bug(permissions): AskUserQuestion should be allowed by default — agents block mid-loop on user-prompt UI

[namastex888]

Symptom

Agents using AskUserQuestion to clarify ambiguous requirements get blocked mid-loop. The tool prompts the user, but the prompt UI is currently bugged/inconsistent (sometimes returns immediate "user doesn't want to proceed" without showing the prompt), forcing agents to fall back to inline text questions.

Manifested today in a felipe session — agent called AskUserQuestion to disambiguate task scope, got auto-rejected, had to retry inline. Felipe's words: "its genie bugged right now".

Root cause (suspected)

AskUserQuestion is not in the default permissions.allow list seeded by the genie installer. Without explicit allow, the harness falls back to permission prompt, and the prompt UI has the bug above.

Cross-references in genie repo confirming the tool is already known/expected:

• skills/report/SKILL.md:44 — instructs agents to use AskUserQuestion for clarifications
• src/term-commands/history.ts:159,228 — TUI already renders AskUserQuestion as 'question' entry type
• .genie/wishes/agent-yaml-permissions-wireup/WISH.md:27 — listed under full preset

So the tool is wired into genie's UX, but not enabled by default.

Workaround

Add to ~/.claude/settings.json:

```json
{
"permissions": {
"allow": ["AskUserQuestion"]
}
}
```

Proposed fix

Add AskUserQuestion to the default permissions.allow list seeded by genie init / install onboarding. Likely candidates:

1. Wherever the installer writes the initial ~/.claude/settings.json
2. Default permission preset in agent.yaml schema (the full preset already lists it — make sure that preset is the install default)
3. If genie ships a hook that observes AskUserQuestion, ensure it never returns deny for the bare tool call (only inspect/log)

Acceptance

• Fresh genie install → first session where agent calls AskUserQuestion → prompt renders cleanly, no auto-rejection
• No regression for the existing report SKILL flow (which depends on this tool)

Severity

P2 — workflow degradation. Agents fall back to inline-text questions, which works but breaks structured-choice UX (preview rendering, multi-select, header chips).