fix(daemon): close stdin on AskUserQuestion tool_use so runs don't hang (#4277)

* fix(daemon): close stdin on AskUserQuestion tool_use so runs don't hang (#4273)

Remove the dead AskUserQuestion detection branch from
applyClaudeStreamJsonRunBookkeeping. The branch (left over from PR #4114
cleanup) added the tool_use id to pendingHostAnswers and returned early
when the CLI emitted an AskUserQuestion tool_use event. This blocked the
subsequent non-tool_use turn_end from closing stdin, leaving the run
stdinOpen indefinitely — the cleanTerminalTurn condition could never be
satisfied because pendingHostAnswers.size was always > 0. The run stayed
`running` forever and the web Continue button stayed disabled.

The AskUserQuestion tool wiring and POST /api/runs/:id/tool-result
endpoint were intentionally removed in PR #4114; per AGENTS.md the
stream-json stdin skeleton is retained only as generic infrastructure.
The detection branch had no live counterpart to answer it and was
therefore permanently blocking.

Fix: remove the branch, the pendingHostAnswers type field, and the
pendingHostAnswers guard in the cleanTerminalTurn condition. A
non-tool_use turn_end now correctly closes stdin regardless of what
tool_use events preceded it.

* docs(spec): reconcile stale AskUserQuestion/tool-result references

Address @PerishCode review on #4277: the 'Why OD diverged' section and the
'Interactive intact' acceptance test still described the removed host-answer
contract (stream-json tool_result / POST /api/runs/:id/tool-result) as live.
Rewrite both to the current <question-form> -> POST /api/chat path, citing the
PR #4114 removal already stated earlier in the spec, so future maintainers
cannot reintroduce the dead path.

* docs(spec): correct stdin lifecycle direction in stream-json note

Address @PerishCode second review on #4277: prior cleanup wrote that stdin
closes on a tool_use stop reason, inverting the invariant the daemon code and
regression test rely on. Per applyClaudeStreamJsonRunBookkeeping, stdin stays
open across tool_use pauses and closes only on a terminal non-tool_use
turn_end/usage. Reword the clause to match.

Author: cbeaulieu-gt

apps/daemon/src/server.ts

@@ -4574,7 +4574,6 @@ export function classifyChatRunCloseStatus(params: {
 
 type ClaudeStreamJsonBookkeepingRun = {
   stdinOpen?: boolean;
-  pendingHostAnswers?: Set<string>;
   turnCompletedCleanly?: boolean;
   child?: {
     stdin?: {
@@ -4601,30 +4600,18 @@ export function applyClaudeStreamJsonRunBookkeeping(
     stopReason?: unknown;
   };
 
-  if (
-    run.stdinOpen &&
-    event.type === 'tool_use' &&
-    (event.name === 'AskUserQuestion' || event.name === 'ask_user_question') &&
-    typeof event.id === 'string'
-  ) {
-    if (!run.pendingHostAnswers) run.pendingHostAnswers = new Set();
-    run.pendingHostAnswers.add(event.id);
-    return;
-  }
-
   const cleanTerminalTurn =
-    ((event.type === 'turn_end' &&
+    (event.type === 'turn_end' &&
       // `stop_reason: tool_use` means the model paused to wait for tool
-      // execution (claude-code is about to run an internal tool, or we owe a
-      // host tool_result). Either way the conversation is still in flight.
+      // execution (claude-code is about to run an internal tool). The
+      // conversation is still in flight.
       event.stopReason !== 'tool_use') ||
-      (event.type === 'usage' && event.stopReason !== 'tool_use')) &&
-    (!run.pendingHostAnswers || run.pendingHostAnswers.size === 0);
+    (event.type === 'usage' && event.stopReason !== 'tool_use');
   if (!cleanTerminalTurn) return;
 
-  // Record clean completion even if stdin was already closed by the
-  // host-answer path. The close-status classifier reads this to ignore late
-  // SessionEnd hook failures after the final assistant turn completed.
+  // Record clean completion even if stdin was already closed. The
+  // close-status classifier reads this to ignore late SessionEnd hook
+  // failures after the final assistant turn completed.
   run.turnCompletedCleanly = true;
   if (run.stdinOpen) {
     if (run.child?.stdin && !run.child.stdin.destroyed) {

apps/daemon/tests/chat-run-artifact-quiet-period.test.ts

@@ -551,7 +551,6 @@ describe('applyClaudeStreamJsonRunBookkeeping', () => {
   it('keeps stdin open when usage reports a tool_use stop reason', () => {
     const run = {
       stdinOpen: true,
-      pendingHostAnswers: new Set<string>(),
       turnCompletedCleanly: false,
       child: {
         stdin: {
@@ -571,4 +570,35 @@ describe('applyClaudeStreamJsonRunBookkeeping', () => {
     expect(run.stdinOpen).toBe(true);
     expect(run.child.stdin.end).not.toHaveBeenCalled();
   });
+
+  it('closes stdin and records clean completion after an AskUserQuestion tool_use followed by end_turn (#4273)', () => {
+    // Regression test: the dead AskUserQuestion detection branch used to add
+    // the tool_use id to pendingHostAnswers and return early, preventing stdin
+    // from ever closing when the subsequent turn_end arrived.
+    const run = {
+      stdinOpen: true,
+      turnCompletedCleanly: false,
+      child: {
+        stdin: {
+          destroyed: false,
+          end: vi.fn(),
+        },
+      },
+    };
+
+    applyClaudeStreamJsonRunBookkeeping(run, {
+      type: 'tool_use',
+      name: 'AskUserQuestion',
+      id: 'auq_123',
+    });
+
+    applyClaudeStreamJsonRunBookkeeping(run, {
+      type: 'turn_end',
+      stopReason: 'end_turn',
+    });
+
+    expect(run.turnCompletedCleanly).toBe(true);
+    expect(run.stdinOpen).toBe(false);
+    expect(run.child.stdin.end).toHaveBeenCalled();
+  });
 });

specs/change/20260529-claude-session-resume/spec.md

@@ -34,8 +34,10 @@ Constraints:
 - Do not regress any existing behavior. Resume is best-effort: when a guard
   fails, the capability is absent, or `--resume` is rejected at runtime, the
   daemon falls back to today's full-transcript spawn for that turn.
-- Do not break the interactive `stream-json` / `AskUserQuestion` machinery
-  (`pendingHostAnswers`, `POST /api/runs/:id/tool-result`).
+- Do not break the `stream-json` input skeleton (generic mid-turn stdin
+  infrastructure). Note: the `AskUserQuestion` tool wiring and
+  `POST /api/runs/:id/tool-result` endpoint were removed in PR #4114; the
+  `pendingHostAnswers` dead branch was cleaned up in PR #4273.
 - Keep the daemon the source of truth: the stored session pointer is a cache
   keyed on daemon-owned conversation state, never the authoritative history.
 - Claude-only in v1. Other adapters keep `resumesSessionViaCli` unset and the
@@ -110,9 +112,14 @@ OD is a synchronous, interactive design-chat, not an async issue/task runner.
 These traits are why resume must be guarded rather than unconditional, and why
 it stays Claude-first and opt-out-able:
 
-1. **Interactive mid-turn tools.** OD keeps `stream-json` stdin open to answer
-   `AskUserQuestion` with a real `tool_result`. multica disables it
-   (`--disallowedTools AskUserQuestion`). Resume must not disturb this path.
+1. **Interactive mid-turn clarification.** OD routes clarifying questions
+   through the `<question-form>` markdown artifact; answers flow back as the
+   next user message (`POST /api/chat`). The `AskUserQuestion` tool wiring and
+   `POST /api/runs/:id/tool-result` endpoint were removed (PR #4114); the
+   `stream-json` stdin skeleton is retained only as generic mid-turn input
+   infrastructure, with stdin staying open across `tool_use` pauses and closing only after a terminal non-`tool_use` turn/end result. multica
+   disables `AskUserQuestion` entirely (`--disallowedTools AskUserQuestion`).
+   Resume must not disturb the `<question-form>` clarification path.
 2. **Per-turn prompt rewriting.** OD recomposes the system prompt + skills +
    memory + design system into the `# Instructions` block of the stdin user
    message every turn. Because the instructions ride the stdin message (not a
@@ -334,8 +341,12 @@ per `AGENTS.md`:
 - **Capability probe.** A `claude --help` without `--resume` → flag never sent
   AND every turn carries the full transcript (skip-transcript stays off because
   `resolveResumeDecision` returns null without the capability).
-- **Interactive intact.** An `AskUserQuestion` turn still resolves via
-  `POST /api/runs/:id/tool-result` under resume.
+- **Interactive intact.** A clarifying-question turn (`<question-form>` artifact)
+  still resolves through `POST /api/chat` (the user's answer as the next message)
+  under resume; the `<question-form>` path is independent of session state and
+  is unaffected by `--resume`. (The `AskUserQuestion` tool wiring and
+  `POST /api/runs/:id/tool-result` endpoint were removed in PR #4114 and are
+  not part of this feature.)
 
 Human verification (per `AGENTS.md`, two namespaced runtimes: `main` vs branch):
 drive a multi-turn chat through production HTTP only; confirm continuity holds