feat(policy): add @ai-sdk/policy-opa package (OPA adapter for toolApproval)

Introduces a new monorepo package at packages/policy-opa that adapts Open
Policy Agent policies to the AI SDK's public toolApproval callback. The
package sits entirely on top of the published surface; no edits to
packages/ai, packages/provider, or packages/provider-utils.

Everything is exported from the package root. The engine-neutral core is a
PolicyClient interface, shadow() for safe policy rollout with fire-and-forget
telemetry, and wrapMcpTools() for making approval configuration total over a
discovered tool surface. The OPA layer ships opaPolicy / optionalOpaPolicy
(Rego-as-code), the wasmPolicyClient and httpPolicyClient backends
(lazy-loaded optional peer deps), opaCapabilityMiddleware for fail-closed
model-level tool filtering, and normalizeOpaDecision for users who call OPA
themselves.

Transitive enforcement (coarse dispatchers like bash / http.request /
MCP proxies) is handled inside the user's toolApproval by parsing the
dispatcher input and routing to the same Rego rule that gates the
direct tool. README documents two DRY forms: a shared TS predicate and
a shared Rego helper rule.

Includes an end-to-end integration test (generateText +
MockLanguageModelV3) plus a runnable example at examples/mock/basic.ts
that demonstrates allow / deny / transitive-deny paths.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: dnukumamras

.changeset/policy-opa-package.md

@@ -0,0 +1,21 @@
+---
+'@ai-sdk/policy-opa': patch
+---
+
+Introduce `@ai-sdk/policy-opa`, an Open Policy Agent adapter for the
+`toolApproval` callback on `generateText` / `streamText` / `ToolLoopAgent`.
+
+Everything is exported from the package root. The engine-neutral core is a
+`PolicyClient` interface, `shadow()` for safe policy rollout with
+fire-and-forget telemetry, and `wrapMcpTools()` for making approval
+configuration total over a discovered tool surface. The OPA layer ships
+`opaPolicy` / `optionalOpaPolicy` (Rego-as-code authorization),
+`wasmPolicyClient` and `httpPolicyClient` backends (lazy-loaded optional peer
+deps), `opaCapabilityMiddleware` for fail-closed model-level tool filtering,
+and `normalizeOpaDecision` for users who call OPA themselves.
+
+Sits entirely on top of the public SDK surface, with no changes to `ai`,
+`@ai-sdk/provider`, or `@ai-sdk/provider-utils`. Transitive enforcement
+(coarse dispatchers like `bash` / `http.request` / MCP proxies) is handled
+inside the user's `toolApproval` by parsing the dispatcher input and routing
+to the same Rego rule that gates the direct tool.

packages/policy-opa/CHANGELOG.md

@@ -0,0 +1 @@
+# @ai-sdk/policy