OpenCode's per-message token counts exclude cache reads from input and reasoning from output

Fact

OpenCode (v2) stores sessions in a SQLite file, opencode.db under its data directory (~/.local/share/opencode on Linux and macOS), not in JSONL. Each model step is one session_message row with type = 'assistant', whose JSON data holds tokens: {input, output, reasoning, cache: {read, write}}. input is uncached input only and output excludes reasoning, so the four buckets are disjoint. cost is 0 for subscription providers.

Why it matters

Adding cache.read to input as if it were a subset, or reading output as the billed output, undercounts reasoning models: a step can show output 68 beside reasoning 232. Summing session_v2's per-session totals instead of the rows also drifts, because those include aborted steps whose rows carry no tokens at all. A tool that runs a model is not its provider, either: Claude Code can point at a GLM endpoint and OpenCode can drive Claude.

How to apply

• Bill input at the input rate, output + reasoning at the output rate and cache.read at the provider's cached-input rate, which is not always a fixed fraction of input (GLM's is not).
• Open the file read-only (file:…?mode=ro as a URI); it is in WAL mode, so watch opencode.db-wal too when detecting changes.
• Skip rows without tokens; message ids are unique, so no de-duplication.
• Derive the provider from the model id when grouping by subscription.