[thorstenhans]

Why

Frameworks like the Vercel AI SDK define tools with Zod. Turning the SDK's hand-written JSON Schema (the Anthropic / OpenAI shape) into something those frameworks understand is error-prone and loses validation and types. Going the other way is reliable: Zod 4 generates JSON Schema itself (z.toJSONSchema).

What changes

• interpreterTool() returns a strict Zod schema as inputSchema. { description, inputSchema } goes straight into the AI SDK's tool(), and execute receives typed arguments.
• Adapters derive JSON Schema from it: toAnthropicTool(), toOpenAITool(), and the new toToolDefinition(). They still accept hand-written JSON Schema ToolDefinitions (app-builder's create_app is unchanged).
• Validation for free: inputSchema.safeParse(args) rejects unknown fields such as allowedHosts, kvEnabled, and identity, so the model cannot set execution policy. The ai-code-executor example uses it instead of hand-rolled checks.
• Types follow the configuration: interpreterTool({ languages: ["python"] }) infers language: "python" with no dependencies, matching the schema.
• README shows the Interpreter, tool, and execute setup for the Vercel AI SDK ToolLoopAgent and the Anthropic SDK, with the floating-promise rule in the agent system prompt (per AGENTS.md).
• zod@^4.6.7 is a regular dependency, imported only by the ./tools entry point.

Compatibility

What the model sees is unchanged. For every language configuration, the derived JSON Schema and the description are identical to the previous hand-written output (key order aside). A golden fixture (sdk/test/fixtures/interpreter-tool.json) pins this. While building it, the comparison caught two regressions, both fixed here:

• input.body had become required.
• Zod emitted propertyNames: { type: "string" } for string maps; this is now stripped because OpenAI strict mode rejects it.

Breaking (0.1.0): interpreterTool().inputSchema is now a Zod schema. Code that reads JSON Schema from it should use toToolDefinition(tool).inputSchema.

Testing

• sdk: npm run check (typecheck incl. new compile-time type assertions; 45 tests pass, including the golden test, which I confirmed fails when the body regression is put back)
• AI SDK and Anthropic SDK compatibility: the README snippets typecheck against ai@7.0.166, @ai-sdk/anthropic@4.0.99, and @anthropic-ai/sdk@1.6.0 (scratch project, not run against a model)
• Examples: ai-code-executor and app-builder (both providers) and sdk-component build to Wasm; example lockfiles only gain the SDK's zod entry and pass npm ci
• npm audit gate: sdk is clean. The gate still fails for examples/ai-code-executor/ui and examples/app-builder/ui, which this PR does not touch
• Not run locally: the spin integration suites (check-imports.mjs, smoke.hurl, run-examples.sh). They cover the sandbox components, which do not depend on the SDK

Open questions for review

• zod as a regular dependency or a peer dependency (a peer dependency avoids two copies of zod in apps that also use it, but users must install it themselves)
• Bundle-size impact of zod on the example Wasm components has not been measured

🤖 Generated with Claude Code