fix(streaming): record Responses terminal usage instead of zero tokens

[alecchen]

Bug

Streaming spark requests logged and stored 0/0 tokens. The opencode dashboard showed real usage for the same calls.

Why

The Responses streaming path dropped upstream usage. ProxyResponsesStream (internal/transformer/stream.go) built the terminal Anthropic message_delta with usage hardcoded to zero, and processResponsesSSELine never handled response.completed, which is where upstream puts usage. extractUsageFromSSE (internal/handlers/messages.go) only reads message_delta chunks, so the log and history both got the hardcoded zeros. Chat streaming and non-streaming Responses read usage fine. This path was the only one that did not.

Fix

Read the terminal event and pass its usage through:

• pkg/types/zen.go: ResponsesChunk now decodes the nested response.usage shape alongside the flat one.
• internal/transformer/stream.go: stash usage from response.completed (flat first, nested second, last one wins) and emit it in the terminal message_delta through responsesUsageToAnthropic. A stream with no usage event still records 0.
• Both callers (handleResponsesStreaming, StreamProxy.ProxyStream) share ProxyResponsesStream, so one change fixes both.

Validation

• go test ./internal/transformer/ -run TestProxyResponsesStream: 8 passed (5 existing, 3 new covering flat, nested, and missing usage).
• go test ./... green; gofmt, go vet clean.
• Live: streaming spark call now shows nonzero tokens in the log and /api/history. Other models unaffected.

[kilo-code-bot]

Code Review Roast 🔥

Verdict: 1 Issue Found | Recommendation: Add CHANGELOG entry, then merge

Overview

Severity	Count
🚨 critical	0
⚠️ warning	0
💡 suggestion	1
🤏 nitpick	0

Issue Details (click to expand)

File	Line	Roast
CHANGELOG.md	10-16	You fixed a bug that makes spark model tokens actually show up in logs and /api/history instead of 0/0, but forgot to tell the CHANGELOG about it under [Unreleased] > ### Fixed. The git history isn't going to write itself, and future humans debugging this will want to know when the great token-reveal happened.

---

🏆 Best part: The author actually wrote tests for the bug they fixed — three of them covering flat nested, and missing usage on response.completed. I need to sit down. Not only that, but they verified existing tests (which already include bare response.completed lines) still pass, confirming the new handler is a clean no-op when no usage data is present.

💀 Worst part: The CHANGELOG got ghosted. You fixed a user-facing bug that changes what users see in their logs and dashboard, but left the changelog entry field blank. That's the kind of omission that turns a 5-minute blame search into a 3am rabbit hole.

📊 Overall: Like a surgeon's precision on a papercut — the fix is clean, tested, and correct, routing both flat and nested response.usage shapes through the terminal message_delta via responsesUsageToAnthropic, but don't forget to document it for posterity.

Files Reviewed (3 files)

• pkg/types/zen.go - 0 issues (new ResponsesResult type and Response field on ResponsesChunk, properly documented)
• internal/transformer/stream.go - 0 issues (new responsesUsageToAnthropic helper, response.completed capture in processResponsesSSELine, terminal message_delta usage wired up)
• internal/transformer/stream_test.go - 0 issues (3 new tests covering flat nested, and missing usage scenarios)

---

_{Reviewed by free · Input: 236.5K · Output: 35.7K · Cached: 820.4K}

_{Review guidance: REVIEW.md from base branch main (truncated)}