api: anthropic messages schema tracing

[changminbark]

Description
This PR adds more struct definitions and their corresponding JSONMarshal and JSONUnmarshal methods according to the Anthropic Messages API. These are metrics that are provided in the latest Anthropic messages API. Currently, these metrics/metadata are not being tracked for tracing purposes. Eventually, these metrics should be used in openinference tracing as seen in

ai-gateway/internal/tracing/openinference/anthropic/messages.go

Line 154 in fefb039

// TODO: support for other content types.

It does not introduce any tracing as that should be opened in a new issue. This means that another issue should be opened to include tracking these metrics.

Related Issues/PRs (if applicable)

Fixes #1799

Special notes for reviewers (if applicable)
Claude code was used to write most of the code here, but I manually checked all of code that it generated. One thing that I noticed is that some of the previous API schema for Anthropic implemented in this repository were moved to beta status by Anthropic. This is something that the maintainers should discuss.

1. https://platform.claude.com/docs/en/api/messages#body-messages-content
2. https://platform.claude.com/docs/en/api/messages/create
3. https://platform.claude.com/docs/en/api/beta/messages#body-messages-content
4. https://platform.claude.com/docs/en/api/beta/messages/create

[nacx]

Overall LGTM.

Looking at the model I see that for several structs (toolunion, for example, not all the existing properties are added to the structs. Is there any reason why some properties ahve been added and others have been left out?

[changminbark]

@nacx I wanted to confirm with the new Anthropic beta messages API and how we should deal with that. Some of the old fields from the schema were migrated to beta status by Anthropic. Should I go ahead and implement the rest of the structs?

[codecov-commenter]

Codecov Report

❌ Patch coverage is 99.31973% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 84.20%. Comparing base (505bc4d) to head (889ee14).

Files with missing lines	Patch %	Lines
internal/apischema/anthropic/anthropic.go	99.31%	1 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1843      +/-   ##
==========================================
+ Coverage   83.71%   84.20%   +0.49%     
==========================================
  Files         126      126              
  Lines       16657    17075     +418     
==========================================
+ Hits        13944    14378     +434     
+ Misses       1811     1803       -8     
+ Partials      902      894       -8     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[changminbark]

@nacx I went ahead and implemented every schema parameter (mainly using Claude Code since it is simple struct defining and UnmarshalJSON/MarshalJSON). I checked through everything and it looks good. Some of the MarshalJSON functions use value receivers instead of pointer receivers because they are required fields in the parent structs (does not make sense to make them pointers as that would imply that they are optional fields).

I am unsure about how to run the code coverage test but I can add more unit tests if necessary.

[ehfd]

@nacx Can the CI be approved?

[changminbark]

@gavrissh I don't think the failing e2e tests are due to my commits. Or am I wrong?

[nacx]

/retest

[adirags]

I tried hitting an EAIGW with this change from a Claude Code client and was getting

time=2026-02-28T02:01:43.036Z level=ERROR msg="returning user-facing error for malformed request" request_id=9bfc5c6c-1c51-4e8e-aec1-cc5d64d07baa is_upstream_filter=false isUpstreamFilter=false error="malformed request: failed to parse JSON for /v1/messages"

I'm running the claude code CLI from a linux machine

$ claude --version
2.1.62 (Claude Code)
-----
❯ hi
  ⎿  API Error: 400 {"type":"error","error":{"type":"BadRequest","code":"400","message":"malformed request: failed to
     parse JSON for /v1/messages"}}

I added debugs and was able to narrow the issue down to the (*ToolUnion).UnmarshalJSON() function, where we were hitting return errors.New("missing type field in tool").

I haven't read the anthropic spec, but do we need to allow for clients coming in without the "type" field as well? I'm not sure if the behavior differs on Claude Code across versions, or clients
Can we treat the tool as "custom" if the type is not provided?

        if !typ.Exists() {
-               return errors.New("missing type field in tool")
+               // No type field: treat as a custom (user-defined) tool
+               var tool Tool
+               if err := json.Unmarshal(data, &tool); err != nil {
+                       return fmt.Errorf("failed to unmarshal tool: %w", err)
+               }
+               t.Tool = &tool
+               return nil
        }

Or even just fall through the the default case and do nothing?

[ehfd]

@adirags
I think we can follow up in another issue. I need to troubleshoot something else as well.

[changminbark]

@adirags How is your Claude Code set up? What model server/upstream is the EAIGW pointing to? There is a new PR for helping translate Anthropic requests to OpenAI compatible schema and back #1878. It might be best to try it out again after that PR is merged. If it still does not work then, I think it's better to open a new issue.

[adirags]

@changminbark, I had the EAIGW pointing to a Sonnet 4.6 model on Vertex AI.
I can retry once #1878 is merged. Thanks