[fliaping]

Linux message updates waited for a 30-second snapshot loop, while sending required the owner to open the correct chat manually. This change pins the analysis fork with DB/WAL notifications and a two-second fallback, adds exact-recipient chat navigation, and exposes desktop readiness and recovery through HTTP, MCP and the workspace UI.

Navigation searches only accessible Contacts/Groups results and verifies the exact account and session ID against the original WeChat process before composing. Duplicate display names, message/history matches and unresolved layouts cannot authorize a send. Existing drafts and manual search text are preserved. The session-layout helper derives evidence from the installed original ELF and accepts matching first two numeric version components. No original WeChat or Alpha files are modified.

The Qt desktop driver selects the editor in the verified chat pane, uses explicit GI interface dispatch and recognizes WeChat's painted Send button locally with Tesseract. The exact enabled caption requires at least 80% word confidence. X11 geometry corrects Qt high-DPI coordinates; focus changes, held keys or moved windows block input. Expiry is checked again after navigation and before submission.

Scheduled operations retry transient pre-input failures for at most five minutes after their due time. Changed context cancels an operation; expiry prevents sending; attempted composer writes and submitted/uncertain results are never automatically retried. The UI and fourteen MCP tools expose concrete blockers and recovery instructions. A disconnected Tailscale daemon no longer breaks action rendering, and narrow layouts keep table scrolling inside the card. The portable Skill documents the same guards and recovery contract.

DeepSeek Harness is upgraded to official npm latest 0.2.1-rc.2, checked October 20, 2026, with signed browser cookies, prefixed routes and its current settings RPC. Its official JavaScript remains unchanged. The NAS page connects and opens model configuration; model credentials remain for the owner to configure in the UI. External Tailscale and named Cloudflare Tunnel connectivity still require owner credentials/DNS.

Validation:

• 118 integration/source checks and the preceding 70 targeted analysis tests passed. Coverage includes exact-result selection, draft preservation, expiry during navigation, bounded recovery, context changes, authentication, origin protection and null Tailscale IPs.
• On the unlocked NAS, navigation opened the authorized target from Contacts and verified its original-process session ID. An owned synthetic draft remained unchanged while navigation was blocked, then readiness recovered after clearing only that draft. No new messages were sent during this navigation rollout.
• Authenticated HTTP and MCP navigation passed. Two synthetic operations were prepared and cancelled without submission. Cancelled commit replays stayed cancelled; the composer remained empty. Browser checks at 1280px and 390px showed ready status and no horizontal page overflow.
• The preceding three explicitly authorized real messages still occur once each in local self-message records. Replaying completed commits creates no new messages. Their submitted-state-to-API visibility was approximately 1.79, 1.74 and 2.08 seconds; the third can include a receipt refresh. These checks do not establish independent incoming-message latency, recipient delivery or read receipts.

The final image built from application revision c730ca0 passed the full fresh-profile smoke test, including container restart, authentication, fourteen MCP tools, structured logged-out readiness, Harness RPC and original vendor files. It is loaded on the NAS and selected in the persistent override for the next recreation; nine updated runtime files match its hashes. Production runs the new workspace code in place while preserving the owner's unlocked WeChat process. Harness was refreshed separately and its MCP plugin is active. Image, deployment and rollback evidence is recorded in docs/workspace-validation.md. No combined image is published.

Attachments, quoted replies, mentions and forwarding remain outside this driver's implemented actions. Model-driven end-to-end dialogue awaits the owner's model configuration. This PR targets the integrated workspace branch and preserves unrelated working-tree changes.