sandbox-agent/research/acp/missing-features-spec/01-questions.md

Feature 1: Questions Subsystem

Implementation approach: ACP extension (_sandboxagent/session/request_question)

Summary

v1 had a full question subsystem: agent requests a question from the user, client replies with an answer or rejection, and the system tracks question status. v1 has partial stub implementation in mock only.

Current v1 State

• _sandboxagent/session/request_question is declared as a constant in acp_runtime/mod.rs:33
• Advertised in capability injection (extensions.sessionRequestQuestion: true)
• Mock agent (acp_runtime/mock.rs:174-203) emits questions when prompt contains "question"
• No real agent handler in the runtime for routing question requests/responses between real agent processes and clients
• Mock response handling exists (mock.rs:377-415) but the runtime lacks the general forwarding path

v1 Types (exact, from universal-agent-schema/src/lib.rs)

#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema, ToSchema)]
pub struct QuestionEventData {
    pub question_id: String,
    pub prompt: String,
    pub options: Vec<String>,
    pub response: Option<String>,
    pub status: QuestionStatus,
}

#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema, ToSchema)]
#[serde(rename_all = "snake_case")]
pub enum QuestionStatus {
    Requested,
    Answered,
    Rejected,
}

v1 HTTP Types (exact, from router.rs)

#[derive(Debug, Clone, Serialize, Deserialize, ToSchema, JsonSchema)]
#[serde(rename_all = "camelCase")]
pub struct QuestionReplyRequest {
    pub answers: Vec<Vec<String>>,
}

#[derive(Debug, Clone)]
pub(crate) struct PendingQuestionInfo {
    pub session_id: String,
    pub question_id: String,
    pub prompt: String,
    pub options: Vec<String>,
}

#[derive(Debug, Clone)]
struct PendingQuestion {
    prompt: String,
    options: Vec<String>,
}

Legacy Session REST Endpoints (from router.rs)

session question reply endpoint   -> 204 No Content
session question reject endpoint  -> 204 No Content

reply_question handler

async fn reply_question(
    State(state): State<Arc<AppState>>,
    Path((session_id, question_id)): Path<(String, String)>,
    Json(request): Json<QuestionReplyRequest>,
) -> Result<StatusCode, ApiError> {
    state.session_manager
        .reply_question(&session_id, &question_id, request.answers)
        .await?;
    Ok(StatusCode::NO_CONTENT)
}

reject_question handler

async fn reject_question(
    State(state): State<Arc<AppState>>,
    Path((session_id, question_id)): Path<(String, String)>,
) -> Result<StatusCode, ApiError> {
    state.session_manager
        .reject_question(&session_id, &question_id)
        .await?;
    Ok(StatusCode::NO_CONTENT)
}

v1 SessionManager Methods (exact)

reply_question

Key flow:

1. Look up session, take pending question by question_id
2. Extract first answer: answers.first().and_then(|inner| inner.first()).cloned()
3. Per-agent forwarding:
   • OpenCode: opencode_question_reply(&agent_session_id, question_id, answers)
   • Claude: If linked to a permission (AskUserQuestion/ExitPlanMode), send claude_control_response_line with "allow" and updatedInput; otherwise send claude_tool_result_line
   • Others: TODO
4. Emit QuestionResolved event with status: Answered

reject_question

Key flow:

1. Look up session, take pending question
2. Per-agent forwarding:
   • OpenCode: opencode_question_reject(&agent_session_id, question_id)
   • Claude: If linked to permission, send claude_control_response_line with "deny"; otherwise send claude_tool_result_line with is_error: true
   • Others: TODO
3. Emit QuestionResolved event with status: Rejected

v1 Event Flow

1. Agent emits question.requested event with QuestionEventData { status: Requested, question_id, prompt, options }
2. Client renders question UI
3. Client calls the legacy session question reply or reject endpoint with { answers: [["selected"]] }
4. System emits question.resolved event with QuestionEventData { status: Answered, response: Some("...") } or { status: Rejected }

v1 Agent Capability

AgentId::Claude => AgentCapabilities { questions: true, ... },
AgentId::Codex => AgentCapabilities { questions: false, ... },
AgentId::Opencode => AgentCapabilities { questions: false, ... },
AgentId::Amp => AgentCapabilities { questions: false, ... },
AgentId::Mock => AgentCapabilities { questions: true, ... },

Implementation Plan

ACP Extension Design

The question flow maps to ACP's bidirectional request/response:

1. Agent -> Runtime: Agent process sends _sandboxagent/session/request_question as a JSON-RPC request
2. Runtime -> Client: Runtime forwards as a client-directed request in the SSE stream
3. Client -> Runtime: Client POSTs a JSON-RPC response (answered/rejected)
4. Runtime -> Agent: Runtime forwards the response back to the agent process stdin

Payload Shape

Agent request:

{
  "jsonrpc": "2.0",
  "id": "q-1",
  "method": "_sandboxagent/session/request_question",
  "params": {
    "sessionId": "...",
    "questionId": "uuid",
    "prompt": "Which option?",
    "options": [["option-a", "Option A"], ["option-b", "Option B"]]
  }
}

Client response (answered):

{
  "jsonrpc": "2.0",
  "id": "q-1",
  "result": {
    "status": "answered",
    "answers": [["option-a"]]
  }
}

Client response (rejected):

{
  "jsonrpc": "2.0",
  "id": "q-1",
  "result": {
    "status": "rejected"
  }
}

Files to Modify

File	Change
server/packages/sandbox-agent/src/acp_runtime/mod.rs	Add real request/response forwarding for _sandboxagent/session/request_question (currently only mock)
server/packages/sandbox-agent/src/acp_runtime/mock.rs	Already has mock implementation; verify alignment with final payload shape
sdks/typescript/src/client.ts	Add onQuestion() callback and replyQuestion() / rejectQuestion() methods
frontend/packages/inspector/	Add question rendering in inspector UI

Docs to Update

Doc	Change
docs/openapi.json	N/A (ACP extension, not HTTP endpoint)
docs/sdks/typescript.mdx	Document onQuestion / replyQuestion / rejectQuestion SDK methods
docs/inspector.mdx	Document question rendering in inspector
research/acp/spec.md	Update extension methods list