[Task 08] Wire streaming chat completions

[lubobill1990]

Task 08 — Wire streaming chat completions

Depends on: 07
Unblocks: 11

Goal

Streaming OpenAI requests record one row each, with usage extracted from
the final include_usage chunk. Aborts produce a status = 'aborted' row
with whatever was accumulated.

Scope

In create-chat-completions.ts:

• Before forwarding payload upstream, when payload.stream === true,
  ensure payload.stream_options = { ...payload.stream_options, include_usage: true }.
• Wrap the events() async iterator with a StreamUsageAccumulator:
  • feed(chunk) for every event.
  • On normal completion: recordUsage(... finalize() ...) with status = 'ok'.
  • On c.req.raw.signal.aborted: stop forwarding, recordUsage(...) with
    status = 'aborted' and the partial usage.
  • On thrown error after streaming started: status = 'error' with partial
    usage.

Frame stripping (optional, gated by --strip-usage-frame, default off):
the final choices: [] chunk gets dropped before flushing to the client.

Edge cases

• Some clients send their own stream_options.include_usage. Override with
  true if they sent false — they likely just don't know the proxy needs
  it. Log at debug level.
• Tool-call streams: token accounting is identical; the usage final chunk
  still arrives.
• If the upstream closes without ever sending the usage chunk:
  accumulator.finalize() throws; record a status = 'error' event with
  zero usage and log a warn.

Definition of Done

• Manual smoke: streaming curl request produces 1 row with non-zero
  output_tokens.
• Manual smoke: ctrl-C mid-stream → row with status = 'aborted'.
• Verify with a client that does NOT send include_usage — proxy still
  records.
• --strip-usage-frame removes the final empty-choices chunk from the
  client response.

---

• Source: docs/tasks/08-wire-streaming-chat.md
• Design: docs/design/
• Phase: wiring
• Depends on tasks: 7