Kimi Code Documentation

Kimi Code CLI provides a set of built-in tools managed directly by the runtime. These tools do not require MCP server installation and are automatically selected by the Agent.

Approval Mechanism:

• Read-only tools (e.g., Read, Grep, Glob) are auto-allowed by default.
• Write and execution tools (e.g., Write, Edit, Bash) require user approval by default.
• In YOLO mode, approval for regular tool calls is skipped.
• Plan mode exit approval is always required regardless of mode.

The agent-core-v2 unified read-write model is designed to provide a single source of truth for views, topics, write operations, and subscriptions. It aims to resolve issues like event visibility confusion, inconsistent definition methods, and feedback loops.

Key Design Constraints:

• Persistence is frozen: The design does not change the physical storage layout. wire.jsonl paths, the 'one physical log file per agent' rule, and the PersistedWireRecord data structure remain unchanged.
• API-level unification: Unification occurs only at the in-process API level, specifically for write entry points, read models, subscriptions, phases, and the type registry.
• Compatibility: The design maintains existing protocol_version and migration chain (1.0→1.5) mechanisms and does not introduce new log format migrations.

The Kimi Code CLI operates in a non-sandboxed environment on the user's machine. Actions taken by the agent (such as file reads, writes, or command executions) immediately affect the host system.

Key Environment Variables & Context:

• {{ KIMI_OS }}: The operating system being used.
• {{ KIMI_SHELL }}: The shell used for executing Bash tool commands.
• {{ KIMI_WORK_DIR }}: The current working directory, which serves as the project root. Use absolute paths for tool parameters when required.
• {{ KIMI_NOW }}: The ISO format date/time captured at session start. Note that this value is static and may become stale; for time-sensitive tasks, use the date command via a shell tool to get the current time.

Kimi Code provides several capabilities for autonomous coding and context management:

• Autonomous Actions: Explores codebases, reads/writes code, and runs terminal commands (requires permission).
• Thinking Controls: Toggle reasoning or select specific model-supported thinking effort levels.
• Provider-aware Models: Select specific models across different configured providers even if they share the same name.
• Native Editor Integration: Review AI changes directly using the VS Code diff viewer.
• MCP Support: Extend capabilities using Model Context Protocol servers.
• Slash Commands:
  • /init: Analyzes your current project.
  • /compact: Manages and optimizes the current context.

The loop package provides a stateless agent loop. It is designed to be decoupled from host-layer responsibilities. It does not manage:

• Sessions
• Wire transport
• Compaction execution
• Permissions UI
• Durable protocol bridging

These responsibilities must be handled by the host layer implementing the loop.

Starting with version 0.6.0, the Kimi Code VS Code extension migrated from a Python-based CLI architecture to an in-process Node SDK architecture using @moonshot-ai/kimi-code-sdk.

Key Changes:

• Runtime: Replaced the Python/stdio host with the TypeScript v1 engine running directly in the VS Code Extension Host.
• Identity: The extension identifies itself with the User-Agent kimi-code-vscode/[version].
• Data Sharing: The extension and the TUI (Terminal User Interface) now share the same configuration, authentication, MCP configuration, and sessions, provided they resolve to the same Kimi Code home directory.
• Environment Variables: The legacy kimi.environmentVariables setting has been removed. Provider-specific variables must now be placed in config.toml, and MCP server variables in mcp.json.