AITURK IDE 1.0.0-beta.18: gorunmeyen onay beklemesi yerine aninda "onay gerekli"

A one-shot run (hermes chat --query-file ... -Q, stdin closed, no window)
still installs the prompt_toolkit approval callback but never starts the
prompt UI. The protected instruction-file gate (AGENTS.md / CLAUDE.md /
SOUL.md ...) took the registered callback for a reachable human and waited
out approvals.timeout on a panel painted nowhere. With the profile value of
300000 s the stage job hung for good.

- Gate: in a one-shot run, or when only an automatic approver
  (subagent_auto_approve) is attached, return BLOCKED / APPROVAL REQUIRED
  immediately instead of prompting. Human consent is still required every
  time; yolo and approvals.single_query_mode still do not open this gate.
- One-shot exit prints an "approval_required: {json}" line on stderr
  (session_id, tool_call_id, request_id, targets, reason, applied=false)
  so the launcher can tell "blocked, needs a human" from "still working".
- execute_code: an approval that arrives after the calling script stopped
  waiting (300 s reply timeout, script run timeout, retired kernel cell)
  is not applied.
- Same invisible wait removed from the computer_use and memory-write
  prompts in a one-shot run. The one-shot marker is no longer inherited by
  child processes.
- BLOCKED text no longer breaks on a target path containing braces.

Long-running tools are untouched: no general tool timeout was shortened.

Kart: turkservis-20261007-1421-2

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EwSpve9HuXMy7dHf4RDMzX

tools/approval.py

@@ -78,6 +78,115 @@ _hermes_interactive_ctx: contextvars.ContextVar[Optional[str]] = contextvars.Con
 )
 
 
+# Absolute ``time.monotonic()`` instant after which whoever issued the
+# current tool call has stopped waiting for its result. Set by the
+# execute_code RPC dispatchers, whose sandbox client gives up after a fixed
+# reply timeout. An approval that resolves after this instant must NOT be
+# acted on: the caller already reported the call as failed, so applying it
+# would be a write the agent believes never happened (and may redo).
+# The reply timeout is not the only way a caller goes away: the script can
+# hit its own run timeout or be interrupted first. ``gone`` is a probe the
+# dispatcher supplies for that ("has my script/cell already ended?").
+# Thread-local on purpose: the dispatcher and the tool handler it calls run
+# on the same thread, whatever context the handler is re-bound to.
+_approval_reply_deadline = threading.local()
+
+
+def set_approval_reply_deadline(deadline: Optional[float], gone=None):
+    """Bind the caller's reply window for this thread; return the prior binding.
+
+    ``deadline`` is monotonic seconds (None = unbounded); ``gone`` is an
+    optional zero-argument callable that turns true once the caller has
+    ended. Pass the returned value to :func:`restore_approval_reply_deadline`.
+    """
+    previous = (
+        getattr(_approval_reply_deadline, "value", None),
+        getattr(_approval_reply_deadline, "gone", None),
+    )
+    _approval_reply_deadline.value = deadline
+    _approval_reply_deadline.gone = gone
+    return previous
+
+
+def restore_approval_reply_deadline(previous) -> None:
+    """Undo :func:`set_approval_reply_deadline` / :func:`also_gone_when`."""
+    deadline, gone = previous if previous else (None, None)
+    _approval_reply_deadline.value = deadline
+    _approval_reply_deadline.gone = gone
+
+
+def also_gone_when(probe):
+    """Add another "caller has ended" probe, keeping the current window."""
+    deadline = getattr(_approval_reply_deadline, "value", None)
+    current = getattr(_approval_reply_deadline, "gone", None)
+
+    def gone() -> bool:
+        return bool(probe()) or bool(current and current())
+
+    return set_approval_reply_deadline(deadline, gone)
+
+
+def approval_reply_window_remaining() -> Optional[float]:
+    """Seconds left before the caller stops waiting, or None when unbounded."""
+    deadline = getattr(_approval_reply_deadline, "value", None)
+    if deadline is None:
+        return None
+    return deadline - time.monotonic()
+
+
+def approval_arrived_too_late() -> bool:
+    """True when the caller of the current tool call has stopped waiting."""
+    gone = getattr(_approval_reply_deadline, "gone", None)
+    if gone is not None:
+        try:
+            if gone():
+                return True
+        except Exception:
+            pass
+    remaining = approval_reply_window_remaining()
+    return remaining is not None and remaining <= 0
+
+
+# Approvals that could not be put in front of a human because the run has
+# no interactive channel (single-query ``chat -Q`` stage runs, scripts).
+# The single-query CLI reports them on stderr as ``approval_required:``
+# lines so the coordinating process gets a machine-readable outcome
+# instead of inferring one from prose. Bounded: long-lived processes that
+# never drain it must not grow without limit.
+_human_approval_required: list[dict] = []
+_HUMAN_APPROVAL_REQUIRED_MAX = 50
+_human_approval_required_lock = threading.Lock()
+
+
+def note_human_approval_required(kind: str, targets: list[str], reason: str) -> dict:
+    """Record an approval the run could not present to a human.
+
+    Captures the active session/tool-call identity so the coordinator can
+    tie the outcome to the exact job and resume the same session.
+    """
+    record = {
+        "kind": kind,
+        "targets": list(dict.fromkeys(str(t) for t in targets)),
+        "reason": reason,
+        "request_id": uuid.uuid4().hex,
+        "tool_call_id": _approval_tool_call_id.get() or "",
+        "session_id": _approval_session_id.get() or os.getenv("HERMES_SESSION_ID", ""),
+        "applied": False,
+    }
+    with _human_approval_required_lock:
+        _human_approval_required.append(record)
+        del _human_approval_required[:-_HUMAN_APPROVAL_REQUIRED_MAX]
+    return dict(record)
+
+
+def drain_human_approval_required() -> list[dict]:
+    """Return and clear the recorded approval-required outcomes."""
+    with _human_approval_required_lock:
+        records = list(_human_approval_required)
+        _human_approval_required.clear()
+    return records
+
+
 def set_hermes_interactive_context(interactive: bool) -> contextvars.Token:
     """Bind interactive mode for the current context (thread or asyncio task).
 

tools/code_execution_tool.py

@@ -716,6 +716,36 @@ def _call(tool_name, args):
 # RPC server (runs in a thread inside the parent process)
 # ---------------------------------------------------------------------------
 
+# How long the generated sandbox client waits for one tool-call reply
+# (``_sock.settimeout(300)`` / the file transport's 300s poll deadline).
+_RPC_REPLY_TIMEOUT_SECONDS = 300
+
+
+def _bind_rpc_reply_deadline(call_start: float, stop_event=None):
+    """Tell the approval gates when this call's requester stops waiting.
+
+    The requester is gone at the reply timeout, or earlier when the script
+    itself ended (run timeout, interrupt): ``stop_event`` is set then.
+    """
+    try:
+        from tools.approval import set_approval_reply_deadline
+        gone = stop_event.is_set if stop_event is not None else None
+        return (True, set_approval_reply_deadline(
+            call_start + _RPC_REPLY_TIMEOUT_SECONDS, gone))
+    except Exception:
+        return (False, None)
+
+
+def _restore_rpc_reply_deadline(previous) -> None:
+    if not previous or not previous[0]:
+        return
+    try:
+        from tools.approval import restore_approval_reply_deadline
+        restore_approval_reply_deadline(previous[1])
+    except Exception:
+        pass
+
+
 # Terminal parameters that must not be used from ephemeral sandbox scripts
 _TERMINAL_BLOCKED_PARAMS = {"background", "pty", "notify", "notify_on_complete", "watch_patterns"}
 
@@ -826,12 +856,18 @@ def _rpc_server_loop(
                 # Dispatch through the standard tool handler.
                 # Suppress stdout/stderr from internal tool handlers so
                 # their status prints don't leak into the CLI spinner.
+                # The sandbox client stops waiting for this reply after
+                # _RPC_REPLY_TIMEOUT_SECONDS; an approval granted later
+                # must not apply a write the script already saw fail.
+                _prev_deadline = _bind_rpc_reply_deadline(call_start, stop_event)
                 try:
                     with thread_scoped_silence():
                         result = dispatch(tool_name, tool_args)
                 except Exception as exc:
                     logger.error("Tool call failed in sandbox: %s", exc, exc_info=True)
                     result = tool_error(str(exc))
+                finally:
+                    _restore_rpc_reply_deadline(_prev_deadline)
 
                 tool_call_counter[0] += 1
                 call_duration = time.monotonic() - call_start
@@ -1101,6 +1137,7 @@ def _rpc_poll_loop(
                             tool_args.pop(param, None)
 
                     # Dispatch through the standard tool handler
+                    _prev_deadline = _bind_rpc_reply_deadline(call_start, stop_event)
                     try:
                         with thread_scoped_silence():
                             tool_result = handle_function_call(
@@ -1110,6 +1147,8 @@ def _rpc_poll_loop(
                         logger.error("Tool call failed in remote sandbox: %s",
                                      exc, exc_info=True)
                         tool_result = tool_error(str(exc))
+                    finally:
+                        _restore_rpc_reply_deadline(_prev_deadline)
 
                     tool_call_counter[0] += 1
                     call_duration = time.monotonic() - call_start

tools/code_kernel.py

@@ -322,9 +322,25 @@ class CellAuthority:
                 set_sudo(self._sudo_cb)
             except Exception:
                 previous = None
+        # A cell that settled (timeout, interrupt, error) no longer waits for
+        # this call: an approval granted afterwards must not be applied.
+        reply_window = None
+        try:
+            from tools.approval import also_gone_when
+
+            reply_window = (also_gone_when(lambda: not self.active),)
+        except Exception:
+            reply_window = None
         try:
             return handle_function_call(tool_name, tool_args, task_id=self.task_id)
         finally:
+            if reply_window is not None:
+                try:
+                    from tools.approval import restore_approval_reply_deadline
+
+                    restore_approval_reply_deadline(reply_window[0])
+                except Exception:
+                    pass
             if previous is not None and self._callback_setters is not None:
                 set_approval, set_sudo = self._callback_setters
                 try:

tools/computer_use/tool.py

@@ -613,6 +613,22 @@ def _request_approval(action: str, args: Dict[str, Any],
         # No CLI approval wired — default allow. Gateway approval is handled
         # one layer out via the normal tool-approval infra.
         return None
+    try:
+        from tools.approval import _is_single_query_approval_context
+        unattended = _is_single_query_approval_context()
+    except Exception:
+        unattended = False
+    if unattended:
+        # One-shot run: the prompt callback is installed but its panel is
+        # never shown, so asking would only wait out approvals.timeout.
+        return json.dumps({
+            "error": (
+                "approval required but this is a one-shot run with no "
+                "interactive user to approve it. Silence is not consent; "
+                "do not retry without the user."
+            ),
+            "action": action,
+        })
     summary = _summarize_action(action, args)
     try:
         verdict = cb(action, args, summary)

tools/delegate_tool.py

@@ -102,6 +102,11 @@ def _subagent_auto_approve(command: str, description: str, **kwargs) -> str:
     return "once"
 
 
+# Not a person: gates that require real human consent every time (protected
+# instruction files) must not accept this callback's answer.
+_subagent_auto_approve.unattended = True
+
+
 def _get_subagent_approval_callback():
     """Return the callback to install into subagent worker threads.
 

tools/environments/local.py

@@ -1546,6 +1546,10 @@ def _make_run_env(env: dict) -> dict:
             run_env[real_key] = v
         elif _is_hermes_internal_secret(k):
             continue
+        elif k == "HERMES_SINGLE_QUERY_SESSION":
+            # Describes this process only. A child (an interactive hermes
+            # started from a one-shot run's terminal) sets its own.
+            continue
         else:
             first_party = _is_terminal_first_party_env(k)
             passthrough = _is_passthrough(k)

tools/file_tools.py

@@ -861,12 +861,18 @@ def _request_protected_instruction_approval(
         "required (not bypassed by auto-approve)."
     )
     display = f"<write to {targets}>"
-    blocked = (
-        f"BLOCKED: write to protected agent-instruction file(s) ({targets}) "
-        "{why} The user has NOT consented to this write. Do NOT retry it or "
-        "attempt the same edit via another path (terminal, execute_code, "
-        "etc.)."
-    )
+    class _Blocked:
+        # Not str.format: a target path may itself contain braces.
+        @staticmethod
+        def format(why: str) -> str:
+            return (
+                "BLOCKED: write to protected agent-instruction file(s) "
+                f"({targets}) {why} The user has NOT consented to this "
+                "write. Do NOT retry it or attempt the same edit via "
+                "another path (terminal, execute_code, etc.)."
+            )
+
+    blocked = _Blocked
 
     try:
         import tools.approval as _approval
@@ -874,6 +880,41 @@ def _request_protected_instruction_approval(
         return blocked.format(why="requires approval but the approval "
                                   "subsystem is unavailable.")
 
+    def _needs_human(reason: str, why: str) -> str:
+        # No human could be shown this request. Record it so a one-shot
+        # run can hand its coordinator a machine-readable outcome, and tell
+        # the agent where a human CAN approve it instead of waiting on a
+        # prompt nobody sees.
+        try:
+            _approval.note_human_approval_required(
+                "protected_instruction_file", list(reasons), reason)
+        except Exception:
+            pass
+        return blocked.format(
+            why=why + " APPROVAL REQUIRED: a human must approve this from "
+                      "an interactive session (resume this same session in "
+                      "the desktop app or an interactive terminal and "
+                      "request the write again there).")
+
+    def _too_late() -> str | None:
+        # The approval came back after whoever issued this call stopped
+        # waiting (execute_code RPC reply timeout). The caller already saw
+        # a failure, so applying the write now would be an edit the agent
+        # believes never happened.
+        try:
+            late = _approval.approval_arrived_too_late()
+        except Exception:
+            late = False
+        if not late:
+            return None
+        logger.warning(
+            "Approval for protected instruction file(s) %s arrived after "
+            "the caller stopped waiting; write not applied.", targets)
+        return blocked.format(
+            why="was approved only after the calling script had stopped "
+                "waiting for the result, so the write was NOT applied. "
+                "The approval is spent; a new request needs a new approval.")
+
     # Gateway surface: block on the button round-trip when a notify callback
     # is registered for this session (Telegram/Discord/Slack). One-operation
     # only — no session/permanent buttons are offered.
@@ -898,14 +939,15 @@ def _request_protected_instruction_approval(
             session_key, notify_cb, approval_data, surface="gateway",
         )
         if decision.get("notify_failed"):
-            return blocked.format(
-                why="requires approval but the approval request could not "
-                    "be delivered.")
+            return _needs_human(
+                "approval_not_delivered",
+                "requires approval but the approval request could not "
+                "be delivered.")
         choice = decision.get("choice")
         if decision.get("resolved") and choice in {"once", "session", "always"}:
             # One-operation grant regardless of the tapped scope — nothing
             # is persisted for this gate.
-            return None
+            return _too_late()
         if not decision.get("resolved"):
             return blocked.format(
                 why="approval prompt timed out without a user response. "
@@ -921,6 +963,28 @@ def _request_protected_instruction_approval(
         callback = None
 
     if callback is not None:
+        # A one-shot run (``chat -q`` / ``--query-file``) installs the
+        # prompt callback but never starts the prompt UI: the panel would
+        # be painted nowhere and the wait would last the whole
+        # approvals.timeout with nobody able to answer. A registered
+        # callback is not a reachable human. Fail closed NOW;
+        # approvals.single_query_mode does not apply to this gate.
+        try:
+            unattended = _approval._is_single_query_approval_context()
+        except Exception:
+            unattended = False
+        if unattended:
+            return _needs_human(
+                "no_interactive_user",
+                "requires approval but this is a one-shot run with no "
+                "interactive user to approve it.")
+        if getattr(callback, "unattended", False):
+            # An auto-answering callback (subagent auto-approve) is not a
+            # person; this gate takes consent from a human only.
+            return _needs_human(
+                "no_interactive_user",
+                "requires approval but only an automatic approver is "
+                "attached to this run, not a person.")
         choice = _approval.prompt_dangerous_approval(
             display, description,
             allow_permanent=False,
@@ -929,7 +993,7 @@ def _request_protected_instruction_approval(
         )
         if choice in {"once", "session", "always"}:
             # One-operation grant; never persisted (see docstring).
-            return None
+            return _too_late()
         if choice == "timeout":
             return blocked.format(
                 why="approval prompt timed out without a user response. "
@@ -938,9 +1002,10 @@ def _request_protected_instruction_approval(
 
     # No human channel at all (script, cron, background thread): fail
     # closed. Auto-approving here would recreate the persistence vector.
-    return blocked.format(
-        why="requires approval but no interactive user or gateway is "
-            "present to approve it.")
+    return _needs_human(
+        "no_approval_channel",
+        "requires approval but no interactive user or gateway is "
+        "present to approve it.")
 
 
 def _check_protected_instruction_write(paths: list[str],

tools/write_approval.py

@@ -327,6 +327,14 @@ def _interactive_approval_available() -> bool:
       review affordance (``/memory pending``) instead.
     * Scripts, cron, and background threads — no user present.
     """
+    try:
+        from tools.approval import _is_single_query_approval_context
+        if _is_single_query_approval_context():
+            # One-shot run: the callback exists but its panel is never
+            # shown, so an inline prompt would wait with nobody to answer.
+            return False
+    except Exception:
+        pass
     try:
         from tools.terminal_tool import _get_approval_callback
         return _get_approval_callback() is not None