fix: route AskUserQuestion/ExitPlanMode replies via can_use_tool

[hilr]

Summary

The user's reply to an AskUserQuestion (and the Yes/No on an ExitPlanMode plan) was silently ignored — the LLM proceeded with the default option as if the user had cancelled.

Root cause: the agent runs in bypassPermissions, but that mode does not bypass requiresUserInteraction() tools. In the CLI's permission check the requiresUserInteraction short-circuit (the "1e" step in utils/permissions/permissions.ts) runs before the bypass check, so AskUserQuestion / ExitPlanMode still emit a can_use_tool request. With no callback registered, the SDK raised "canUseTool callback is not provided" (claude_agent_sdk/_internal/query.py); the CLI read that error as a user cancellation and fed the LLM a cancelled tool_result — hence "user cancelled, use defaults".

Fix: register a can_use_tool callback that:

• allows ordinary tools outright (preserving bypass semantics);
• for the two user-interaction tools, bridges the CLI permission request to the IM channel via a two-phase handshake on an asyncio.Future keyed by tool_use_id:
  • run() surfaces the question/plan to the channel and returns; the callback awaits the future, holding the CLI's request open;
  • the user's reply arrives as the next run() call, which resolves the future instead of issuing a fresh query; the callback then returns the decision — AskUserQuestion injects parsed answers into updated_input.answers (the CLI formats them into the tool_result), ExitPlanMode maps to allow / deny-with-feedback.

Drops the stdin-hack _send_tool_result and _format_answers_for_tool_result; _build_plan_approval_result → _build_exit_plan_permission (returns a PermissionResult). Detailed writeup added as an in-code comment at the callback.

Test plan

• pytest — 186 passed
• replaced the flaky LLM-dependent test_ask_user_question_resume with a deterministic future-mechanism test
• added _can_use_tool callback tests (bypass / ExitPlanMode allow+deny / AskUserQuestion answer injection / parse-failure fallback)
• validate live IM→agent round-trip after deploy (diagnostic logging from chore: add diagnostic logging for AskUserQuestion / ExitPlanMode reply flow #66 will confirm the can_use_tool: awaiting user reply → got user reply chain)

🤖 Generated with Claude Code