v2.1.0
The measurement release. OpenWolf now proves its numbers instead of asserting them.
Measurement (ground truth)
- Project-wide measured usage:
openwolf reportand the daemon scan every harness transcript for the project, subagent sidechains and headless runs included, deduplicated by message and request id, per model. - Injection accounting: every digest, hint, warning, and reminder OpenWolf injects is counted per source and shown as overhead next to the savings it claims.
- A/B benchmark harness (repository only): OpenWolf vs bare clones over a fixed task set, medians, cache dimensions reported separately. Methodology in the README.
Index quality
- Tree-sitter symbol extraction on scan: exact AST line ranges and nested methods for ts/tsx/js, python, go, rust, java, ruby, php, plus signature skeletons for large files. Hooks keep the dependency-free regex fast path; wasm failures fall back silently.
- Import-graph importance (PageRank) ranks hints and search results.
openwolf find <query>: ranked symbol/file shortlist from the index, ~1k token output cap.
Platform alignment (Claude Code)
- CLAUDE.md becomes a lean stub; the operating protocol ships as a Claude Code skill loaded on demand. Legacy snippets are swapped byte-identically; customized files are never touched.
- cerebrum syncs into Claude Code auto-memory; the session digest stops double-injecting Do-Not-Repeat on Claude. cerebrum.md stays canonical for other agents.
/handoffregenerates .wolf/STATUS.md on demand; the STATUS nag is gone.- Bash output filter (suggest mode): a once-per-session recipe to cap verbose test/build output. Rewrite mode stays inert: OpenWolf will not auto-approve tool calls.
- Context-health audit endpoint and dashboard card.
Buglog
- SQLite FTS (node:sqlite, Node 22.5+) keyed by normalized error signature powers
openwolf bug searchand the pre-write hook's cross-file fix recall, with fallback on older Node.
Run openwolf update to migrate existing projects (backup taken first).