Carry additionalProperties through the Anthropic tool input_schema so it matches functool strict validation

[PratikDhanave]

What

The Anthropic provider builds anthropic.ToolInputSchemaParam from a func tool's schema by copying only the properties and required keywords and discarding everything else. In particular it drops additionalProperties:false, which functool's strict schema always emits.

This change copies the remaining schema keywords into ToolInputSchemaParam.ExtraFields (which the anthropic-sdk-go marshals via MarshalWithExtras), so the model-facing schema stays in sync with the Go-side validation.

Why

funcTool.Call runs handler -> inputFormat.Unmarshal -> applySchema -> validate -> resolved.Validate, and the resolved schema rejects extra properties. Because the Anthropic path told the model nothing about additionalProperties, a hallucinated extra argument would pass the model but then fail functool validation on the Go side — an asymmetry that only affected Anthropic.

This restores cross-SDK parity: the OpenAI provider forwards the whole schema map (schemaToMap returns the full map) and the Gemini provider sets decl.ParametersJsonSchema = schema. Only Anthropic was stripping keywords. Matching the other providers keeps behavior aligned with the .NET/Python semantics where the declared tool schema and the validated schema are the same.

How it is tested

Added TestToolInputSchemaCarriesAdditionalProperties in agent_test.go. It builds a tool via functool.MustNew over a struct argument, runs the agent through the existing fake-transport harness, captures the request body, and asserts tools[0].input_schema.additionalProperties == false. The test fails before the fix (keyword omitted) and passes after.

[Copilot]

Pull request overview

This PR updates the Anthropic provider’s tool schema construction to preserve full JSON Schema semantics (notably additionalProperties: false) so the schema sent to the model matches the stricter Go-side functool validation behavior, aligning Anthropic with the existing OpenAI/Gemini providers.

Changes:

• Forward non-properties/required JSON Schema keywords into anthropic.ToolInputSchemaParam.ExtraFields so they are included in the serialized input_schema.
• Add a regression test that captures the outbound Anthropic request and asserts tools[0].input_schema.additionalProperties == false.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
provider/anthropicprovider/agent.go	Preserves additional JSON Schema keywords by copying them into ToolInputSchemaParam.ExtraFields when building Anthropic tool params.
provider/anthropicprovider/agent_test.go	Adds a request-capture test asserting additionalProperties: false is present in the tool input_schema.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

API Consistency Review — PR #724

Scope: Internal bug fix in provider/anthropicprovider — no exported Go APIs changed.

Parity verdict: ✅ This change restores cross-provider consistency rather than introducing divergence.

What was checked

• Go internal change only: The affected code is in the unexported buildMessageParams function. No exported types, methods, functions, or option structs were added, removed, or changed. The public-api-change label is not warranted.

• Cross-provider alignment (Go): The OpenAI provider already forwards the full schema map via schemaToMap, and the Gemini provider sets decl.ParametersJsonSchema = schema. The Anthropic provider was the only one stripping keywords like additionalProperties: false. This PR brings Anthropic in line with the other two Go providers.

• Cross-SDK alignment (.NET / Python): In .NET (Microsoft.Agents.AI) and the Python SDK, the declared tool input schema and the schema used for server-side validation are the same object — there is no stripping of keywords before sending to the model. The fix matches that expectation: what the model sees now equals what funcTool.Call validates against.

• No new behavior introduced: This is a correctness fix, not a feature addition. The schema forwarded to Anthropic now matches what OpenAI and Gemini already receive and what .NET/Python expose.

Conclusion: No parity issues. The parity-approved label remains appropriate.

Warning

Firewall blocked 1 domain

The following domain was blocked by the firewall during workflow execution:

• awmgmcpg

To allow these domains, add them to the network.allowed list in your workflow frontmatter:

network:
  allowed:
    - defaults
    - "awmgmcpg"

See Network Configuration for more information.

Generated by Go API Consistency Review Agent · 39.1 AIC · ⌖ 4.95 AIC · ⊞ 5.9K · ◷