You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
[0.1.11] - 2026-06-27
Changed
Route-by-task steering (grove init CLAUDE.md block and the grove
skill's SKILL.md). Both surfaces previously framed code navigation as an
"INVARIANT — grove or it's a steering violation", relegating grep/rg/read/cat/sed to fallbacks allowed only after grove was
tried. With grove's current 7-tool surface (no text-search tool yet), that
over-rigid framing pushed the model into costly one-symbol-at-a-time source
fan-outs for things the shell does cheaply, and gave no guidance for
text/non-code/quick-fact work or for combining grove with the shell. The
steering now routes by task: grove for named symbols and structural
relationships (where-defined, who-calls, what's-in-a-file, how-a-dir-connects,
post-edit check), the shell for text / non-code files / quick facts ("the
right tool, not a fallback"), and an explicit combine path (grep a
literal's line → definition --at; outline → bounded read; map/symbols
→ grep a constant inside). The useful procedure (outline→source chains, stable
symbol-ids, map breadth control, the shape-slice, profile gate, recovery,
setup) is preserved.
Notes
Validated in the nav-3way testbench: on L4-grove-redis the reworked
steering cut context ~41% (560k → 329k tokens) with no loss of answer
completeness — the model combined 20 grove calls with 3 greps + 2 bounded
reads instead of 36 source calls.