v3.32.8 β CI unblock, memory_search fix, version lockstep
v3.32.8 β CI unblock, memory_search fix, version lockstep
Fixed
- CI-breaking lockfile drift (#2719, #2717):
v3/pnpm-lock.yamlwas pinned to@claude-flow/security@^3.0.0-alpha.10whilepackage.jsonhad moved to^3.0.0-alpha.12, failing every--frozen-lockfileCI job (Type Check V3and others) on every PR since the drift landed. Regenerated viapnpm install --lockfile-only. - MCP
memory_searchnamespace regression (#2646, third occurrence of #1123/#1131): omitting the optionalnamespaceparameter returned 0 results even when data existed across multiple namespaces, because the tool handler coerced an omitted namespace to the literal string'default'instead of leaving itundefinedso the search layer's ownnamespace || 'all'fallback could fire. Fixed with a regression test that asserts what the tool handler forwards to the search layer directly, so this class of regression is caught at the tool boundary going forward. - Package version lockstep (surfaced by #2703):
claude-flowhad drifted to 3.32.2 while@claude-flow/cli/ruflowere at 3.32.7, failing thePlugin package install-safetyCI check. All three packages are now back in lockstep at3.32.8.
Verified as already fixed (no code change this release)
- [P0] Worktree daemon fanout quota exhaustion (#2661): confirmed already fixed same-day via #2662/#2663 (shipped v3.27.4/v3.28.0) β repository-scoped supervisor election, workspace leases, and a global AI-launch budget with a circuit breaker replaced the old per-worktree dedup that let every Git worktree spin up its own set of scheduled Claude-powered workers. The tracking issue had stayed open because the fix commits didn't use GitHub's auto-close keyword; closed now with verification (40/43 tests passing; the 3 failures are a Windows symlink-privilege sandbox limitation, not real bugs).
Registry
latest,alpha, andv3alphadist-tags are synced to3.32.8across@claude-flow/cli,claude-flow, andruflo(#2703).
Full Changelog: v3.32.7...v3.32.8