[aizerin]

Summary

When OpenCode uses a custom or proxy provider (such as LiteLLM) without per-model cost configuration in opencode.json, OpenCode reports cost: 0 on completed assistant messages / generation steps.

Currently, traceGeneration() unconditionally sets:

"langfuse.observation.cost_details": JSON.stringify({ total: input.cost })

which exports {"total": 0} as provided_cost_details on the opencode.generation span.

In Langfuse's ingestion pipeline (IngestionService.calculateUsageCosts), any non-empty provided_cost_details takes precedence over model-based price calculation (if (providedCostKeys.length) return ...), causing total_cost to remain 0 even when a matching model definition with pricing tiers exists in Langfuse.

This change only sets langfuse.observation.cost_details when input.cost > 0, allowing Langfuse to compute generation costs from its configured model prices when OpenCode reports 0.

Verification

• Added an integration test in test/integration/v1.test.ts covering both step-started and fallback generation spans with cost: 0
• Ran pnpm run format, pnpm run build, pnpm run lint, pnpm run knip, and pnpm run test:integration

[CLAassistant]

All committers have signed the CLA.

[bezbac]

Hi @aizerin, thanks for your contribution!