[github-actions]

Automated update of Lingua provider types.

Provider: google

Publication mode: needs-review draft

Continuation required

This draft preserves the automation's partial work. One or more validation or advisory checks did not complete successfully; continue from this branch instead of restarting the provider update.

Feedback: comment /bt good or /bt bad to log review feedback to the Braintrust trace.

Human decisions required

The agent completed every unblocked item. These decisions require human input before the update can be finished:

• FunctionResponsePart.fileData / FunctionResponseFileData / FunctionResponseBlob.displayName (google-function-response-media-parts)
  • Question: What is the canonical, non-lossy universal representation for media carried in a Google FunctionResponse.parts (inline bytes or fileData URI, plus mimeType and displayName) together with the {"$ref": "DISPLAY_NAME"} indirection that FunctionResponse.response uses to point at it? Specifically: should Lingua (a) reject any functionResponse that carries parts with an explicit unsupported-mapping error until a universal contract exists, (b) add a universal tool-result media representation that preserves the payload and the name-based linkage as a unit, or (c) resolve the $ref inline so that the universal tool-result output embeds the media directly?
  • Evidence: crates/lingua/src/providers/google/convert.rs:439-458 imports a functionResponse by reading only name, id, and response; FunctionResponse.parts (crates/lingua/src/providers/google/generated.rs:397) is never read. crates/lingua/src/providers/google/convert.rs:708-713 exports with ..Default::default(), so parts is never written. Because convert.rs:441-443 copies fr.response verbatim into ToolResultContentPart.output while the parts carrying the referent are discarded, a Google request with a multimodal function response transforms to Anthropic or chat-completions today with no error and emits a structurally valid but semantically broken tool result whose $ref points at nothing. This spec update makes the problem materially worse by adding a second referenceable carrier: FunctionResponsePart.fileData with the new FunctionResponseFileData { display_name, file_uri, mime_type } (generated.rs:435,452-468) and FunctionResponseBlob.display_name (generated.rs:481-482), both documented as referenceable via {"$ref": "DISPLAY_NAME"} from FunctionResponse.response. crates/lingua/src/universal/message.rs:241-249 shows ToolResultContentPart is { tool_call_id, tool_name, output: Value, custom_tool_call, caller, provider_options } with no media member, and crates/lingua/src/providers/anthropic/convert.rs:490-510 shows the Anthropic importer flattens its tool-result blocks into an opaque Value, so no provider-neutral tool-result media contract exists to reuse. AGENTS.md:19 and AGENTS.md:65 both forbid the current partial conversion.
  • Recommended option: Take option (a) now: extend the Google provider-only rejection so that any functionResponse whose parts array is non-empty returns ConvertError::UnsupportedMapping { from: "Google functionResponse.parts", to: "Lingua universal message" } on the request and response import paths, and leave native google->google passthrough byte-preserving. This is the smallest change that stops the dangling-$ref silent loss immediately, it matches the existing explicit-rejection precedent at convert.rs:60-76, it requires no universal model change, and it does not foreclose option (b) later. Option (b) should be pursued as a separate, deliberately designed universal tool-result media contract covering Anthropic, OpenAI, and Google together, not as a Google-driven addition.
  • Alternatives:
    • Option (b): add a universal tool-result media part (for example a ToolResultContentPart media list carrying { data | uri, media_type, name }) and map Google's parts onto it, preserving display_name as the member name. Upside: genuinely portable multimodal tool results, and the chat-completions and Anthropic importers could later converge on it. Downside: the $ref indirection has no analogue in Anthropic or OpenAI tool results, so the universal model must either keep the raw $ref string (which only Google can resolve, making the representation provider-coupled) or resolve it; it also expands the universal surface based on one provider's encoding, and it requires payload cases plus regenerated transform snapshots for every provider pair.
    • Option (c): resolve the $ref at import time by substituting the referenced part's media into FunctionResponse.response, then map the result onto an existing universal shape. Upside: no new universal type, and the output becomes self-contained. Downside: this is a lossy, semantics-changing rewrite of the provider's payload (the distinction between an inline payload and a named reference is destroyed, the export direction cannot reconstruct the original, and an unresolvable or duplicated display_name has no defined behavior); it contradicts AGENTS.md:19 and AGENTS.md:27.
    • Option (d): leave the current behavior and accept the silent drop. Rejected: it produces a dangling $ref in the transformed output, which is a correctness bug rather than a documented limitation, and it directly violates AGENTS.md:65.
  • Likely files:
    • crates/lingua/src/providers/google/convert.rs
    • crates/lingua/src/providers/google/generated.rs
    • crates/lingua/src/providers/google/adapter.rs
    • crates/lingua/src/universal/message.rs
    • crates/lingua/tests/google_server_tool_call.rs
    • crates/lingua/tests/openai_function_call_output_tool_name.rs
    • payloads/cases/params.ts
    • crates/coverage-report/src/requests_expected_differences.json
  • Validation commands:
    • cargo test -p lingua --lib providers::google
    • cargo test -p lingua --test google_server_tool_call
    • cargo test -p lingua --test openai_function_call_output_tool_name
    • cargo test -p lingua --test import_fixtures
    • make test-payloads
    • cargo test -p coverage-report --test cross_provider_test cross_provider_transformations_have_no_unexpected_failures
    • make typed-boundary-check
    • make typed-boundary-check-branch BASE=main

Provider-only changes

These native wire features are accepted and passed through unchanged, but cross-provider transformation is intentionally unsupported:

• FunctionResponsePart.mediaProcessing (google-function-response-part-media-processing)
  • Native contract: mediaProcessing anywhere inside a Google part, including nested under functionResponse.parts, must be accepted by the generated types, preserved byte-for-byte on google->google passthrough, and rejected with ConvertError::UnsupportedMapping naming mediaProcessing when transformed to any other provider format, on the request, response, and streaming paths alike.
  • Scope decision: mediaProcessing selects how Google's own model consumes video inside provider-hosted inference; AGENTIC is explicitly model-driven dynamic navigation. It has no meaning outside Google and no target provider has an equivalent knob. The repository already treats the top-level field this way at convert.rs:63, so this item only extends the existing provider-only boundary to the newly reachable nesting level.
• ToolType.MEDIA_PROCESSING (google-tool-type-media-processing)
  • Native contract: Both halves of a Google server-tool exchange must be handled consistently: toolCall and toolResponse, for every toolType including MEDIA_PROCESSING, must be accepted by the generated types, preserved byte-for-byte on google->google passthrough, and rejected with ConvertError::UnsupportedMapping naming the offending field when transformed to any other provider format.
  • Scope decision: ToolCall and ToolResponse with a ToolType describe Google-hosted search, maps, URL context, file search, and now media processing: provider-executed invisible tools whose results are produced by Google infrastructure. These are not caller-defined function tools and have no portable contract; the repository already classifies the toolCall half this way at convert.rs:67. This item extends the same boundary to the toolResponse half that the new variant makes reachable.
• AudioTranscriptionConfig.sourceText (google-audio-transcription-source-text)
  • Native contract: A Google request whose generationConfig.audioTranscriptionConfig carries sourceText together with mode SMART must parse into the generated types, be detected as Google format, pass through google->google byte-for-byte, and be rejected with ConvertError::UnsupportedMapping naming generationConfig.audioTranscriptionConfig on any cross-provider transform.
  • Scope decision: sourceText configures Google's hosted SMART transcription pass, changing whether the candidate carries a complete replacement or an append delta. It is provider-executed audio understanding with no target-provider equivalent, and it is already inside an established provider-only boundary; this item keeps the native type current and pins the boundary with the new field present rather than widening it.

Validation

• ./pipelines/generate-provider-types.sh google: success
• make generate-types PROVIDER=google: success
• Braintrust workflow trace: success
• Claude integration plan: success
• Structured plan validation: success
• Human design blockers: true
• Claude focused implementation: success
• Immutable plan revalidation: success
• Initial post-implementation Rust regeneration: success
• Initial provider update path policy: success
• Initial provider semantic policy: success
• Initial post-implementation TypeScript regeneration: success
• Initial formatting: success
• Initial focused provider tests: success
• Initial conditional generator tests: success
• Initial build: failure
• Initial clippy: failure
• Bounded Claude mechanical repair: success
• Effective mechanical validation source: post-repair
• Effective mechanical validation: failure
• make lingua-wasm: skipped
• Unblocked payload capture cases: ``
• Live capture (OpenAI): skipped
• Live capture (Anthropic): skipped
• Live capture (Google): skipped
• Payload cross-provider transform capture: skipped
• Payload fixture sync: skipped
• make test-payloads: skipped
• make typed-boundary-check: skipped
• cargo test -p coverage-report --test cross_provider_test cross_provider_transformations_have_no_unexpected_failures: skipped
• Claude read-only verification: skipped
• Verification report validation: skipped
• Verification verdict:
• Recoverable binary patch archive: success
• Patch artifact upload: success

Only safely scoped patches with a locally archived binary diff are published. Any incomplete AI phase or failed deterministic check produces a needs-review draft so work can continue from the PR branch.