:459).
• Evidence: The vendor measured a similar task (skill mis-loading fell from 16.8% to 7.3%); nothing is measured on ganja. Measuring accuracy is part of the work.

Depends on

Blocked by the judge seam and the [evaluate] table of #59. #85 reopens once this has landed.

Verdict

• Verdict: do
• Size: M — one plan and one decision number.
• Exception to the probe rule: yes. #36 says an idea that needs a measurement or a probe first is not do; this one is an explicit exception: measuring accuracy is part of the work.
• Date: 2026-09-19
• Verdict ticket: #43

Theme: #55

Start with crates/ganja-tool/src/deferral.rs:439 and :459 to understand today's literal matching, then read the judge seam and [evaluate] table from #59. The work is done when tool_search semantically reranks candidates using a Jev judgment with a "none fits" option, while an accuracy measurement demonstrates the result.