Skip to content

v2.11.0

Choose a tag to compare

@VelimirMueller VelimirMueller released this 11 Oct 03:50
· 137 commits to develop since this release
Immutable release. Only release title and notes can be modified.
v2.11.0
3fcf2ec

Added

  • Dashboard Benchmark page shows real ccc-bench runs: the latest run per kind with its date and age, the run list, and per-task detail. New /api/bench/index and /api/bench/run/:id read from CODE_CONTEXT_BENCH_DIR, then <db dir>/bench/, then origin/bench-results in the package's own git checkout; otherwise an explicit empty state. The June simulated data stays as a "Simulated (2026-06)" tab. (#114)
  • CODE_CONTEXT_TOOLSETS=code: a profile that lists only the 11 code-index tools (index_directory, find_symbol, get_file_context, search_files, get_changes, query, execute, health, set_description, set_directory_description, set_change_reason). Tool definitions drop from about 12,300 to about 1,300 tokens per model call. For sessions that read code but never run the sprint process. all still wins. The default does not change. (#109)
  • Fair benchmark arms: task prompts name no tools, a third cc-guided arm adds the "use code-context first" rule to the system prompt, both code-context arms run CODE_CONTEXT_TOOLSETS=code, and sessions record the cost the CLI reports. (#111)

Changed

  • bench:publish copies the merged data/ to CODE_CONTEXT_BENCH_DIR when set (never in Actions), so the dashboard gets new runs without a manual copy. (#114)
  • README token claim quotes the fair agent runs; the 46.3 % figure is labelled as simulated in the stats and bench panels. The site notes that agent runs before 2026-10-11 used grep-steering prompts. (#113)

Fixed

  • Repo-relative paths work. get_file_context and set_description accept src/server/schema.ts as well as absolute paths: relative to the server's working directory first, then a unique suffix match in the index. With several repos in one database, the match under the working directory wins, otherwise the tool lists the candidates. Before, a relative path answered "not in index", and agents fell back to grep. (#109)
  • A provider error or a session with 0 model calls voids a live benchmark run: no run file, the artefacts stay for diagnosis. (#112)
  • The dashboard build clears old hashed chunks in dist/dashboard/assets; the 2.10.0 pack held three copies of each. (#110)