Fix: decision cards never appear in the Claude desktop app (AskUserQuestion parked by permissionDecision 'defer')

[habassa5]

Summary

Fixes #2466 — AskUserQuestion decision cards never appear when gstack runs in the Claude desktop app. The question is silently swallowed, the user sees nothing, and the agent gets back [Tool result missing due to internal error].

hosts/claude/hooks/question-preference-hook.ts emitted permissionDecision: 'defer' on its no-enforcement path. In Claude Code, defer does not mean "no opinion" — it means park this tool call so a later resume can pick it up.

From the Claude Code 2.1.221 binary:

case "defer": {
  if (!n.options.isNonInteractiveSession) {
    C(`Hook ${ce.hookName} returned permissionDecision=defer in interactive mode; ignoring (defer is print-mode only)`, {level:"warn"});
    break;                                   // ignored → card renders normally
  }
  ...
  return T.push({ message: al({ type: "hook_deferred_tool", ... }) })   // parked, never runs
}

defer is honored only when the session is flagged isNonInteractiveSession — and the Claude desktop app is flagged non-interactive even with a human present. So the AskUserQuestion call is parked, never executes, and never returns.

This is why it went unreported: in the terminal CLI defer is ignored and cards work fine, which is how most people run gstack. It reproduces only in the desktop app.

The fix

Omit permissionDecision on the no-enforcement path. It is .optional() in the PreToolUse hookSpecificOutput schema:

E.object({
  hookEventName: E.literal("PreToolUse"),
  permissionDecision: DM_().optional(),
  permissionDecisionReason: E.string().optional(),
  updatedInput: E.record(E.string(), E.unknown()).optional(),
  additionalContext: E.string().optional()
})

Omitting it is also what this path already means. The hook's own docstring states:

never-ask + one-way → defer (safety override; one-way always asks)

If defer parked the tool, a one-way door would never reach the human — the exact inverse of the documented safety override. So the intent is "no enforcement, proceed normally, let the user be asked."

Only allow / deny / ask are safe to emit when the tool is expected to actually run.

What is preserved

• additionalContext still flows → plan-tune memory-nugget injection unaffected
• deny paths untouched → never-ask auto-decide and the Conductor prose redirect both still work (deny is a valid, honored value)
• PostToolUse logging hooks unaffected

Verification

Bisected on a real desktop-app session with all three AUQ hooks registered:

Hooks live	Result
none	card renders ✅
question-preference-hook only (emits defer)	[Tool result missing due to internal error] ❌
all three, with this patch	card renders ✅

Across that machine's full transcript history before the fix: 3,700+ tool calls of every other kind resolved 100%; AskUserQuestion went 0-for-8. After the fix, decision cards render and return normally, including multi-question and multiSelect cards.

Tests

14 assertions across three suites encoded the old contract and are updated to toBeUndefined(). Added a regression guard (never emits permissionDecision:defer) covering the no-preference, no-marker, empty-stdin, and non-AUQ paths, plus the one-way-door safety override.

93/93 passing across the five hook suites against current main (v1.60.2.0):

bun test test/question-preference-hook.test.ts test/memory-cache-injection.test.ts \
         test/auq-error-fallback-hook.test.ts test/question-log-hook.test.ts \
         test/hook-scripts.test.ts
# 93 pass, 0 fail

🤖 Generated with Claude Code

[trunk-io]

Merging to main in this repository is managed by Trunk.

• To merge this pull request, check the box to the left or comment /trunk merge below.

After your PR is submitted to the merge queue, this comment will be automatically updated with its status. If the PR fails, failure details will also be posted here