claude-code-api-streaming

Emit SSE frames with Content-Type text/event-stream in this order: message_start, content_block_start, one or more content_block_delta events, content_block_stop, message_delta with final stop_reason and usage, then message_stop. Each frame needs both an event: line and a JSON data: payload.

Start with content_block_start containing a tool_use block with id, name, and empty input, then stream arguments as input_json_delta partial_json chunks that concatenate into valid JSON. Set message_delta stop_reason to tool_use so the client executes the tool and returns a tool_result.

Claude Code flags streams with zero events, missing message_start, incomplete content blocks without a stop reason, SSE creation returning 404, mid-stream disconnects, timeouts, or non-Anthropic JSON payloads. Ensure the full event sequence completes and errors return HTTP status codes instead of partial SSE.

Yes, Claude Code may issue non-streaming requests after stream failures or for background tasks. The stream:false response must return a complete message JSON whose structure matches the aggregated final message from the streaming path.

Common values are end_turn for normal completion, tool_use when the client must run a tool, max_tokens for truncation, stop_sequence when a stop sequence is hit, refusal, and model_context_window_exceeded. Omitting message_delta breaks final usage and stop_reason reporting.