feat(acp): LLM-based message rewrite middleware with custom prompts

[zhangxy-zju]

Motivation

In ACP scenarios, raw model output often contains technical details (file paths, SQL, code snippets, tool call info, etc.) that are not user-friendly for business users. We need a configurable mechanism that allows users to define custom rewrite rules to transform model output into formats and styles suited for their target audience.

Solution

Introduce a Message Rewrite Middleware that, at the end of each model turn, collects accumulated output (thoughts + messages) and passes it to an LLM rewriter. The rewritten result is then emitted as a separate assistant message.

Key Capabilities

• Configurable: Controlled via messageRewrite field in settings.json — toggle on/off, set custom prompt, choose rewrite target (message/thought/both)
• Custom prompts: Users can write rewrite prompts tailored to their business scenario (e.g., data analysis, code generation, document writing)
• Full path coverage: Supported in both interactive mode (Session) and non-interactive mode (CLI/eval), including cron-triggered scenarios
• Safe degradation: Rewrite failures or timeouts (30s) are silently skipped without affecting the original output

Configuration Example

{
  "messageRewrite": {
    "enabled": true,
    "target": "both",
    "prompt": "Your custom rewrite prompt..."
  }
}

Modules Involved

• packages/cli/src/acp-integration/session/rewrite/ — TurnBuffer, LlmRewriter, MessageRewriteMiddleware, config, types
• packages/cli/src/acp-integration/session/Session.ts — Interactive mode integration
• packages/cli/src/nonInteractiveCli.ts — Non-interactive mode integration