openai-responses provider drops entire SSE stream when gateway omits event: lines (empty content, zero usage)

[initializ-mk]

Problem

The openai-responses provider's SSE parser (readStream, forge-core/llm/providers/responses.go) routes events solely on the SSE event: line:

if after, ok := strings.CutPrefix(line, "event: "); ok {
    currentEvent = after      // ← only ever set here
    continue
}
data, ok := strings.CutPrefix(line, "data: ")
...
switch currentEvent { case "response.output_text.delta": ... case "response.completed": ... }

The OpenAI Responses API sends both an event: line and a type field inside the data: payload. But event: is optional per the SSE spec, and some gateways omit it — notably the Bedrock openai-sigv4 shim (Kong route /bedrock/openai-sigv4/v1/responses), which streams spec-legal SSE carrying the type only as a JSON field:

data: {"content_index":0,"delta":"Loading","type":"response.output_text.delta"}
...
data: {"response":{...,"usage":{"input_tokens":76,"output_tokens":73,"total_tokens":149}},"type":"response.completed"}

With no event: lines, currentEvent stays "", the switch matches no case, and every frame is dropped — text deltas and the terminal response.completed (which carries usage). Result: empty assistant message, zero tokens, empty finish reason.

Field evidence

Observed on a staging agent (openai-responses / openai.gpt-5.4 via the Bedrock gateway):

• llm_call audit: input_tokens:0, output_tokens:0, tokens_unavailable:true, total_input_tokens:0 on a 619ms call.
• Runtime log: {"finish_reason":"","msg":"llm response"} — empty finish reason is the smoking gun, since response.completed is the only handler that sets one.
• invocation_complete: all token totals zero; the user got an empty answer.

The raw stream (captured via curl) confirms the gateway returns correct content + usage:{input_tokens:76,output_tokens:73,total_tokens:149} — forge simply never parses it.

Fix

Dispatch on the payload type field, falling back to the event: line only when the payload has none (authoritative + present in both dialects; matches how the OpenAI SDKs route):

eventType := currentEvent
var typed struct{ Type string `json:"type"` }
if json.Unmarshal([]byte(data), &typed) == nil && typed.Type != "" {
    eventType = typed.Type
}
switch eventType { ... }

Also bump bufio.Scanner's 64KB line cap — the terminal response.completed frame embeds the full output[] on one data: line and overruns it on large answers, aborting the stream mid-parse.

Acceptance

• A stream with no event: lines (type in data.type) parses text + usage correctly.
• The dual-dialect stream (both event: and type) still works (real OpenAI) — no regression.
• A >64KB terminal frame parses without a scanner overflow.

Related / out of scope

• The model emits read_skill("weather") as plain output_text in a phase:"commentary" message rather than a function_call item, so the skill never loads — a separate model/gateway-prompt behavior issue, not this parse bug.
• Optional follow-up: capture usage.input_tokens_details.cached_tokens / cache_write_tokens for cache-usage parity with Anthropic llm_call undercounts tokens: capture cache_read/creation + emit total_input_tokens (parity with initializ-sdk#10) #431.