AI编程编辑器

Installation

Go install

go install github.com/spachava753/cpe@latest

Quick Start

CPE requires a YAML configuration file with at least one model profile. There is no zero-config mode.

Full Example

```yaml

3. Configure an ACP client

CPE communicates over stdio JSON-RPC. The client launches cpe acp serve and hosts the chat/thread UI.

Zed supports external ACP agents through agent_servers. See Zed External Agents and the Zed ACP page.

A minimal Zed settings entry looks like this:

{
  "agent_servers": {
    "CPE": {
      "type": "custom",
      "command": "cpe",
      "args": ["acp", "serve", "--config", "/absolute/path/to/cpe.yaml"],
      "env": {
        "ANTHROPIC_API_KEY": "your-api-key"
      }
    }
  }
}

If Zed cannot find cpe on its PATH, set command to an absolute path such as /Users/me/.local/bin/cpe.

Configuration

Config Discovery

CPE searches for configuration in this order:

1. --config explicit path
2. ./cpe.yaml or ./cpe.yml
3. User config directory, usually ~/.config/cpe/cpe.yaml on Linux and ~/Library/Application Support/cpe/cpe.yaml on macOS

yaml-language-server: $schema=https://raw.githubusercontent.com/spachava753/cpe/refs/heads/main/schema/cpe-config-schema.json

version: "1.0"

models: - ref: sonnet display_name: "Claude Sonnet" id: claude-sonnet-4-5-20250929 type: anthropic api_key_env: ANTHROPIC_API_KEY context_window: 200000 max_output: 64000 input_cost_per_million: 3 output_cost_per_million: 15 systemPromptPath: ./agent_instructions.md timeout: 5m thinkingValues: - name: "Fast" value: "1024" description: "Lower reasoning budget" - name: "Deep" value: "8192" description: "Higher reasoning budget" generationParams: temperature: 0.2 maxGenerationTokens: 12000 codeMode: enabled: true maxTimeout: 3600 largeOutputCharLimit: 20000 mcpServers: search: type: http url: https://search.example.com/mcp headers: Authorization: "Bearer ${SEARCH_API_KEY}" compaction: autoTriggerThreshold: 0.8 maxAutoCompactionRestarts: 2 toolDescription: "Compact the current session into a concise continuation summary." inputSchema: type: object properties: summary: type: string required: [summary] initialMessageTemplate: | Continue from this compacted CPE session. Compaction arguments: {{ .ToolArgumentsJSON }}

- ref: gpt display_name: "GPT" id: gpt-5.1 type: responses auth_method: oauth context_window: 400000 max_output: 128000 ```

Before rendering initialMessageTemplate, CPE validates each compact_conversation argument object against inputSchema. Compaction must be the only tool call in its assistant response; mixed responses reject every call without executing siblings. Mixed-call and schema-validation failures allow at most three recoverable retries for each compaction cycle. A successful compaction resets the retry budget. Template execution failures are terminal configuration or runtime errors. Failed attempts do not reset session-scoped tool state or consume a successful compaction restart. Successful attempts persist the completed tool result and replacement root before resetting compaction-scoped state or publishing completion, so session loading can replay both the call and its completion. Because the two branches cannot currently be saved atomically, failure to persist the replacement root after the successful result is an invariant panic rather than a returned false-success branch.

Environment Variables

API key variables are configured per model profile through api_key_env. OAuth-backed profiles use auth_method: oauth and provider account commands where supported. Anthropic Vertex AI profiles use Google Application Default Credentials instead of api_key_env.

`configuration file not found`

Create cpe.yaml in the current directory or user config directory, or pass --config /path/to/cpe.yaml from the ACP client command args.

Command Reference

cpe [command]

Root flags:
  --config string    Path to YAML configuration file
  --db-path string   ACP session SQLite database path (env: CPE_DB_PATH)
  -v, --version      Print the version number and exit

Commands:
  acp              Serve ACP and manage persisted sessions
    serve           Start the stdio ACP server
    list, ls        List sessions by last activity, newest first
                    --page uint       Page number, starting at 1 (default 1)
                    --page-size uint  Sessions per page (default 20, maximum 1000)
    show <id>       Render complete session history as Markdown
    delete <id>     Delete a session
    fork <id>       Fork a session and print the new session ID

  model, models     Inspect configured model profiles
    list, ls        List configured model refs
    info <ref>      Show model profile details
    system-prompt   Render the selected model profile's system prompt
                    -m, --model string  Model profile ref for profile-specific inspection

  mcp               Inspect MCP servers for a selected model profile
    list-servers, ls-servers
    list-tools, ls-tools <server>
    info <server>
    call-tool --server <server> --tool <tool> --args '{}'
    code-desc       Print the starlark_repl tool description
                    -m, --model string  Model profile ref whose MCP servers should be inspected

  account           Manage provider account credentials and usage
    login <provider>
    logout <provider>
    usage <provider> [--watch | --raw]

  completion        Generate shell completion scripts

API key missing

Ensure the environment variable named by api_key_env is visible to the ACP server process. For editor-launched processes, put required variables in the client's agent server env block or in the environment that launches the editor.

For anthropic_vertex profiles, CPE does not use api_key_env; configure Google Application Default Credentials or GOOGLE_APPLICATION_CREDENTIALS for the ACP server process instead.

MCP Tool Integration

CPE is an MCP client inside each ACP session. Configure MCP servers on a model profile with mcpServers:

CPE also accepts MCP servers forwarded by the ACP client and merges them with configured servers. Duplicate server names fail fast so tool behavior is not ambiguous.

The bundled text_edit file editing tool is registered directly by CPE and does not require MCP configuration. Set disable_edit_tool: true on a model profile to omit it.

Use the MCP inspection commands to inspect and test configured servers:

cpe mcp list-servers --model sonnet
cpe mcp list-tools search --model sonnet
cpe mcp list-tools search --show-all --model sonnet
cpe mcp info search --model sonnet
cpe mcp call-tool --server search --tool web_search --args '{"query":"golang"}' --model sonnet

Troubleshooting