feat(ai): expose inputMessages and partialText in onAbort callback

[Hovo-Dev]

Background

When a streamText call is aborted — via AbortSignal, the user stopping generation, or a dropped connection — the onAbort callback fires but receives no usage data. onFinish never fires on abort at all. This makes it impossible to track token consumption for billing or quota purposes when streams don't complete naturally.

Summary

packages/ai/src/generate-text/stream-text.ts

• Added usage?: LanguageModelUsage and totalUsage?: LanguageModelUsage to StreamTextOnAbortCallback
• Added inputMessages: ModelMessage[] — the full message list passed to the model for the aborted step (initial messages + completed prior step tool
  call/result pairs)
• Added partialText?: string — the text actively being streamed at abort time; undefined if no text was generated; resets at each step boundary
• Introduced currentStepUsage, completedStepsUsage, currentStepInputMessages, and pullLevelTextContent as outer-scope tracking variables
• Wired all fields into the existing abort() function — if the provider had not yet sent usage data before the abort, usage is undefined (honest, no
  estimation)

packages/openai/src/chat/openai-chat-language-model.ts

• OpenAI's finish chunk (which carries real usage) was only emitted in flush(), which never runs on abort
• Now emits the finish chunk as soon as usage data arrives in the stream transform(), with a finishSent guard to prevent double-emission in flush()

content/docs/07-reference/01-ai-sdk-core/02-stream-text.mdx U

• Documented usage, totalUsage, inputMessages, and partialText fields in the onAbort callback parameters

content/docs/03-ai-sdk-core/50-error-handling.mdx

• Updated onAbort section to describe all four new fields

Usage

onAbort({ totalUsage, usage, inputMessages, partialText }) {                                      
  const completedTokens = totalUsage?.totalTokens ?? 0;                                                                                                     

  // Use real usage if provider sent it, otherwise estimate with your own tokenizer:
  const abortedStepTokens = usage?.totalTokens                         
    ?? (countTokens(inputMessages) + countTokens(partialText ?? ''));                                                                                       
                                                                                                                                                            
  trackUsage(completedTokens + abortedStepTokens);                                                                                                          
}                                                                                                                                                           

Manual Verification

Verified via automated tests that exercise the full stream processing pipeline:

1. Abort fires before any usage chunk → usage and totalUsage are undefined
2. Abort fires after the provider sent a usage chunk → usage and totalUsage contain the real token counts
3. Multi-step stream, abort in step 2 → totalUsage correctly sums step 1 and step 2 usage
4. Abort before first chunk → inputMessages contains the step's input, partialText is undefined
5. Abort mid-text → partialText contains the streamed text up to abort
6. Multi-step abort → inputMessages includes tool call/result pairs from prior steps

All 2200+ existing tests continue to pass.

Checklist

• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
• I have reviewed this pull request (self-review)

Related Issues

Fixes #7628
Fixes #7805

[Hovo-Dev]

Hi @gr2m,

Took a look at this and managed to get it working cleanly — happy to share the approach.

The SDK was already holding the token data we needed, it just wasn't being handed off when a stream was aborted. We also noticed that usage from OpenAI arrives mid-stream but was only forwarded at natural close, so we made sure it gets captured as soon as it arrives. That way, whether the stream finishes or gets cut short, the data is there.

Looking forward to your feedback!

[Hovo-Dev]

@lgrammel - Could you please take a look?

[reactsaas]

can someone explain is token usage returend by the model on abort or is like usage or token info avaible mid stream ??

currently having problem with abortsignal getting no usage at all with gemini model

[Hovo-Dev]

@reactsaas Hey good to see you here! Actually, abort signal neither returns usage nor any data (history or partial generated text) to count the tokens manually, that's why you have to merge the generated stream data on fly with OnChunk method and then count the tokens accordingly. Given that, my current PR is especially designed to return the conversation list and partially generated text, so the client can at least count the tokens with ease without merging the mid-generated stream data on the fly and get rid of boilerplate code.

Example of a usage with the new release.

onAbort({ totalUsage, usage, inputMessages, partialText }) {                                      
  const completedTokens = totalUsage?.totalTokens ?? 0;                                                                                                     

  // Use real usage if provider sent it, otherwise estimate with your own tokenizer:
  const abortedStepTokens = usage?.totalTokens                         
    ?? (countTokens(inputMessages) + countTokens(partialText ?? ''));                                                                                       
                                                                                                                                                            
  trackUsage(completedTokens + abortedStepTokens);                                                                                                          
}                                                                                                                                                           

[reactsaas]

@reactsaas Hey good to see you here! Actually, abort signal neither returns usage nor any data (history or partial generated text) to count the tokens manually, that's why you have to merge the generated stream data on fly with OnChunk method and then count the tokens accordingly. Given that, my current PR is especially designed to return the conversation list and partially generated text, so the client can at least count the tokens with ease without merging the mid-generated stream data on the fly and get rid of boilerplate code.

Example of a usage with the new release.

onAbort({ totalUsage, usage, inputMessages, partialText }) {                                      
  const completedTokens = totalUsage?.totalTokens ?? 0;                                                                                                     

  // Use real usage if provider sent it, otherwise estimate with your own tokenizer:
  const abortedStepTokens = usage?.totalTokens                         
    ?? (countTokens(inputMessages) + countTokens(partialText ?? ''));                                                                                       
                                                                                                                                                            
  trackUsage(completedTokens + abortedStepTokens);                                                                                                          
}                                                                                                                                                           

Hey thanks for the answer. I took your aproach now counting the tokens. 🍪

[Hovo-Dev]

@reactsaas Glad to hear that, but I believe you can use it once this PR is merged.