Together AI: token usage always 0 — bundled @ai-sdk/togetherai doesn't send stream_options.include_usage

Summary

Token usage and cost are recorded as 0 for all Together AI models (tested: zai-org/GLM-5.2, zai-org/GLM-5.3, moonshotai/Kimi-K3) on opencode 1.18.29. The UI shows 0 tokens / $0 cost, and the values stored on assistant messages in opencode.db are all zero — this is not a display bug.

Root cause

1. opencode's outbound streaming requests to Together do not include stream_options: {include_usage: true} (verified by capturing the actual request body through a local proxy).
2. Together's current serving stack only includes usage in streaming responses when that flag is set. Verified with curl against https://api.together.xyz/v1/chat/completions:
   • Without stream_options.include_usage: GLM-5.2/5.3 streams contain no usage at all; Kimi-K3 emits usage only in a non-standard location (inside choices[0], which the AI SDK parser ignores).
   • With the flag: all three models return a standard top-level usage chunk.
3. opencode force-sets includeUsage: true only for providers whose npm package is @ai-sdk/openai-compatible (packages/opencode/src/provider/provider.ts, resolveSDK). models.dev routes togetherai through the custom @ai-sdk/togetherai package, so the force-set never applies, and the bundled version of that package does not request usage.
4. This used to work: messages recorded real usage through 2026-07-17. Together changed their serving stack around 2026-07-18→28 (SSE chunk IDs changed from chatcmpl-... to e.g. 9ac28c1f-aws_uw2), after which usage stopped being included by default. Same opencode version, same models — behavior changed server-side, so this affects all opencode versions, not just recent ones.

Suggested fix

@ai-sdk/togetherai@3.0.45 hardcodes includeUsage: true on its chat models (it wraps OpenAICompatibleChatLanguageModel with includeUsage: true). Bumping the bundled @ai-sdk/togetherai dependency to >= 3.0.45 should restore usage reporting with no other changes.

Note: routing togetherai through @ai-sdk/openai-compatible instead is not a good fix — opencode's transform layer gates Together's reasoning_effort parameter on the npm string being exactly @ai-sdk/togetherai (packages/opencode/src/provider/transform.ts), so switching packages silently drops reasoning-effort control.

Workaround attempted (for completeness)

Overriding the provider in config with "npm": "@ai-sdk/openai-compatible" + options.baseURL + explicitly declaring each model in provider.togetherai.models does restore token recording (verified: usage and cost land in the DB), but reasoning_effort is no longer sent, so it's a tradeoff most users shouldn't have to make.

Environment

• opencode 1.18.29 (linux x64)
• Provider: togetherai (API key via /connect)
• Models: moonshotai/Kimi-K3, zai-org/GLM-5.3, zai-org/GLM-5.2