Repository navigation
Conversation
|
The following comment was made by an LLM, it may be inaccurate: Related PRs Found#12721 - feat(tui): add tokens per second to response footer
#42372 - feat(app): show tokens-per-second in context usage indicator
The PR description explicitly acknowledges these relationships and indicates that #12721 should take precedence if rebased. The current PR (#46108) is a cleaner implementation of the same design on a current branch. |
|
Tested this PR on Windows 11 (native, Windows Terminal, bun 1.3.14) from
Example from a quick session — same model, back-to-back responses:
That per-response variance visibility is exactly the use case described in #43857 — the slower response was same-model, so it was invisible until now. Works as expected on Windows; hope this can get merged. |
|
ae0e69a to
8a662fd
Compare
8a662fd to
7d2cd35
Compare
7d2cd35 to
dbf02b8
Compare
Includes: - feat(tui): show tokens per second in the response footer - refactor(core): drop the unused isValidForTPS export - fix(core): report tokens per second on tool-call turns Co-authored-by: João Carlos Magalhães de Castro <[email protected]>
dbf02b8 to
1de08bd
Compare
|
This branch has been upgraded to V2. We're no longer taking PRs for V1. If you think this is still relevant, please port it over and open it for V2. — from 𝕺𝖕𝖊𝖓𝕮𝖔𝖉𝖊 |
Issue for this PR
Closes #6096
Type of change
Credit
This is @JohnC0de's #12721 rebased onto current
dev— same design (TPS calc incore/so non-TUI consumers can use it,firstTokenrecorded inprocessor.ts), which is itself a clean take on @edlsh's #5497. #12721 only went stale — it'sdirtyagainstdevnow. If @JohnC0de rebases it, theirs should land and I'll close this. Take anything from here that helps.What does this PR do?
Adds a
tok/sreadout to the assistant footer, after the duration:18.3s · 131 tok/s.processor.tsstampstime.firstTokenon the first streamed event (tool-only turns included). TPS isgeneratedTokens / ((completed − firstToken) / 1000), output plus reasoning. The calc lives inpackages/core/src/session/tokens.tsso the SDK can reuse it;firstTokenis optional on the message schema and persisted, so a session-level average can come later without a migration.One departure from #12721: tool-call turns are not filtered out. A turn that ends in a tool call is one provider step with its own usage and timestamps, and in practice it's most turns — on my box over three days, 83% of assistant turns in top-level sessions and 98% in subagent sessions end in
tool-calls, so filtering them made the readout look intermittent. Sampled tool-call turns give ordinary rates (47–183 tok/s). Errored turns, summaries, and unfinished turns are still skipped, and the 250 ms floor stays — it's what suppresses non-streamed responses that arrive in one chunk (a 228-token turn landing in 43 ms would otherwise read as 5,000 tok/s).How did you verify your code works?
The timestamp guard is compiler-enforced, not test-enforced: delete the narrowing and it's a type error (
'completed' is possibly 'undefined'). The zero-token, sub-threshold, rounding, and finish-reason guards each go red when reverted.packages/coreTPS tests: 17 pass / 0 failpackages/opencodefull suite: 3396 pass / 0 failbun typecheckclean inopencode,core,tuiScreenshots / recordings
Footer gains a
· NN tok/ssegment; no new surface.Checklist