[pcbeingused333]

What

ChatResult.CompletionUsage.PromptTokensDetails now decodes with decodeIfPresent, defaulting a missing audio_tokens or cached_tokens to 0. A regression test decodes a response whose prompt_tokens_details contains only cached_tokens.

Why

audioTokens and cachedTokens were decoded as required Ints. OpenAI-compatible providers (DeepSeek, Gemini's compat endpoint) often send only cached_tokens. Decoding prompt_tokens_details then threw, and the try? that guards usage in ChatResult and ChatStreamResult swallowed the error, so the entire usage (including prompt_tokens, completion_tokens, total_tokens) came back nil instead of just lacking the missing detail.

Affected Areas

ChatResult.CompletionUsage.PromptTokensDetails decoding, which is used by both ChatResult and ChatStreamResult. The property types stay Int, so there is no public API change. (Making them Int?, like CompletionTokensDetails, would be an alternative but source-breaking.)

Tests: full swift test in a swift:6.2 container (223 tests) passes; the new test fails without the change (XCTUnwrap gets a nil usage). No linked issue.

---

This change was prepared with the help of an AI assistant (Claude Code), and the author is responsible for the change.

🤖 Generated with Claude Code

[nezhyborets]

Thanks for the PR, please see the comment

[nezhyborets]

Hi, thanks! Please see other uses of decodeIfPresent, the usage of parsingOptions. Looks something like this

public init(from decoder: any Decoder) throws {
    let container = try decoder.container(keyedBy: CodingKeys.self)
    let parsingOptions = decoder.userInfo[.parsingOptions] as? ParsingOptions ?? []
    self.id = try container.decodeString(forKey: .id, parsingOptions: parsingOptions)
    ...
}

[pcbeingused333]

Done, now via decodeInt(forKey:parsingOptions:) like the rest of ChatResult; the test uses .relaxed.

This comment was drafted with Claude Code.

[nezhyborets]

Sorry, just one more change before merging. Let me also see if Copilot (balanced) wants something else

Edit: Copilot didn't provide anything meaningful so let's just add the public init

[nezhyborets]

Declaring init(from:) in the struct body stops Swift from synthesizing the memberwise init(audioTokens:cachedTokens:), so PromptTokensDetails can now only be created by decoding, even from @testable tests. Could you either move init(from:) into an extension, or add a public init(audioTokens:cachedTokens:) like CompletionTokensDetails has? A public one would also make the promptTokensDetails: parameter of CompletionUsage.init usable for anyone mocking usage.

[Copilot]

🟡 Changes recommended

Missing detail fields still cause default-configured clients to discard the entire usage object, and the regression test’s relaxed parsing setting does not cover this default behavior.

1 open issue

• Default missing usage details to zero during decoding · New

What changed in this PR

This PR aims to preserve chat usage data when OpenAI-compatible providers omit prompt-token detail fields, without changing the public Int property types.

Changes:

• Adds custom decoding for prompt-token details using parsing options, shared by chat and streaming responses.
• Adds a regression test for a response containing only cached_tokens, with relaxed parsing enabled.

File	Description
Tests/​OpenAITests/​ChatResultTests.swift	Tests usage decoding when audio_tokens is omitted.
Sources/​OpenAI/​Public/​Models/​ChatResult.swift	Adds parsing-option-aware decoding for prompt-token details.

_{Review effort: Balanced}

[nezhyborets]

Also just ran the checks to see if they find anything that needs fixing

[pcbeingused333]

Added public init(audioTokens:cachedTokens:) to PromptTokensDetails (same shape as CompletionTokensDetails), plus a small test that builds CompletionUsage with it. ChatResultTests pass.

This comment was drafted with Claude Code.