Gemini CLI run_shell_command tool fails in ACP mode, missing tool results

[hzyu-hub]

Description

When using Gemini CLI via ACP protocol (spawned by Alma app), the run_shell_command tool consistently fails to return results, causing the AI SDK to throw AI_MissingToolResultsError.

Steps to Reproduce

1. Run Gemini CLI with ACP flags: gemini --acp --experimental-acp
2. Ask for a shell command execution (e.g., 'run ls command')
3. Observe the error

Expected Behavior

The run_shell_command tool should execute successfully and return results to the ACP client.

Actual Behavior

• run_shell_command tool call is initiated but never returns a result
• ACP client receives no tool result
• Error: AI_MissingToolResultsError: Tool results are missing for tool calls run_shell_command-xxx

Environment

• Gemini CLI version: 0.34.0
• macOS Sonnet
• ACP Client: Alma v0.0.723
• Flags used: --acp --experimental-acp

Additional Context

This appears to be related to existing issues #14561 and #5382. The issue is that in ACP/non-interactive mode, the run_shell_command tool cannot render user confirmation prompts and fails silently without returning a result to the ACP stream.

Suggested Fix

• Handle ACP/non-interactive mode for run_shell_command tool
• Either auto-approve commands in ACP mode or return an error result instead of hanging indefinitely
• Ensure tool results are always returned to the ACP stream, even on failure

[gemini-cli]

Bug in tool execution during ACP mode.

Effort Analysis:
The run_shell_command hangs in ACP mode because it expects user confirmation but cannot render the prompt. Resolving this involves modifying the asynchronous flow to handle ACP mode specifically (e.g., auto-approve or error out). Debugging this timeout/hang in non-interactive mode is effort/medium.

[gemini-cli]

Tool failure in ACP mode is a critical workflow block.