Feature: batch/merge queued messages into single API request (like Claude Code)

Problem

When using Kimi Code TUI with a slow or thinking-intensive model, users often queue multiple messages while the agent is processing. Currently, each queued message is sent as a separate API request once the current turn completes.

This differs from Claude Code behavior, where queued messages are merged into a single request when the agent becomes available. The current behavior causes:

• Loss of conversational context between queued messages
• Multiple API round-trips instead of one
• Fragmented reasoning when user intended to provide multi-part input

Current Behavior

1. User sends message A (agent starts processing)
2. User queues messages B, C, D while agent is busy
3. Agent finishes A → sends B as separate request → finishes B → sends C as separate request → etc.

Expected Behavior (Claude Code style)

1. User sends message A (agent starts processing)
2. User queues messages B, C, D while agent is busy
3. Agent finishes A → merges B+C+D into single request → processes combined input

Use Case

This is especially important for:

• Complex multi-step instructions that user types incrementally
• Corrections/additions to previous queued message
• Users on high-latency connections who want to minimize round-trips

Proposed Solution

Add a config option (e.g., in or ) that:

1. Accumulates all pending queued messages when current turn completes
2. Concatenates them with appropriate separators (e.g., double newline)
3. Sends as single user message to the API

Alternative: make this the default behavior with opt-out, since it matches Claude Code UX that many users expect.

Environment

• Kimi Code version: 0.31.1
• Platform: macOS (but affects all)
• Provider: any (tested with OpenCode proxy → Kimi K3)

The issue names no files, tests, or entry points. Start by locating the TUI's queued-message handling and the configuration path for request behavior. Done means pending messages are combined into one API request after the current turn, with the chosen default or opt-out behavior covered by tests.