v4.14.1

Bug Fixes

• Fix indexing crashes on deep or non-ASCII sources

What's Changed

• Fix indexing crashes on deep or non-ASCII sources by @deepskyblue86 in #1

New Contributors

• @deepskyblue86 made their first contribution in #1

Full Changelog: v4.14.0...v4.14.1

v4.14.0

• Add crates.io publish metadata (description, license, repository, keywords, categories)
• Exclude bench/, specs/, homebrew/, docs/ from crate package

Install via cargo:

cargo install wonk

Full Changelog: v4.13.0...v4.14.0

v4.13.0

What's new

• Agent benchmarks: New benchmark suite measuring Claude Code token consumption with vs without wonk across 25 code-understanding tasks on 5 real-world repos. Results: 37.4% total token reduction with quality maintained (0.90 vs 0.91 baseline).
• Python class fields: Annotated assignments in classes (e.g. pydantic title: str) are now indexed as variables.
• Granular Python imports: from foo import bar, baz now creates separate references for bar and baz instead of one blob.

Full benchmark report: bench/results/agent_report.md

Full Changelog: v4.12.1...v4.13.0

v4.12.1

Fix

• Fix cargo fmt formatting in mcp.rs that caused CI failure

Full Changelog: v4.12.0...v4.12.1

v4.12.0

What's New

• File filters everywhere: --file/-f on search, ref, show; --reference-file, --callers-file, --callees-file, --destination-file for disambiguation in call graph commands
• Pagination: --page N for paginated budget output across all commands
• Test exclusion: --include-tests global flag (tests excluded by default)
• Ref output modes: --output files for compact unique-file-paths-only output
• Sym limit: --limit N to cap result count
• Show file-only mode: wonk show --file path --shallow without requiring a symbol name
• Skip embed: --skip-embed on wonk update for structural-only index refreshes
• Incremental MCP init: wonk_init now checks wonk_version and uses the incremental path when the index is current, avoiding full re-embedding
• Reduced Ollama activity: SessionStart hook uses --skip-embed to prevent unnecessary embedding work on every session start

Full Changelog: v4.11.0...v4.12.0

v4.11.0

Qualified Paths, Cfg Macro Expansion, Subclass Refs, Auto-Regex

Changes

• Qualified name resolution — wonk_sym tokio::spawn and wonk_show tokio::spawn now parse :: paths, using the prefix as a file-path hint for Rust module navigation
• Rust cfg macro pre-expansion — cfg_rt!{}, cfg_if!{} etc. are stripped before tree-sitter parsing so inner items (functions, structs) are properly indexed
• Subclass/implementor refs — wonk_ref now includes types that implement/extend the target via type_edges, shown with kind=subclass
• Auto-regex detection — search queries containing backslash escapes (\w, \d) or character classes ([A-Z]) automatically enable regex mode
• --file filter on wonk_sym — restrict symbol lookup to files matching a path substring
• Test/bench path deprioritization — symbols from test, bench, mock, and example directories are sorted to the end of results
• Spec updates — added PRD-REF-REQ-004 (output=files), synced architecture tool manifest (18→22 tools), added v4.10.0 changelog

Full Changelog: v4.10.0...v4.11.0

Full Changelog: v4.10.0...v4.11.0

v4.10.0

MCP Token Efficiency Fixes

Benchmark-driven improvements to reduce token waste when AI agents use wonk MCP tools.

Changes

• Soft truncation hints — "Showing N of M matches" instead of "Increase budget" (prevents agent distrust → Read fallback)
• wonk_search clarified as keyword/regex — description now explicitly says "Do NOT use natural language" and directs to wonk_ask
• RRF fusion moved to wonk_ask — single smart search entry point that blends structural + semantic results; NL queries gracefully degrade to pure semantic
• wonk_ref output=files — output="files" returns just unique file paths (like grep --files-with-matches), 8× less verbose for "list files" queries
• .d.ts deprioritization in wonk_sym — type definition files pushed to end of results so actual source appears first
• TS interface shallow mode fixed — method_signature and property_signature now indexed with scope, so shallow mode shows member signatures for interfaces
• wonk_show description updated — "Returns complete symbol source code — you do not need to Read the same file afterward"
• wonk_ask added to MCP instructions — discoverable by agents

Full Changelog: v4.9.0...v4.10.0

v4.9.0

Auto-shallow fallback in wonk_show on budget overflow.

When a container type's full body exceeds the token budget, automatically retry in shallow mode (signature + child signatures) instead of returning empty results. Results include auto_shallow: true so agents know the body was downgraded. Hints now suggest shallow:true instead of Read.

Full Changelog: v4.8.0...v4.9.0

v4.8.0

What's New

• Doc comment extraction — Tree-sitter extracts doc comments for all 12 supported languages (Rust ///, Go //, Python docstrings, JSDoc /**, etc.), truncated to 200 chars, stored in the index and surfaced in wonk summary output.

• Simplified summary detail levels — Replaced rich/light/symbols with two levels:

  • outline (new default): top-level types + functions with doc comments, no methods — compact and token-efficient
  • rich: all symbols in tree hierarchy including scoped methods
  • Backward compat: light and symbols still parse as outline

• Auto LLM descriptions — wonk summary now automatically attempts LLM description generation (no more --semantic flag). Gracefully degrades when Ollama is unreachable. New directory overview prompt generates descriptions from child summaries.

• MCP wonk_summary simplified — Removed detail, semantic, and tree parameters. Always uses outline mode for optimal token efficiency.

Breaking Changes

• --semantic and --tree CLI flags removed from wonk summary
• MCP wonk_summary no longer accepts detail, semantic, or tree parameters
• Default detail level changed from rich to outline

Full Changelog: v4.7.0...v4.8.0

v4.7.0

Full Changelog: v4.6.0...v4.7.0