mason-context 0.17.3 → 0.18.0

package/CHANGELOG.md

@@ -1,5 +1,26 @@
 # Changelog
 
+## 0.18.0 — 2026-09-16
+
+- Fix overlapping snapshot writes restoring old descriptions and deleted flows. Snapshot saves, verdicts, and partial writes share a per-checkout cross-process lock; reads, replacement, and consolidation cleanup stay within the protected operation.
+- Bind map verification to entry kind, content, and bounded source evidence through review tokens. Changed or deleted entries return conflicts; older calls without tokens request a fresh review without recording a verdict. Existing snapshot files remain compatible; callers must include the returned entry kind and review token to record a verdict. Restart all Mason MCP processes for a checkout after upgrading; older binaries do not honor the new snapshot lock.
+- Recheck saved map verification evidence on retrieval. Changed evidence reports stale verification; unavailable evidence and legacy verdicts without tokens report unknown. Preserve historical verdicts and failure reasons, including when reviewed local edits are reverted to a clean checkout.
+- Diagnose incomplete lock owners and interrupted recovery guards, document safe manual cleanup after writers stop, and cover both interruption windows with controlled process-termination tests.
+- Cover the reported overwrite with a controlled regression, exercise both call orders through one MCP session and separate processes, and test interrupted writers, contention, partial cleanup, and stale review evidence. Run the snapshot regressions in the existing macOS, Linux, and Windows distribution matrix before publishing.
+
+Run `mason upgrade` for standalone installations, or `npm install -g mason-context@0.18.0` for npm installations. Restart **every Mason MCP process** sharing a checkout after upgrading: older processes do not participate in the new snapshot lock. Existing decisions and snapshot files are preserved. Older snapshot verdicts remain visible but report unknown verification until reviewed again with an evidence token. Existing integrations do not need setup again.
+
+## 0.17.4 — 2026-09-15
+
+- Share Git document/source inventories within each automation inspection, with a separate fresh inventory before publication. Batch module ignore checks by directory level and reuse only identical scoped history queries.
+- Let automation's complete input validation replace overlapping repair discovery while retaining standalone repair guards, original evidence, skipped-check retries and per-tool receipts.
+- Include automation's Git subprocess counts and timings in opt-in profiles. Extend hook benchmarks with many top-level modules, incomplete checks, selectable Node runtimes and optional delayed Git reads.
+- Clarify that each developer runs setup once per checkout; upgrading an existing integration normally requires restarting the assistant.
+
+Synthetic measurements reduced a warm pre/post pair from 0.83 s to 0.60 s with native Git, and from 2.93 s to 1.19 s with deliberately delayed inventories. These are local benchmarks, not workplace latency guarantees; see [method and limitations](docs/hook-performance.md#git-inventory-follow-up-2026-09-15).
+
+Run `mason upgrade` for standalone installations, or `npm install -g mason-context@0.17.4` for npm installations, then restart running assistants. Existing integrations do not need setup again for this release. Decisions and retained repair evidence are preserved.
+
 ## 0.17.3 — 2026-09-15
 
 - Reduce repeated hook work by sharing the current audit across repair baselines and reusing complete current audits on unchanged sequential calls. Keep original history, review assessments, skipped-check retries and per-tool evidence receipts independent.