mirror of
https://github.com/colbymchenry/codegraph.git
synced 2026-09-19 07:34:57 +08:00
Colby McHenry
50401a6deb
The n=4 README A/B showed explore had re-inflated to 35–42K on big repos — which (a) the host externalizes to a file the agent Reads back, (b) costs cache-writes (cost went WITH-more-expensive on vscode/okhttp/gin), and (c) spreads thin across 12 files, dropping the agent-named flow methods. Ground truth from the WITHOUT arm: the agent natively reads ~6–9 files as ~100-line windows (median 100, 51–250 dominant), located by signature greps. Shape explore to match: - Cap every tier's output at the ~24K inline ceiling (was 28/35/38K) with a 25K absolute hard-stop — a bigger single response just externalizes. Repo size now scales the CALL budget (getExploreBudget), not the response size. - One ~150–250 line WINDOW per file: per-symbol bodyCap 2x->1.5x and the spine is windowed too (a big-spine god-file like tokio worker.rs no longer eats the whole response and starves the co-flow file harness.rs); central whole-file 4x->1.5x / 400->280 lines. Concentration onto the flow then emerges from the cap + the named-file-first sort dropping peripheral files. - Explore's named-symbol injection now enumerates overloads via the direct name index (getNodesByName), not FTS — so a 50+-overload name (tokio `poll`) surfaces the wanted def (`Harness::poll`) for the type-token bias to pick, same fix already applied to codegraph_node. Probe-validated: vscode explore 35K->16K; tokio now returns worker.rs + harness.rs(poll) + mod.rs windows in 24.5K (harness.rs was dropped before); all 7 README repos <=~25K. Full suite green (1112). Agent A/B re-bench pending. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>