Whats New

lwc v0.19.3

English · 简体中文

Local-first team memory: automatic core-memory synchronization, external Agent conflict resolution, personal-key sign-in, explicit space access, cloud-only reads, compensating recovery and a bilingual console. Includes member provisioning with initial grants, permission-checked page sharing and knowledge-link navigation.

Team guide · Full release notes

Core schema upgrades to 20 and server control schema to 6; back up before upgrading. External identity providers require deployment configuration; real provider login was not exercised in this acceptance run.

The v0.19.3 is available on GitHub, crates.io, Homebrew and the public multi-architecture container registry. All six native targets, final container smoke checks and the release workflow passed. All 27 public assets were downloaded and all 25 archive checksums matched. npm is also published: both the public version and latest tag resolve to 0.19.3; its checksum matches the uploaded package. Failed tags v0.19.0 through v0.19.2 remain unchanged and unpublished.

---

Earlier release notes

lwc v0.18.5

Language: English · 简体中文

This release collects all changes since the last published version, v0.17.18, including the work carried by the unsuccessful v0.18.0–v0.18.4 tags. Those tags remain unchanged; this description covers the complete upgrade rather than only the final corrective commit.

Highlights

Durable Discussion

• Preserve visible brainstorming and clarification across sessions, independently of which Skill or prompt initiated it. Record a title and description, questions, options, grouped answers, unassigned replies and explicit capture gaps in the existing SQLite store.
• Address questions, answer fragments and summary conclusions by stable IDs. Correct, reorder, reassociate, withdraw and restore individual items while retaining original text and revision history; grouped changes commit atomically.
• Recover a bounded checkpoint after interruption or compaction. Read individual items or paginated records/history; export the complete record explicitly.
• Write detailed summaries with evidence references, separate user confirmation and transitive local invalidation when cited information changes. Closing checks answered questions, fresh summary coverage and unresolved replies; it does not grant implementation permission.
• Use shared CLI/MCP contracts, context ownership, CAS revisions and stable idempotency keys. Archives and Sync preserve discussion history without transferring local context bindings.

Native CodeGraph and clearer state

• Forward native CLI stdout, stderr and exit codes, and retain the native MCP result instead of wrapping the query payload. Keep diagnostics and freshness checks separate from native results.
• Expose runtime query contracts and command discovery. Unified doctor diagnostics identify the checkout, logical repository and selected index owner; missing indexes do not silently select another owner.
• Retain workspace boundaries and distinguish index readiness from current-code evidence. Scoped file fingerprints and dirty/untracked checks report their limits; empty results are not proof that code can be safely removed.
• Use operating-system path canonicalization and Windows file identity for ordinary/verbatim paths, preserving namespace and workspace guards.

Agent workflow and Plan

• Share machine-readable contracts, examples and field errors between CLI and MCP; reduce default receipts and Hook context while retaining explicit expanded reads.
• Revise Plan scope and individual steps while preserving identities, CAS and history. Distinguish waived or replaced work from completed verification. Reconcile provides read-only candidate evidence rather than silently advancing a plan.
• Label recalled historical state as latest-known rather than live truth, and clarify the primary owner for durable knowledge to avoid duplicate maintenance.

Codex-specific native plugin

• Provide a native LWC plugin with Hooks, Skills, MCP metadata and square black/blue logos for light and dark interfaces.
• The native installer archives duplicate bundled Skill names from global $CODEX_HOME/skills (default ~/.codex/skills) into recoverable backups. It handles directories and symlinks without modifying link targets, unrelated Skills or any other Agent's installation.
• This adaptation does not depend on a particular local Skill manager. Claude, Pi and other Agent installation routes remain unchanged.

Safety and compatibility

• SQLite advances to schema 19. Upgrade the core before using Discussion; v0.17.18 cannot provide the new commands. Keep recoverable backups when upgrading persistent stores, and do not alternate older binaries against migrated stores.
• Visible Discussion records are a dedicated archive, not hidden reasoning or automatic curated Wiki ingestion. Existing sensitive-input detection rejects suspect content; excluded content must be represented by a nonsensitive gap.
• Portable collection relies on the Discussion Skill protocol. Exact host-message capture supplements an already bound discussion only when raw input and native message identity are available; it is not a universal transcript or delivery guarantee.
• Windows CLI binaries reserve an explicit 8 MiB stack to support the full command parser in debug builds; this addresses the observed startup overflow under the Windows default 1 MiB stack.
• Work cancellation is durable across Worker startup: a cancel file written before the Worker claims its state now terminates the Work instead of being overwritten by its initial running state.
• Windows graph queue locking retries temporary access denial while a deleted lock still has an open handle. Retries remain bounded; persistent permission errors still fail closed.
• Concurrent schema upgrades recheck versions inside the migration transaction. The v17 and v18 steps accept an already completed, supported upgrade without downgrading or repeating it; unknown future versions remain rejected.
• The Sync continuity regression includes the new Discussion object kind. Release preparation preserves existing dependency versions and checksums.

Verification

• Targeted BDD/TDD covers retained originals, CAS, idempotency, atomic rollback, granular corrections, summary invalidation and closure, migration, capture deduplication, compaction/context isolation and changeset preservation.
• Archive import/export/merge, MCP boundaries, Skill packaging and Codex-only duplicate migration have dedicated checks. Native plugin validation and local installation were exercised.
• Earlier Agent UX verification covered macOS, Linux and Windows/Wine, plus real CodeGraph worktree/freshness scenarios. See benchmarks/results/2026-09-11-agent-ux-acceptance.md and docs/discussion-acceptance.md for exact scope and limitations; these are not claims of universal host capture or performance improvement.
• Publication is gated by CI, release acceptance, six native platform targets, combined checksums, and the exact-release npm installer smoke. The linked Actions runs provide the final execution results; npm-registry, crates.io and Homebrew publication are separate from GitHub Release creation.

Upgrade

Upgrade the LWC core and learning runtimes together using the published distribution channel. For Codex native installation and migration, follow integrations/codex-lwc/README.md; do not enable duplicate direct Hooks, and start a new task after changing plugin discovery. Other Agents retain their existing installation procedure.