You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Fixed
Attach memory and rules to every fresh OpenCode model request, including later steps and other sessions. Detect config, index, directory, tombstone, and sentinel changes before cache reuse; do not consume cross-process notifications.
Queue automatic consolidation with noReply: true, preserve agent/model/variant, retain native continuation on API failure, and resume pending work in the fallback prompt.
Fail closed on local and shared contention; treat EPERM as alive, capture per-acquisition release tokens, and bound stale reclaim when deletion fails.
Validate safe regular topic files and no-follow reads, reserve the index filename, preserve unrelated removed-topic slug collisions, and enforce pinned TUI removal inside the lock.
Refuse FIFO reads without blocking; detect trusted config-symlink target edits before cache reuse.
Flush and clean atomic temporary writes, persist removal intent before dropping discoverability, roll back ordinary topic/index write failures, and preserve racing config creation and existing legacy backups.
Lock local-to-shared migration in consistent order, copy atomically, respect both removal lists, and retry failed migrations only marking success after completion.
Preserve JSONC literal commas, refresh topic metadata, compare full duplicate timestamps while retaining pins, heal stale flags, correct repair totals and malformed-field recovery, and surface TUI refusal/filesystem results.
Changed
inject_every_n_turns now controls forced disk refreshes in model requests; memory is always injected and system tokens are not inherently free.
Documentation records current host compatibility, co-tenant coordination, and per-file recovery boundaries.
Tests
92 to 120 smoke tests, plus real-process OpenCode/OpenCode and OpenCode/pi writer checks.
npm run test:host verifies native tool execution, subsequent requests, another session, and automatic consolidation with a real OpenCode 1.18.34 server and local fake model.