bug: call_model_stream_raw never requests stream_options.include_usage; cost tracking silent in production

[rdwj]

Summary

LLMClient.call_model_stream_raw in fipsagents.baseagent.llm does not set stream_options={\"include_usage\": True} when issuing streaming chat completions. vLLM (and other OpenAI-compat servers) only emit a usage chunk on streaming responses when this flag is explicitly opted into. Without the flag, StreamMetrics.prompt_tokens / completion_tokens stay None for the entire stream, and OpenAIChatServer._persist_cost_data (added in #117 / #116) returns early — meaning cost tracking is a no-op in real deployments.

The unit tests in tests/test_server_openai.py synthesise StreamMetrics with populated prompt_tokens / completion_tokens, so they exercise the persistence path correctly but bypass the gap in the LLM client.

Reproduction

Cluster smoke against fipsagents 0.14.0 + fipsagents-platform 0.2.0, gpt-oss-20b on RHPDS:

1. POST two /v1/chat/completions with the same session_id.
2. Both calls return content correctly; the response's usage block is {prompt_tokens: null, completion_tokens: null, total_tokens: null}.
3. Inspect the platform's session row: messages populated (204 bytes), cost_data empty {}.

Direct curl against the same vLLM endpoint with stream_options.include_usage=true confirms the model happily emits a final usage chunk: {prompt_tokens:70, total_tokens:75, completion_tokens:5, prompt_tokens_details:{cached_tokens:64}}. The endpoint isn't broken — the client never asks.

Fix

In LLMClient.call_model_stream_raw, after call_kwargs[\"stream\"] = True:

call_kwargs.setdefault(\"stream_options\", {\"include_usage\": True})

setdefault so callers can opt out by passing stream_options={\"include_usage\": False} (or pass other stream options).

Tests

• Regression: assert _client.chat.completions.create is called with stream_options={\"include_usage\": True} when streaming. Mock the OpenAI client; check the kwargs.
• E2E coverage already exists for the persistence path; once the flag is forwarded, the existing usage-driven tests prove the loop closes.

Impact

• Affects every fipsagents 0.14.0 deployment that depends on cost_data accumulation.
• No data loss — sessions persistence is unaffected; only the cost accumulator is empty.
• Patch-level fix: cut 0.14.1.

Related

• Surfaced during cluster smoke for #116 / #104.
• HTTP backend's last-write-wins gap (fips-agents/fipsagents-platform#4) is a separate, unrelated issue.