[Bug]: Claude run fails with "provider event stream closed unexpectedly" after t3_worktree_handoff (still on 0.1.2-nightly.20261021.3473) #22863

Split out of #15136 as requested.

Environment

• T3 Code server 0.1.2-nightly.20261021.3473 (macOS arm64, desktop app)
• Claude Code 2.1.306, claudeAgent / claude-opus-5-5, full access

Steps to reproduce

1. In a Claude thread attached to the project root, ask: "can you switch this thread to a worktree off main?"
2. The agent calls t3_worktree_handoff with { branch: "bramley/wip-2026-10-21", baseRef: "main" } (no continuationPrompt).

Expected behavior

The thread moves into the worktree and the turn ends normally.

Actual behavior

• The worktree is created and the thread is rebound to it.
• The run ends failed with "The provider event stream closed unexpectedly. Retry the turn; if it keeps failing, check the provider and server logs."
• The handoff tool item stays running, so the composer shows "Waiting on background task" ([Bug]: Stale "Waiting on background task T3 Worktree Handoff" banner persists after a Claude worktree handoff #15993). A later server restart reports it as cancelled background work.

The provider event log shows the handoff tool_use at 12:01:23.395Z, then T3's own outgoing query.close at 12:01:27.436Z, and the run marked failed 2 ms later. So the session is ended by T3's planned detach, not by Claude Code crashing.

Screenshot:

Likely cause

The thread.metadata.update that changes worktreePath detaches every session on the thread (apps/server/src/orchestration-v2/Orchestrator.ts, detachSessionIds) without settling the active run. For single-thread providers, ProviderSessionManager.detach then releases with manual_shutdown, and failSubscribers hands the running turn a ProviderAdapterEventStreamError.

Related: #15993, #14245, #16000, #22748. Open PRs: #15203, #15989.

🤖 Generated with Claude Code