Files
pipelines/docs/notes/chat-usage-metrics.md

1.9 KiB

Reported chat metrics — 2026-10-21

The chat header displays output tokens/sec, reported token totals, last measured prompt/context usage and model context limit, and harness-reported USD cost. Missing data remains unavailable; reported zeros remain zero. No text-length tokenizer, default context limit, subscription-price guess, or local billing estimate is used. Session snapshots replace previous values and never admit or complete queued work or create transcript messages.

OpenCode V2 counters use mutually exclusive native input/output/reasoning/cache components. Context and rate come from a bounded recent matching-model assistant read, so an in-flight message with no usage does not erase the latest known measurement. A compaction boundary prevents reviving pre-compaction prompt usage. Request-to-stream timing includes request latency and excludes tool settlement. Codex's native totals/context snapshots are used directly without double-counting cached-input or reasoning-output subsets; rate and cost remain unknown where the native protocol provides no compatible measurements.

Full Go race/build/vet checks and web regressions passed. Browser fixtures cover 390/420/1100px and a short viewport, known/partial/unknown/zero values, duplicate snapshots, and preserved draft, focus, selection and transcript DOM. Existing compaction, tool-summary and composer browser checks also passed.

Clean runtime code 0b8595f16e6352baec8ccc8e389a8783d3af4f52 was deployed with a SQLite/binary rollback backup after checking pipeline work and both native chat backends. Only the Pipelines server restarted; other services were untouched. An authenticated live browser received and formatted all four OpenCode metrics for this existing owned chat, with shell cache v12. Verification initiated no new model turn, session, pipeline run, or provider generation. Private numeric usage and credentials are not copied into this note.