/v1/responses streaming SSE missing fields required by Vercel AI SDK (output_index, id on function_call items, created_at/model)

[MidnightOnMars]

Summary

The /v1/responses streaming SSE implementation (from PR #18486) is missing three fields that the Vercel AI SDK's @ai-sdk/openai package requires via Zod schema validation. This prevents tools like OpenCode, Cursor, and other AI SDK-based clients from using the Responses API with llama-server for tool calling.

Chat Completions (/v1/chat/completions) works correctly. The model produces perfect structured tool calls. This is specifically about the Responses API streaming event format.

Missing fields

1. output_index (number) on multiple event types

Affected events: response.output_item.added, response.output_item.done, response.function_call_arguments.delta, response.output_text.delta

The AI SDK uses output_index to track which output item each event belongs to. Without it, the Zod schema validation fails (chunk.success = false) and the event is silently dropped.

Fix: Add a sequential counter starting at 0, incremented each time a new output item is added.

2. id (string) on function_call items

Affected events: response.output_item.added and response.output_item.done where item.type === "function_call"

llama-server sends call_id but not id on function_call items. The AI SDK schema requires both fields. When id is missing, the Zod discriminated union matches type === "function_call" but fails inner validation. The stream handler skips the event, hasFunctionCall never becomes true, and finishReason is always "stop" regardless of the model's actual output.

Fix: Copy call_id to id (or generate a unique id) on function_call items.

3. created_at (number) and model (string) on response.created

The AI SDK schema requires response.{id, created_at, model} but llama-server sends only response.{id, status}.

Fix: Add created_at (unix timestamp) and model (the model name from the request) to the response object in response.created events.

What does NOT need fixing

The AI SDK has a catch-all that transforms unknown event types into { type: "unknown_chunk" }. These events are silently ignored and do NOT cause errors:

• response.reasoning_text.delta (vs AI SDK's expected response.reasoning_summary_text.delta)
• response.in_progress
• response.content_part.added/done
• response.output_text.done

Evidence

Tested with GPT-OSS-120B on llama.cpp build 8305 (d63aa39) using OpenCode v1.2.26 with @ai-sdk/openai. The model handles 11 tools + 48KB system prompts correctly at the API level (verified via curl). An 80-line Node.js SSE proxy that patches only these three fields makes tool calling work end-to-end through the AI SDK.

Reproduction

1. Start llama-server with a tool-calling model and --jinja
2. Configure OpenCode with "npm": "@ai-sdk/openai" pointing at the server
3. Run opencode run "Create a file called test.txt" --format json
4. Result: reason: "stop", no tool execution. Model produced correct function_call but AI SDK silently dropped it.

Related

• PR server: /v1/responses (partial) #18486 (Responses API implementation)
• Issue Misc. bug: OpenAI API v1/responses llama-server #14702 (original feature request)
• Issue Feature Request: Support OpenAI Responses API (/v1/responses) in llama.cpp server #19138 (Responses API feature request)