feat(agentic-server): relay streaming chat completions and meter their usage

[pyramation]

Summary

The gateway assumed every successful chat response was JSON (await upstream.json()), so any client asking for stream: true got SyntaxError: Unexpected token 'd', "data: {"id".... That is what a real harness does by default — found while pointing DeepSeek Harness at agentic-server as its OpenAI-compatible provider, where every turn failed.

/v1/chat/completions now branches on req.body.stream === true:

const body = streaming ? withUsageStreamOptions(transformed) : transformed;   // stream_options.include_usage
...
if (streaming) {
  const usage = await relayEventStream(upstream, res);   // frames forwarded verbatim
  sink?.logInference({ ...identity, inputTokens: usage?.prompt_tokens ?? 0, ... });
  return;
}

Three things worth knowing:

• Usage is scanned out of the relayed bytes, not parsed from a body. UsageScanner buffers across chunk boundaries, splits on \n, skips [DONE]/keep-alives/unparseable frames, and keeps the last usage object — so metering still fires on a streamed turn. stream_options: { include_usage: true } is added only when the caller expressed no preference.
• Non-OpenAI streaming is rejected 501, not silently downgraded. resolveUpstreamUrl rewrites ollama/anthropic to their native endpoints, whose stream frames are not OpenAI chunks; answering a whole JSON body to a client that asked for a stream would be worse than failing loud.
• Disconnects abort upstream via res.on('close'), not req. The request stream's close fires as soon as the body is consumed, so aborting on it kills every stream mid-flight (AbortError: This operation was aborted — observed before the fix). The error path also end()s instead of writing JSON once res.headersSent.

Verified end to end beyond the unit tests: DeepSeek Harness, running Constructive's 18 neutral @agentic-kit/db-tools through a local adapter, completed a tool-calling turn routed through this gateway, with the injected sink recording promptTokens/completionTokens for the streamed turn.

Link to Devin session: https://app.devin.ai/sessions/2d292e43fd2e4f34bb84605b5aa2a250
Requested by: @pyramation

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring