[Bug] Kimi K3 via OpenCode Go returns empty responses through the Responses API (missing output_item events in SSE stream)

What happened

Selecting kimi-k3 through OpenCode Go (https://opencode.ai/zen/go/v1) with the Responses API produces a turn with zero visible output, zero tool calls, and no token usage. The model actually replies — the text arrives in the stream — but it is dropped because the stream never emits the item scaffolding the Responses API requires.

Environment

• Client: OpenAI Codex CLI 0.146.0 (custom provider: wire_api = "responses", base_url = https://opencode.ai/zen/go/v1)
• Model: kimi-k3 via OpenCode Go
• Auth: standard bearer API key (valid; other models work)

Reproduction

codex exec --yolo -m kimi-k3 "say hi"

Expected: the model replies (e.g. "Hi!").

Actual: no assistant output at all. The session transcript shows token_count with info: None and task_complete with no error; the command exits successfully. Codex logs this for every streamed delta:

ERROR codex_core::util: OutputTextDelta without active item

Root cause evidence

The raw SSE from POST https://opencode.ai/zen/go/v1/responses for kimi-k3 contains only these event types:

response.output_text.delta   (10x)
response.completed           (1x)
ping                         (1x)

Example payload:

{"id":"chatcmpl-6a71034d30b9ed1c67db2906","type":"response.output_text.delta","delta":"Hi","response":{"id":"chatcmpl-6a71034d30b9ed1c67db2906","model":"kimi-k3"}}

There are no response.output_item.added, response.content_part.added, or response.output_item.done events. Per the Responses API spec, text deltas must be scoped to an item (via item_id/output_index) created by an output_item.added event. Without that, Responses-API clients (Codex, and likely others) cannot attach the deltas, so the whole reply is discarded.

Additional data

• The same setup with deepseek-v4-flash, deepseek-v4-pro, gpt-5.6-luna, and glm-5.2 works correctly (real output, token usage reported).
• A tool-using prompt with kimi-k3 also completes with no tool calls executed and no output.
• Tried with reasoning effort low and high — identical behavior.

Suggested fix

Emit proper item events in the streamed Responses response for kimi-k3 (and any other models routed through the chat-completions translator): response.output_item.added → response.content_part.added → response.output_text.delta (with item_id and output_index) → response.output_item.done → response.completed.