Bridge RequestUserInput events to ACP permission prompts

[c0st1nus]

Summary

This fixes a deadlock in codex-acp when codex_core requests pre-tool user input for MCP approvals.

Previously, EventMsg::RequestUserInput was treated as an unexpected event in the ACP bridge. When that happened, core would wait indefinitely for Op::UserInputAnswer, so the turn could stall before any McpToolCallBegin event was emitted.

This change bridges option-based RequestUserInput events into ACP request_permission(...) prompts, then maps the selected ACP option back into the RequestUserInputResponse shape expected by core.

What changed

• Added handling for EventMsg::RequestUserInput in src/thread.rs.
• Reused the existing ACP permission-request flow instead of introducing a separate interaction channel.
• Added a new pending interaction variant for user-input requests so the response can be resolved through the same async path already used for approvals.
• Converted single-question, option-based RequestUserInput prompts into ACP request_permission(...) requests.
• Mapped the selected ACP option back to the original answer label expected by codex_core, then submitted Op::UserInputAnswer.
• Added a fail-closed fallback for unsupported RequestUserInput shapes:
  if the request cannot be represented through ACP permission options, codex-acp now returns an empty response instead of hanging indefinitely.
• Updated the thread test stub to accept Op::UserInputAnswer.

Why this fixes the issue

For MCP approvals, codex_core may request user input before it emits McpToolCallBegin. If the ACP bridge drops that request, core waits forever and the client only sees early events such as usage_update.

With this patch:

• RequestUserInput is surfaced to the ACP client as a permission prompt.
• The user's selection is translated back into Op::UserInputAnswer.
• Core can resume the approval flow and continue to the actual MCP tool call.

Verification

• Added a focused unit test covering the happy path:
  test_request_user_input_uses_permission_request
• Added a focused unit test covering unsupported input shape fallback:
  test_request_user_input_without_options_cancels_without_permission_prompt
• Ran:
  cargo test test_request_user_input_
• Ran:
  cargo fmt --check

Notes

• This change intentionally targets the option-based RequestUserInput shape used by the MCP approval flow.
• Unsupported request shapes now fail closed instead of deadlocking.
• This PR does not address the separate StreamError silent-hang path.

[Copilot]

Pull request overview

This PR fixes a deadlock in codex-acp by handling EventMsg::RequestUserInput events, translating option-based user-input prompts into ACP request_permission(...) prompts, and mapping the selected option back into Op::UserInputAnswer so codex_core can resume MCP approval flow.

Changes:

• Added a new pending-interaction variant to track user-input requests and resolve them through the existing async permission-request flow.
• Implemented RequestUserInput event handling that converts option-based prompts into ACP permission requests and submits Op::UserInputAnswer on resolution.
• Added unit tests for the supported option-based path and the fail-closed unsupported-shape fallback.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[c0st1nus]

Manual verification update:

I tested this patch in Zed against the original failing scenario using a locally built codex-acp and ChatGPT account auth (no API key), with Playwright MCP enabled.

Before this change, Codex ACP would accept the prompt, emit usage_update, and then stall indefinitely without emitting any tool_call, message chunk, or stop reason.

After this change, the same flow now proceeds correctly:

• the pre-tool approval path is surfaced through ACP
• the pending user-input request is resolved
• the Playwright tool call is emitted normally
• the minimal reproduction completes successfully

I verified it with the original repro shape:
Do not explain steps. Do not write a plan. Immediately use the Playwright tool. Open about:blank, set document.body.innerText = "ok" and return only the result.

Result: Playwright now runs successfully in Zed and returns the expected output instead of hanging before tool invocation.