Task agent resume fails silently when transcript ends with API error

Bug: Task agent resume fails silently when transcript ends with API error

Summary

When a subagent's conversation ends with an API error (e.g., 400 due to tool use concurrency), attempting to resume that agent fails silently - returning 0 tool uses · 0 tokens · 1s with no error message.

Steps to Reproduce

1. Spawn a subagent via Task() tool
2. Let the agent work until it hits an API 400 error (e.g., "tool use concurrency issues")
3. The agent returns to parent with error state
4. Try to resume the agent with Task(resume: "agent-id")
5. Result: Resume completes instantly with 0 tool uses · 0 tokens · 1s
6. Expected: Error message like "Cannot resume agent: transcript ended with error"

Evidence

Agent transcript (agent-aa2fcb9.jsonl) ends with:

{
  "error": "invalid_request",
  "isApiErrorMessage": true,
  "content": [{"text": "API Error: 400 due to tool use concurrency issues. Run /rewind to recover the conversation."}]
}

Screenshot shows the silent failure:

● batman(Apply font edit fix) resuming aa2fcb9
└ Done (0 tool uses · 0 tokens · 1s)

● Agent resume failed. I'll apply the fix directly.

Environment

• Claude Code version: 2.1.0
• Model: claude-opus-4-5-20251101
• OS: macOS

Impact

• Users get no feedback about why resume failed
• Claude detects failure only by observing 0 tokens returned
• Makes debugging agent issues difficult
• Wastes user time trying to resume unrecoverable agents

Suggested Fix

When Task(resume: "id") is called:

1. Check if transcript's last message has "error" or "isApiErrorMessage": true
2. If so, return clear error: "Cannot resume agent {id}: transcript ended with API error: {error_message}"
3. Optionally suggest: "Start a new agent or use /rewind"

Related

This may also affect the general "tool use concurrency" error handling - perhaps there should be automatic recovery instead of corrupting the transcript.