Unhandled case: [object Object] shown when SDK stream ends with had_error after API stall — breaks conversation continuity

Summary

When the Anthropic API SSE stream stalls mid-tool-call, the next bash tool dispatch aborts with a generic "Shell command failed" error and the stream ends with had_error=true / no result. The VSCode UI fails to render this state gracefully and shows Unhandled case: [object Object] in the chat. The conversation cannot be resumed normally — on reload, the in-progress turn is lost.

Environment

• Claude Code extension: cc_version=2.1.141.149 (cc_entrypoint=claude-vscode)
• VSCode: 1.120.0 (commit 0958016b2af9f09bb4257e0df4a95e2f90590f9f)
• macOS Darwin 25.4.0
• Model: Opus 4.7 (1M context)

Expected behavior

When the SDK stream errors out, the UI should:

1. Render a friendly error in chat (e.g. "Connection interrupted — tap to retry") instead of the raw Unhandled case: [object Object] string.
2. Preserve the conversation so reloading the panel doesn't lose the prior turn.

Actual behavior

Sequence from the extension log (relevant excerpts):

\\\
10:06:58.759 [WARN] [Stall] tool_dispatch_end tool=Bash outcome=error durationMs=24
10:06:58.759 [DEBUG] Bash tool error (24ms): Shell command failed
10:08:26.531 sdk_stream_ended_no_result {"had_error":true, "message_count":177}
\\\

\\\
10:12:43.017 [WARN] Streaming stall detected: 43.5s gap between events (stall #1)
10:12:46.045 [WARN] [Stall] tool_dispatch_end tool=Bash outcome=error durationMs=27
10:12:46.045 [DEBUG] Bash tool error (27ms): Shell command failed
10:12:29.490 sdk_stream_ended_no_result {"had_error":true, "message_count":199}
\\\

Repro pattern observed: a 5-call parallel Bash batch where the first call sits in the permission UI for ~30s, then calls 2–5 fire back-to-back within ~70ms. The 3rd shell spawn fails with a sub-100ms error. Also reproduces on a single bash call whose output is large (~333KB image extraction via \magick\) — looks like API SSE backpressure or response size correlates.

Impact

The UI's \Unhandled case: [object Object]\ is a dead end: no retry button, no error context, and reloading the panel loses the in-flight turn entirely. For long planning / coding sessions this is severe — a 90-minute session can be lost to one stall.

Suggested fixes (priority order)

1. Catch the unhandled error in the chat renderer and show a friendly message + retry affordance.
2. Persist the in-flight turn before reload so resume works after a stream error.
3. Investigate the underlying stream stall — large tool outputs or parallel bash batches seem to correlate.