fix(session): surface empty aborted turns - #17909

…n, external watchdog

- MCP shared map now tracks lastUsed timestamp; periodic sweep closes
  idle clients after OPENCODE_MCP_IDLE_MS (default 10 min)
- Instance disposal force-closes MCP clients immediately
- Session sweep in serve mode archives sessions idle beyond
  OPENCODE_SESSION_IDLE_MS (default 30 min)
- SIGTERM/SIGINT in serve mode captures final memory snapshot
- External watchdog script + launchd plist for crash-surviving
  process-tree RSS monitoring

Addresses child-process memory accumulation that caused overnight
macOS kernel watchdog panic (8.5 GB tree RSS, 238 active sessions).
Primary leak source identified as chrome-devtools-mcp --autoConnect
growing ~13 MB/min.

Ref: anomalyco#16697

Co-authored-by: Copilot <[email protected]>

Shutdown handler was missing Instance.disposeAll() and MCP.closeAll(),
causing MCP child processes (chrome-devtools-mcp, etc.) to become
orphans on SIGTERM/SIGINT. Now the shutdown sequence is:

1. Stop session sweep
2. Capture final memory snapshot
3. Stop memory monitor
4. Dispose all instances (triggers force-release of MCP refs)
5. Close any remaining shared MCP clients
6. Stop HTTP server

Also added MCP.stopMcpSweep() and MCP.closeAll() exports.

Co-authored-by: Copilot <[email protected]>

Question tool and permission ask() both use Deferred.await() which
blocks indefinitely with no timeout or abort handling. When the AI
SDK issues multiple tool calls in a single step, one stuck tool
(e.g., unanswered question, unapproved permission) blocks the
entire stream, preventing the cleanup code from marking tools as
errored.

Fix: race both Question.ask() and PermissionNext.ask() against the
abort signal from the AI SDK. This ensures tool execution unblocks
when the session is cancelled, allowing the processor cleanup to
mark stuck tools as errored.

Co-authored-by: Copilot <[email protected]>

When the server restarts while tools are executing, their parts remain
stuck in "running" or "pending" state forever. Session.recover() runs
at startup and marks all such parts as "error" with a descriptive
message, unblocking affected sessions.

Co-authored-by: Copilot <[email protected]>

…cal'

- Add GET /global/session endpoint using Session.listGlobal()
- Create Recent page component with search, project labels, diff stats
- Add /recent route and navigation button on Home page

Co-authored-by: Copilot <[email protected]>

Adds a status icon in the sidebar bottom section so the dashboard
is accessible from any project view, not just the home page.

Co-authored-by: Copilot <[email protected]>

- New rename tool lets the agent set a descriptive session title
- System prompt instructs agents to rename early once task is understood
- No permission required — lightweight metadata update

Co-authored-by: Copilot <[email protected]>

When packages/app/dist exists, serve static files from it directly.
Falls back to proxy to app.opencode.ai if local build is missing.
Enables local web UI customizations without cloud deployment.

Co-authored-by: Copilot <[email protected]>

Use same Button components, typography tokens, and layout patterns
as the Home page for a consistent native look.

Co-authored-by: Copilot <[email protected]>

30-minute default was archiving all sessions, hiding them from sidebar.

Co-authored-by: Copilot <[email protected]>

Co-authored-by: Copilot <[email protected]>

Assistant messages without time.completed cause permanent spinner
in sidebar after server restart. Now Session.recover() also completes
these orphaned messages, matching the existing tool part recovery.

Co-authored-by: Copilot <[email protected]>

When session status is idle, the pending memo now returns undefined
instead of searching for incomplete assistant messages. This prevents
permanent Thinking spinners caused by:
- Orphaned messages from interrupted streams
- Stale in-memory store data after server-side recovery

Fixes anomalyco#17680

Co-authored-by: Copilot <[email protected]>