VS Code: Corrupted session file causes unrecoverable crash loop (exit code 1)

Bug Description

A corrupted/large session .jsonl file causes Claude Code to crash immediately on startup in VS Code, creating an unrecoverable crash loop. The extension auto-resumes the last session on every restart, hitting the same bug repeatedly with no way for the user to escape without manually renaming the session file.

Environment

• Claude Code: v2.1.71 (VS Code extension)
• OS: Windows 11 Home 10.0.26200
• Node.js: v24.14.0
• 16 MCP servers configured

Steps to Reproduce

1. Have a large session file (~2.7MB .jsonl)
2. Open VS Code — extension auto-resumes the session
3. Send any message
4. Extension crashes with TypeError: undefined is not an object (evaluating 'gH(M).length')
5. VS Code shows error: Claude Code process exited with code 1
6. Re-opening Claude Code panel or restarting VS Code always resumes the same broken session → crash loop

Error from Logs

2026-03-08T14:06:16.459Z [ERROR] "TypeError: TypeError: undefined is not an object (evaluating 'gH(M).length')
    at <anonymous> (B:/~BUN/root/src/entrypoints/cli.js:1696:12340)
    at reduce (native:1:11)
    at Z4 (B:/~BUN/root/src/entrypoints/cli.js:1696:12317)
    at processTicksAndRejections (native:7:39)"

Followed immediately by:

[error] Error from Claude (on channel uids0di0cdj): Error: Claude Code process exited with code 1

Crash Loop Evidence

The same session (68a02f9d-...) was auto-resumed across 5 VS Code restarts and 11+ channel crashes, all hitting the identical TypeError:

Session 1 (15:05): 7 channel crashes, all exit code 1
Session 2 (15:39): crash on resume
Session 3 (15:48): crash on resume  
Session 4 (15:58): crash on resume
Session 5 (16:10): crash on resume

Secondary Effect: API Rate Limiting

Each crash triggers MCP server reconnections and API calls. After 11+ rapid crashes, Anthropic's own API started returning 429:

[clientData] fetch failed: 429
AxiosError: [url=https://api.anthropic.com/api/oauth/claude_cli/client_data,status=429,body=Rate limited.]

This made subsequent startups even worse — settings failed to load on top of the session bug.

Workaround

Manually rename the corrupted session file:

mv ~/.claude/projects/<project>/68a02f9d-*.jsonl ~/.claude/projects/<project>/68a02f9d-*.jsonl.bak

After this, VS Code starts a fresh session and works normally.

Suggested Fixes

1. Graceful session recovery: If resuming a session throws, catch the error and start a fresh session instead of crashing
2. Crash loop detection: If the same session crashes N times (e.g., 3), stop auto-resuming it
3. Backoff on API calls: Don't hammer client_data endpoint on every restart — use exponential backoff when rate-limited
4. User-visible error: Show "Session could not be resumed" in the UI instead of silently crashing

Diagnostic Context

• The session file was 2.7MB (large but not extreme)
• The crash happens in minified code (cli.js:1696) at what appears to be a .length access on an undefined value during message processing
• All MCP servers connected successfully before the crash — the bug is in session/message handling, not MCP initialization