Matt

This extensive commit significantly **enhances and stabilizes the experimental Session API's memory compaction system** and addresses numerous edge cases. It introduces **new capabilities** like token-threshold-based auto-compaction via `compactAfter()` and `onCompaction()` builder methods, along with real-time status broadcasting of compaction phases and token estimates to the client. Key **bug fixes** include refining token budget calculations for summary and tail messages, preventing summary hallucination with a content-agnostic prompt, and correctly handling virtual message IDs during iterative compaction. These changes improve the **reliability, configurability, and user feedback** of the session memory, particularly within the `session-memory` example client which now features streaming chat and token usage displays.

This commit introduces a new **experimental Session API** within the `agents` package, providing robust **stateful memory management** for AI agents. It features a **chainable builder** for `Session` configuration, supporting **tree-structured messages**, **context blocks**, and **iterative compaction** with FTS5 search. The `AgentSessionProvider` and `AgentContextProvider` offer SQLite-backed durable storage, significantly enhancing how agents manage conversational history and context. This **new capability** enables more complex and persistent agent interactions, including tool calls, and is demonstrated with an updated `session-memory` example.

This commit introduces a **new capability** by integrating `isomorphic-git` into the `@cloudflare/shell` package, enabling **pure JavaScript Git operations** directly within the workspace filesystem. It provides a custom `fs-adapter` to bridge the shell's `FileSystem` with `isomorphic-git`'s API, exposing CLI-matching commands like `gitInit`, `gitStatus`, and `gitCommit` via a new `createGit` function and a `ToolProvider` for codemode sandboxes with secure token injection. This significantly enhances the **WorkspaceChatAgent** by wiring up these new git tools, while also including **bug fixes** for `git status` mappings and preserving `fs` error codes in `workspace.ts` for robust error handling. Comprehensive unit tests are added for the new git functionality, and experimental examples are relocated to `examples/`, alongside a minor **maintenance** update to increase test delays in `ai-chat` to reduce flakiness.

This commit **updates the documentation** and **adds a changelog entry** for **`@cloudflare/codemode`** versions v0.2.0–v0.2.1. It details **new capabilities** including `codeMcpServer` and `openApiMcpServer` for MCP server wrappers, and the `modules` option for `DynamicWorkerExecutor`. The **installation instructions** are clarified to reflect optional `ai`/`zod` peer dependencies, and the `generateTypes` import path is updated to `@cloudflare/codemode/ai`. This **documentation update** provides users with accurate and comprehensive information regarding the latest features and changes in the **`@cloudflare/codemode` library**.

This commit introduces a **new capability** to the `@cloudflare/codemode` package by adding a `./mcp` export, providing two crucial functions for integrating with the codemode ecosystem. It exports `codeMcpServer`, which wraps an existing MCP server to expose its tools as a typed `code` tool for LLMs, and `openApiMcpServer`, which dynamically generates search and execute MCP tools from an OpenAPI specification. This significantly enhances **codemode integration** by enabling seamless exposure of existing MCP services and external APIs to AI agents with full type safety. The change includes comprehensive tests, updated examples (`codemode-mcp`, `codemode-mcp-openapi`), and refinements to internal logic and type definitions.

This commit introduces a **new feature** to the **`codemode` package**, enhancing the `DynamicWorkerExecutor` by internalizing code normalization and tool name sanitization. The `DynamicWorkerExecutor.execute()` method in `packages/codemode/src/executor.ts` now automatically calls `normalizeCode()` on input code and `sanitizeToolName()` on function keys. This **improves developer experience** by removing the need for users to manually pre-process raw code and tool names before execution. The `README.md` and tests have been updated to reflect this significant change in the executor's responsibilities.

This commit introduces a **new feature** to the **`codemode` package**, enabling the injection of custom JavaScript modules into the `DynamicWorkerExecutor`'s sandbox environment. Users can now provide an optional `modules` configuration within `DynamicWorkerExecutorOptions`, allowing sandbox code to `import` and utilize these external libraries. This enhancement to the `executor.ts` module significantly increases the flexibility and power of codemode operations by allowing more complex and customized logic within the isolated worker context. New tests in `executor.test.ts` validate this functionality, ensuring robust support for extended sandbox capabilities.

This commit **refactors** the `@cloudflare/codemode` package to introduce an **AI-free main entry point**, significantly reducing its dependency footprint. It **extracts core utilities** like `sanitizeToolName`, `generateTypesFromJsonSchema`, and `jsonSchemaToType` into new modules (`utils.ts`, `json-schema-types.ts`) that do not depend on `ai` or `zod`. Consequently, `ai` and `zod` are now **optional peer dependencies**, only required when importing AI SDK-dependent features from the new `@cloudflare/codemode/ai` entry point, which now houses `createCodeTool` and `generateTypes`. This **breaking change** improves modularity and allows for lighter installations for users not leveraging AI SDK integrations, while also fixing duplicate type declarations.

This commit **exports** the `normalizeCode` utility from the **`@codemode` package**, making it publicly available for use. This **new capability** significantly **hardens code snippets against common LLM output quirks** by stripping markdown fences, unwrapping `export default` statements, handling named function auto-calls, and correctly wrapping anonymous function/class declarations. The utility was extracted from `tool.ts` into its own `normalize.ts` module, accompanied by extensive unit tests. A previously attempted brittle regex-based TypeScript annotation stripping was removed to prevent code corruption, ensuring more reliable processing. This provides a robust utility for standardizing generated code, improving downstream reliability.

This commit primarily **refactors** the **`codemode`** package by **removing the `zod-to-ts` dependency**, replacing it with the AI SDK's `asSchema()` for Zod-to-JSON Schema conversion. It significantly **refactors** the `jsonSchemaToTypeString` utility in `packages/codemode/src/types.ts` for improved robustness and recursion handling, while also **fixing** `extractDescriptions` to correctly unwrap Zod types for JSDoc generation. The schema conversion tests are also **refactored** with a `testBoth()` helper to provide dual JSON Schema and Zod coverage, enhancing test comprehensiveness. This work improves the reliability and maintainability of **type string generation from schemas** and ensures more accurate JSDoc output within the `codemode` module.

This commit introduces an experimental session memory API with automatic micro-compaction for agents, providing a new core capability for managing conversation history. It significantly impacts agent efficiency by automatically truncating older messages, and its modular design allows future memory providers to easily integrate this compaction.

Fixes a runtime error in `AIChatAgent` where `jsonSchema` was not initialized, preventing `getAITools()` from working correctly during message processing. This ensures the AI chat agent can properly utilize its tools when handling messages.

This commit significantly enhances the `codemode` package, enabling robust and performant TypeScript type generation for AI SDK tools and various JSON schemas. It introduces a production-hardened JSON Schema to TypeScript converter, ensuring accurate types for developers while maintaining fast schema handling in the `mcp/client`.

This commit **enhances the documentation** by adding a new entry for the **Cloudflare API server** to the list of available Model Context Protocol (MCP) servers. It introduces a new documentation page, `mcp-servers-for-cloudflare.mdx`, within the **agents and MCP server documentation** section. This **documentation update** improves user guidance by making the Cloudflare API server discoverable for those seeking to integrate with Cloudflare via MCP. The change ensures that users have comprehensive information on supported MCP server options.

This commit fixes a reliability issue within the `ai-chat` package, ensuring the `/get-messages` endpoint functions correctly for all users, especially those who customize `onRequest` without calling `super.onRequest()`. This improves the robustness and compatibility of custom `ai-chat` implementations without introducing any breaking changes.

Fixes an issue within the `codemode` package where the tool's default prompt could lead to TypeScript syntax errors. This ensures the `codemode` tool consistently generates correct JavaScript output.

This commit introduces a major new RPC transport capability for the `agents` package, allowing Agents to communicate directly via Durable Objects within the same Worker, enhancing performance and reliability by persisting connections across hibernation. It extends the `addMcpServer` API to support this new transport type and includes significant refactoring, documentation, and example updates.

This commit provides a crucial **documentation fix** for the **Agents Codemode** feature, clarifying that **Dynamic Workers** exclusively execute generated **JavaScript**, not TypeScript. It updates the `codemode.mdx` API reference to accurately reflect this execution model, while also noting that TypeScript type definitions are still generated for LLM context. Additionally, the `changelog` entry for the Codemode SDK rewrite is adjusted to use the more general term 'code' instead of 'JavaScript' when describing LLM output. This **clarification** ensures developers have precise information regarding the runtime environment and prevents potential confusion about supported languages within the **Agents** platform.

This commit **documents** a significant **architectural rewrite** for the `@cloudflare/codemode` package by adding a **new changelog entry** for version `v0.1.0`. The entry details the transition to a **modular, runtime-agnostic SDK architecture**, which represents a major **refactoring and maintenance effort**. This update to the **`@cloudflare/codemode` SDK** enhances its flexibility and adaptability across various runtime environments, laying a new foundation for future development and usage of the codemode tools.

Fixes a critical issue in the `codemode` example's `AIChatAgent` that prevented message persistence on page reload by resolving a 404 error on the `/get-messages` endpoint. This restores expected functionality for the example.