`streamResponse` delivers only a single final snapshot — no partial content

[artemnovichkov]

Description

session.streamResponse(to:generating:) with ClaudeLanguageModel never yields partial snapshots. The for try await loop receives exactly one element containing the complete response, after the full generation time. Same behavior on the Simulator and on a physical device. With SystemLanguageModel, the same code streams partials as expected.

Environment

• ClaudeForFoundationModels 0.1.1
• Xcode 27 beta, iOS 27 beta (Simulator and device)
• Model: .sonnet4_6, auth: .apiKey

Minimal reproduction

import FoundationModels
import ClaudeForFoundationModels

@Generable
struct Ideas: Equatable {
    @Guide(description: "Gift ideas", .count(5))
    var ideas: [String]
}

let model = ClaudeLanguageModel(
    name: .sonnet4_6,
    auth: .apiKey("YOUR_API_KEY")
)
let session = LanguageModelSession(model: model)

var snapshots = 0
let stream = session.streamResponse(
    to: "Suggest gift ideas for a sci-fi fan.",
    generating: Ideas.self
)
for try await partial in stream {
    snapshots += 1
    print("snapshot \(snapshots):", partial.content)
}
print("total snapshots:", snapshots) // 1 — arrives only when generation completes

Expected

Multiple cumulative PartiallyGenerated snapshots as content streams in (as with SystemLanguageModel).

Actual

A single snapshot with the complete result, delivered at the end.

Notes

I'm not sure the issue related to the framework or to Foundation Models, but maybe you know the solution to fix it. Thanks!

[trilorez]

Thanks for the report! The Claude api does not provide token counts on delta responses but FoundationModels appears to require a token count greater than zero to yield a partial snapshot. I've opened this fix for now that emits a placeholder token count on deltas and will track any FoundationModels updates to handle this more cleanly.

[artemnovichkov]

Thank you! Feel free to ping me if you need any testing!