Merge pull request #63104 from NousResearch/bb/active-turn-steering

feat: redirect active turns when users correct the agent

acp_adapter/server.py

@@ -1218,12 +1218,19 @@ class HermesACPAgent(acp.Agent):
             with state.runtime_lock:
                 if state.is_running and state.current_prompt_text:
                     state.interrupted_prompt_text = state.current_prompt_text
-            state.cancel_event.set()
-            try:
-                if getattr(state, "agent", None) and hasattr(state.agent, "interrupt"):
-                    state.agent.interrupt()
-            except Exception:
-                logger.debug("Failed to interrupt ACP session %s", session_id, exc_info=True)
+                # Publish cancellation and hard-stop the agent before another
+                # prompt can acquire this lock and mistake the turn for
+                # redirectable work.
+                state.cancel_event.set()
+                try:
+                    if getattr(state, "agent", None) and hasattr(state.agent, "interrupt"):
+                        state.agent.interrupt()
+                except Exception:
+                    logger.debug(
+                        "Failed to interrupt ACP session %s",
+                        session_id,
+                        exc_info=True,
+                    )
             logger.info("Cancelled session %s", session_id)
 
     async def fork_session(
@@ -1352,6 +1359,26 @@ class HermesACPAgent(acp.Agent):
             elif rewrite_idle:
                 user_text = steer_text
                 user_content = steer_text
+        elif (
+            text_only_prompt
+            and isinstance(user_content, str)
+            and not user_text.startswith("/")
+        ):
+            # Some ACP clients implement "stop and send" as two protocol calls:
+            # cancel the active prompt, then submit plain correction text. Keep
+            # the cancelled request attached so deictic follow-ups ("not that
+            # file") still have an explicit target.
+            interrupted_prompt = ""
+            with state.runtime_lock:
+                if not state.is_running and state.interrupted_prompt_text:
+                    interrupted_prompt = state.interrupted_prompt_text
+                    state.interrupted_prompt_text = ""
+            if interrupted_prompt:
+                user_text = (
+                    f"{interrupted_prompt}\n\n"
+                    f"User correction/guidance after interrupt: {user_text}"
+                )
+                user_content = user_text
 
         # Intercept slash commands — handle locally without calling the LLM.
         # Slash commands are text-only; if the client included images/resources,
@@ -1366,23 +1393,54 @@ class HermesACPAgent(acp.Agent):
                     await self._send_usage_update(state)
                 return PromptResponse(stop_reason="end_turn")
 
-        # If Zed sends another regular prompt while the same ACP session is
-        # still running, queue it instead of racing two AIAgent loops against
-        # the same state.history. /steer and /queue are handled above and can
-        # land immediately.
+        # If the client sends another regular text prompt while this ACP session
+        # is running, route it through the core active-turn redirect. Rich media
+        # and older runtimes retain the proven next-turn queue fallback.
+        redirected = False
+        queued_depth: int | None = None
         with state.runtime_lock:
             if state.is_running:
-                queued_text = user_text or "[Image attachment]"
-                state.queued_prompts.append(queued_text)
-                depth = len(state.queued_prompts)
-                if self._conn:
-                    update = acp.update_agent_message_text(
-                        f"Queued for the next turn. ({depth} queued)"
+                if (
+                    text_only_prompt
+                    and isinstance(user_content, str)
+                    and getattr(
+                        state.agent,
+                        "_supports_active_turn_redirect",
+                        False,
                     )
-                    await self._conn.session_update(session_id, update)
-                return PromptResponse(stop_reason="end_turn")
-            state.is_running = True
-            state.current_prompt_text = user_text or "[Image attachment]"
+                    is True
+                    and hasattr(state.agent, "redirect")
+                ):
+                    try:
+                        redirected = bool(state.agent.redirect(user_content))
+                    except Exception:
+                        logger.debug(
+                            "ACP active-turn redirect failed for %s",
+                            session_id,
+                            exc_info=True,
+                        )
+                if not redirected:
+                    queued_text = user_text or "[Image attachment]"
+                    state.queued_prompts.append(queued_text)
+                    queued_depth = len(state.queued_prompts)
+            else:
+                state.is_running = True
+                state.current_prompt_text = user_text or "[Image attachment]"
+
+        if redirected:
+            if self._conn:
+                update = acp.update_agent_message_text(
+                    "Redirected the active turn with your correction."
+                )
+                await self._conn.session_update(session_id, update)
+            return PromptResponse(stop_reason="end_turn")
+        if queued_depth is not None:
+            if self._conn:
+                update = acp.update_agent_message_text(
+                    f"Queued for the next turn. ({queued_depth} queued)"
+                )
+                await self._conn.session_update(session_id, update)
+            return PromptResponse(stop_reason="end_turn")
 
         logger.info("Prompt on session %s: %s", session_id, user_text[:100])
 

agent/agent_init.py

@@ -720,6 +720,8 @@ def init_agent(
     agent._execution_thread_id: int | None = None  # Set at run_conversation() start
     agent._interrupt_thread_signal_pending = False
     agent._client_lock = threading.RLock()
+    agent._model_request_active = threading.Event()
+    agent._supports_active_turn_redirect = True
 
     # /steer mechanism — inject a user note into the next tool result
     # without interrupting the agent. Unlike interrupt(), steer() does
@@ -731,6 +733,13 @@ def init_agent(
     agent._pending_steer: Optional[str] = None
     agent._pending_steer_lock = threading.Lock()
 
+    # Active-turn redirect mechanism. A regular follow-up sent while the model
+    # is generating is different from a hard /stop: preserve the valid turn
+    # prefix, cancel only the in-flight model request, and rebuild its tail with
+    # the correction. The loop drains this slot at a role-safe boundary.
+    agent._pending_redirect: Optional[str] = None
+    agent._pending_redirect_lock = threading.Lock()
+
     # Concurrent-tool worker thread tracking.  `_execute_tool_calls_concurrent`
     # runs each tool on its own ThreadPoolExecutor worker — those worker
     # threads have tids distinct from `_execution_thread_id`, so
@@ -897,6 +906,12 @@ def init_agent(
     agent._stream_writer_tls = threading.local()
     agent._stream_writer_dropped = 0
 
+    # Displayed reasoning text streamed during the current model response,
+    # captured only when a surface consumed it via a reasoning callback. Used
+    # by active-turn redirect to checkpoint what the user actually saw without
+    # ever persisting hidden provider reasoning.
+    agent._current_streamed_reasoning_text = ""
+
     # Optional current-turn user-message override used when the API-facing
     # user message intentionally differs from the persisted transcript
     # (e.g. CLI voice mode adds a temporary prefix for the live call only).

agent/codex_runtime.py

@@ -702,6 +702,16 @@ def run_codex_app_server_turn(
         except Exception:
             pass
         agent._codex_session = None
+        _user_interrupted = bool(
+            getattr(agent, "_interrupt_requested", False)
+        )
+        _interrupt_message = (
+            getattr(agent, "_interrupt_message", None)
+            if _user_interrupted
+            else None
+        )
+        if _user_interrupted:
+            agent.clear_interrupt()
         return {
             "final_response": (
                 f"Codex app-server turn failed: {exc}. "
@@ -711,9 +721,27 @@ def run_codex_app_server_turn(
             "api_calls": 0,
             "completed": False,
             "partial": True,
+            "interrupted": _user_interrupted,
+            **(
+                {"interrupt_message": _interrupt_message}
+                if _interrupt_message
+                else {}
+            ),
             "error": str(exc),
         }
 
+    # This runtime bypasses the normal conversation-loop finalizer. Mirror its
+    # interrupt handoff/cleanup so a hard stop cannot poison the next turn and a
+    # message-bearing compatibility interrupt can still be replayed by callers.
+    _user_interrupted = bool(
+        turn.interrupted and getattr(agent, "_interrupt_requested", False)
+    )
+    _interrupt_message = (
+        getattr(agent, "_interrupt_message", None) if _user_interrupted else None
+    )
+    if _user_interrupted:
+        agent.clear_interrupt()
+
     # If the turn signalled the underlying client is wedged (deadline
     # blown, post-tool watchdog tripped, OAuth refresh died, subprocess
     # exited), retire the session so the next turn respawns codex
@@ -819,6 +847,12 @@ def run_codex_app_server_turn(
         "api_calls": api_calls,
         "completed": not turn.interrupted and turn.error is None,
         "partial": turn.interrupted or turn.error is not None,
+        "interrupted": _user_interrupted,
+        **(
+            {"interrupt_message": _interrupt_message}
+            if _interrupt_message
+            else {}
+        ),
         "error": turn.error,
         # The codex app-server runtime IS an early-return path that bypasses
         # conversation_loop, but we flush the projected assistant/tool messages

agent/conversation_loop.py

@@ -103,6 +103,53 @@ _API_CALL_MODULES = frozenset({
 })
 
 
+def _apply_active_turn_redirect(agent: Any, messages: List[Dict[str, Any]], text: str) -> None:
+    """Append a provider-safe checkpoint and correction to the live turn.
+
+    Incomplete provider reasoning blocks are not valid replay items (Anthropic
+    signs them; Responses reasoning items require their following output).
+    Preserve only what Hermes actually displayed, demoted to ordinary text,
+    then add the correction as a real user message. This keeps role alternation
+    valid and leaves every previously cached message byte-for-byte unchanged.
+    """
+    reasoning = str(
+        getattr(agent, "_current_streamed_reasoning_text", "") or ""
+    ).strip()
+    visible = agent._strip_think_blocks(
+        getattr(agent, "_current_streamed_assistant_text", "") or ""
+    ).strip()
+
+    checkpoint_parts = ["[This response was interrupted by a user correction.]"]
+    if reasoning:
+        checkpoint_parts.extend(
+            ["Reasoning shown before the interruption:", reasoning]
+        )
+    if visible:
+        checkpoint_parts.extend(
+            ["Visible response before the interruption:", visible]
+        )
+    checkpoint = "\n\n".join(checkpoint_parts)
+
+    # The normal live tail is user or tool, so an assistant checkpoint followed
+    # by the correction preserves strict alternation. If a transport already
+    # committed an assistant item, attribute the checkpoint inside the user
+    # correction instead of creating assistant→assistant.
+    if messages and messages[-1].get("role") == "assistant":
+        correction = (
+            "[Context from the interrupted assistant response]\n"
+            f"{checkpoint}\n\n"
+            f"{text}"
+        )
+        messages.append({"role": "user", "content": correction})
+    else:
+        messages.append({"role": "assistant", "content": checkpoint})
+        messages.append({"role": "user", "content": text})
+
+    agent._current_streamed_assistant_text = ""
+    agent._current_streamed_reasoning_text = ""
+    agent._stream_needs_break = True
+
+
 def _image_error_max_dimension(error: Exception) -> Optional[int]:
     """Extract a provider-reported image dimension ceiling, if present."""
     parts = []
@@ -733,6 +780,16 @@ def run_conversation(
         )
 
     while (api_call_count < agent.max_iterations and agent.iteration_budget.remaining > 0) or agent._budget_grace_call:
+        _redirect_text = agent._drain_pending_redirect()
+        if _redirect_text:
+            _apply_active_turn_redirect(agent, messages, _redirect_text)
+            if isinstance(original_user_message, str):
+                original_user_message = (
+                    f"{original_user_message}\n\n"
+                    f"User correction during the turn: {_redirect_text}"
+                )
+            agent._persist_session(messages, conversation_history)
+
         # Reset per-turn checkpoint dedup so each iteration can take one snapshot
         agent._checkpoint_mgr.new_turn()
 
@@ -1532,22 +1589,59 @@ def run_conversation(
 
                 from hermes_cli.middleware import run_llm_execution_middleware
 
-                response = run_llm_execution_middleware(
-                    api_kwargs,
-                    _perform_api_call,
-                    original_request=_original_api_kwargs,
-                    task_id=effective_task_id,
-                    turn_id=turn_id,
-                    api_request_id=api_request_id,
-                    session_id=agent.session_id or "",
-                    platform=agent.platform or "",
-                    model=agent.model,
-                    provider=agent.provider,
-                    base_url=agent.base_url,
-                    api_mode=agent.api_mode,
-                    api_call_count=api_call_count,
-                    middleware_trace=list(_llm_middleware_trace),
-                )
+                _model_request_active = getattr(agent, "_model_request_active", None)
+                _redirect_lock = getattr(agent, "_pending_redirect_lock", None)
+                if _redirect_lock is not None:
+                    with _redirect_lock:
+                        if _model_request_active is not None:
+                            _model_request_active.set()
+                elif _model_request_active is not None:
+                    _model_request_active.set()
+                _redirect_crossed_response = False
+                try:
+                    response = run_llm_execution_middleware(
+                        api_kwargs,
+                        _perform_api_call,
+                        original_request=_original_api_kwargs,
+                        task_id=effective_task_id,
+                        turn_id=turn_id,
+                        api_request_id=api_request_id,
+                        session_id=agent.session_id or "",
+                        platform=agent.platform or "",
+                        model=agent.model,
+                        provider=agent.provider,
+                        base_url=agent.base_url,
+                        api_mode=agent.api_mode,
+                        api_call_count=api_call_count,
+                        middleware_trace=list(_llm_middleware_trace),
+                    )
+                finally:
+                    if _redirect_lock is not None:
+                        with _redirect_lock:
+                            if _model_request_active is not None:
+                                _model_request_active.clear()
+                            _redirect_crossed_response = bool(
+                                agent._pending_redirect
+                            )
+                    else:
+                        if _model_request_active is not None:
+                            _model_request_active.clear()
+                        _redirect_crossed_response = agent._has_pending_redirect()
+                if _redirect_crossed_response:
+                    # The response and redirect can cross on different threads:
+                    # redirect() observed the request as active just before this
+                    # call returned. Discard that now-stale response and rebuild
+                    # from the correction rather than silently losing it.
+                    if thinking_spinner:
+                        thinking_spinner.stop("")
+                        thinking_spinner = None
+                    if agent.thinking_callback:
+                        agent.thinking_callback("")
+                    if agent.clear_interrupt(preserve_redirect=True):
+                        _retry.restart_with_redirected_messages = True
+                    else:
+                        interrupted = True
+                    break
                 
                 api_duration = time.time() - api_start_time
                 
@@ -2509,6 +2603,15 @@ def run_conversation(
                     thinking_spinner = None
                 if agent.thinking_callback:
                     agent.thinking_callback("")
+                if agent._has_pending_redirect():
+                    # redirect() deliberately used the interrupt machinery to
+                    # cancel only this provider request. Keep its correction
+                    # queued, clear the cancellation bit, and let the outer
+                    # loop rebuild a clean request tail. Never materialize
+                    # incomplete signed/encrypted reasoning items.
+                    if agent.clear_interrupt(preserve_redirect=True):
+                        _retry.restart_with_redirected_messages = True
+                        break
                 api_elapsed = time.time() - api_start_time
                 agent._vprint(f"{agent.log_prefix}⚡ Interrupted during API call.", force=True)
                 interrupted = True
@@ -4457,6 +4560,15 @@ def run_conversation(
                             f"{int(sleep_end - time.time())}s remaining"
                         )
         
+        if _retry.restart_with_redirected_messages:
+            # The cancelled request produced no valid assistant item. Reuse the
+            # same logical iteration after the outer loop appends the displayed
+            # partial context and correction to ``messages``.
+            api_call_count -= 1
+            agent.iteration_budget.refund()
+            _retry.restart_with_redirected_messages = False
+            continue
+
         # If the API call was interrupted, skip response processing
         if interrupted:
             _turn_exit_reason = "interrupted_during_api_call"

agent/onboarding.py

@@ -52,6 +52,13 @@ def busy_input_hint_gateway(mode: str) -> str:
             "Send `/busy interrupt` or `/busy queue` to change this, or "
             "`/busy status` to check. This notice won't appear again."
         )
+    if mode == "redirect":
+        return (
+            "💡 First-time tip — I redirected the current run using your message. "
+            "Completed work stays in context, and `/stop` still cancels the task. "
+            "Send `/busy queue` to wait for a separate turn, or `/busy status` "
+            "to check. This notice won't appear again."
+        )
     return (
         "💡 First-time tip — I just interrupted my current task to answer you. "
         "Send `/busy queue` to queue follow-ups for after the current task instead, "
@@ -74,6 +81,12 @@ def busy_input_hint_cli(mode: str) -> str:
             "after the next tool call. Use /busy interrupt or /busy queue to "
             "change this. This tip only shows once."
         )
+    if mode == "redirect":
+        return (
+            "(tip) Your correction redirected the current run without discarding "
+            "completed work. Use /stop to cancel or /busy queue to wait for a "
+            "separate turn. This tip only shows once."
+        )
     return (
         "(tip) Your message interrupted the current run. "
         "Use /busy queue to queue messages for the next turn instead, "

agent/transports/codex_app_server_session.py

@@ -300,6 +300,8 @@ class CodexAppServerSession:
         self._client: Optional[CodexAppServerClient] = None
         self._thread_id: Optional[str] = None
         self._interrupt_event = threading.Event()
+        self._active_turn_id: Optional[str] = None
+        self._active_turn_lock = threading.Lock()
         # Pending file-change items, keyed by item id. Populated on
         # item/started for fileChange items; consumed by the approval
         # bridge when codex sends item/fileChange/requestApproval. The
@@ -374,6 +376,8 @@ class CodexAppServerSession:
         if self._closed:
             return
         self._closed = True
+        with self._active_turn_lock:
+            self._active_turn_id = None
         if self._client is not None:
             try:
                 self._client.close()
@@ -395,6 +399,33 @@ class CodexAppServerSession:
         and unwind. Called by AIAgent's _interrupt_requested path."""
         self._interrupt_event.set()
 
+    def request_steer(self, text: str) -> bool:
+        """Append user guidance to the active Codex turn via ``turn/steer``."""
+        cleaned = str(text or "").strip()
+        if not cleaned:
+            return False
+        with self._active_turn_lock:
+            turn_id = self._active_turn_id
+            thread_id = self._thread_id
+            client = self._client
+        if not turn_id or not thread_id or client is None:
+            return False
+        try:
+            response = client.request(
+                "turn/steer",
+                {
+                    "threadId": thread_id,
+                    "input": [{"type": "text", "text": cleaned}],
+                    "expectedTurnId": turn_id,
+                },
+                timeout=10,
+            )
+        except (CodexAppServerError, TimeoutError):
+            logger.debug("turn/steer rejected for active Codex turn", exc_info=True)
+            return False
+        accepted_turn_id = response.get("turnId") if isinstance(response, dict) else None
+        return accepted_turn_id in {None, turn_id}
+
     # ---------- diagnostics ----------
 
     def _format_error_with_stderr(
@@ -469,11 +500,18 @@ class CodexAppServerSession:
             # Subprocess almost certainly unhealthy — retire so the next
             # turn re-spawns cleanly.
             result.should_retire = True
+            self._interrupt_event.clear()
             return result
         assert self._client is not None and self._thread_id is not None
         result.thread_id = self._thread_id
 
-        self._interrupt_event.clear()
+        # Do not clear here: a hard stop can arrive while ensure_started() is
+        # spawning/initializing the subprocess. Honor it before launching a
+        # Codex turn instead of erasing the signal.
+        if self._interrupt_event.is_set():
+            result.interrupted = True
+            self._interrupt_event.clear()
+            return result
         projector = CodexEventProjector()
 
         user_input_text = _coerce_turn_input_text(user_input)
@@ -505,6 +543,7 @@ class CodexAppServerSession:
                 result.error = self._format_error_with_stderr(
                     "turn/start failed", exc
                 )
+            self._interrupt_event.clear()
             return result
         except TimeoutError as exc:
             # turn/start hanging is a strong signal the subprocess is wedged.
@@ -514,9 +553,12 @@ class CodexAppServerSession:
                 "turn/start timed out", exc
             )
             result.should_retire = True
+            self._interrupt_event.clear()
             return result
 
         result.turn_id = (ts.get("turn") or {}).get("id")
+        with self._active_turn_lock:
+            self._active_turn_id = result.turn_id
         deadline = time.monotonic() + turn_timeout
         turn_complete = False
         # Post-tool watchdog state. last_tool_completion_at is set whenever
@@ -741,6 +783,9 @@ class CodexAppServerSession:
                 )
             result.should_retire = True
 
+        with self._active_turn_lock:
+            self._active_turn_id = None
+        self._interrupt_event.clear()
         return result
 
     def compact_thread(

agent/turn_retry_state.py

@@ -73,6 +73,10 @@ class TurnRetryState:
     # was rolled back off ``messages`` and the loop should re-issue the API
     # call against the newly-activated provider (#32421).
     restart_with_rebuilt_messages: bool = False
+    # A user correction cancelled the in-flight provider request. The outer
+    # loop must append a role-safe checkpoint + user message, rebuild the API
+    # payload, and retry the same logical iteration.
+    restart_with_redirected_messages: bool = False
 
     def __iter__(self):
         # Convenience for debugging / tests: iterate (name, value) pairs.

apps/desktop/src/app/chat/composer/controls.tsx

@@ -1,11 +1,9 @@
 import { Button } from '@/components/ui/button'
 import { Codicon } from '@/components/ui/codicon'
-import { KbdCombo } from '@/components/ui/kbd'
 import { Tip } from '@/components/ui/tooltip'
 import { useI18n } from '@/i18n'
 import { triggerHaptic } from '@/lib/haptics'
-import { AudioLines, iconSize, Layers3, Loader2, Square, SteeringWheel, Volume2, VolumeX } from '@/lib/icons'
-import { formatCombo } from '@/lib/keybinds/combo'
+import { AudioLines, iconSize, Layers3, Loader2, Square, Volume2, VolumeX } from '@/lib/icons'
 import { cn } from '@/lib/utils'
 
 import type { ConversationStatus } from './hooks/use-voice-conversation'
@@ -42,7 +40,6 @@ export function ComposerControls({
   autoSpeak,
   busy,
   busyAction,
-  canSteer,
   canSubmit,
   compactModelPill = false,
   conversation,
@@ -51,13 +48,11 @@ export function ComposerControls({
   state,
   voiceStatus,
   onDictate,
-  onSteer,
   onToggleAutoSpeak
 }: {
   autoSpeak: boolean
   busy: boolean
-  busyAction: 'queue' | 'stop'
-  canSteer: boolean
+  busyAction: 'steer' | 'queue' | 'stop'
   canSubmit: boolean
   compactModelPill?: boolean
   conversation: ConversationProps
@@ -66,49 +61,24 @@ export function ComposerControls({
   state: ChatBarState
   voiceStatus: VoiceStatus
   onDictate: () => void
-  onSteer: () => void
   onToggleAutoSpeak: () => void
 }) {
   const { t } = useI18n()
   const c = t.composer
-  const steerCombo = formatCombo('mod+enter')
-  const steerLabel = `${c.steer} (${steerCombo})`
-
-  const steerTip = (
-    <span className="inline-flex items-center gap-1.5">
-      {c.steer}
-      <KbdCombo combo="mod+enter" size="sm" variant="inverted" />
-    </span>
-  )
 
   if (conversation.active) {
     return <ConversationPill {...conversation} disabled={disabled} />
   }
 
   const showVoicePrimary = !busy && !hasComposerPayload
+  // steer + stop both interrupt the live turn (the difference is whether a
+  // correction rides along), so they share the stop glyph; only queue differs.
+  const busyLabel = busyAction === 'queue' ? c.queueMessage : busyAction === 'steer' ? c.steer : c.stop
 
   return (
     <div className="ml-auto flex shrink-0 items-center gap-(--composer-control-gap)">
       <ModelPill compact={compactModelPill} disabled={disabled} model={state.model} />
-      {/* While the agent runs and the user is typing, steer takes over the mic's
-          slot rather than crowding the row with an extra button. */}
-      {canSteer ? (
-        <Tip label={steerTip}>
-          <Button
-            aria-label={steerLabel}
-            className={GHOST_ICON_BTN}
-            disabled={disabled}
-            onClick={onSteer}
-            size="icon"
-            type="button"
-            variant="ghost"
-          >
-            <SteeringWheel className={iconSize.sm} />
-          </Button>
-        </Tip>
-      ) : (
-        <DictationButton disabled={disabled} onToggle={onDictate} state={state.voice} status={voiceStatus} />
-      )}
+      <DictationButton disabled={disabled} onToggle={onDictate} state={state.voice} status={voiceStatus} />
       <AutoSpeakButton active={autoSpeak} disabled={disabled} onToggle={onToggleAutoSpeak} />
       {showVoicePrimary ? (
         <Tip label={c.startVoice}>
@@ -127,9 +97,9 @@ export function ComposerControls({
           </Button>
         </Tip>
       ) : (
-        <Tip label={busy ? (busyAction === 'queue' ? c.queueMessage : c.stop) : c.send}>
+        <Tip label={busy ? busyLabel : c.send}>
           <Button
-            aria-label={busy ? (busyAction === 'queue' ? c.queueMessage : c.stop) : c.send}
+            aria-label={busy ? busyLabel : c.send}
             className={PRIMARY_ICON_BTN}
             disabled={disabled || !canSubmit}
             type="submit"

apps/desktop/src/app/chat/composer/hooks/use-composer-submit.ts

@@ -17,7 +17,6 @@ interface UseComposerSubmitArgs {
   activeQueueSessionKeyRef: RefObject<string | null>
   attachments: ComposerAttachment[]
   busy: boolean
-  canSteer: boolean
   clearDraft: () => void
   disabled: boolean
   draftRef: RefObject<string>
@@ -43,7 +42,7 @@ interface UseComposerSubmitArgs {
  * queue meet. `submitDraft` is the one decision tree (queue-edit save · slash-
  * now-while-busy · queue · drain · send · stop); `dispatchSubmit` is the shared
  * send-with-restore primitive (re-loads + re-stashes the draft if the gateway
- * rejects, so nothing is ever lost); `steerDraft` nudges the live turn. Reads
+ * rejects, so nothing is ever lost); `steerDraft` redirects the live turn. Reads
  * the draft + queue APIs; owns no state of its own beyond the stable
  * external-submit listener ref.
  */
@@ -52,7 +51,6 @@ export function useComposerSubmit({
   activeQueueSessionKeyRef,
   attachments,
   busy,
-  canSteer,
   clearDraft,
   disabled,
   draftRef,
@@ -151,7 +149,14 @@ export function useComposerSubmit({
         triggerHaptic('submit')
         clearDraft()
         dispatchSubmit(text)
+      } else if (!attachments.length && text.trim()) {
+        // Cursor-style stop-and-correct: interrupt the live turn and redirect
+        // it with this text. redirect() preserves the shown reasoning/work; if
+        // the turn already ended, steerDraft re-queues so nothing is lost.
+        steerDraft()
       } else if (payloadPresent) {
+        // Attachments can't ride a redirect (no tool-result image carriage) —
+        // queue the whole payload for the next turn.
         queueCurrentDraft()
       } else {
         // Stop button (the only way to reach here while busy with an empty
@@ -173,16 +178,18 @@ export function useComposerSubmit({
     focusInput()
   }
 
-  // Steer the live turn (nudge without interrupting). Clears the draft up front
-  // for snappy feedback; if the gateway rejects (no live tool window) the words
-  // are re-queued so nothing is lost — same safety net as a plain queue.
+  // Redirect the live turn with a correction. The gateway either restarts the
+  // active model request with its displayed context or waits for the current
+  // tool boundary. If the turn already ended, queue the words instead.
   const steerDraft = () => {
-    if (!onSteer || !canSteer) {
+    const text = draftRef.current.trim()
+
+    // Guard on live editor state, not the render-lagged `canSteer`: a redirect
+    // fired on a fast Enter must not be dropped because state hasn't synced.
+    if (!onSteer || !text || attachments.length > 0 || SLASH_COMMAND_RE.test(text)) {
       return
     }
 
-    const text = draftRef.current.trim()
-
     triggerHaptic('submit')
     clearDraft()
 

apps/desktop/src/app/chat/composer/index.tsx

@@ -227,12 +227,15 @@ export function ChatBar({
   const { compactPill, stacked } = useComposerMetrics({ composerRef, composerSurfaceRef, editorRef, poppedOut })
   const hasComposerPayload = hasText || attachments.length > 0
   const canSubmit = busy || hasComposerPayload
-  const busyAction = busy && hasComposerPayload ? 'queue' : 'stop'
 
   // Steer only makes sense mid-turn, text-only (the gateway can't carry images
   // into a tool result) and never for a slash command (those execute inline).
   const canSteer = busy && !!onSteer && attachments.length === 0 && isSteerableText
 
+  // While busy: text redirects the live turn (Cursor-style stop-and-correct),
+  // attachments queue for the next turn, an empty composer stops.
+  const busyAction: 'steer' | 'queue' | 'stop' = canSteer ? 'steer' : hasComposerPayload ? 'queue' : 'stop'
+
   // The submit engine — the orchestration seam where draft + queue meet. Owns
   // the submit decision tree, the send-with-restore primitive, and steer.
   const { steerDraft, submitDraft } = useComposerSubmit({
@@ -240,7 +243,6 @@ export function ChatBar({
     activeQueueSessionKeyRef,
     attachments,
     busy,
-    canSteer,
     clearDraft,
     disabled,
     draftRef,
@@ -721,7 +723,6 @@ export function ChatBar({
       autoSpeak={autoSpeak}
       busy={busy}
       busyAction={busyAction}
-      canSteer={canSteer}
       canSubmit={canSubmit}
       compactModelPill={poppedOut || compactPill}
       conversation={{
@@ -737,7 +738,6 @@ export function ChatBar({
       disabled={disabled}
       hasComposerPayload={hasComposerPayload}
       onDictate={dictate}
-      onSteer={steerDraft}
       onToggleAutoSpeak={handleToggleAutoSpeak}
       state={state}
       voiceStatus={voiceStatus}

apps/desktop/src/app/session/hooks/use-prompt-actions/index.test.tsx

@@ -61,6 +61,8 @@ interface HarnessHandle {
   activeSessionIdRef: MutableRefObject<string | null>
   cancelRun: () => Promise<void>
   restoreToMessage: (messageId: string, target?: { text?: string; userOrdinal?: number | null }) => Promise<void>
+  redirectPrompt: (text: string) => Promise<boolean>
+  /** @deprecated Use `redirectPrompt`. */
   steerPrompt: (text: string) => Promise<boolean>
   submitText: (text: string, options?: SubmitTextOptions) => Promise<boolean>
 }
@@ -160,6 +162,8 @@ function Harness({
         act(async () => actions.cancelRun(...args)) as Promise<void>,
       restoreToMessage: (...args: Parameters<typeof actions.restoreToMessage>) =>
         act(async () => actions.restoreToMessage(...args)) as Promise<void>,
+      redirectPrompt: (...args: Parameters<typeof actions.redirectPrompt>) =>
+        act(async () => actions.redirectPrompt(...args)) as Promise<boolean>,
       steerPrompt: (...args: Parameters<typeof actions.steerPrompt>) =>
         act(async () => actions.steerPrompt(...args)) as Promise<boolean>,
       submitText: (...args: Parameters<typeof actions.submitText>) =>
@@ -168,6 +172,7 @@ function Harness({
   }, [
     actions.cancelRun,
     actions.restoreToMessage,
+    actions.redirectPrompt,
     actions.steerPrompt,
     actions.submitText,
     activeSessionIdRef,
@@ -703,32 +708,41 @@ describe('usePromptActions submit / queue drain semantics', () => {
   })
 })
 
-describe('usePromptActions steerPrompt', () => {
+describe('usePromptActions redirectPrompt', () => {
   afterEach(() => {
     cleanup()
     vi.restoreAllMocks()
   })
 
-  it('injects the trimmed text via session.steer and reports acceptance on a queued status', async () => {
-    const requestGateway = vi.fn(async () => ({ status: 'queued' }) as never)
+  it('redirects the live turn with trimmed correction text', async () => {
+    const requestGateway = vi.fn(async () => ({ status: 'redirected' }) as never)
 
     let handle: HarnessHandle | null = null
+    const capturedStates: Record<string, unknown>[] = []
     await actRender(
-      <Harness onReady={h => (handle = h)} refreshSessions={async () => undefined} requestGateway={requestGateway} />
+      <Harness
+        onReady={h => (handle = h)}
+        onSeedState={state => capturedStates.push(state)}
+        refreshSessions={async () => undefined}
+        requestGateway={requestGateway}
+      />
     )
 
-    const accepted = await handle!.steerPrompt('  nudge the run  ')
+    const accepted = await handle!.redirectPrompt('  nudge the run  ')
 
     expect(accepted).toBe(true)
-    // Steer never starts a turn — it rides the live run via session.steer only.
-    expect(requestGateway).toHaveBeenCalledWith('session.steer', {
+    expect(requestGateway).toHaveBeenCalledWith('session.redirect', {
       session_id: RUNTIME_SESSION_ID,
       text: 'nudge the run'
     })
     expect(requestGateway).not.toHaveBeenCalledWith('prompt.submit', expect.anything())
+    expect((capturedStates.at(-1)?.messages as unknown[]).at(-1)).toMatchObject({
+      role: 'user',
+      parts: [{ type: 'text', text: 'nudge the run' }]
+    })
   })
 
-  it('reports rejection (so the caller queues) when the gateway has no live tool window', async () => {
+  it('reports rejection so the caller queues when the turn already ended', async () => {
     const requestGateway = vi.fn(async () => ({ status: 'rejected' }) as never)
 
     let handle: HarnessHandle | null = null
@@ -736,12 +750,12 @@ describe('usePromptActions steerPrompt', () => {
       <Harness onReady={h => (handle = h)} refreshSessions={async () => undefined} requestGateway={requestGateway} />
     )
 
-    expect(await handle!.steerPrompt('too late')).toBe(false)
+    expect(await handle!.redirectPrompt('too late')).toBe(false)
   })
 
-  it('reports rejection (never throws) when the steer RPC errors', async () => {
+  it('reports rejection without throwing when the redirect RPC errors', async () => {
     const requestGateway = vi.fn(async () => {
-      throw new Error('agent does not support steer')
+      throw new Error('agent does not support redirect')
     })
 
     let handle: HarnessHandle | null = null
@@ -749,20 +763,93 @@ describe('usePromptActions steerPrompt', () => {
       <Harness onReady={h => (handle = h)} refreshSessions={async () => undefined} requestGateway={requestGateway} />
     )
 
-    expect(await handle!.steerPrompt('boom')).toBe(false)
+    expect(await handle!.redirectPrompt('boom')).toBe(false)
   })
 
   it('skips the RPC entirely for empty text', async () => {
-    const requestGateway = vi.fn(async () => ({ status: 'queued' }) as never)
+    const requestGateway = vi.fn(async () => ({ status: 'redirected' }) as never)
 
     let handle: HarnessHandle | null = null
     await actRender(
       <Harness onReady={h => (handle = h)} refreshSessions={async () => undefined} requestGateway={requestGateway} />
     )
 
-    expect(await handle!.steerPrompt('   ')).toBe(false)
+    expect(await handle!.redirectPrompt('   ')).toBe(false)
     expect(requestGateway).not.toHaveBeenCalled()
   })
+
+  it('accepts a queued redirect during the agent-build window and records the correction', async () => {
+    // running=True but the agent is still building: the gateway queues the
+    // correction instead of rejecting, so the composer must NOT re-queue it.
+    const requestGateway = vi.fn(async () => ({ status: 'queued' }) as never)
+
+    let handle: HarnessHandle | null = null
+    const capturedStates: Record<string, unknown>[] = []
+    await actRender(
+      <Harness
+        onReady={h => (handle = h)}
+        onSeedState={state => capturedStates.push(state)}
+        refreshSessions={async () => undefined}
+        requestGateway={requestGateway}
+      />
+    )
+
+    expect(await handle!.redirectPrompt('build-window nudge')).toBe(true)
+    expect(requestGateway).toHaveBeenCalledWith('session.redirect', {
+      session_id: RUNTIME_SESSION_ID,
+      text: 'build-window nudge'
+    })
+    expect(requestGateway).not.toHaveBeenCalledWith('prompt.submit', expect.anything())
+    expect((capturedStates.at(-1)?.messages as unknown[]).at(-1)).toMatchObject({
+      role: 'user',
+      parts: [{ type: 'text', text: 'build-window nudge' }]
+    })
+  })
+
+  it('resumes the stored session and retries once when session.redirect reports "session not found"', async () => {
+    const STORED_SESSION_ID = 'stored-db-xyz789'
+    const RECOVERED_SESSION_ID = 'rt-recovered-456'
+    const calls: { method: string; params?: Record<string, unknown> }[] = []
+    let redirectAttempts = 0
+
+    const requestGateway = vi.fn(async (method: string, params?: Record<string, unknown>) => {
+      calls.push({ method, params })
+
+      if (method === 'session.redirect') {
+        redirectAttempts += 1
+
+        if (redirectAttempts === 1) {
+          throw new Error('session not found')
+        }
+
+        return { status: 'redirected' } as never
+      }
+
+      if (method === 'session.resume') {
+        return { session_id: RECOVERED_SESSION_ID } as never
+      }
+
+      return {} as never
+    })
+
+    let handle: HarnessHandle | null = null
+    await actRender(
+      <Harness
+        onReady={h => (handle = h)}
+        refreshSessions={async () => undefined}
+        requestGateway={requestGateway}
+        storedSessionId={STORED_SESSION_ID}
+      />
+    )
+    await waitFor(() => expect(handle).not.toBeNull())
+
+    expect(await handle!.redirectPrompt('reconnect nudge')).toBe(true)
+    expect(calls.map(c => c.method)).toEqual(['session.redirect', 'session.resume', 'session.redirect'])
+    expect(calls[0]?.params).toEqual({ session_id: RUNTIME_SESSION_ID, text: 'reconnect nudge' })
+    expect(calls[1]?.params).toEqual({ session_id: STORED_SESSION_ID, source: 'desktop' })
+    expect(calls[2]?.params).toEqual({ session_id: RECOVERED_SESSION_ID, text: 'reconnect nudge' })
+    expect(handle!.activeSessionIdRef.current).toBe(RECOVERED_SESSION_ID)
+  })
 })
 
 describe('usePromptActions restoreToMessage', () => {

apps/desktop/src/app/session/hooks/use-prompt-actions/index.ts

@@ -41,7 +41,7 @@ import type {
   HandoffRequestResponse,
   HandoffStateResponse,
   ImageAttachResponse,
-  SessionSteerResponse
+  SessionRedirectResponse
 } from '../../../types'
 
 import {
@@ -599,11 +599,11 @@ export function usePromptActions({
     }
   }, [activeSessionIdRef, busyRef, copy.stopFailed, requestGateway, selectedStoredSessionIdRef, updateSessionState])
 
-  // Steer = nudge the live turn without interrupting: the gateway appends the
-  // text to the next tool result so the model reads it on its next iteration
-  // (desktop parity with `/steer`). Returns false on reject (no live tool
-  // window) so the caller can fall back to queueing the words for the next turn.
-  const steerPrompt = useCallback(
+  // The desktop steering action is an immediate correction: the core cancels
+  // model generation and rebuilds the live turn with displayed reasoning and
+  // completed work intact. During a tool it waits for the safe result boundary.
+  // Returns false when the turn raced to completion so the composer can queue.
+  const redirectPrompt = useCallback(
     async (rawText: string): Promise<boolean> => {
       const text = sanitizeComposerInput(rawText).trim()
       const sessionId = activeSessionId || activeSessionIdRef.current
@@ -612,25 +612,54 @@ export function usePromptActions({
         return false
       }
 
-      try {
-        const result = await requestGateway<SessionSteerResponse>('session.steer', { session_id: sessionId, text })
+      // Accepted whether the live turn was redirected in place or queued for
+      // the next turn (the build window, before the agent is wired) — either
+      // way the correction reaches the model, so record it once as a real user
+      // message after the interrupted checkpoint, matching the durable core
+      // transcript rather than a system note that changes role after reload.
+      const send = async (id: string): Promise<boolean> => {
+        const result = await requestGateway<SessionRedirectResponse>('session.redirect', { session_id: id, text })
 
-        if (result?.status === 'queued') {
+        if (result?.status === 'redirected' || result?.status === 'queued') {
           triggerHaptic('submit')
-          // Inline note (not a toast) so the nudge lives in the transcript next
-          // to the turn it steered. The `steer:` prefix is rendered as a codicon
-          // row by SystemMessage (see STEER_NOTE_RE), same style as slash output.
-          appendSessionTextMessage(sessionId, 'system', `steer:${text}`)
+          appendSessionTextMessage(id, 'user', text)
 
           return true
         }
-      } catch {
+
+        return false
+      }
+
+      try {
+        return await send(sessionId)
+      } catch (err) {
+        // A stale runtime id after reconnect 404s ("session not found"): resume
+        // the stored session and retry once, mirroring stopPrompt so a
+        // correction right after a reconnect isn't lost to the race.
+        if (isSessionNotFoundError(err) && selectedStoredSessionIdRef.current) {
+          try {
+            const resumed = await requestGateway<{ session_id: string }>('session.resume', {
+              session_id: selectedStoredSessionIdRef.current,
+              source: 'desktop'
+            })
+
+            const recoveredId = resumed?.session_id
+
+            if (recoveredId) {
+              activeSessionIdRef.current = recoveredId
+
+              return await send(recoveredId)
+            }
+          } catch {
+            // fall through — caller queues so nothing is lost
+          }
+        }
         // Swallow — caller queues the text so nothing is lost.
       }
 
       return false
     },
-    [activeSessionId, activeSessionIdRef, appendSessionTextMessage, requestGateway]
+    [activeSessionId, activeSessionIdRef, appendSessionTextMessage, requestGateway, selectedStoredSessionIdRef]
   )
 
   const reloadFromMessage = useCallback(
@@ -806,7 +835,9 @@ export function usePromptActions({
     handoffSession,
     reloadFromMessage,
     restoreToMessage,
-    steerPrompt,
+    redirectPrompt,
+    /** @deprecated Use `redirectPrompt` — this is an active-turn redirect, not tool steer. */
+    steerPrompt: redirectPrompt,
     submitText,
     transcribeVoiceAudio
   }

apps/desktop/src/app/types.ts

@@ -60,6 +60,11 @@ export interface SessionSteerResponse {
   text?: string
 }
 
+export interface SessionRedirectResponse {
+  status?: 'redirected' | 'queued' | 'rejected'
+  text?: string
+}
+
 export interface SessionTitleResponse {
   title?: string
   // True when the session row isn't persisted yet and the title was queued

cli.py

@@ -3816,7 +3816,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
             enabled=CLI_CONFIG["display"].get("persistent_output", True),
             max_lines=CLI_CONFIG["display"].get("persistent_output_max_lines", 200),
         )
-        # busy_input_mode: "interrupt" (Enter interrupts current run),
+        # busy_input_mode: "interrupt" (Enter redirects current run),
         # "queue" (Enter queues for next turn), or "steer" (Enter injects
         # mid-run via /steer, arriving after the next tool call).
         _bim = str(CLI_CONFIG["display"].get("busy_input_mode", "interrupt")).strip().lower()
@@ -13479,6 +13479,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                 payload = (text, images) if images else text
                 if self._agent_running and not (text and _looks_like_slash_command(text)):
                     _effective_mode = self.busy_input_mode
+                    redirected = False
                     if _effective_mode == "steer":
                         # Route Enter through /steer — inject mid-run after the
                         # next tool call.  Images can't ride along (steer only
@@ -13506,15 +13507,35 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                         preview = text if text else f"[{len(images)} image{'s' if len(images) != 1 else ''} attached]"
                         _cprint(f"  Queued for the next turn: {preview[:80]}{'...' if len(preview) > 80 else ''}")
                     elif _effective_mode == "interrupt":
-                        self._interrupt_queue.put(payload)
-                        # Debug: log to file when message enters interrupt queue
-                        try:
-                            _dbg = _hermes_home / "interrupt_debug.log"
-                            with open(_dbg, "a", encoding="utf-8") as _f:
-                                _f.write(f"{time.strftime('%H:%M:%S')} ENTER: queued interrupt msg={str(payload)[:60]!r}, "
-                                         f"agent_running={self._agent_running}\n")
-                        except Exception:
-                            pass
+                        if not images and text:
+                            try:
+                                if (
+                                    self.agent is not None
+                                    and getattr(
+                                        self.agent,
+                                        "_supports_active_turn_redirect",
+                                        False,
+                                    )
+                                    is True
+                                    and hasattr(self.agent, "redirect")
+                                ):
+                                    redirected = bool(self.agent.redirect(text))
+                            except Exception:
+                                redirected = False
+                        if redirected:
+                            preview = text[:80] + ("..." if len(text) > 80 else "")
+                            _cprint(f"  {_ACCENT}↪ Redirected current turn: '{preview}'{_RST}")
+                        else:
+                            # Compatibility path for older agents, multimodal
+                            # follow-ups, or a turn that finished in the race.
+                            self._interrupt_queue.put(payload)
+                            try:
+                                _dbg = _hermes_home / "interrupt_debug.log"
+                                with open(_dbg, "a", encoding="utf-8") as _f:
+                                    _f.write(f"{time.strftime('%H:%M:%S')} ENTER: queued interrupt msg={str(payload)[:60]!r}, "
+                                             f"agent_running={self._agent_running}\n")
+                            except Exception:
+                                pass
                     # First-touch onboarding: on the very first busy-while-running
                     # event for this install, print a one-line tip explaining the
                     # /busy knob.  Flag persists to config.yaml and never fires
@@ -13528,7 +13549,8 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                             mark_seen,
                         )
                         if not is_seen(CLI_CONFIG, BUSY_INPUT_FLAG):
-                            _cprint(f"  {_DIM}{busy_input_hint_cli(self.busy_input_mode)}{_RST}")
+                            _hint_mode = "redirect" if redirected else _effective_mode
+                            _cprint(f"  {_DIM}{busy_input_hint_cli(_hint_mode)}{_RST}")
                             mark_seen(_hermes_home / "config.yaml", BUSY_INPUT_FLAG)
                             CLI_CONFIG.setdefault("onboarding", {}).setdefault("seen", {})[BUSY_INPUT_FLAG] = True
                     except Exception:

gateway/run.py

@@ -6030,10 +6030,14 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
             )
             effective_mode = "queue"
         steered = False
+        redirected = False
         if effective_mode == "steer":
             steer_text = (event.text or "").strip()
             can_steer = (
                 steer_text
+                and event.message_type == MessageType.TEXT
+                and not event.media_urls
+                and not event.media_types
                 and running_agent is not None
                 and running_agent is not _AGENT_PENDING_SENTINEL
                 and hasattr(running_agent, "steer")
@@ -6047,6 +6051,21 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
             if not steered:
                 # Fall back to queue (merge into pending messages, no interrupt)
                 effective_mode = "queue"
+        elif (
+            effective_mode == "interrupt"
+            and event.message_type == MessageType.TEXT
+            and not event.media_urls
+            and not event.media_types
+            and running_agent is not None
+            and running_agent is not _AGENT_PENDING_SENTINEL
+            and getattr(running_agent, "_supports_active_turn_redirect", False) is True
+            and hasattr(running_agent, "redirect")
+        ):
+            try:
+                redirected = bool(running_agent.redirect((event.text or "").strip()))
+            except Exception as exc:
+                logger.warning("Gateway redirect failed for session %s: %s", session_key, exc)
+                redirected = False
 
         # Store the message so it's processed as the next turn after the
         # current run finishes (or is interrupted).  Skip this for a
@@ -6063,16 +6082,22 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
         # turn (#43066 sub-bug 2). The FIFO path gives each text its own
         # turn in arrival order while still preserving photo-burst / album
         # merge semantics for media.
-        if not steered:
+        if not steered and not redirected:
             self._queue_or_replace_pending_event(session_key, event)
 
         is_queue_mode = effective_mode == "queue"
         is_steer_mode = effective_mode == "steer"
+        is_redirect_mode = effective_mode == "interrupt" and redirected
 
         # If not in queue/steer mode, interrupt the running agent immediately.
         # This aborts in-flight tool calls and causes the agent loop to exit
         # at the next check point.
-        if effective_mode == "interrupt" and running_agent and running_agent is not _AGENT_PENDING_SENTINEL:
+        if (
+            effective_mode == "interrupt"
+            and not redirected
+            and running_agent
+            and running_agent is not _AGENT_PENDING_SENTINEL
+        ):
             try:
                 _interrupt_text = event.text
                 _media_urls = getattr(event, "media_urls", None) or []
@@ -6170,6 +6195,11 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                 f"⏩ Steered into current run{status_detail}. "
                 f"Your message arrives after the next tool call."
             )
+        elif is_redirect_mode:
+            message = (
+                f"↪ Redirected current run{status_detail}. "
+                f"I'll adjust using your correction."
+            )
         elif is_queue_mode and demoted_for_subagents:
             # #30170 — explain the demotion so the user knows their
             # follow-up didn't accidentally kill the subagent and
@@ -6211,6 +6241,8 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                     _hint_mode = "steer"
                 elif is_queue_mode:
                     _hint_mode = "queue"
+                elif is_redirect_mode:
+                    _hint_mode = "redirect"
                 else:
                     _hint_mode = "interrupt"
                 message = (
@@ -10952,7 +10984,13 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                 # is empty, the agent lacks steer(), or steer() rejects.
                 steer_text = (event.text or "").strip()
                 steered = False
-                if steer_text and hasattr(running_agent, "steer"):
+                if (
+                    event.message_type == MessageType.TEXT
+                    and not event.media_urls
+                    and not event.media_types
+                    and steer_text
+                    and hasattr(running_agent, "steer")
+                ):
                     try:
                         steered = bool(running_agent.steer(steer_text))
                     except Exception as exc:
@@ -10996,6 +11034,27 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                 )
                 self._queue_or_replace_pending_event(_quick_key, event)
                 return None
+            # Text-only corrections redirect the live turn (preserving
+            # displayed context) when the runtime supports it; media/voice and
+            # older runtimes fall back to the proven interrupt path below.
+            if (
+                event.message_type == MessageType.TEXT
+                and not event.media_urls
+                and not event.media_types
+                and getattr(running_agent, "_supports_active_turn_redirect", False)
+                is True
+                and hasattr(running_agent, "redirect")
+            ):
+                try:
+                    if running_agent.redirect((event.text or "").strip()):
+                        logger.debug("PRIORITY redirect for session %s", _quick_key)
+                        return None
+                except Exception as exc:
+                    logger.warning(
+                        "PRIORITY redirect failed for session %s: %s",
+                        _quick_key,
+                        exc,
+                    )
             logger.debug("PRIORITY interrupt for session %s", _quick_key)
             _interrupt_text = event.text
             _media_urls = getattr(event, "media_urls", None) or []

hermes_cli/cli_commands_mixin.py

@@ -2620,7 +2620,7 @@ class CLICommandsMixin:
             /busy status        Show current busy input mode
             /busy queue         Queue input for the next turn instead of interrupting
             /busy steer         Inject Enter mid-run via /steer (after next tool call)
-            /busy interrupt     Interrupt the current run on Enter (default)
+            /busy interrupt     Redirect the current run on Enter (default)
         """
         from cli import _ACCENT, _DIM, _RST, _cprint, save_config_value
         parts = cmd.strip().split(maxsplit=1)
@@ -2631,7 +2631,7 @@ class CLICommandsMixin:
             elif self.busy_input_mode == "steer":
                 _behavior = "steers into current run (after next tool call)"
             else:
-                _behavior = "interrupts current run"
+                _behavior = "redirects current run immediately"
             _cprint(f"  {_DIM}Enter while busy: {_behavior}{_RST}")
             _cprint(f"  {_DIM}Usage: /busy [queue|steer|interrupt|status]{_RST}")
             return
@@ -2649,7 +2649,7 @@ class CLICommandsMixin:
             elif arg == "steer":
                 behavior = "Enter will steer your message into the current run (after the next tool call)."
             else:
-                behavior = "Enter will interrupt the current run while Hermes is busy."
+                behavior = "Enter will redirect the current run while Hermes is busy; /stop still cancels it."
             _cprint(f"  {_ACCENT}✓ Busy input mode set to '{arg}' (saved to config){_RST}")
             _cprint(f"  {_DIM}{behavior}{_RST}")
         else:

run_agent.py

@@ -2809,8 +2809,33 @@ class AIAgent:
             if session_has_running_agent:
                 running_agent.interrupt(new_message.text)
         """
-        self._interrupt_requested = True
-        self._interrupt_message = message
+        # A hard stop and redirect share one lock so /stop cannot race with an
+        # accepted correction and accidentally turn itself into a retry.
+        _redirect_lock = getattr(self, "_pending_redirect_lock", None)
+        if _redirect_lock is not None:
+            with _redirect_lock:
+                self._interrupt_requested = True
+                self._interrupt_message = message
+                self._pending_redirect = None
+        else:
+            self._interrupt_requested = True
+            self._interrupt_message = message
+            self._pending_redirect = None
+
+        # Codex app-server owns its model/tool loop and watches a private
+        # interrupt event rather than Hermes' per-thread flag.
+        if getattr(self, "api_mode", None) == "codex_app_server":
+            _codex_session = getattr(self, "_codex_session", None)
+            _request_interrupt = getattr(_codex_session, "request_interrupt", None)
+            if callable(_request_interrupt):
+                try:
+                    _request_interrupt()
+                except Exception:
+                    logger.debug(
+                        "Failed to interrupt Codex app-server turn",
+                        exc_info=True,
+                    )
+
         # A cron turn performs its API request on the conversation thread to
         # avoid the nested interrupt-worker deadlock.  Unlike the normal worker
         # path, its client is registered here so this cross-thread interrupt can
@@ -2863,10 +2888,29 @@ class AIAgent:
         if not self.quiet_mode:
             print("\n⚡ Interrupt requested" + (f": '{message[:40]}...'" if message and len(message) > 40 else f": '{message}'" if message else ""))
 
-    def clear_interrupt(self) -> None:
-        """Clear any pending interrupt request and the per-thread tool interrupt signal."""
-        self._interrupt_requested = False
-        self._interrupt_message = None
+    def clear_interrupt(self, *, preserve_redirect: bool = False) -> bool:
+        """Clear the interrupt request and per-thread tool signal.
+
+        ``preserve_redirect`` is used only by the conversation loop after it
+        intentionally cancels a model request to rebuild that same logical
+        turn. Public hard-stop paths keep the default and clear everything.
+        """
+        _redirect_lock = getattr(self, "_pending_redirect_lock", None)
+        if _redirect_lock is not None:
+            with _redirect_lock:
+                if preserve_redirect and not self._pending_redirect:
+                    return False
+                self._interrupt_requested = False
+                self._interrupt_message = None
+                if not preserve_redirect:
+                    self._pending_redirect = None
+        else:
+            if preserve_redirect and not getattr(self, "_pending_redirect", None):
+                return False
+            self._interrupt_requested = False
+            self._interrupt_message = None
+            if not preserve_redirect:
+                self._pending_redirect = None
         self._interrupt_thread_signal_pending = False
         if self._execution_thread_id is not None:
             _set_interrupt(False, self._execution_thread_id)
@@ -2895,6 +2939,7 @@ class AIAgent:
         if _steer_lock is not None:
             with _steer_lock:
                 self._pending_steer = None
+        return True
 
     def steer(self, text: str) -> bool:
         """
@@ -2932,6 +2977,118 @@ class AIAgent:
                 self._pending_steer = cleaned
         return True
 
+    def redirect(self, text: str) -> bool:
+        """Redirect the active turn without converting it into a new task.
+
+        During a normal Hermes model request this cancels only that request;
+        the conversation loop retains completed messages/tool results, records
+        the displayed partial reasoning as plain assistant context, appends the
+        correction as a real user message, and retries. During tool execution
+        it degrades to ``steer()`` so the tool can finish at a safe boundary.
+        Codex app-server has a native ``turn/steer`` operation and uses it
+        directly instead of cancelling.
+
+        Returns ``False`` when there is no live turn or the text is empty, so
+        surfaces can fall back to their existing next-turn queue.
+        """
+        if not text or not text.strip():
+            return False
+        cleaned = text.strip()
+
+        # Codex owns its internal reasoning/tool loop, so use its first-class
+        # active-turn steering protocol rather than interrupting the subprocess.
+        if getattr(self, "api_mode", None) == "codex_app_server":
+            _codex_session = getattr(self, "_codex_session", None)
+            _native_steer = getattr(_codex_session, "request_steer", None)
+            if callable(_native_steer):
+                _redirect_lock = getattr(self, "_pending_redirect_lock", None)
+                if _redirect_lock is not None:
+                    with _redirect_lock:
+                        if self._interrupt_requested:
+                            return False
+                elif self._interrupt_requested:
+                    return False
+                try:
+                    return bool(_native_steer(cleaned))
+                except Exception:
+                    logger.debug("Codex app-server turn/steer failed", exc_info=True)
+                    return False
+
+        # Never kill a tool merely to deliver conversational guidance. The
+        # existing steer drain puts it on the final tool result before the next
+        # model decision, including delegate_task children.
+        if getattr(self, "_executing_tools", False):
+            return self.steer(cleaned)
+
+        _model_active = getattr(self, "_model_request_active", None)
+        _redirect_lock = getattr(self, "_pending_redirect_lock", None)
+        if _redirect_lock is None:
+            if _model_active is None or not _model_active.is_set():
+                return False
+            existing = getattr(self, "_pending_redirect", None)
+            if self._interrupt_requested and not existing:
+                return False
+            self._pending_redirect = (
+                f"{existing}\n\n[Additional user correction]\n{cleaned}"
+                if existing
+                else cleaned
+            )
+            self._interrupt_requested = True
+            self._interrupt_message = None
+        else:
+            with _redirect_lock:
+                if _model_active is None or not _model_active.is_set():
+                    # The response completed before we acquired the state lock.
+                    # Reject so the surface queues a new turn.
+                    return False
+                if self._interrupt_requested and not self._pending_redirect:
+                    return False
+                if self._pending_redirect:
+                    self._pending_redirect = (
+                        f"{self._pending_redirect}\n\n"
+                        f"[Additional user correction]\n{cleaned}"
+                    )
+                else:
+                    self._pending_redirect = cleaned
+                self._interrupt_requested = True
+                self._interrupt_message = None
+
+        # Interrupt only the model request. Do not fan out to tool workers or
+        # child agents as interrupt() does.
+        _execution_thread_id = getattr(self, "_execution_thread_id", None)
+        if _execution_thread_id is not None:
+            _set_interrupt(True, _execution_thread_id)
+            self._interrupt_thread_signal_pending = False
+        else:
+            self._interrupt_thread_signal_pending = True
+        _abort_active_request = getattr(self, "_active_request_abort", None)
+        if callable(_abort_active_request):
+            try:
+                _abort_active_request("redirect_abort")
+            except Exception:
+                logger.debug("Failed to abort request for redirect", exc_info=True)
+        return True
+
+    def _has_pending_redirect(self) -> bool:
+        """Return whether an active-turn redirect is waiting to be applied."""
+        _redirect_lock = getattr(self, "_pending_redirect_lock", None)
+        if _redirect_lock is None:
+            return bool(getattr(self, "_pending_redirect", None))
+        with _redirect_lock:
+            return bool(self._pending_redirect)
+
+    def _drain_pending_redirect(self) -> Optional[str]:
+        """Return and clear pending active-turn correction text."""
+        _redirect_lock = getattr(self, "_pending_redirect_lock", None)
+        if _redirect_lock is None:
+            text = getattr(self, "_pending_redirect", None)
+            self._pending_redirect = None
+            return text
+        with _redirect_lock:
+            text = self._pending_redirect
+            self._pending_redirect = None
+        return text
+
     def _drain_pending_steer(self) -> Optional[str]:
         """Return the pending steer text (if any) and clear the slot.
 
@@ -4918,6 +5075,7 @@ class AIAgent:
                         pass
                 self._record_streamed_assistant_text(tail)
         self._current_streamed_assistant_text = ""
+        self._current_streamed_reasoning_text = ""
 
     def _record_streamed_assistant_text(self, text: str) -> None:
         """Accumulate visible assistant text emitted through stream callbacks."""
@@ -5258,6 +5416,15 @@ class AIAgent:
                 cb(text)
             except Exception:
                 pass
+            else:
+                # Only checkpoint reasoning that a surface actually displayed.
+                # show_reasoning=false leaves the callback unset, so hidden
+                # provider thinking never becomes visible transcript content.
+                if isinstance(text, str) and text:
+                    self._current_streamed_reasoning_text = (
+                        getattr(self, "_current_streamed_reasoning_text", "")
+                        + text
+                    )
 
     def _fire_tool_gen_started(self, tool_name: str) -> None:
         """Notify display layer that the model is generating tool call arguments.

tests/acp_adapter/test_acp_commands.py

@@ -16,13 +16,19 @@ class FakeAgent:
         self.disabled_toolsets = []
         self.tools = []
         self.valid_tool_names = set()
+        self._supports_active_turn_redirect = True
         self.steers = []
+        self.redirects = []
         self.runs = []
 
     def steer(self, text):
         self.steers.append(text)
         return True
 
+    def redirect(self, text):
+        self.redirects.append(text)
+        return True
+
     def run_conversation(self, *, user_message, conversation_history, task_id, **kwargs):
         self.runs.append(user_message)
         messages = list(conversation_history or [])
@@ -147,6 +153,62 @@ async def test_acp_steer_after_zed_interrupt_replays_interrupted_prompt_with_gui
     assert state.interrupted_prompt_text == ""
 
 
+@pytest.mark.asyncio
+async def test_acp_plain_correction_redirects_running_turn():
+    acp_agent, state, fake, _conn = make_agent_and_state()
+    state.is_running = True
+
+    response = await acp_agent.prompt(
+        session_id=state.session_id,
+        prompt=[TextContentBlock(type="text", text="No, use Postgres instead")],
+    )
+
+    assert response.stop_reason == "end_turn"
+    assert fake.redirects == ["No, use Postgres instead"]
+    assert state.queued_prompts == []
+    assert fake.runs == []
+
+
+@pytest.mark.asyncio
+async def test_acp_plain_correction_after_cancel_replays_original_prompt():
+    acp_agent, state, fake, _conn = make_agent_and_state()
+    state.interrupted_prompt_text = "implement it with SQLite"
+
+    response = await acp_agent.prompt(
+        session_id=state.session_id,
+        prompt=[TextContentBlock(type="text", text="No, use Postgres instead")],
+    )
+
+    assert response.stop_reason == "end_turn"
+    assert fake.runs == [
+        "implement it with SQLite\n\n"
+        "User correction/guidance after interrupt: No, use Postgres instead"
+    ]
+    assert state.interrupted_prompt_text == ""
+
+
+@pytest.mark.asyncio
+async def test_acp_cancel_publishes_hard_stop_while_holding_runtime_lock():
+    acp_agent, state, fake, _conn = make_agent_and_state()
+    state.is_running = True
+    state.current_prompt_text = "original request"
+    observed = {}
+
+    def interrupt():
+        acquired = state.runtime_lock.acquire(blocking=False)
+        observed["lock_held"] = not acquired
+        if acquired:
+            state.runtime_lock.release()
+
+    fake.interrupt = interrupt
+
+    await acp_agent.cancel(state.session_id)
+
+    assert observed["lock_held"] is True
+    assert state.cancel_event.is_set()
+    assert state.interrupted_prompt_text == "original request"
+
+
 @pytest.mark.asyncio
 async def test_acp_steer_on_idle_session_runs_as_regular_prompt():
     # /steer on an idle session (no running turn, nothing to salvage) should

tests/agent/test_codex_app_server_persist.py

@@ -76,6 +76,34 @@ def test_codex_success_flushes_and_reports_persisted():
     assert result["agent_persisted"] is True
 
 
+def test_codex_user_interrupt_is_reported_and_cleared():
+    agent = _make_agent(session_db=None)
+    turn = _make_turn()
+    turn.interrupted = True
+    turn.final_text = ""
+    agent._codex_session.run_turn.return_value = turn
+    agent._interrupt_requested = True
+    agent._interrupt_message = "new correction"
+
+    def clear_interrupt():
+        agent._interrupt_requested = False
+        agent._interrupt_message = None
+
+    agent.clear_interrupt.side_effect = clear_interrupt
+    result = run_codex_app_server_turn(
+        agent,
+        user_message="hello",
+        original_user_message="hello",
+        messages=[{"role": "user", "content": "hello"}],
+        effective_task_id="task-1",
+    )
+
+    assert result["interrupted"] is True
+    assert result["interrupt_message"] == "new correction"
+    agent.clear_interrupt.assert_called_once_with()
+    assert agent._interrupt_requested is False
+
+
 def test_codex_turn_persists_each_message_exactly_once():
     """The user turn (flushed at turn start) must not be duplicated; the
     projected assistant message must land once.  Uses a real SessionDB and the

tests/agent/test_turn_retry_state.py

@@ -32,6 +32,7 @@ EXPECTED_FIELDS = {
     "restart_with_compressed_messages",
     "restart_with_length_continuation",
     "restart_with_rebuilt_messages",
+    "restart_with_redirected_messages",
 }
 
 

tests/agent/transports/test_codex_app_server_session.py

@@ -57,6 +57,8 @@ class FakeClient:
             return {"turn": {"id": "turn-fake-001"}}
         if method == "turn/interrupt":
             return {}
+        if method == "turn/steer":
+            return {"turnId": (params or {}).get("expectedTurnId")}
         return {}
 
     def notify(self, method: str, params=None):
@@ -562,30 +564,53 @@ class TestRunTurn:
         assert r.should_retire is True
         assert r.final_text == ""
 
-    def test_interrupt_during_turn_issues_turn_interrupt(self):
+    def test_interrupt_during_startup_skips_turn_start(self):
         client = FakeClient()
-        # Don't queue turn/completed — the loop has to interrupt out
-        client.queue_notification(
-            "item/completed",
-            item={"type": "commandExecution", "id": "x", "command": "sleep 60",
-                  "cwd": "/", "status": "inProgress",
-                  "aggregatedOutput": None, "exitCode": None,
-                  "commandActions": []},
-            threadId="t", turnId="tu1",
-        )
         s = make_session(client)
         s.ensure_started()
-        # Trip the interrupt before run_turn even consumes the notification.
-        # The loop will see interrupt set on its first iteration and bail.
         s.request_interrupt()
         r = s.run_turn("loop forever", turn_timeout=2.0)
+
+        assert r.interrupted is True
+        assert not any(method == "turn/start" for method, _params in client.requests)
+
+    def test_interrupt_after_turn_start_issues_turn_interrupt(self):
+        client = FakeClient()
+        s = make_session(client)
+
+        def request_handler(method, params):
+            if method == "thread/start":
+                return {"thread": {"id": "thread-fake-001"}}
+            if method == "turn/start":
+                s.request_interrupt()
+                return {"turn": {"id": "turn-fake-001"}}
+            return {}
+
+        client._request_handler = request_handler
+        r = s.run_turn("loop forever", turn_timeout=2.0)
+
         assert r.interrupted is True
-        # turn/interrupt was requested with the right turnId
         assert any(
             method == "turn/interrupt" and params.get("turnId") == "turn-fake-001"
             for (method, params) in client.requests
         )
 
+    def test_steer_appends_input_to_active_turn(self):
+        client = FakeClient()
+        s = make_session(client)
+        s.ensure_started()
+        with s._active_turn_lock:
+            s._active_turn_id = "turn-live-123"
+
+        assert s.request_steer("Use Postgres instead") is True
+        method, params = client.requests[-1]
+        assert method == "turn/steer"
+        assert params == {
+            "threadId": "thread-fake-001",
+            "input": [{"type": "text", "text": "Use Postgres instead"}],
+            "expectedTurnId": "turn-live-123",
+        }
+
     def test_deadline_exceeded_records_error(self):
         client = FakeClient()
         # No notifications and no completion → must hit deadline

tests/gateway/test_busy_session_ack.py

@@ -211,6 +211,54 @@ class TestBusySessionAck:
         # Verify agent interrupt was called
         agent.interrupt.assert_called_once_with("Are you working?")
 
+    @pytest.mark.asyncio
+    async def test_interrupt_mode_redirects_capable_core_agent(self):
+        runner, _sentinel = _make_runner()
+        runner._busy_input_mode = "interrupt"
+        adapter = _make_adapter()
+        event = _make_event(text="No, use Postgres")
+        sk = build_session_key(event.source)
+
+        agent = MagicMock()
+        agent._supports_active_turn_redirect = True
+        agent.redirect.return_value = True
+        agent._active_children = []
+        agent.get_activity_summary.return_value = {}
+        runner._running_agents[sk] = agent
+        runner.adapters[event.source.platform] = adapter
+
+        assert await runner._handle_active_session_busy_message(event, sk) is True
+
+        agent.redirect.assert_called_once_with("No, use Postgres")
+        agent.interrupt.assert_not_called()
+        assert sk not in adapter._pending_messages
+        content = adapter._send_with_retry.call_args.kwargs.get("content", "")
+        assert "Redirected current run" in content
+
+    @pytest.mark.asyncio
+    async def test_text_event_with_attachment_is_queued_not_redirected(self):
+        runner, _sentinel = _make_runner()
+        runner._busy_input_mode = "interrupt"
+        adapter = _make_adapter()
+        event = _make_event(text="use this attachment")
+        # QQBot and other adapters may retain unknown attachment MIME types on
+        # a TEXT event, so message_type alone is not a safe redirect gate.
+        event.media_urls = ["https://example.invalid/attachment.bin"]
+        event.media_types = ["application/octet-stream"]
+        sk = build_session_key(event.source)
+
+        agent = MagicMock()
+        agent._supports_active_turn_redirect = True
+        agent._active_children = []
+        runner._running_agents[sk] = agent
+        runner.adapters[event.source.platform] = adapter
+
+        assert await runner._handle_active_session_busy_message(event, sk) is True
+
+        agent.redirect.assert_not_called()
+        assert adapter._pending_messages[sk] is event
+        assert adapter._pending_messages[sk].media_urls == event.media_urls
+
     @pytest.mark.asyncio
     async def test_queue_mode_suppresses_interrupt_and_updates_ack(self):
         """When busy_input_mode is 'queue', message is queued WITHOUT interrupt."""

tests/run_agent/test_run_agent.py

@@ -12,6 +12,7 @@ import json
 import logging
 import re
 import threading
+import time
 import uuid
 from logging.handlers import RotatingFileHandler
 from pathlib import Path
@@ -4850,6 +4851,136 @@ class TestRunConversation:
             "content": "Sure, here's how to do it: first",
         }
 
+    def test_redirect_during_thinking_retries_same_turn_with_context(self, agent):
+        """A corrective follow-up keeps displayed reasoning and does not end the turn."""
+        self._setup_agent(agent)
+        agent.reasoning_callback = lambda _text: None
+        final = _mock_response(content="Using Postgres instead.", finish_reason="stop")
+        requests = []
+        persisted = []
+
+        def _fake_api_call(api_kwargs):
+            requests.append(api_kwargs)
+            if len(requests) == 1:
+                agent._fire_reasoning_delta("I should implement this with SQLite.")
+                assert agent.redirect("No, use Postgres instead.") is True
+                raise InterruptedError("redirect cancelled the first request")
+            return final
+
+        with (
+            patch.object(agent, "_interruptible_api_call", side_effect=_fake_api_call),
+            patch.object(
+                agent,
+                "_persist_session",
+                side_effect=lambda messages, *_a, **_k: persisted.append(
+                    [dict(message) for message in messages]
+                ),
+            ),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+        ):
+            result = agent.run_conversation("Choose a database and implement it.")
+
+        assert result["completed"] is True
+        assert result["interrupted"] is False
+        assert result["final_response"] == "Using Postgres instead."
+        assert len(requests) == 2
+
+        replay = requests[1]["messages"]
+        assert [m["role"] for m in replay[-3:]] == [
+            "user",
+            "assistant",
+            "user",
+        ]
+        checkpoint = replay[-2]["content"]
+        assert "interrupted by a user correction" in checkpoint
+        assert "I should implement this with SQLite." in checkpoint
+        assert replay[-1]["content"] == "No, use Postgres instead."
+        assert agent._pending_redirect is None
+        assert any(
+            snapshot[-1].get("content") == "No, use Postgres instead."
+            and snapshot[-2].get("role") == "assistant"
+            for snapshot in persisted
+            if len(snapshot) >= 2
+        )
+
+    def test_redirect_wins_race_with_response_completion(self, agent):
+        """If the provider returns as redirect lands, discard the stale answer."""
+        self._setup_agent(agent)
+        stale = _mock_response(content="Using SQLite.", finish_reason="stop")
+        corrected = _mock_response(content="Using Postgres.", finish_reason="stop")
+        calls = 0
+
+        def _fake_api_call(_api_kwargs):
+            nonlocal calls
+            calls += 1
+            if calls == 1:
+                assert agent.redirect("Use Postgres instead.") is True
+                return stale
+            return corrected
+
+        with (
+            patch.object(agent, "_interruptible_api_call", side_effect=_fake_api_call),
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+        ):
+            result = agent.run_conversation("Choose a database.")
+
+        assert calls == 2
+        assert result["final_response"] == "Using Postgres."
+        assert all(
+            message.get("content") != "Using SQLite."
+            for message in result["messages"]
+        )
+
+    def test_redirect_from_input_thread_cancels_live_model_request(self, agent):
+        """Exercise the real cross-thread path used by CLI and gateways."""
+        self._setup_agent(agent)
+        agent.reasoning_callback = lambda _text: None
+        entered = threading.Event()
+        results = {}
+        calls = 0
+        final = _mock_response(content="Corrected answer.", finish_reason="stop")
+
+        def _fake_api_call(_api_kwargs):
+            nonlocal calls
+            calls += 1
+            if calls == 1:
+                agent._fire_reasoning_delta("Following the original approach.")
+                entered.set()
+                deadline = time.time() + 2
+                while not agent._interrupt_requested and time.time() < deadline:
+                    time.sleep(0.01)
+                raise InterruptedError("request cancelled by redirect")
+            return final
+
+        with (
+            patch.object(agent, "_interruptible_api_call", side_effect=_fake_api_call),
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+        ):
+            worker = threading.Thread(
+                target=lambda: results.update(
+                    result=agent.run_conversation("Take the original approach.")
+                )
+            )
+            worker.start()
+            assert entered.wait(timeout=2)
+            assert agent.redirect("Use the corrected approach.") is True
+            worker.join(timeout=5)
+
+        assert worker.is_alive() is False
+        assert calls == 2
+        assert results["result"]["completed"] is True
+        assert results["result"]["final_response"] == "Corrected answer."
+        checkpoint = results["result"]["messages"][-3]
+        assert "Following the original approach." in checkpoint["content"]
+        assert results["result"]["messages"][-2]["content"] == (
+            "Use the corrected approach."
+        )
+
     def test_interrupt_before_any_stream_keeps_sentinel(self, agent):
         """An interrupt with no streamed text falls back to the metadata sentinel."""
         from agent.conversation_loop import INTERRUPT_WAITING_FOR_MODEL_PREFIX

tests/run_agent/test_steer.py

@@ -23,6 +23,24 @@ def _bare_agent() -> AIAgent:
     agent = object.__new__(AIAgent)
     agent._pending_steer = None
     agent._pending_steer_lock = threading.Lock()
+    agent._pending_redirect = None
+    agent._pending_redirect_lock = threading.Lock()
+    agent._model_request_active = threading.Event()
+    agent._executing_tools = False
+    agent._execution_thread_id = None
+    agent._interrupt_thread_signal_pending = False
+    agent._interrupt_requested = False
+    agent._interrupt_message = None
+    agent._active_children = []
+    agent._active_children_lock = threading.Lock()
+    agent._tool_worker_threads = None
+    agent._tool_worker_threads_lock = None
+    agent._current_streamed_reasoning_text = ""
+    agent._current_streamed_assistant_text = ""
+    agent._stream_needs_break = False
+    agent._strip_think_blocks = lambda content: content
+    agent.quiet_mode = True
+    agent.api_mode = "chat_completions"
     return agent
 
 
@@ -72,6 +90,161 @@ class TestSteerDrain:
         assert agent._drain_pending_steer() is None
 
 
+class TestActiveTurnRedirect:
+    def test_rejects_when_no_turn_is_active(self):
+        agent = _bare_agent()
+        assert agent.redirect("change course") is False
+        assert agent._pending_redirect is None
+
+    def test_cancels_only_an_active_model_request(self):
+        agent = _bare_agent()
+        agent._model_request_active.set()
+
+        assert agent.redirect("use Postgres") is True
+        assert agent._pending_redirect == "use Postgres"
+        assert agent._interrupt_requested is True
+        assert agent._interrupt_message is None
+
+    def test_multiple_redirects_preserve_message_boundaries(self):
+        agent = _bare_agent()
+        agent._model_request_active.set()
+
+        assert agent.redirect("first correction") is True
+        assert agent.redirect("second correction") is True
+        assert agent._pending_redirect == (
+            "first correction\n\n"
+            "[Additional user correction]\n"
+            "second correction"
+        )
+
+    def test_hard_interrupt_wins_over_new_redirect(self):
+        agent = _bare_agent()
+        agent._model_request_active.set()
+        agent._interrupt_requested = True
+
+        assert agent.redirect("too late") is False
+        assert agent._pending_redirect is None
+
+    def test_hidden_reasoning_is_not_checkpointed(self):
+        agent = _bare_agent()
+        agent.reasoning_callback = None
+        agent._current_streamed_reasoning_text = ""
+
+        agent._fire_reasoning_delta("private provider thinking")
+
+        assert agent._current_streamed_reasoning_text == ""
+
+    def test_response_completion_before_redirect_lock_rejects_correction(self):
+        agent = _bare_agent()
+        agent._model_request_active.set()
+        started = threading.Event()
+        outcome = {}
+
+        def redirect():
+            started.set()
+            outcome["accepted"] = agent.redirect("late correction")
+
+        with agent._pending_redirect_lock:
+            worker = threading.Thread(target=redirect)
+            worker.start()
+            assert started.wait(timeout=1)
+            # Mirrors conversation_loop clearing the request-active marker
+            # under this same lock before redirect can commit its slot.
+            agent._model_request_active.clear()
+        worker.join(timeout=1)
+
+        assert outcome["accepted"] is False
+        assert agent._pending_redirect is None
+
+    def test_hard_stop_wins_concurrent_redirect(self):
+        agent = _bare_agent()
+        agent._model_request_active.set()
+        start = threading.Barrier(3)
+        outcome = {}
+
+        def redirect():
+            start.wait()
+            outcome["redirect"] = agent.redirect("change course")
+
+        def hard_stop():
+            start.wait()
+            agent.interrupt("stop requested")
+
+        redirect_thread = threading.Thread(target=redirect)
+        stop_thread = threading.Thread(target=hard_stop)
+        redirect_thread.start()
+        stop_thread.start()
+        start.wait()
+        redirect_thread.join(timeout=1)
+        stop_thread.join(timeout=1)
+
+        assert redirect_thread.is_alive() is False
+        assert stop_thread.is_alive() is False
+        assert agent._interrupt_requested is True
+        assert agent._interrupt_message == "stop requested"
+        assert agent._pending_redirect is None
+
+    def test_codex_app_server_hard_stop_reaches_native_session(self):
+        agent = _bare_agent()
+        calls = []
+        agent.api_mode = "codex_app_server"
+        agent._codex_session = type(
+            "_CodexSession",
+            (),
+            {"request_interrupt": lambda self: calls.append("interrupt")},
+        )()
+
+        agent.interrupt()
+
+        assert calls == ["interrupt"]
+
+    def test_codex_app_server_redirect_rejects_after_hard_stop(self):
+        agent = _bare_agent()
+        calls = []
+        agent.api_mode = "codex_app_server"
+        agent._interrupt_requested = True
+        agent._codex_session = type(
+            "_CodexSession",
+            (),
+            {"request_steer": lambda self, text: calls.append(text) or True},
+        )()
+
+        assert agent.redirect("too late") is False
+        assert calls == []
+
+    def test_redirect_during_tool_execution_uses_safe_steer_boundary(self):
+        agent = _bare_agent()
+        agent._executing_tools = True
+
+        assert agent.redirect("also check migrations") is True
+        assert agent._pending_redirect is None
+        assert agent._pending_steer == "also check migrations"
+        assert agent._interrupt_requested is False
+
+
+class TestActiveTurnRedirectCheckpoint:
+    def test_assistant_tail_puts_correction_last(self):
+        from agent.conversation_loop import _apply_active_turn_redirect
+
+        agent = _bare_agent()
+        agent._current_streamed_reasoning_text = "Shown reasoning."
+        agent._current_streamed_assistant_text = "Visible draft."
+        messages = [
+            {"role": "user", "content": "start"},
+            {"role": "assistant", "content": "committed assistant item"},
+        ]
+
+        _apply_active_turn_redirect(agent, messages, "Use Postgres instead.")
+
+        assert [m["role"] for m in messages] == ["user", "assistant", "user"]
+        assert messages[-1]["role"] == "user"
+        assert messages[-1]["content"].endswith("Use Postgres instead.")
+        assert sum(1 for m in messages if m["role"] == "assistant") == 1
+        assert "Shown reasoning." in messages[-1]["content"]
+        assert "Visible draft." in messages[-1]["content"]
+        assert "Context from the interrupted assistant response" in messages[-1]["content"]
+
+
 class TestSteerInjection:
     def test_appends_to_last_tool_result(self):
         agent = _bare_agent()
@@ -196,10 +369,12 @@ class TestSteerClearedOnInterrupt:
         agent._tool_worker_threads_lock = None
 
         agent.steer("will be dropped")
+        agent._pending_redirect = "also drop this"
         assert agent._pending_steer == "will be dropped"
 
         agent.clear_interrupt()
         assert agent._pending_steer is None
+        assert agent._pending_redirect is None
 
 
 class TestPreApiCallSteerDrain:

tests/test_tui_gateway_queue_on_busy.py

@@ -1,12 +1,12 @@
-"""A prompt that lands mid-turn is interrupted + queued, never dropped.
+"""A prompt that lands mid-turn is redirected or queued, never dropped.
 
 Before this, ``prompt.submit`` on a running session returned ``session busy``,
 forcing clients into a deadline-bounded busy-retry. When turn teardown outlived
 the deadline — e.g. a slow, non-interruptible tool (``web_search``) still
 running when the user hit stop — the resubmitted message was silently dropped
 ("it just doesn't listen"). The gateway now applies the ``busy_input_mode``
-policy: interrupt the live turn (default) and queue the message to run as the
-next turn, drained in ``run``'s tail.
+policy: redirect the live turn by default, with the legacy interrupt + queue
+path retained as a compatibility fallback.
 """
 
 import threading
@@ -49,7 +49,26 @@ def test_enqueue_merges_second_arrival_losslessly():
 
 # ── _handle_busy_submit (policy) ───────────────────────────────────────────
 
-def test_busy_interrupt_mode_interrupts_and_queues(monkeypatch):
+def test_busy_interrupt_mode_redirects_active_turn(monkeypatch):
+    monkeypatch.setattr(server, "_load_busy_input_mode", lambda: "interrupt")
+    seen = []
+    agent = types.SimpleNamespace(
+        _supports_active_turn_redirect=True,
+        redirect=lambda text: seen.append(text) or True,
+        interrupt=lambda *a, **k: (_ for _ in ()).throw(
+            AssertionError("redirect must not hard-interrupt")
+        ),
+    )
+    session = _session(agent=agent, running=True)
+
+    resp = server._handle_busy_submit("r1", "sid", session, "redirect", "ws-1")
+
+    assert resp["result"]["status"] == "redirected"
+    assert seen == ["redirect"]
+    assert session.get("queued_prompt") is None
+
+
+def test_busy_interrupt_mode_falls_back_for_legacy_agent(monkeypatch):
     monkeypatch.setattr(server, "_load_busy_input_mode", lambda: "interrupt")
     calls = {"interrupt": 0}
     agent = types.SimpleNamespace(interrupt=lambda *a, **k: calls.__setitem__("interrupt", calls["interrupt"] + 1))
@@ -134,6 +153,67 @@ def test_busy_helper_retries_when_turn_finished(monkeypatch):
     assert session.get("queued_prompt") is None
 
 
+def test_busy_interrupt_mode_normalizes_rich_text_before_redirect(monkeypatch):
+    monkeypatch.setattr(server, "_load_busy_input_mode", lambda: "interrupt")
+    seen = []
+    agent = types.SimpleNamespace(
+        _supports_active_turn_redirect=True,
+        redirect=lambda text: seen.append(text) or True,
+        interrupt=lambda *a, **k: None,
+    )
+    session = _session(agent=agent, running=True)
+    rich = [{"type": "text", "text": "  redirect me  "}]
+
+    resp = server._handle_busy_submit(
+        "r1",
+        "sid",
+        session,
+        rich,
+        "ws-1",
+    )
+
+    assert resp["result"]["status"] == "redirected"
+    assert seen == ["redirect me"]
+    assert session.get("queued_prompt") is None
+
+
+def test_busy_queue_fallback_preserves_original_structured_text(monkeypatch):
+    monkeypatch.setattr(server, "_load_busy_input_mode", lambda: "interrupt")
+    rich = [{"type": "text", "text": "  keep me  "}]
+    agent = types.SimpleNamespace(
+        _supports_active_turn_redirect=True,
+        redirect=lambda text: False,
+        interrupt=lambda *a, **k: None,
+    )
+    session = _session(agent=agent, running=True)
+
+    resp = server._handle_busy_submit("r1", "sid", session, rich, "ws-1")
+
+    assert resp["result"]["status"] == "queued"
+    assert session["queued_prompt"]["text"] == rich
+
+
+def test_busy_interrupt_mode_queues_multimodal_payload_instead_of_redirect(monkeypatch):
+    monkeypatch.setattr(server, "_load_busy_input_mode", lambda: "interrupt")
+    seen = []
+    rich = [
+        {"type": "text", "text": "caption"},
+        {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc"}},
+    ]
+    agent = types.SimpleNamespace(
+        _supports_active_turn_redirect=True,
+        redirect=lambda text: seen.append(text) or True,
+        interrupt=lambda *a, **k: None,
+    )
+    session = _session(agent=agent, running=True)
+
+    resp = server._handle_busy_submit("r1", "sid", session, rich, "ws-1")
+
+    assert resp["result"]["status"] == "queued"
+    assert seen == []
+    assert session["queued_prompt"]["text"] == rich
+
+
 # ── _drain_queued_prompt ───────────────────────────────────────────────────
 
 def test_drain_fires_queued_prompt_and_claims_running(monkeypatch):

tests/test_tui_gateway_server.py

@@ -6835,6 +6835,78 @@ def test_session_steer_errors_when_agent_has_no_steer_method():
     assert resp["error"]["code"] == 4010
 
 
+def test_session_redirect_calls_capable_core_agent(monkeypatch):
+    calls = []
+    agent = types.SimpleNamespace(
+        _supports_active_turn_redirect=True,
+        redirect=lambda text: calls.append(text) or True,
+    )
+    session = _session(agent=agent)
+    server._sessions["sid"] = session
+    try:
+        before = session.get("last_active")
+        resp = server.handle_request(
+            {
+                "id": "1",
+                "method": "session.redirect",
+                "params": {"session_id": "sid", "text": "use Postgres"},
+            }
+        )
+    finally:
+        server._sessions.pop("sid", None)
+
+    assert resp["result"] == {
+        "status": "redirected",
+        "text": "use Postgres",
+    }
+    assert calls == ["use Postgres"]
+    assert session.get("last_active") is not None
+    assert before is None or session["last_active"] >= before
+
+
+def test_session_redirect_queues_during_agent_build_window(monkeypatch):
+    # A fresh turn flips running=True and builds the agent asynchronously, so
+    # session["agent"] is briefly None. A correction landing here must queue
+    # (lossless, reaches the model next turn), not hard-reject as unsupported.
+    session = _session(running=True)
+    session["agent"] = None
+    server._sessions["sid"] = session
+    try:
+        resp = server.handle_request(
+            {
+                "id": "1",
+                "method": "session.redirect",
+                "params": {"session_id": "sid", "text": "wait, use SQLite"},
+            }
+        )
+    finally:
+        server._sessions.pop("sid", None)
+
+    assert resp["result"] == {"status": "queued", "text": "wait, use SQLite"}
+    assert session["queued_prompt"]["text"] == "wait, use SQLite"
+
+
+def test_session_redirect_rejects_when_idle_without_agent(monkeypatch):
+    # No live turn and no agent: nothing to redirect, and we must not queue a
+    # phantom turn — keep the explicit unsupported rejection.
+    session = _session(running=False)
+    session["agent"] = None
+    server._sessions["sid"] = session
+    try:
+        resp = server.handle_request(
+            {
+                "id": "1",
+                "method": "session.redirect",
+                "params": {"session_id": "sid", "text": "hi"},
+            }
+        )
+    finally:
+        server._sessions.pop("sid", None)
+
+    assert resp["error"]["code"] == 4010
+    assert session.get("queued_prompt") is None
+
+
 def test_session_info_includes_mcp_servers(monkeypatch):
     fake_status = [
         {"name": "github", "transport": "http", "tools": 12, "connected": True},

tui_gateway/server.py

@@ -5553,6 +5553,38 @@ def _coerce_message_text(content: Any) -> str:
     return str(content)
 
 
+_TEXT_ONLY_BUSY_PART_KINDS = frozenset({"text", "input_text", "output_text"})
+
+
+def _is_text_only_busy_payload(content: Any) -> bool:
+    """True when a busy submit carries only plain text, not attachments/media."""
+    if content is None:
+        return False
+    if isinstance(content, (str, int, float)):
+        return True
+    if isinstance(content, list):
+        if not content:
+            return False
+        for part in content:
+            if isinstance(part, str):
+                continue
+            if not isinstance(part, dict):
+                return False
+            kind = part.get("type")
+            if kind in _TEXT_ONLY_BUSY_PART_KINDS:
+                continue
+            if kind is None and isinstance(part.get("text"), str):
+                continue
+            return False
+        return True
+    if isinstance(content, dict):
+        kind = content.get("type")
+        if kind in _TEXT_ONLY_BUSY_PART_KINDS:
+            return True
+        return kind is None and isinstance(content.get("text"), str)
+    return False
+
+
 def _history_to_messages(history: list[dict]) -> list[dict]:
     messages = []
     tool_call_args = {}
@@ -5761,15 +5793,13 @@ def _handle_busy_submit(
     a turn is in flight, instead of rejecting it with ``session busy``.
 
     The old rejection forced clients into a deadline-bounded busy-retry that
-    silently dropped the send when turn teardown outlived the deadline (e.g. a
-    slow, non-interruptible tool like ``web_search`` running when the user hits
-    stop). The message is instead queued to run as the next turn — and, for the
-    default ``interrupt`` policy, the live turn is interrupted so it winds down
-    promptly. Drained in ``run``'s tail (see ``_run_prompt_submit``).
+    silently dropped the send when turn teardown outlived the deadline. The
+    default policy now redirects a capable core agent in place; older agents
+    retain the proven interrupt-and-queue path drained from ``run``'s tail.
 
-    Modes: ``interrupt`` (default) → interrupt + queue; ``queue`` → queue
-    without interrupting; ``steer`` → inject into the live turn if accepted,
-    else queue.
+    Modes: ``interrupt`` (default) → redirect the live turn, falling back to
+    hard interrupt + queue for older agents; ``queue`` → queue without
+    interrupting; ``steer`` → inject after the current atomic action.
     """
     mode = _load_busy_input_mode()
     agent = session.get("agent")
@@ -5778,14 +5808,34 @@ def _handle_busy_submit(
             # The turn ended between prompt.submit's first busy check and this
             # helper. Let the caller retry and claim the now-idle session.
             return None
-    if mode == "steer" and agent is not None and hasattr(agent, "steer"):
+    text_only = _is_text_only_busy_payload(text)
+    plain_text = _coerce_message_text(text).strip() if text_only else ""
+    if mode == "steer" and text_only and plain_text and agent is not None and hasattr(agent, "steer"):
         try:
-            if agent.steer(text):
+            if agent.steer(plain_text):
                 with session["history_lock"]:
                     session["last_active"] = time.time()
                 return _ok(rid, {"status": "steered"})
         except Exception:
             pass  # fall through to queue
+    # Text-only corrections redirect the live turn in place when the runtime
+    # supports it; media/attachment payloads and older agents fall through to
+    # the proven interrupt + queue path below.
+    if (
+        mode == "interrupt"
+        and text_only
+        and plain_text
+        and agent is not None
+        and getattr(agent, "_supports_active_turn_redirect", False) is True
+        and hasattr(agent, "redirect")
+    ):
+        try:
+            if agent.redirect(plain_text):
+                with session["history_lock"]:
+                    session["last_active"] = time.time()
+                return _ok(rid, {"status": "redirected"})
+        except Exception:
+            pass  # preserve the proven interrupt + queue fallback below
     # Queue before asking the live turn to stop. In particular, never call a
     # provider or compute-host method while holding history_lock: an interrupt
     # can wait behind the very operation it is trying to cancel.
@@ -9585,6 +9635,43 @@ def _(rid, params: dict) -> dict:
     return _ok(rid, {"status": "queued" if accepted else "rejected", "text": text})
 
 
+@method("session.redirect")
+def _(rid, params: dict) -> dict:
+    """Redirect the active model turn while preserving valid work/context."""
+    text = (params.get("text") or "").strip()
+    if not text:
+        return _err(rid, 4002, "text is required")
+    session, err = _sess_nowait(params, rid)
+    if err:
+        return err
+    agent = session.get("agent")
+    # Turn-build window: a fresh turn flips running=True and kicks off an async
+    # agent build, so session["agent"] is briefly None. That is not an
+    # unsupported runtime — queue the correction server-side so it reaches the
+    # model as the next turn, instead of a misleading 4010 the client silently
+    # swallows into a lost follow-up.
+    if agent is None and session.get("running"):
+        _enqueue_prompt(session, text, current_transport() or _stdio_transport)
+        session["last_active"] = time.time()
+        return _ok(rid, {"status": "queued", "text": text})
+    if (
+        agent is None
+        or getattr(agent, "_supports_active_turn_redirect", False) is not True
+        or not hasattr(agent, "redirect")
+    ):
+        return _err(rid, 4010, "agent does not support active-turn redirect")
+    try:
+        accepted = agent.redirect(text)
+    except Exception as exc:
+        return _err(rid, 5000, f"redirect failed: {exc}")
+    if accepted:
+        session["last_active"] = time.time()
+    return _ok(
+        rid,
+        {"status": "redirected" if accepted else "rejected", "text": text},
+    )
+
+
 @method("terminal.resize")
 def _(rid, params: dict) -> dict:
     session, err = _sess_nowait(params, rid)

ui-tui/src/app/useSubmission.ts

@@ -158,9 +158,9 @@ export function useSubmission(opts: UseSubmissionOptions) {
   //   - 'steer'     : inject into the current turn via session.steer; falls
   //                   back to queue when steer is rejected (no agent / no
   //                   tool window).
-  //   - 'interrupt' (default): queue the text + interrupt with `keepBusy`; the
-  //                   busy→false settle edge drains it once (desktop parity).
-  //                   No optimistic send → no duplicate bubble / race note.
+  //   - 'interrupt' (default): submit immediately; the backend redirects the
+  //                   active model request (or safely steers after a tool),
+  //                   with legacy interrupt + queue as its compatibility path.
   //
   // `opts.fallbackToFront` re-inserts at the queue head (queue-edit picks keep
   // their position); the mainline submit path appends.
@@ -201,14 +201,13 @@ export function useSubmission(opts: UseSubmissionOptions) {
         return
       }
 
-      // 'interrupt': queue + interrupt(keepBusy); the settle edge drains it once.
-      enqueueText()
-
-      if (live.sid) {
-        turnController.interruptTurn({ appendMessage, gw, sid: live.sid, sys }, { keepBusy: true })
-      }
+      // The gateway owns the atomic redirect decision because it knows whether
+      // the agent is in model generation, tool execution, or an older runtime.
+      // Reuse the normal submit pipeline so the correction gets its user bubble
+      // and file-drop interpolation exactly once.
+      send(full)
     },
-    [appendMessage, composerActions, composerRefs, gw, sys]
+    [composerActions, composerRefs, gw, send, sys]
   )
 
   const dispatchSubmission = useCallback(

website/docs/user-guide/cli.md

@@ -241,14 +241,14 @@ Most terminals send the same byte sequence for `Enter` and `Shift+Enter` by defa
 
 Where the terminal cannot distinguish them, `Alt+Enter` and `Ctrl+J` continue to work everywhere. **On Windows Terminal specifically, `Alt+Enter` is captured by the terminal (toggles fullscreen) and never reaches Hermes — use `Ctrl+Enter` (delivered as `Ctrl+J`) or `Ctrl+J` directly for a newline.**
 
-## Interrupting the Agent
+## Redirecting the Agent Mid-Turn
 
-You can interrupt the agent at any point:
+While the agent is working, you can send a correction without starting a new turn:
 
-- **Type a new message + Enter** while the agent is working — it interrupts and processes your new instructions
+- **Type a new message + Enter** — redirects the active turn using your correction
 - **`Ctrl+C`** — interrupt the current operation (press twice within 2s to force exit)
-- In-progress terminal commands are killed immediately (SIGTERM, then SIGKILL after 1s)
-- Multiple messages typed during interrupt are combined into one prompt
+- Completed tool work and reasoning already shown stay in context
+- A running tool reaches its safe boundary before the correction is applied
 
 ### Busy Input Mode
 
@@ -256,7 +256,7 @@ The `display.busy_input_mode` config key controls what happens when you press En
 
 | Mode | Behavior |
 |------|----------|
-| `"interrupt"` (default) | Your message interrupts the current operation and is processed immediately |
+| `"interrupt"` (default) | Your message redirects the active turn. Model generation restarts with displayed reasoning and completed work preserved; running tools finish first |
 | `"queue"` | Your message is silently queued and sent as the next turn after the agent finishes |
 | `"steer"` | Your message is injected into the current run via `/steer`, arriving at the agent after the next tool call — no interrupt, no new turn |
 
@@ -266,7 +266,7 @@ display:
   busy_input_mode: "steer"   # or "queue" or "interrupt" (default)
 ```
 
-`"queue"` mode is useful when you want to prepare follow-up messages without accidentally canceling in-flight work. `"steer"` mode is useful when you want to redirect the agent mid-task without interrupting — e.g. "actually, also check the tests" while it's still editing code. Unknown values fall back to `"interrupt"`.
+`"queue"` mode prepares a separate follow-up turn. `"steer"` always waits for the next tool-result boundary. The default `"interrupt"` mode responds sooner during model generation while avoiding cancellation of a running tool. Use `/stop` when you want to cancel the turn and its foreground work. Unknown values fall back to `"interrupt"`.
 
 `"steer"` has two automatic fallbacks: if the agent hasn't started yet, or if images are attached, the message falls back to `"queue"` behavior so nothing is lost.
 
@@ -280,7 +280,7 @@ You can also change it inside the CLI:
 ```
 
 :::tip First-touch hint
-The very first time you press Enter while Hermes is working, Hermes prints a one-line reminder explaining the `/busy` knob (`"(tip) Your message interrupted the current run…"`). It only fires once per install — a flag in `config.yaml` under `onboarding.seen.busy_input_prompt` latches it. Delete that key to see the tip again.
+The first time you press Enter while Hermes is working, Hermes prints a one-line reminder explaining the `/busy` knob. It only fires once per install; `onboarding.seen.busy_input_prompt` in `config.yaml` records that it was shown. Delete that key to see the tip again.
 :::
 
 ### Suspending to Background

website/docs/user-guide/messaging/index.md

@@ -352,18 +352,18 @@ gateway:
 
 Use `/whoami` from any platform to see the active scope, your tier (admin / user / unrestricted), and which slash commands you can run. See the [Telegram](/user-guide/messaging/telegram#slash-command-access-control) and [Discord](/user-guide/messaging/discord#slash-command-access-control) pages for platform-specific examples.
 
-## Interrupting the Agent
+## Redirecting the Agent
 
-Send any message while the agent is working to interrupt it. Key behaviors:
+Send a message while the agent is working to correct the active turn:
 
-- **In-progress terminal commands are killed immediately** (SIGTERM, then SIGKILL after 1s)
-- **Tool calls are cancelled** — only the currently-executing one runs, the rest are skipped
-- **Multiple messages are combined** — messages sent during interruption are joined into one prompt
-- **`/stop` command** — interrupts without queuing a follow-up message
+- **Model generation restarts with context** — reasoning already shown and visible partial text are retained as an ordinary assistant checkpoint
+- **Completed work stays available** — prior tool calls and results remain in the turn
+- **Running tools finish safely** — the correction is applied at the next tool-result boundary instead of killing the tool
+- **`/stop` remains a hard stop** — use it to cancel the active turn and foreground work
 
 ### Queue vs interrupt vs steer (busy-input mode)
 
-By default, messaging a busy agent interrupts it. Two other modes are available:
+By default, messaging a busy agent redirects its active turn. Two other modes are available:
 
 - `queue` — follow-up messages wait and run as the next turn after the current task finishes.
 - `steer` — follow-up messages are injected into the current run via `/steer`, arriving at the agent after the next tool call. No interrupt, no new turn. Falls back to `queue` behavior if the agent hasn't started yet.
@@ -376,7 +376,7 @@ display:
 
 The first time you message a busy agent on any platform, Hermes appends a one-line reminder to the busy-ack explaining the knob (`"💡 First-time tip — …"`). The reminder fires once per install — a flag under `onboarding.seen.busy_input_prompt` latches it. Delete that key to see the tip again.
 
-If you find the busy-ack noisy — especially with voice input or rapid-fire messages — set `display.busy_ack_enabled: false`. Your input is still queued/steered/interrupts as normal, only the chat reply is silenced.
+If you find the busy acknowledgment noisy, set `display.busy_ack_enabled: false`. Input handling is unchanged; only the confirmation message is hidden.
 
 ## Tool Progress Notifications