Repository navigation
V2: align Anthropic streaming usage aggregation with upstream SDKs #40713
Description
Problem
The native Anthropic Messages parser currently normalizes usage independently on message_start and message_delta, then merges normalized Usage values. Anthropic usage is split across those events, and newer responses can include usage.iterations for compaction, message, advisor, and fallback inference.
This makes correct aggregation harder than it needs to be and risks losing or undercounting partial usage details. Vercel AI SDK instead accumulates raw Anthropic usage across stream events and normalizes once when message_stop arrives.
Follow-up scope
- Accumulate typed raw Anthropic usage in parser state across
message_startandmessage_delta. - Normalize once at
message_stoprather than merging partially normalizedUsageobjects. - Match Anthropic and AI SDK nullability for streamed usage fields.
- Define and test
usage.iterationsaccounting:- include executor
compactionandmessagework - keep
advisor_messageseparately attributable - avoid double-counting fallback responses
- handle cache-read and cache-creation counters consistently
- include executor
- Preserve the complete provider payload in
providerMetadata.anthropic. - Keep normalized
Usageinvariants intact when only partial counters are available.
Acceptance criteria
- Fixtures cover split start/delta usage, compaction, advisor, and fallback cases.
- Normalized totals match the current Anthropic/AI SDK semantics.
- No provider metadata has to be reparsed to reconstruct normalized usage.
Activity
Zaczero commentedon Aug 18, 2026
The Anthropic merge-normalized-
Usagerisk named here already has a worse sibling on the OpenAI Chat path: usage is replaced, not merged.opencode/packages/ai/src/protocols/openai-chat.ts
Line 692 in 9e22c40
The last chunk that carries any usage object wins outright. A stream that emits
prompt_tokens_detailson one chunk and bare totals on the next drops the cache read/write. Every OpenAI-compatible provider reuses this protocol, so litellm / DeepSeek-compat / GLM-compat all have the same exposure.Anthropic is the better of the two — it folds per-field:
opencode/packages/ai/src/protocols/anthropic-messages.ts
Lines 709 to 716 in 9e22c40
But it still merges already-normalized
Usage, which is exactly the gap this issue describes. Accumulating the raw events and normalizing once atmessage_stop/ the final Chat chunk fixes both parsers the same way.mapUsagealso stashesproviderMetadata: { openai: usage }. That object is not on the persisted message (105k assistant rows, none carry the key), so a later replay cannot recover the dropped details. The fold has to happen while the stream is still in hand. #34022 is the “cache read is zero” report this replace would produce on a multi-usage stream.neriousy commentedon Oct 21, 2026
Partly done: #46431 added typed
AnthropicUsagewithiterations.mapUsagesums the iterations and takes context tokens from the last one. The parser still merges usage normalized separately at message_start and message_delta, advisor usage isn't counted separately, and nothing de-duplicates fallback responses. Closing this tracker.