v1.3.0

What's Changed

• Skill audit: freshness, pagination, graph-coverage honesty + vec_cache by @denfry in #6
• Audit follow-ups and release 1.3.0 by @denfry in #7

Full Changelog: v1.2.1...v1.3.0

v1.0.0

codebase-index 1.0.0

First stable release. The headline: multi-language symbol extraction now actually works, and retrieval is symbol-aware.

Fixed

• Multi-language tree-sitter symbol extraction. Previously a repo of 303 Java files produced 0 symbols, silently disabling symbol / refs / impact. Java now yields 3,543 symbols; Go / Rust / C / C++ / C# / Ruby / PHP / Kotlin plus a Tier-B generic path are covered.

Added

• Symbol-aware retrieval ranking — candidates scored by how many query terms their camelCase/underscore-split name covers. recall@3 against objective ground truth improved 20% -> 70% (a disciplined grep agent scores 40%) while using ~13x fewer tokens to answer.
• Parse guardrails (parse_failed / treesitter_zero_symbols counters) with doctor reporting, locking symbol extraction against silent regression.
• Multi-CLI installer for Claude Code / Codex / OpenCode.
• Honest benchmark harness (tests/benchmark_honest.py) comparing the index against a no-skill grep agent on a real repository.

See CHANGELOG.md for full details.