Skills API: Multiple skills with streaming doesn't return message_delta/message_stop events

[basewakamaru]

When using multiple Skills with streaming enabled in the Foundry API, the stream ends prematurely without returning
message_delta and message_stop events. This prevents retrieval of the container ID and causes incomplete stream
handling.

Environment

• SDK: @anthropic-ai/foundry-sdk
• Platform: Microsoft Foundry
• Model: claude-haiku-4-5
• Stream: true
• Betas: code-execution-2025-08-25, skills-2025-10-02

Steps to Reproduce

1. Initialize Foundry client with 2 or more Skills in the container parameter
2. Send a message request with stream: true
3. Observe the stream event sequence

Expected Behavior

• Stream should include: message_start → content_block_* events → message_delta → message_stop
• Container ID should be available in message_delta.delta.container.id
• All events should complete before stream ends

Actual Behavior

• Stream ends after content_block_stop events
• message_delta event is never received
• message_stop event is never received
• Stream completes without final events, leaving stream in incomplete state

Logs

Request Configuration:

{
  "model": "claude-haiku-4-5",
  "stream": true,
  "betas": ["code-execution-2025-08-25", "skills-2025-10-02"],
  "container": {
    "skills": [
      {
        "type": "anthropic",
        "skill_id": "xlsx",
        "version": "latest"
      },
      {
        "type": "anthropic",
        "skill_id": "pptx",
        "version": "latest"
      }
    ]
  }
}
Stream Event Sequence (Actual):
1. ✅ message_start (no container field)
2. ✅ content_block_start: text
3. ✅ content_block_delta: text_delta (multiple events)
4. ✅ content_block_stop
5. ✅ content_block_start: server_tool_use (first skill)
6. ✅ content_block_delta: input_json_delta (multiple events)
7. ✅ content_block_stop
8. ✅ content_block_start: server_tool_use (second skill)
9. ✅ content_block_delta: input_json_delta (multiple events)
10. ✅ content_block_stop
11. ❌ Stream ends here - message_delta and message_stop never arrive
Workaround (Non-streaming mode works):
// stream: false で実行するとコンテナIDが正常に返却される
{
  "stream": false,
  "container": {
    "id": "container_011CXbQCtRNCswtwatctJJwk",
    ...
  }
}
Additional Notes
- Single skill with streaming: ✅ Works correctly
- Multiple skills added sequentially (1st skill, then add 2nd): ✅ Works correctly (message_delta received)
- Multiple skills on initial request: ❌ message_delta never received
- Non-streaming mode with multiple skills: ✅ Container ID returned correctly
This appears to be specific to initial requests with multiple Skills using streaming mode.
Related Issues
- #842 - Streaming responses consistently interrupted mid-transmission
- liteLLM #8066 - Tool calls not working with streaming
---

[tarik0]

This is connected to PPTX & DOCX skills #893