"question" tool hangs session

[toojays]

Checklist

• I've read the README's Issues section.
• I'm running the latest agent-shell version (20260124.2145).
• I'm running the latest agent version (OpenCode 1.1.35).
• I'm running the latest agent ACP package (N/A?).
• For requesting new agent support, I'm including a link to the ACP-capable agent or related ACP package.
• For issues, I'm including ACP traffic (as per README).

Problem

When the LLM uses the "question" tool to ask me a question, I don't see the question, nor have any way to respond, and so the session appears to hang:

FWIW I'm using Kimi K2 Instruct 0905 via OpenRouter.

Here's the prompt followed by two ACP replies:

:direction outgoing
:kind      request
:object
           jsonrpc 2.0
           method  session/prompt
           id      10
           params
                   sessionId ses_40c8c4113ffe9pEIgpcBRduICy
                   prompt
                             type text
                             text Use the question tool.

:direction incoming
:kind      notification
:object
           jsonrpc 2.0
           method  session/update
           params
                   sessionId ses_40c8c4113ffe9pEIgpcBRduICy
                   update
                             sessionUpdate tool_call
                             toolCallId    functions.question:0
                             title         question
                             kind          other
                             status        pending
                             locations
                             rawInput

:direction incoming
:kind      notification
:object
           jsonrpc 2.0
           method  session/update
           params
                   sessionId ses_40c8c4113ffe9pEIgpcBRduICy
                   update
                             sessionUpdate tool_call_update
                             toolCallId    functions.question:0
                             status        in_progress
                             kind          other
                             title         question
                             locations
                             rawInput
                                           questions
                                                     question You're analyzing a legacy codebase with inconsistent naming conventions. You need to standardize all function names from camelCase to snake_case, but only in files that haven't been modified in the last 30 days. What's your approach using the available tools?
                                                     header   Legacy Codebase Analysis
                                                     options
                                                              label       Use Git log to find old files
                                                              description Use bash git log to identify files not modified in 30 days, then use Grep to find camelCase functions, then Edit to rename them

                                                              label       Use Task agent exploration
                                                              description Use Task agent with explore subagent to find all camelCase functions and analyze file modification dates

                                                              label       Use Glob and Grep combo
                                                              description Use Glob to find all code files, Grep to find camelCase patterns, then git log to filter by modification date

                                                              label       Manual file inspection
                                                              description Use Read tool to manually inspect each file for camelCase functions and check git history individually

[xenodium]

Thank you for filing and including all details.

Peeking at the incoming notifications, this is looking like it needs work on OpenCode side of things.

AFAIK, there is no generic question mechanism beyond session/request_permission in Agent Client Protocol to enable this flow (please correct me if otherwise).

While we could attempt special handling for OpenCode in agent-shell, there isn't enough information about how this should be handled. For example, how are responses to be packaged (whats the schema) and sent back? Unlike session/request_permission, there are no option identifiers.

Ideally, questions are exposed through existing Agent Client Protocol schema, so all clients benefit out ofthe box without OpenCode-specific logic.

The closest I've seen to question-handling recently is in Claude Code and the options are presented as a regular prompt response, which finishes the session/prompt request. This enables the user to type the selection number and send it like a regular prompt.

Maybe worth investigating if they are looking into question handling in the OpenCode repo/project?