fix: enable token usage tracking and configurable stream timeout for Ollama provider

[kasjens]

Summary

Three related fixes for the Ollama provider:

1. Token usage tracking: The provider was unconditionally stripping stream_options: {"include_usage": true} from requests (fix: prevent Ollama provider from hanging on tool-calling requests #7723), preventing Ollama from returning token counts in streaming responses. This is now gated behind OLLAMA_STREAM_USAGE (default: true) so modern Ollama builds get usage tracking while older builds can opt out with OLLAMA_STREAM_USAGE=false. Invalid values are handled safely — a warning is logged and stream_options is disabled.

2. Fallback usage parsing: Added fallback for Ollama-native token fields (prompt_eval_count, eval_count) in get_usage(), so token tracking works even when Ollama doesn't translate to standard OpenAI field names (prompt_tokens, completion_tokens). OpenAI fields take precedence when both are present. Null OpenAI fields (e.g. "completion_tokens": null) correctly fall through to the Ollama-native fields instead of silently dropping usage.

3. Configurable stream timeout: The hardcoded 30s per-chunk timeout was too aggressive for slower models (CPU inference, large parameter counts, complex reasoning). The timeout is now configurable via a resolution chain: OLLAMA_STREAM_TIMEOUT > GOOSE_STREAM_TIMEOUT > OLLAMA_TIMEOUT > default (120s). Zero values are treated as invalid and skipped to prevent immediate stall errors on every chunk.

Testing

• All 29 related tests pass (zero failures)
• 6 timeout resolution tests: verify the fallback chain, default behavior, and zero-value rejection
• 3 usage parsing tests: verify Ollama-native field fallback, OpenAI field precedence, and null OpenAI field fallthrough
• 2 stream_options gating tests: verify OLLAMA_STREAM_USAGE default-on and opt-out behavior

Related Issues

Relates to #8479
Relates to #8476
Relates to #7723

[jamadeo]

Nice one, thank you @kasjens . Is it possible we can avoid the config parameter by looking at the ollama version?

[jamadeo]

Can we do without a config flag? Maybe https://docs.ollama.com/api-reference/get-version and flag this on a minimum version?

[kasjens]

Thanks @jamadeo! I considered that but went with the config flag because:

1. Ollama isn't always direct — users behind proxies or compatible API servers (LiteLLM, LocalAI) may not expose /api/version.
2. Version ≠ capability — custom builds/forks may support stream_options without a recognizable version string.
3. It defaults to enabled — so modern installs work out of the box. Only users on older builds need to set OLLAMA_STREAM_USAGE=false.

Happy to add version detection as a best-effort first pass with the config flag as fallback if you'd prefer that approach!

[jamadeo]

Makes sense, and seems like most of the time you'd never add it at all

[michaelneale]

I think this is worth having in, thanks!