fix(llm): Anthropic rejects top-level anyOf/oneOf/allOf in tool input_schema

[ajhochy]

Summary

Anthropic's Messages API rejects tool definitions whose input_schema has anyOf, oneOf, or allOf at the top level of the object, returning:

Error: tools.2.custom.input_schema: input_schema does not support oneOf, allOf, or anyOf at the top level

This crashes the entire session — no response is returned, and the agent errors out immediately.

Root Cause

ProviderTransform.schema() in packages/opencode/src/provider/transform.ts has per-provider schema sanitization for Moonshot/Kimi and Google/Gemini, but no sanitization for Anthropic. When an MCP tool produces a JSON Schema with a top-level combiner (e.g. { anyOf: [...] }), it is passed unmodified to the Anthropic Messages API, which rejects it.

PR #1100 (fix(llm): emit anyOf as sole schema key for Vertex tool defs) fixed the same class of problem for Vertex/Gemini but did not cover @ai-sdk/anthropic.

Note: json-schema.ts's normalize() already collapses single-element anyOf arrays and flattens other common cases — but this runs at the tool-definition build time for opencode-native tools. MCP tools that arrive with their schemas pre-formed (tool.inputSchema set directly, not derived via ToolJsonSchema.fromTool()) can bypass that normalization and still carry a top-level combiner into lowerTool() → input_schema: tool.inputSchema.

Affected Code

packages/opencode/src/provider/transform.ts — the schema() export function (currently ~line 1339). The Gemini branch (lines ~1377–1453) rewrites integer enums and enforces items on arrays but does not strip top-level combiners. There is no Anthropic branch at all.

Proposed Fix

Add an Anthropic-specific pass in schema() that:

1. Detects @ai-sdk/anthropic (and @ai-sdk/google-vertex/anthropic) via model.api.npm
2. Recursively walks the schema and, at any level where anyOf/oneOf/allOf is the only key (or the sole meaningful key), unwraps the single-element case to the element itself — mirroring what normalize() in json-schema.ts already does
3. For multi-element top-level combiners that cannot be safely collapsed (e.g. a true union of two different object shapes), falls back to { type: "object" } to avoid the API rejection — a lossy but non-crashing result

The fix should be consistent with the existing normalize() logic in json-schema.ts lines 56–76.

Repro

1. Configure an agent with @ai-sdk/anthropic as the provider
2. Connect an MCP server whose tools expose a schema with a top-level anyOf (e.g. an optional/nullable parameter encoded as { anyOf: [{ type: "string" }, { type: "null" }] } at the root of input_schema)
3. Send any message — the first API call immediately fails with the error above

Environment

• opencode fork (Rhythm), observed on claude-opus-4-8 and claude-sonnet-4-6 via @ai-sdk/anthropic
• The error is provider-side validation, not a transient failure — it reproduces on every call

[ajhochy]

Filed in wrong repo. Re-filing in ajhochy/Rhythm.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Anthropic API rejects MCP tools with anyOf in input_schema #35516: Anthropic API rejects MCP tools with anyOf in input_schema
• Bug: Anthropic models error Input_schema does not support oneOf, allOf, or anyOf at the top level #37916: Bug: Anthropic models error Input_schema does not support oneOf, allOf, or anyOf at the top level

Both describe the same root cause (Anthropic's Messages API rejecting top-level anyOf/oneOf/allOf in input_schema) with MCP tools.