Skip to content

Your lightweight open-source personal agent.

fix(agent): preserve streamed replies after model failures (#6494)

chengyongru219c8bb2bcaa
diff --git a/docs/webui.md b/docs/webui.md
index 8865ba183..529404681 100644
--- a/docs/webui.md
+++ b/docs/webui.md
@@ -133,6 +133,8 @@ File previews follow the active topic's access mode. Restricted workspace access
### Model failures
+If a model stream fails after producing reply text, the text is retained for reading and follow-up questions. A reply can look complete even when the request ended with a connection error.
+
After a model request fails and automatic retries stop, the WebUI shows the failure category. Use **Continue** to resume from the saved conversation, or open model settings to change the model or correct its authorization. Review completed work before continuing: saved tool results are retained, but the model can choose to call tools again. Continuing does not resend the original message. Sending a new message or dismissing the notice clears the pending decision. Saved recovery decisions remain available after a refresh or gateway restart; temporary chats still follow their normal lifetime.
### Delegated work
diff --git a/nanobot/agent/hook.py b/nanobot/agent/hook.py
index fa6dab3f1..a20bbec41 100644
--- a/nanobot/agent/hook.py
+++ b/nanobot/agent/hook.py
@@ -26,6 +26,8 @@ class AgentHookContext:
tool_results: list[Any] = field(default_factory=list)
tool_events: list[dict[str, str]] = field(default_factory=list)
streamed_content: bool = False
+ # Raw segments observed this round, retained only if its terminal response fails.
+ streamed_content_parts: list[str] = field(default_factory=list, repr=False)
streamed_reasoning: bool = False
stream_continues_current_message: bool = False
final_content: str | None = None
diff --git a/nanobot/agent/runner.py b/nanobot/agent/runner.py
index bdac12f70..fa8782ecd 100644
--- a/nanobot/agent/runner.py
+++ b/nanobot/agent/runner.py
@@ -726,6 +726,23 @@ class AgentRunner:
response,
)
+ if response.finish_reason == "error":
+ interrupted_parts = [
+ content
+ for part in context.streamed_content_parts
+ if (content := extract_reasoning(None, None, part)[1])
+ ]
+ interrupted_content = hook.finalize_content(
+ context, "\n\n".join(interrupted_parts),
+ )
+ if not is_blank_text(interrupted_content):
+ # This is unsynchronized output, not a completed response or
+ # provider-owned projection. Keep it in the next native delta.
+ messages.append({
+ **build_assistant_message(interrupted_content),
+ "_interrupted": True,
+ })
+
# Check for mid-turn injections BEFORE signaling stream end.
# If injections are found we keep the stream alive (resuming=True)
# so streaming channels don't prematurely finalize the card.
@@ -934,6 +951,13 @@ class AgentRunner:
first_output_at: float | None = None
generation_started_at: float | None = None
generation_elapsed_s = 0.0
+ stream_buf = ""
+
+ def _end_stream_segment() -> None:
+ nonlocal stream_buf
+ if stream_buf:
+ context.streamed_content_parts.append(stream_buf)
+ stream_buf = ""
def _generation_delta(delta: str) -> None:
nonlocal first_output_at, generation_started_at
@@ -981,6 +1005,8 @@ class AgentRunner:
if event.get("kind") != "hosted_tool":
return
await _close_native_reasoning()
+ if event.get("phase") == "start":
+ _end_stream_segment()
await hook.on_provider_tool_event(context, event)
call_id = event.get("call_id")
if not call_id:
@@ -995,6 +1021,8 @@ class AgentRunner:
thinking_buf = ""
async def _stream(delta: str) -> None:
+ nonlocal stream_buf
+ stream_buf += delta
_generation_delta(delta)
if delta:
context.streamed_content = True
@@ -1018,6 +1046,7 @@ class AgentRunner:
async def _stream_recover() -> None:
_pause_generation()
await _close_native_reasoning()
+ _end_stream_segment()
await hook.on_stream_end(context, resuming=True)
kwargs.update(
@@ -1039,6 +1068,7 @@ class AgentRunner:
raise
_pause_generation()
await _close_native_reasoning()
+ _end_stream_segment()
if first_output_at is not None:
response.ttft_ms = max(0, round((first_output_at - request_started_at) * 1000))
if generation_elapsed_s > 0:
diff --git a/nanobot/events.py b/nanobot/events.py
index 39dfd8d75..c4be93f93 100644
--- a/nanobot/events.py
+++ b/nanobot/events.py
@@ -70,6 +70,7 @@ class RecoveryStateEvent(AgentEvent):
attempts: int = 0
can_continue: bool | None = None
failure_error_kind: str | None = None
+ reply_preserved: bool | None = None
@dataclass(frozen=True)
diff --git a/nanobot/providers/base.py b/nanobot/providers/base.py
index de886f772..429dadbd3 100644
--- a/nanobot/providers/base.py
+++ b/nanobot/providers/base.py
@@ -1120,7 +1120,7 @@ class LLMProvider(ABC):
),
timeout=resolve_stream_idle_timeout_s(),
)
- if on_content_delta and response.content:
+ if on_content_delta and response.content and response.finish_reason != "error":
await on_content_delta(response.content)
return response
diff --git a/nanobot/session/recovery.py b/nanobot/session/recovery.py
index bea08000a..7c2c5429e 100644
--- a/nanobot/session/recovery.py
+++ b/nanobot/session/recovery.py
@@ -452,4 +452,6 @@ def recovery_state_from_metadata(metadata: Mapping[str, Any] | None) -> dict[str
"authentication", "billing", "connection", "timeout", "rate_limit", "server", "unknown",
}:
payload["failure_error_kind"] = failure_error_kind
+ if state.get("reply_preserved") is True:
+ payload["reply_preserved"] = True
return payload
diff --git a/nanobot/webui/outbound_wire.py b/nanobot/webui/outbound_wire.py
index cf89fd1a3..316f7c968 100644
--- a/nanobot/webui/outbound_wire.py
+++ b/nanobot/webui/outbound_wire.py
@@ -73,6 +73,7 @@ class RecoveryStateWirePayload(_ChatWirePayload):
reason: NotRequired[str]
can_continue: NotRequired[bool]
failure_error_kind: NotRequired[str]
+ reply_preserved: NotRequired[bool]
class RetryStatusWirePayload(_ChatWirePayload):
@@ -139,6 +140,8 @@ def encode_recovery_state(
payload["can_continue"] = event.can_continue
if event.failure_error_kind is not None:
payload["failure_error_kind"] = event.failure_error_kind
+ if event.reply_preserved is not None:
+ payload["reply_preserved"] = event.reply_preserved
return payload
diff --git a/nanobot/webui/recovery.py b/nanobot/webui/recovery.py
index dca69e7e1..15c246d3a 100644
--- a/nanobot/webui/recovery.py
+++ b/nanobot/webui/recovery.py
@@ -187,6 +187,8 @@ class RecoveryCoordinator:
can_continue=any(message.get("role") == "user" for message in session.messages),
)
waiting["failure_error_kind"] = failure_error_kind or "unknown"
+ if session.messages and session.messages[-1].get("_interrupted") is True:
+ waiting["reply_preserved"] = True
await self.sessions.state.restore_interruption(session)
await self._publish(route[1], waiting)
return
@@ -449,6 +451,7 @@ class RecoveryCoordinator:
attempts=cast(int, state.get("attempts", 0)),
can_continue=cast(bool | None, state.get("can_continue")),
failure_error_kind=cast(str | None, state.get("failure_error_kind")),
+ reply_preserved=cast(bool | None, state.get("reply_preserved")),
),
)
)
diff --git a/packages/client-events/fixtures.json b/packages/client-events/fixtures.json
index d8fb4d418..611a0ce97 100644
--- a/packages/client-events/fixtures.json
+++ b/packages/client-events/fixtures.json
@@ -5,7 +5,7 @@
{"event":"context_compaction","chat_id":"chat","compaction_id":"c1","phase":"cancelled"},
{"event":"recovery_state","chat_id":"chat","status":"resuming","recovery_id":"r1","attempts":0},
{"event":"recovery_state","chat_id":"chat","status":"awaiting_user","recovery_id":"r1","attempts":1,"reason":"tool_state_unknown","can_continue":false},
- {"event":"recovery_state","chat_id":"chat","status":"awaiting_user","recovery_id":"r2","attempts":0,"reason":"model_request_failed","failure_error_kind":"connection"},
+ {"event":"recovery_state","chat_id":"chat","status":"awaiting_user","recovery_id":"r2","attempts":0,"reason":"model_request_failed","failure_error_kind":"connection","reply_preserved":true},
{"event":"recovery_state","chat_id":"chat","status":"recovered","recovery_id":"r1","attempts":1},
{"event":"recovery_state","chat_id":"chat","status":"failed","recovery_id":"r1","attempts":2},
{"event":"retry_status","chat_id":"chat","state":"waiting","attempt":1,"max_attempts":4,"error_kind":"connection","retry_after_s":2},
diff --git a/packages/client-events/notifications.ts b/packages/client-events/notifications.ts
index d5312fa6d..ea8ba6f8a 100644
--- a/packages/client-events/notifications.ts
+++ b/packages/client-events/notifications.ts
@@ -8,6 +8,7 @@ export interface RecoveryState {
attempts?: number
can_continue?: boolean
failure_error_kind?: string
+ reply_preserved?: boolean
}
export type CompactionReason = "idle" | "context_pressure" | "manual" | "provider_native"
@@ -63,6 +64,7 @@ export function isRecoveryState(value: unknown): value is RecoveryState {
&& optional(value.attempts, "number")
&& optional(value.can_continue, "boolean")
&& optional(value.failure_error_kind, "string")
+ && optional(value.reply_preserved, "boolean")
}
/** Undefined means another protocol family; null means a malformed notification. */
diff --git a/tests/agent/test_prompt_responsiveness.py b/tests/agent/test_prompt_responsiveness.py
index 3e53e5735..e012ba435 100644
--- a/tests/agent/test_prompt_responsiveness.py
+++ b/tests/agent/test_prompt_responsiveness.py
@@ -454,23 +454,35 @@ async def test_local_token_estimation_leaves_loop_available(
loop._extra_hooks.append(Lifecycle())
- def count(*_args, **_kwargs):
- captured.append(current_request_context())
+ from nanobot.utils.helpers import estimate_prompt_tokens_chain
+
+ def count(*args, **kwargs):
+ captured.append((current_request_context(), threading.get_ident()))
runtime_loop.call_soon_threadsafe(entered.set)
try:
assert release.wait(5), "test did not release local token estimation"
- return 10, "test-counter"
+ return estimate_prompt_tokens_chain(*args, **kwargs)
finally:
runtime_loop.call_soon_threadsafe(completed.set)
module = "context_governance" if phase == "request" else "runner"
monkeypatch.setattr(f"nanobot.agent.{module}.estimate_prompt_tokens_chain", count)
task = asyncio.create_task(loop.process_direct("Hello", session_key="cli:token-work"))
+ entry = asyncio.create_task(entered.wait())
try:
- await asyncio.wait_for(entered.wait(), timeout=5)
- assert captured[0] is not None
- assert captured[0].session_key == "cli:token-work"
- assert captured[0].workspace_scope.project_path == tmp_path.resolve()
+ # Admission and prompt I/O precede estimation. Its worker gate, not turn
+ # startup, owns the five-second responsiveness deadline.
+ done, _ = await asyncio.wait((task, entry), return_when=asyncio.FIRST_COMPLETED)
+ if task in done:
+ task.result()
+ assert entry in done, "turn ended before local token estimation started"
+ request, worker_thread = captured[0]
+ assert request is not None
+ assert request.session_key == "cli:token-work"
+ assert request.workspace_scope.project_path == tmp_path.resolve()
+ assert worker_thread != threading.get_ident()
+ assert not completed.is_set()
+ assert not release.is_set()
assert current_request_context() is None
if phase == "request":
loop.provider.chat_stream_with_retry.assert_not_awaited()
@@ -493,9 +505,10 @@ async def test_local_token_estimation_leaves_loop_available(
assert [public_history_message(message)["content"] for message in saved.messages] == ["Hello", "Done"]
finally:
release.set()
- if not task.done():
- task.cancel()
- await asyncio.gather(task, return_exceptions=True)
+ for pending in (task, entry):
+ if not pending.done():
+ pending.cancel()
+ await asyncio.gather(task, entry, return_exceptions=True)
if captured:
await asyncio.wait_for(completed.wait(), timeout=5)
await loop.aclose()
diff --git a/tests/agent/test_runner_interrupted_output.py b/tests/agent/test_runner_interrupted_output.py
new file mode 100644
index 000000000..38bce9d42
--- /dev/null
+++ b/tests/agent/test_runner_interrupted_output.py
@@ -0,0 +1,287 @@
+"""Model failures retain the assistant text already delivered to the user."""
+
+from __future__ import annotations
+
+from copy import deepcopy
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from agent.runner_helpers import make_run_spec
+from nanobot.agent.progress_hook import AgentProgressHook
+from nanobot.agent.runner import _PERSISTED_MODEL_ERROR_PLACEHOLDER, AgentRunner
+from nanobot.config.schema import ModelPresetConfig
+from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
+from nanobot.providers.fallback_provider import FallbackProvider
+from nanobot.providers.model_api import ModelAPICapabilities
+from nanobot.providers.openai_compat_provider import OpenAICompatProvider
+from nanobot.providers.openai_responses import parse_response_output
+from nanobot.providers.unconfigured_provider import UnconfiguredProvider
+from nanobot.utils.progress_events import output_events
+
+
[email protected]("text", ["The answer is 42.", "The answer starts with"])
+async def test_failed_stream_retains_real_reply_for_the_next_request(text):
+ provider = MagicMock(spec=LLMProvider)
+ tools = MagicMock()
+ tools.get_definitions.return_value = []
+ tools.execute = AsyncMock()
+
+ async def fail_after_output(**kwargs):
+ await kwargs["on_thinking_delta"]("private reasoning")
+ await kwargs["on_content_delta"]("<think>hidden thought</think>" + text)
+ return LLMResponse(
+ content="network diagnostic, not assistant text",
+ finish_reason="error",
+ error_kind="connection",
+ tool_calls=[ToolCallRequest(id="unfinished", name="exec", arguments={})],
+ )
+
+ provider.chat_stream_with_retry = AsyncMock(side_effect=fail_after_output)
+ continue_goal = AsyncMock(return_value="continue the goal")
+ checkpoints = []
+ result = await AgentRunner().run(make_run_spec(
+ provider, initial_messages=[{"role": "user", "content": "question"}],
+ tools=tools, model="test", max_iterations=2, max_tool_result_chars=1000,
+ hook=AgentProgressHook(streaming=True),
+ continuation_callback=continue_goal,
+ checkpoint_callback=AsyncMock(side_effect=checkpoints.append),
+ ))
+
+ assert result.stop_reason == "error"
+ assert result.failure_error_kind == "connection"
+ assert result.messages[-1]["content"] == text
+ assert result.messages[-1]["_interrupted"] is True
+ assert result.final_content == result.error == "network diagnostic, not assistant text"
+ assert checkpoints == []
+ continue_goal.assert_not_awaited()
+ tools.execute.assert_not_awaited()
+
+ provider.chat_stream_with_retry = AsyncMock(return_value=LLMResponse(content="follow-up"))
+ await AgentRunner().run(make_run_spec(
+ provider,
+ initial_messages=[*result.messages, {"role": "user", "content": "explain that"}],
+ tools=tools, model="test", max_iterations=1, max_tool_result_chars=1000,
+ ))
+ sent = provider.chat_stream_with_retry.await_args.kwargs["messages"]
+ assert [message["content"] for message in sent if message.get("role") == "assistant"] == [text]
+
+
+async def test_failed_stream_keeps_visible_segments_across_hosted_tools_and_recovery():
+ provider = MagicMock(spec=LLMProvider)
+ tools = MagicMock()
+ tools.get_definitions.return_value = []
+ visible = []
+
+ async def show(delta):
+ visible.append(delta)
+
+ async def fail_after_segments(**kwargs):
+ await kwargs["on_content_delta"]("<thi")
+ await kwargs["on_content_delta"]("nk>hidden</think>Before search.")
+ await kwargs["on_tool_call_delta"]({
+ "kind": "hosted_tool", "phase": "start", "call_id": "search-1",
+ "name": "web_search", "arguments": {},
+ })
+ await kwargs["on_content_delta"]("<think>unfinished hidden thought")
+ await kwargs["on_stream_recover"]()
+ await kwargs["on_content_delta"]("<thinking>new hidden thought</thinking>After search.")
+ return LLMResponse(content="connection lost", finish_reason="error", error_kind="connection")
+
+ provider.chat_stream_with_retry = AsyncMock(side_effect=fail_after_segments)
+ result = await AgentRunner().run(make_run_spec(
+ provider, initial_messages=[{"role": "user", "content": "search"}],
+ tools=tools, model="test", max_iterations=1, max_tool_result_chars=1000,
+ hook=AgentProgressHook(output_events(on_stream=show), streaming=True),
+ ))
+
+ assert visible == ["Before search.", "After search."]
+ assert result.messages[-1]["content"] == "Before search.\n\nAfter search."
+ assert result.messages[-1]["_interrupted"] is True
+ assert result.stop_reason == "error"
+
+
+async def test_reasoning_only_failure_does_not_claim_a_preserved_reply():
+ provider = MagicMock(spec=LLMProvider)
+ tools = MagicMock()
+ tools.get_definitions.return_value = []
+
+ async def fail_after_reasoning(**kwargs):
+ await kwargs["on_thinking_delta"]("native private reasoning")
+ await kwargs["on_content_delta"]("<think>unfinished private reasoning")
+ return LLMResponse(content="connection lost", finish_reason="error", error_kind="connection")
+
+ provider.chat_stream_with_retry = AsyncMock(side_effect=fail_after_reasoning)
+ result = await AgentRunner().run(make_run_spec(
+ provider, initial_messages=[{"role": "user", "content": "question"}],
+ tools=tools, model="test", max_iterations=1, max_tool_result_chars=1000,
+ hook=AgentProgressHook(streaming=True),
+ ))
+
+ assert result.messages[-1]["content"] == _PERSISTED_MODEL_ERROR_PLACEHOLDER
+ assert result.messages[-1].get("_interrupted") is not True
+ assert result.stop_reason == "error"
+
+
+async def test_nonstreaming_provider_error_is_not_observed_as_assistant_body():
+ provider = UnconfiguredProvider("test")
+ tools = MagicMock()
+ tools.get_definitions.return_value = []
+ show = AsyncMock()
+ result = await AgentRunner().run(make_run_spec(
+ provider, initial_messages=[{"role": "user", "content": "question"}],
+ tools=tools, model="test", max_iterations=1, max_tool_result_chars=1000,
+ hook=AgentProgressHook(output_events(on_stream=show), streaming=True),
+ ))
+
+ assert result.stop_reason == "error"
+ assert result.failure_error_kind == "unknown"
+ assert result.error == result.final_content
+ assert "Open Settings → Models" in result.final_content
+ assert result.messages[-1]["content"] == _PERSISTED_MODEL_ERROR_PLACEHOLDER
+ assert result.messages[-1].get("_interrupted") is not True
+ show.assert_not_awaited()
+
+
+class _StreamingProvider(LLMProvider):
+ def __init__(self, steps):
+ super().__init__(provider_name="test")
+ self.steps = iter(steps)
+ self.calls = 0
+
+ def get_default_model(self):
+ return "test"
+
+ async def chat(self, **kwargs):
+ raise AssertionError("expected a streaming request")
+
+ async def chat_stream(self, **kwargs):
+ self.calls += 1
+ text, response = next(self.steps)
+ await kwargs["on_content_delta"](text)
+ return response
+
+
[email protected]("fallback_fails", [False, True])
+async def test_retry_and_fallback_preserve_failed_segments_only_on_terminal_error(
+ monkeypatch, fallback_fails,
+):
+ monkeypatch.setattr(LLMProvider, "_CHAT_RETRY_DELAYS", (0,))
+ timeout = LLMResponse(content="stream timeout diagnostic", finish_reason="error", error_kind="timeout")
+ primary = _StreamingProvider([
+ ("First attempt.", timeout), ("Second attempt.", timeout),
+ ])
+ final = LLMResponse(
+ content="fallback connection diagnostic" if fallback_fails else "Backup answer.",
+ finish_reason="error" if fallback_fails else "stop",
+ error_kind="connection" if fallback_fails else None,
+ )
+ fallback = _StreamingProvider([("Backup answer.", final)])
+ provider = FallbackProvider(
+ primary, [ModelPresetConfig(model="backup", provider="custom")],
+ lambda _preset: fallback,
+ )
+ tools = MagicMock()
+ tools.get_definitions.return_value = []
+ result = await AgentRunner().run(make_run_spec(
+ provider, initial_messages=[{"role": "user", "content": "question"}],
+ tools=tools, model="test", max_iterations=1, max_tool_result_chars=1000,
+ hook=AgentProgressHook(streaming=True),
+ ))
+
+ assert primary.calls == 2
+ assert fallback.calls == 1
+ assistant = [message for message in result.messages if message["role"] == "assistant"]
+ assert len(assistant) == 1
+ if fallback_fails:
+ assert assistant[0]["content"] == "First attempt.\n\nSecond attempt.\n\nBackup answer."
+ assert assistant[0]["_interrupted"] is True
+ assert result.stop_reason == "error"
+ assert result.final_content == "fallback connection diagnostic"
+ else:
+ assert assistant[0]["content"] == "Backup answer."
+ assert assistant[0].get("_interrupted") is not True
+ assert result.stop_reason == "completed"
+ assert result.final_content == "Backup answer."
+
+
[email protected]("midturn", [False, True])
+async def test_native_next_request_contains_preserved_reply_without_repeating_completed_tool(
+ monkeypatch, midturn,
+):
+ provider = OpenAICompatProvider(
+ api_key="test", default_model="test", model_api=ModelAPICapabilities(
+ supported_apis=("responses",), preferred_api="responses",
+ ),
+ )
+ monkeypatch.setattr(provider, "_ensure_client", AsyncMock(return_value=MagicMock()))
+ requests = []
+ injections = []
+ checkpoints = []
+ followup = {"role": "user", "content": "explain that"}
+
+ async def send(_client, body, **kwargs):
+ requests.append(deepcopy(body["input"]))
+ if len(requests) == 1:
+ return parse_response_output({
+ "status": "completed", "output": [{
+ "type": "function_call", "id": "fc_1", "call_id": "call_1",
+ "name": "read_file", "arguments": "{}",
+ }],
+ }, state_provider=kwargs["provider"], state_model=body["model"],
+ state_input_items=body["input"])
+ if len(requests) == 2:
+ await kwargs["on_content_delta"]("Preserved conclusion.")
+ if midturn:
+ injections.append(followup)
+ raise ConnectionError("connection lost after reply")
+ return LLMResponse(content="follow-up answer")
+
+ async def drain():
+ batch = list(injections)
+ injections.clear()
+ return batch
+
+ monkeypatch.setattr(provider._responses, "sdk_request", send)
+ tools = MagicMock()
+ tools.get_definitions.return_value = []
+ tools.execute = AsyncMock(return_value="completed tool output")
+ result = await AgentRunner().run(make_run_spec(
+ provider, initial_messages=[{"role": "user", "content": "question"}],
+ tools=tools, model="test", max_iterations=3, max_tool_result_chars=1000,
+ hook=AgentProgressHook(streaming=True), injection_callback=drain,
+ checkpoint_callback=AsyncMock(side_effect=checkpoints.append),
+ ))
+ if not midturn:
+ assert result.stop_reason == "error"
+ assert result.messages[-1]["_interrupted"] is True
+ assert result.provider_state is not None
+ await AgentRunner().run(make_run_spec(
+ provider, initial_messages=[*result.messages, followup], tools=tools,
+ model="test", max_iterations=1, max_tool_result_chars=1000,
+ provider_state=result.provider_state.with_pending_messages([
+ *result.provider_state.pending_messages, followup,
+ ]),
+ ))
+ else:
+ assert result.stop_reason == "completed"
+
+ assert len(requests) == 3
+ assert requests[2][-2:] == [{
+ "type": "message", "role": "assistant", "status": "completed",
+ "content": [{"type": "output_text", "text": "Preserved conclusion."}],
+ }, {
+ "role": "user", "content": [{"type": "input_text", "text": "explain that"}],
+ }]
+ assert [item for item in requests[2] if item.get("type") == "function_call_output"] == [{
+ "type": "function_call_output", "call_id": "call_1", "output": "completed tool output",
+ }]
+ assert not any(checkpoint["assistant_message"].get("_interrupted") for checkpoint in checkpoints)
+ tools.execute.assert_awaited_once_with("read_file", {})
+ await provider.aclose()
diff --git a/tests/session/test_recovery.py b/tests/session/test_recovery.py
index f12238aba..a715c7659 100644
--- a/tests/session/test_recovery.py
+++ b/tests/session/test_recovery.py
@@ -739,10 +739,15 @@ async def test_dismiss_does_not_queue_work(tmp_path: Path) -> None:
@pytest.mark.asyncio
-async def test_model_failure_waits_across_restart_and_rejects_duplicate_continue(tmp_path: Path) -> None:
[email protected]("reply_preserved", [False, True])
+async def test_model_failure_waits_across_restart_and_rejects_duplicate_continue(
+ tmp_path: Path, reply_preserved: bool,
+) -> None:
coordinator, bus, sessions = _coordinator(tmp_path)
session = sessions.get_or_create("websocket:chat")
session.messages = [{"role": "user", "content": "finish this"}]
+ if reply_preserved:
+ session.messages.append({"role": "assistant", "content": "saved reply", "_interrupted": True})
_persist(sessions, session)
await coordinator.turn_completed("websocket:chat", failure_kind="model", failure_error_kind="connection")
@@ -752,6 +757,11 @@ async def test_model_failure_waits_across_restart_and_rejects_duplicate_continue
assert bus.inbound.empty()
state = sessions.get_or_create("websocket:chat").metadata[RECOVERY_METADATA_KEY]
assert state["failure_error_kind"] == "connection"
+ assert state.get("reply_preserved", False) is reply_preserved
+ from nanobot.session.recovery import recovery_state_from_metadata
+ public_state = recovery_state_from_metadata({RECOVERY_METADATA_KEY: state})
+ assert public_state is not None
+ assert public_state.get("reply_preserved", False) is reply_preserved
assert state["status"] == "awaiting_user"
payload = {"chat_id": "chat", "recovery_id": state["recovery_id"]}
await coordinator.handle_action("continue", payload)
diff --git a/tests/webui/test_outbound_wire.py b/tests/webui/test_outbound_wire.py
index 09c783703..a5910c264 100644
--- a/tests/webui/test_outbound_wire.py
+++ b/tests/webui/test_outbound_wire.py
@@ -161,3 +161,18 @@ def test_encode_turn_end_projects_complete_wire_contract() -> None:
"failure_attempts": 4,
"failure_message": "Connection refused",
}
+
+
[email protected]("preserved", [False, True])
+def test_encode_recovery_state_reports_reply_preservation(preserved: bool) -> None:
+ payload = encode_recovery_state(
+ "chat-1",
+ RecoveryStateEvent(
+ status="awaiting_user", recovery_id="recovery-1",
+ reason="model_request_failed", failure_error_kind="connection",
+ reply_preserved=preserved,
+ ),
+ )
+ assert payload["reply_preserved"] is preserved
+ assert payload["status"] == "awaiting_user"
+ assert payload["failure_error_kind"] == "connection"
diff --git a/webui/src/components/thread/RecoveryNotice.tsx b/webui/src/components/thread/RecoveryNotice.tsx
index f7ad9af2f..b3824b5fa 100644
--- a/webui/src/components/thread/RecoveryNotice.tsx
+++ b/webui/src/components/thread/RecoveryNotice.tsx
@@ -45,7 +45,9 @@ export function RecoveryNotice({ state, onContinue, onDismiss, onOpenSettings }:
? t("recovery.failedHelp", {
defaultValue: "The saved task could not be restored safely. Review it before continuing.",
})
- : modelFailure?.body ?? t("recovery.review", { defaultValue: "Review the task before continuing. Tools will not be replayed automatically." });
+ : modelFailure && state.reply_preserved === true
+ ? t("recovery.replyPreserved", { defaultValue: "Reply saved. You can continue." })
+ : modelFailure?.body ?? t("recovery.review", { defaultValue: "Review the task before continuing. Tools will not be replayed automatically." });
const run = (action: "continue" | "dismiss") => {
setPending(action);
setError(null);
diff --git a/webui/src/hooks/useNanobotStream.ts b/webui/src/hooks/useNanobotStream.ts
index 6b1c176dd..8a749eb98 100644
--- a/webui/src/hooks/useNanobotStream.ts
+++ b/webui/src/hooks/useNanobotStream.ts
@@ -870,6 +870,9 @@ export function useNanobotStream(
...(typeof ev.failure_error_kind === "string"
? { failure_error_kind: ev.failure_error_kind }
: {}),
+ ...(typeof ev.reply_preserved === "boolean"
+ ? { reply_preserved: ev.reply_preserved }
+ : {}),
};
setRecoveryState(next);
if (ev.reason === "model_request_failed" || ev.status === "resuming") {
diff --git a/webui/src/i18n/locales/en/common.json b/webui/src/i18n/locales/en/common.json
index 7989aaba2..7a50e6019 100644
--- a/webui/src/i18n/locales/en/common.json
+++ b/webui/src/i18n/locales/en/common.json
@@ -2546,6 +2546,7 @@
"failedHelp": "The saved task could not be restored safely. Review it before continuing.",
"resuming": "Restoring interrupted task…",
"review": "Review the task before continuing. Tools will not be replayed automatically.",
+ "replyPreserved": "Reply saved. You can continue.",
"safeResume": "Continuing from saved conversation context.",
"dismiss": "Dismiss",
"continue": "Continue"
@@ -2670,7 +2671,7 @@
"authenticationBody": "Sign in again or check the provider’s API key in model settings.",
"billingTitle": "Model provider quota is unavailable",
"billingBody": "Add credit or check billing for the provider account, then try again.",
- "connectionTitle": "Could not connect to the model provider",
+ "connectionTitle": "Model connection failed",
"timeoutTitle": "Model provider request timed out",
"rateLimitTitle": "Model provider rate limit reached",
"serverTitle": "Model provider service error",
diff --git a/webui/src/i18n/locales/es/common.json b/webui/src/i18n/locales/es/common.json
index 99988a5f5..e6a0d6bf3 100644
--- a/webui/src/i18n/locales/es/common.json
+++ b/webui/src/i18n/locales/es/common.json
@@ -2341,6 +2341,7 @@
"failedHelp": "La tarea guardada no se pudo restaurar de forma segura. Revísala antes de continuar.",
"resuming": "Restaurando la tarea interrumpida…",
"review": "Revisa la tarea antes de continuar. Las herramientas no se repetirán automáticamente.",
+ "replyPreserved": "La respuesta está guardada. Puedes continuar.",
"safeResume": "Continuando desde el contexto guardado de la conversación.",
"dismiss": "Descartar",
"continue": "Continuar"
@@ -2465,7 +2466,7 @@
"authenticationBody": "Vuelve a iniciar sesión o revisa la clave API en los ajustes del modelo.",
"billingTitle": "La cuota del proveedor del modelo no está disponible",
"billingBody": "Añade crédito o revisa la facturación de la cuenta del proveedor y vuelve a intentarlo.",
- "connectionTitle": "No se pudo conectar con el proveedor del modelo",
+ "connectionTitle": "Error de conexión con el modelo",
"timeoutTitle": "La solicitud al proveedor del modelo agotó el tiempo de espera",
"rateLimitTitle": "Se alcanzó el límite de solicitudes del proveedor del modelo",
"serverTitle": "Error del servicio del proveedor del modelo",
diff --git a/webui/src/i18n/locales/fr/common.json b/webui/src/i18n/locales/fr/common.json
index 49203f74c..e45998b45 100644
--- a/webui/src/i18n/locales/fr/common.json
+++ b/webui/src/i18n/locales/fr/common.json
@@ -2340,6 +2340,7 @@
"failedHelp": "La tâche enregistrée n’a pas pu être restaurée en toute sécurité. Vérifiez-la avant de continuer.",
"resuming": "Restauration de la tâche interrompue…",
"review": "Vérifiez la tâche avant de continuer. Les outils ne seront pas relancés automatiquement.",
+ "replyPreserved": "La réponse est enregistrée. Vous pouvez continuer.",
"safeResume": "Reprise depuis le contexte de conversation enregistré.",
"dismiss": "Ignorer",
"continue": "Continuer"
@@ -2464,7 +2465,7 @@
"authenticationBody": "Reconnectez-vous ou vérifiez la clé API dans les paramètres du modèle.",
"billingTitle": "Quota du fournisseur du modèle indisponible",
"billingBody": "Ajoutez du crédit ou vérifiez la facturation du compte fournisseur, puis réessayez.",
- "connectionTitle": "Impossible de se connecter au fournisseur du modèle",
+ "connectionTitle": "Échec de connexion au modèle",
"timeoutTitle": "La requête au fournisseur du modèle a expiré",
"rateLimitTitle": "Limite de requêtes du fournisseur du modèle atteinte",
"serverTitle": "Erreur du service du fournisseur du modèle",
diff --git a/webui/src/i18n/locales/id/common.json b/webui/src/i18n/locales/id/common.json
index 3d51359fd..9191dfad6 100644
--- a/webui/src/i18n/locales/id/common.json
+++ b/webui/src/i18n/locales/id/common.json
@@ -2340,6 +2340,7 @@
"failedHelp": "Tugas tersimpan tidak dapat dipulihkan dengan aman. Tinjau sebelum melanjutkan.",
"resuming": "Memulihkan tugas yang terputus…",
"review": "Tinjau tugas sebelum melanjutkan. Alat tidak akan dijalankan ulang secara otomatis.",
+ "replyPreserved": "Balasan tersimpan. Anda dapat melanjutkan.",
"safeResume": "Melanjutkan dari konteks percakapan yang tersimpan.",
"dismiss": "Abaikan",
"continue": "Lanjutkan"
@@ -2464,7 +2465,7 @@
"authenticationBody": "Masuk kembali atau periksa kunci API di pengaturan model.",
"billingTitle": "Kuota penyedia model tidak tersedia",
"billingBody": "Tambahkan kredit atau periksa tagihan akun penyedia, lalu coba lagi.",
- "connectionTitle": "Tidak dapat terhubung ke penyedia model",
+ "connectionTitle": "Koneksi model gagal",
"timeoutTitle": "Permintaan ke penyedia model kehabisan waktu",
"rateLimitTitle": "Batas permintaan penyedia model tercapai",
"serverTitle": "Kesalahan layanan penyedia model",
diff --git a/webui/src/i18n/locales/ja/common.json b/webui/src/i18n/locales/ja/common.json
index cb798fff1..16588a142 100644
--- a/webui/src/i18n/locales/ja/common.json
+++ b/webui/src/i18n/locales/ja/common.json
@@ -2340,6 +2340,7 @@
"failedHelp": "保存されたタスクを安全に復元できませんでした。続行前に確認してください。",
"resuming": "中断されたタスクを復元しています…",
"review": "続行する前にタスクを確認してください。ツールは自動的に再実行されません。",
+ "replyPreserved": "返信は保存されています。生成を続けられます。",
"safeResume": "保存された会話コンテキストから続行しています。",
"dismiss": "閉じる",
"continue": "続行"
@@ -2464,7 +2465,7 @@
"authenticationBody": "モデル設定で再度ログインするか、API キーを確認してください。",
"billingTitle": "モデルプロバイダーの利用枠を使用できません",
"billingBody": "プロバイダーアカウントにクレジットを追加するか、請求状況を確認してから、もう一度お試しください。",
- "connectionTitle": "モデルプロバイダーに接続できません",
+ "connectionTitle": "モデル接続エラー",
"timeoutTitle": "モデルプロバイダーへのリクエストがタイムアウトしました",
"rateLimitTitle": "モデルプロバイダーのリクエスト上限に達しました",
"serverTitle": "モデルプロバイダーでサービスエラーが発生しました",
diff --git a/webui/src/i18n/locales/ko/common.json b/webui/src/i18n/locales/ko/common.json
index 79e304dbf..934193efd 100644
--- a/webui/src/i18n/locales/ko/common.json
+++ b/webui/src/i18n/locales/ko/common.json
@@ -2340,6 +2340,7 @@
"failedHelp": "저장된 작업을 안전하게 복구할 수 없습니다. 계속하기 전에 검토하세요.",
"resuming": "중단된 작업을 복구하는 중…",
"review": "계속하기 전에 작업을 검토하세요. 도구는 자동으로 다시 실행되지 않습니다.",
+ "replyPreserved": "답변이 저장되었습니다. 계속 생성할 수 있습니다.",
"safeResume": "저장된 대화 컨텍스트에서 계속합니다.",
"dismiss": "닫기",
"continue": "계속"
@@ -2464,7 +2465,7 @@
"authenticationBody": "모델 설정에서 다시 로그인하거나 API 키를 확인하세요.",
"billingTitle": "모델 제공업체 할당량을 사용할 수 없음",
"billingBody": "제공업체 계정에 크레딧을 추가하거나 결제 상태를 확인한 후 다시 시도하세요.",
- "connectionTitle": "모델 제공업체에 연결할 수 없음",
+ "connectionTitle": "모델 연결 오류",
"timeoutTitle": "모델 제공업체 요청 시간 초과",
"rateLimitTitle": "모델 제공업체 요청 한도 초과",
"serverTitle": "모델 제공업체 서비스 오류",
diff --git a/webui/src/i18n/locales/pt-BR/common.json b/webui/src/i18n/locales/pt-BR/common.json
index b45fd39ef..29a423da0 100644
--- a/webui/src/i18n/locales/pt-BR/common.json
+++ b/webui/src/i18n/locales/pt-BR/common.json
@@ -2354,6 +2354,7 @@
"failedHelp": "Não foi possível restaurar a tarefa salva com segurança. Revise-a antes de continuar.",
"resuming": "Restaurando a tarefa interrompida…",
"review": "Revise a tarefa antes de continuar. As ferramentas não serão executadas novamente de forma automática.",
+ "replyPreserved": "A resposta foi salva. Você pode continuar.",
"safeResume": "Continuando a partir do contexto de conversa salvo.",
"dismiss": "Descartar",
"continue": "Continuar"
@@ -2478,7 +2479,7 @@
"authenticationBody": "Entre novamente ou verifique a chave API nas configurações do modelo.",
"billingTitle": "A cota do provedor do modelo está indisponível",
"billingBody": "Adicione créditos ou verifique o faturamento da conta do provedor e tente novamente.",
- "connectionTitle": "Não foi possível conectar ao provedor do modelo",
+ "connectionTitle": "Falha na conexão com o modelo",
"timeoutTitle": "A solicitação ao provedor do modelo expirou",
"rateLimitTitle": "Limite de solicitações do provedor do modelo atingido",
"serverTitle": "Erro no serviço do provedor do modelo",
diff --git a/webui/src/i18n/locales/vi/common.json b/webui/src/i18n/locales/vi/common.json
index a2e4598d9..c0ead703c 100644
--- a/webui/src/i18n/locales/vi/common.json
+++ b/webui/src/i18n/locales/vi/common.json
@@ -2340,6 +2340,7 @@
"failedHelp": "Không thể khôi phục an toàn tác vụ đã lưu. Hãy xem lại trước khi tiếp tục.",
"resuming": "Đang khôi phục tác vụ bị gián đoạn…",
"review": "Hãy xem lại tác vụ trước khi tiếp tục. Công cụ sẽ không tự động chạy lại.",
+ "replyPreserved": "Câu trả lời đã được lưu. Bạn có thể tiếp tục.",
"safeResume": "Đang tiếp tục từ ngữ cảnh hội thoại đã lưu.",
"dismiss": "Bỏ qua",
"continue": "Tiếp tục"
@@ -2464,7 +2465,7 @@
"authenticationBody": "Đăng nhập lại hoặc kiểm tra khóa API trong cài đặt mô hình.",
"billingTitle": "Hạn mức của nhà cung cấp mô hình không khả dụng",
"billingBody": "Hãy nạp thêm tín dụng hoặc kiểm tra thanh toán của tài khoản nhà cung cấp rồi thử lại.",
- "connectionTitle": "Không thể kết nối với nhà cung cấp mô hình",
+ "connectionTitle": "Lỗi kết nối mô hình",
"timeoutTitle": "Yêu cầu tới nhà cung cấp mô hình đã hết thời gian chờ",
"rateLimitTitle": "Đã đạt giới hạn yêu cầu của nhà cung cấp mô hình",
"serverTitle": "Lỗi dịch vụ nhà cung cấp mô hình",
diff --git a/webui/src/i18n/locales/zh-CN/common.json b/webui/src/i18n/locales/zh-CN/common.json
index 35c4a5856..3b0e003da 100644
--- a/webui/src/i18n/locales/zh-CN/common.json
+++ b/webui/src/i18n/locales/zh-CN/common.json
@@ -2546,6 +2546,7 @@
"failedHelp": "无法安全恢复已保存的任务,继续前请先检查。",
"resuming": "正在恢复中断的任务…",
"review": "继续前请检查任务。工具不会被自动重放。",
+ "replyPreserved": "回复已保留,可继续生成。",
"safeResume": "正在从已保存的对话上下文继续。",
"dismiss": "忽略",
"continue": "继续"
@@ -2670,7 +2671,7 @@
"authenticationBody": "请在模型设置中重新登录或检查 API 密钥。",
"billingTitle": "模型提供商账户额度不足",
"billingBody": "请为提供商账户充值或检查账单状态后重试。",
- "connectionTitle": "无法连接模型提供商",
+ "connectionTitle": "模型连接异常",
"timeoutTitle": "模型提供商请求超时",
"rateLimitTitle": "模型提供商请求受限",
"serverTitle": "模型提供商服务异常",
diff --git a/webui/src/i18n/locales/zh-TW/common.json b/webui/src/i18n/locales/zh-TW/common.json
index 48027b8b9..60a03ac2c 100644
--- a/webui/src/i18n/locales/zh-TW/common.json
+++ b/webui/src/i18n/locales/zh-TW/common.json
@@ -2340,6 +2340,7 @@
"failedHelp": "無法安全復原已儲存的任務,繼續前請先檢查。",
"resuming": "正在復原中斷的任務…",
"review": "繼續前請檢查任務。工具不會自動重播。",
+ "replyPreserved": "回覆已保留,可繼續生成。",
"safeResume": "正在從已儲存的對話上下文繼續。",
"dismiss": "略過",
"continue": "繼續"
@@ -2464,7 +2465,7 @@
"authenticationBody": "請在模型設定中重新登入或檢查 API 金鑰。",
"billingTitle": "模型供應商額度無法使用",
"billingBody": "請為供應商帳戶儲值或檢查帳單狀態後再試。",
- "connectionTitle": "無法連線至模型供應商",
+ "connectionTitle": "模型連線異常",
"timeoutTitle": "模型供應商請求逾時",
"rateLimitTitle": "模型供應商請求受到限制",
"serverTitle": "模型供應商服務發生錯誤",
diff --git a/webui/src/tests/model-request-failure.test.ts b/webui/src/tests/model-request-failure.test.ts
index a940cf4b8..0c76d2f82 100644
--- a/webui/src/tests/model-request-failure.test.ts
+++ b/webui/src/tests/model-request-failure.test.ts
@@ -12,7 +12,7 @@ describe("model request failure copy", () => {
expect(titles).toEqual([
"Model provider authorization failed",
"Model provider quota is unavailable",
- "Could not connect to the model provider",
+ "Model connection failed",
"Model provider request timed out",
"Model provider rate limit reached",
"Model provider service error",
diff --git a/webui/src/tests/recovery-notice.test.tsx b/webui/src/tests/recovery-notice.test.tsx
index 7c2722819..d697b096a 100644
--- a/webui/src/tests/recovery-notice.test.tsx
+++ b/webui/src/tests/recovery-notice.test.tsx
@@ -154,3 +154,32 @@ describe("RecoveryNotice", () => {
expect(screen.getByRole("button", { name: "Continue" })).toBeEnabled();
});
});
+
+
+it("shows a saved reply without treating the failed request as completed", async () => {
+ const onContinue = vi.fn().mockResolvedValue(undefined);
+ const onDismiss = vi.fn().mockResolvedValue(undefined);
+ const { rerender } = render(
+ <RecoveryNotice
+ state={{ ...INTERRUPTED, reason: "model_request_failed", failure_error_kind: "connection", reply_preserved: true }}
+ onContinue={onContinue}
+ onDismiss={onDismiss}
+ />,
+ );
+
+ expect(screen.getByRole("alert")).toHaveTextContent("Model connection failed");
+ expect(screen.getByRole("alert")).toHaveTextContent("Reply saved. You can continue.");
+ fireEvent.click(screen.getByRole("button", { name: "Continue" }));
+ await waitFor(() => expect(screen.queryByRole("alert")).not.toBeInTheDocument());
+ expect(onContinue).toHaveBeenCalledOnce();
+
+ rerender(
+ <RecoveryNotice
+ state={{ ...INTERRUPTED, recovery_id: "no-output", reason: "model_request_failed", failure_error_kind: "connection" }}
+ onContinue={onContinue}
+ onDismiss={onDismiss}
+ />,
+ );
+ expect(screen.getByRole("alert")).toHaveTextContent("Model connection failed");
+ expect(screen.getByRole("alert")).toHaveTextContent("Try again later or choose another model.");
+});
diff --git a/webui/src/tests/shared-proxy-setup.test.tsx b/webui/src/tests/shared-proxy-setup.test.tsx
index ae16dc5e7..1705f1c6b 100644
--- a/webui/src/tests/shared-proxy-setup.test.tsx
+++ b/webui/src/tests/shared-proxy-setup.test.tsx
@@ -140,6 +140,7 @@ describe("Inline shared proxy setup", () => {
expect(input).toBeDisabled();
fireEvent.keyDown(setup, { key: "Escape" });
expect(setup).toBeInTheDocument();
+ await waitFor(() => expect(requestMutationMock).toHaveBeenCalledTimes(2));
await act(async () => { resolveSave(saved); });
await waitFor(() => expect(screen.queryByRole("dialog", { name: "Shared proxy" })).not.toBeInTheDocument());
expect(screen.getByRole("dialog", { name: "Docs MCP" })).toBe(parent);
diff --git a/webui/src/tests/thread-shell.test.tsx b/webui/src/tests/thread-shell.test.tsx
index 5478d5a34..7d837c512 100644
--- a/webui/src/tests/thread-shell.test.tsx
+++ b/webui/src/tests/thread-shell.test.tsx
@@ -5207,7 +5207,7 @@ describe("ThreadShell", () => {
});
const banner = await screen.findByRole("alert");
- expect(banner).toHaveTextContent("无法连接模型提供商");
+ expect(banner).toHaveTextContent("模型连接异常");
expect(banner).toHaveTextContent(
"已尝试 4 次,仍未成功。可稍后继续或切换模型。",
);
@@ -5224,7 +5224,7 @@ describe("ThreadShell", () => {
});
});
expect(screen.getAllByRole("alert")).toHaveLength(1);
- expect(screen.getByRole("alert")).toHaveTextContent("无法连接模型提供商");
+ expect(screen.getByRole("alert")).toHaveTextContent("模型连接异常");
expect(screen.getByRole("button", { name: "继续" })).toBeEnabled();
}
diff --git a/webui/src/tests/useNanobotStream.test.tsx b/webui/src/tests/useNanobotStream.test.tsx
index d22285351..bdbc92a4b 100644
--- a/webui/src/tests/useNanobotStream.test.tsx
+++ b/webui/src/tests/useNanobotStream.test.tsx
@@ -822,6 +822,26 @@ describe("useNanobotStream", () => {
},
);
+ it.each([true, false, undefined])("keeps saved-reply facts consistent live and on attach (%s)", (preserved) => {
+ const fake = fakeClient();
+ const { result } = renderHook(() => useNanobotStream("chat-preserved", EMPTY_MESSAGES), {
+ wrapper: wrap(fake.client),
+ });
+ const state = {
+ recovery_id: "preserved-1", status: "awaiting_user" as const,
+ reason: "model_request_failed", failure_error_kind: "connection",
+ ...(preserved === undefined ? {} : { reply_preserved: preserved }),
+ };
+ act(() => fake.emit("chat-preserved", {
+ event: "recovery_state", chat_id: "chat-preserved", ...state,
+ }));
+ expect(result.current.recoveryState).toEqual(state);
+ act(() => fake.emit("chat-preserved", {
+ event: "attached", chat_id: "chat-preserved", recovery_state: state,
+ }));
+ expect(result.current.recoveryState).toEqual(state);
+ });
+
it("does not let historical recovered state clear a later active turn", () => {
const fake = fakeClient();
const { result } = renderHook(