[Bug] OpenAI-compatible providers never request streaming usage, so Ollama and similar servers report no tokens #1658

Description

createOpenAICompatible(...) in source/ai-sdk-client/providers/provider-factory.ts (lines 532-538) is created without includeUsage: true, so the SDK never sends stream_options: {include_usage: true}. Ollama and some other servers only send token usage in a streamed reply when asked, so Nanocoder gets none. The per-response usage indicator falls back to estimates, and real token counts, cost and cache-hit data are missing. The --json report from run has no usage, and the agent eval harness records 0 tokens for local models (benchmarks/agent/harness.ts:441-443 uses usage.totalTokens ?? 0).

Environment

• OS: Windows 11 (10.0.26300)
• Node version: 22.22.3
• Nanocoder version: 1.31.0 (main at 5c2e693)
• Provider: Ollama 0.35.1 (OpenAI-compatible endpoint)
• Model: qwen2.5:1.5b

Steps to Reproduce

1. Configure an Ollama provider.
2. Put a small logging proxy between Nanocoder and Ollama (or inspect the request another way).
3. Run nanocoder --trust-directory --json run "Reply with the single word OK".
4. Look at the request body and at the JSON report.
5. As a control, send the same request to Ollama by hand, once with "stream_options":{"include_usage":true} and once without.

Expected Behavior

The request contains "stream_options": {"include_usage": true}, and the report includes token usage.

Actual Behavior

The request keys are only model, messages, tools, tool_choice, stream, and the streamed reply has no usage. The report has no usage field:

keys: kind, exitCode, finalText, reasoning, toolCalls, steps, filesChanged
usage: undefined

The control request shows the server does send usage when asked:

with stream_options:    "usage":{"prompt_tokens":31,"prompt_tokens_details":{"cached_tokens":5}, ...
without stream_options: 0 usage chunks

Suggested Fix

• Pass includeUsage: true to createOpenAICompatible in provider-factory.ts (the SDK option exists in @ai-sdk/openai-compatible).
• Consider a per-provider setting to turn it off, for any server that rejects stream_options.
• Add a test in provider-factory.spec.ts: start a small local HTTP server, stream one reply through the provider, and check that the request body has stream_options.include_usage === true and that the usage comes back.
• Related: the harness could store missing usage as null instead of 0.

Additional Context

• Related: feat(usage): show API-reported context usage with an estimate fallback #543 (API-reported usage with an estimate fallback), [Feature] Token Usage & Cost Indicator per Response #756 (usage indicator), open PR feat: (P1)track prompt prefix stability and report cache hit rate #1649 (cache hit rate, which needs this data).
• I have searched existing issues to ensure this is not a duplicate
• I can reproduce this issue consistently
• This issue occurs with the latest version of nanocoder