Commits on Sep 23, 2026

1. fix(explore): weight query terms by reach so a prose question's rare … …

   …word is not drowned by its common ones
   
   On a 136k-node index, `codegraph_explore "how the task.eager setting
   influences subagent delegation in the system prompt"` rendered four wrong
   files and none of the answer. Three compounding causes, all in ranking:
   
   1. findRelevantContext's text channel searches each term on its own and
      merges by raw score. A flat exact-name bonus (+80) lifts a property
      called `system` (700+ nodes match `system*`) to ~100, the same as the
      `EAGER` enum, while `eagerTasks` (44 nodes match `eager*`) scores 45 —
      so the query's one discriminating word never survives the merged cut.
      Each term's hits are now weighted by how much of its match set the
      channel can see: `min(1, max(50, sqrt(N)) / reach)`. Absolute reach,
      not idf: on a 270-node payroll fixture `cycle` reaches 22% of nodes
      and IS the hand-written workflow the query asks for; a relative
      measure discounts it under the generated CRUD (explore-allocation-1500
      pins this). `CODEGRAPH_TERM_RARITY_SHARPNESS=0` restores the old rank.
   
   2. handleExplore's named-symbol seeding lets a bare word tier a file when
      a sibling query token is also a symbol there. Any file large enough —
      a 15k-line session class, a generated protobuf, a platform .d.ts —
      declares a `task`, a `model` and a `role`, so on "how subagent model
      routing picks a model role from the task difficulty" the session class
      was tiered above the routing module that scored 102 to its 50. The
      corroborating sibling must now be a precise token or a name declared
      in at most max(8, 0.25% of files) files (`countFilesDeclaringName`,
      via idx_nodes_lower_name).
   
   3. The `+e` stem the `-ing`/`-er` rules emit (`setting` -> `sette`) is not
      a substring of its base, so the co-occurrence re-rank counted it as a
      second concept and doubled `setText` on a query about settings —
      masked until (1) removed the noise around it. Root grouping is now
      `groupTermsByRoot` in query-utils (shared prefix of max(4, len-1)).
   
   Eight prose questions over that index, ground truth by grep:
     answer rendered 7/8 -> 8/8, answer first 4/8 -> 6/8,
     off-topic files 18 of 53 -> 6 of 44.
   
   Tests: context-term-rarity (weighting, both directions via env),
   explore-ambient-corroboration (guard, both directions via env + diag
   sidecar), query-term-root-groups (grouping contract). Full engine suite
   diffed against pristine HEAD: symmetric difference is only Windows
   spawn/sync flakes that flip both ways and pass on retry; the two
   pre-existing CRLF failures (explore-factory-closure, explore-oversize-
   member) are unchanged. scripts/probe-*.mjs are the tuning harness.

   

   kingkillery committed Sep 23, 2026
   Configuration menu

   • View commit details
   • Copy full SHA for da33657
   • Browse repository at this point

   Copy the full SHA

   da33657 View commit details

   Browse the repository at this point in the history