V4 alignment: vendor package alignment #231

Why. Several vendor packages diverge from their pinned upstream packages. The main case: when layer E1 merged, openai-compatible lost its settings for extension points. Other upstream subpaths and fields are also missing.

Fixed in.

• B1 to B3: layer E1 (and C1 for B2).
• B21: layer B6.
• N3: layer C1.
• The rest: alignment migration.

Existing fixes. For B1, the D line still has the closure implementation (fb0f79e5: aimux-providers/src/openai_compatible/mod.rs:90-92, config.rs:154). Layer E1 (7330a6e) removed it. Restore the D-line version.

Checklist

• B1 (U22): transformRequestBody and metadataExtractor were deleted from openai-compatible.
  • Evidence: v4: aimux-providers/src/openai_compatible/ has neither.
  • Doc: doc:159 (D15), :797, :1571. Upstream: @ai-sdk/openai-compatible@3.0.62 openai-compatible-provider.ts:107, :114, :185-186, openai-compatible-chat-language-model.ts:401, :451-452 (the extractor also covers streams).
  • Fixed in: layer E1. Restore (decided).
• B2: supportsMultiPartToolContent is missing.
  • Evidence: layers B3 to B8 had an internal field (config.rs:133); C1 deleted it, and no layer exposes a setting.
  • Upstream: openai-compatible-provider.ts:100, :183, convert-to-openai-compatible-chat-messages.ts:183, :304.
  • Fixed in: layer C1, exposed as a public setting (decided).
• B3: convertUsage became a closed enum, and supportedUrls became static data.
  • Evidence: v4: aimux-providers/src/openai_compatible/config.rs:57, :130-131, mod.rs:84-86. The D line has closures (fb0f79e5 config.rs:25, :131).
  • Upstream: openai-compatible-provider.ts:119, :125 (functions).
  • Fixed in: layer E1. Keep the closures (decided).
• B21: the Anthropic stream never emits Raw, and the include_raw_chunks comment is stale.
  • Evidence: no StreamPart::Raw in v4: aimux-providers/src/anthropic/stream.rs. v4: aimux-core/src/options.rs:203-213 says only the OpenAI-compatible family honours it; 18 files do.
  • Doc: doc:109, RFC-0036:43. Upstream: @ai-sdk/anthropic@4.0.68 anthropic-language-model.ts:1948-1949.
  • Fixed in: layer B6.
• N3: stale doc comments say the status code is "observed status only", while openai_stream_error infers a status from the error code or type.
  • Evidence: v4: aimux-core/src/error.rs:618-623 (status_code()), v4: aimux-provider-utils/src/response_handler.rs:296-304 (stream_error_api_call), against v4: aimux-providers/src/openai/mod.rs:87-175. Layer C1 (573de7e) introduced the inference.
  • Upstream: @ai-sdk/openai@4.0.80 src/openai-stream-error.ts:27-34, :240-266, :286-290 (the inference is correct).
  • Fixed in: layer C1. Comments only.
• C15: Vertex maas and xai subpaths are missing.
  • Evidence: v4: aimux-providers/src/vertex/ has neither.
  • Doc: doc:769. Upstream: @ai-sdk/google-vertex@5.0.98 src/maas, src/xai.
• C16: Anthropic InvokeModel on Bedrock is missing.
  • Evidence: v4: aimux-providers/src/bedrock/ has no anthropic.
  • Doc: doc:768. Upstream: @ai-sdk/amazon-bedrock@5.0.100 src/anthropic, amazon-bedrock-anthropic-fetch.ts:14-63.
  • Either implement it, or register it as a product difference in doc §5 with layer B6.
• C17: openai_output always reads openai.logprobs, whatever the provider namespace.
  • Evidence: v4: aimux-core/src/openai_output.rs:347.
  • Doc: doc:820.
• C50: the custom provider does not forward discovery.
  • Evidence: v4: aimux-core/src/custom_provider.rs:145-152 forwards files and skills only.
  • Doc: doc:1593. Upstream: ai@7.0.122 src/registry/custom-provider.ts:64 (no discovery; the doc requires forwarding).
• U26: speech results lack usage.
  • Evidence: v4: aimux-core/src/speech_model.rs has no usage field, but V4 pins provider 4.0.21.
  • Upstream: @ai-sdk/provider@4.0.21 src/speech-model/v4/speech-model-v4-result.ts:27 (usage?: JSONObject).
  • The tracing half of U26 is done after the migration, together with C3 (telemetry split).
• Note, doc carry only (no code): A1, A3 and A7 doc edits travel with layers B3, B4 and B3; see the planned alignment-doc PR plan. They are listed here so the vendor layers' PR descriptions pick them up.

Checklist: further items

• N4.1-1: Azure lacks deepseek() and completion().

  • Evidence: upstream azure@4.0.84 src/azure-openai-provider.ts:52-67, :327-341 reuses @ai-sdk/deepseek/internal and the OpenAI completion model. v4: aimux-providers/src/azure/mod.rs:437-502 has only chat, responses, embedding, image, transcription and speech. doc:764.
  • Fixed in: layer C1. If not ported, register it in the doc's known-gaps list.

• NP2-4: OpenAI Responses reasoning is not merged by item id (RSN-13).

  • Evidence: v4: aimux-providers/src/openai/responses/convert.rs:333-380 pushes each reasoning part separately, so store=true repeats item_reference and store=false splits one item into several. Upstream openai@4.0.80 src/responses/convert-to-openai-responses-input.ts:651, :1158-1209 merges summaries per reasoningId and sends item_reference once.
  • Fixed in: layer B4.

• N31-5.2-1: provider-utils exports an OpenAI-shaped shared error handler. Also on master. Low priority.

  • Evidence: v4: aimux-provider-utils/src/response_handler.rs:222 (create_standard_json_error_response_handler), used by 11 providers (groq, deepseek, jev, linkup, parallel_ai, serper, tinyfish, you_com, azure transcription, …). PIPE:261-268 forbids it. Upstream provider-utils has only createJsonErrorResponseHandler; each package defines its own error schema.
  • Fixed in: the layers that own those providers (B4/B5/B7).

• S2-4: Anthropic-AWS differs from upstream @ai-sdk/anthropic-aws 2.0.60.

  • Upstream anthropic-aws-provider.ts: workspaceId is required and read from ANTHROPIC_AWS_WORKSPACE_ID at request time (:219-227); region comes from AWS_REGION (:136-141, :210-217); with no apiKey / ANTHROPIC_AWS_API_KEY it signs with SigV4 from env AWS credentials (:128-134, :160-178).
  • Ours (v4: aimux-providers/src/anthropic_aws/mod.rs:60-76, :130-148): workspace optional and not read from env, region defaults to us-east-1 and never reads env, only x-api-key.
  • Fixed in: layer C1. Add the Anthropic-AWS row to the doc's §3.3 per-package timing table.
    Acceptance. Port the upstream openai-compatible tests: transformRequestBody, metadataExtractor (generate and stream), and multi-part tool content. Add an Anthropic raw-chunk test, plus fixtures for the new subpaths.