Files
CLIProxyAPI/internal/runtime/executor/helps
sususu98 4b2beb3da1 feat(executor): measure effective TTFT with protocol-aware token classification (#5313)
* feat(executor): measure effective TTFT with protocol-aware token classification

- Introduce protocol-aware streaming token classification under internal/runtime/executor/helps
- Implement responses_ttft_helpers.go for OpenAI Responses / Codex WebSocket & SSE streaming
- Filter out container metadata, rate limits, handshake headers, and tool execution outputs
- Implement dual-track firstPacketDuration fallback for abnormal / zero-token stream terminations
- Add TrackHTTPClientRoundTripOnly to bypass false TTFT triggers on first response body reads
- Optimize UsageReporter.ObserveTokenEvent hot path with fast-path RLock short-circuiting (0 B/op)
- Align terminal event completions across SSE, WebSocket streaming, and non-streaming executions
- Add EnsurePublished metric recording fallback for successful terminal turns lacking usage blocks
- Add stub helpers with comprehensive specifications for Chat, Claude, and Gemini protocols

* fix(executor): normalize SSE response.done and terminate incomplete WS streams

* fix(usage): capture first-packet fallback on initial body reads with TrackHTTPClientRoundTripOnly
2026-08-29 01:22:21 +08:00
..