v2.2.0

The reliability release: OpenWolf proves what it does.

Built on an empirical audit of 16 live projects (6,869 API records) that found a hook crashing silently for 3 weeks, self-reported counters drifting ~20x from transcript ground truth, and shipped files no agent ever used.

Added

• Hook health: per-hook heartbeats, crash capture, session-start self-test of hook imports, install verification on openwolf update (fails loudly), dashboard surfacing of failing hooks.
• Transcript-verified measurement: per-session verified counts of hook runs, failures, and injected context that actually entered the conversation, with a schema probe that degrades to labeled estimates if the transcript format drifts.
• Cache-invalidation attribution: openwolf report and the daemon name what broke the prompt cache (model switch, compaction, version change, cache expiry, unattributed) and how many tokens each rebuild rewrote. No other tool attributes this.
• Position-weighted cost model: waste valued as tokens x remaining API calls x cache-read rate.

Fixed

• Session state keyed by harness session id: concurrent sessions no longer cross-contaminate duplicate tracking.
• Ranged reads never make a later full read look like a duplicate.
• Edit-count warning fires once per file per session.
• Removed a counter that was never incremented anywhere.

Removed

• The anatomy staleness banner (fired in 75% of measured digests), unfilled STATUS template injection, and dead shipped weight (reframe-frameworks.md, identity.md, designqc/suggestions stubs) - deleted on update only when verifiably untouched.

Run openwolf update to migrate (backup taken first).