Implement watchdog to capture session termination and gracefully exist from cleint session.

[Agent-Hellboy]

Thanks for implementing this. Could we add a watchdog to prevent the client from hanging and returning unclear error messages?

exit command is working fine but for usability, scenario like this should provide info to user as what has happened exactly.

_models.py", line 1031, in aiter_lines
    async for text in self.aiter_text():
        for line in decoder.decode(text):
            yield line
  File "/home/proshan/explore-TOOL-5406/fast-agent/.venv/lib/python3.13/site-packages/httpx/_models.py", line 1018, in aiter_text
    async for byte_content in self.aiter_bytes():
    ...<2 lines>...
            yield chunk
  File "/home/proshan/explore-TOOL-5406/fast-agent/.venv/lib/python3.13/site-packages/httpx/_models.py", line 997, in aiter_bytes
    async for raw_bytes in self.aiter_raw():
    ...<2 lines>...
            yield chunk
  File "/home/proshan/explore-TOOL-5406/fast-agent/.venv/lib/python3.13/site-packages/httpx/_models.py", line 1055, in aiter_raw
    async for raw_stream_bytes in self.stream:
    ...<2 lines>...
            yield chunk
  File "/home/proshan/explore-TOOL-5406/fast-agent/.venv/lib/python3.13/site-packages/httpx/_client.py", line 176, in __aiter__
    async for chunk in self._stream:
        yield chunk
  File "/home/proshan/explore-TOOL-5406/fast-agent/.venv/lib/python3.13/site-packages/httpx/_transports/default.py", line 270, in __aiter__
    with map_httpcore_exceptions():
         ~~~~~~~~~~~~~~~~~~~~~~~^^
  File "/home/proshan/.local/share/uv/python/cpython-3.13.3-linux-x86_64-gnu/lib/python3.13/contextlib.py", line 162, in __exit__
    self.gen.throw(value)
    ~~~~~~~~~~~~~~^^^^^^^
  File "/home/proshan/explore-TOOL-5406/fast-agent/.venv/lib/python3.13/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
    raise mapped_exc(message) from exc
httpx.RemoteProtocolError: peer closed connection without sending complete message body (incomplete chunked read)

[Agent-Hellboy]

should i implement this?, it would help we understand the arch a bit.

[evalstate]

Thanks @Agent-Hellboy -- i picked up a defect in this area yesterday, but not sure if it will fix this issue. See #345, #342 and #291 for some existing conversations -- would love your input on improving this.

[Agent-Hellboy]

Thanks @Agent-Hellboy -- i picked up a defect in this area yesterday, but not sure if it will fix this issue. See #345, #342 and #291 for some existing conversations -- would love your input on improving this.

sure thing

[Agent-Hellboy]

Hey @evalstate, Could you please share the current status of one of the base capabilities, Cancellation? I can see it’s implemented at the protocol layer(I am doubtful about this as well) but not at the UI/UX layer. Currently, it simply kills the Python process. The UI should receive a cancellation request and forward it; And Ctrl-C should be the cancellation command to cancel the current tool/resource operation and propagate across all server capabilities.

it should do something like.

uv run -m mcp_agent.cli go --stdio "uv run progress_server.py" --model=haiku

Type '/' for commands, '@' to switch agent. Ctrl+T multiline, CTRL+E external editor.

Agent agent: 1 MCP Server, 2 tools, 2 resources available

agent > ***CALL_TOOL progress_task {"steps": 30}

▎▶ agent ────────────────────────────────────────────────[claude-3-5-haiku-latest turn 1]

***CALL_TOOL progress_task {"steps": 30}                                                 

─────────────────────────────────────────────────────────────────────────────────────────


▎◀ agent ────────────────────────────────────────────────────────────────────────────────

I'll help you run the progress_task with 30 steps. I'll invoke the tool as you specified:

─| progress_server |─────────────────────────────────────────────────────────────────────


▎◀ agent ──────────────────────────────────[tool request - progress_server-progress_task]

{'steps': 30}

─| progress_ta… | ping |─────────────────────────────────────────────────────────────────


▎▶ agent ───────────────────────────────────────────────────────────[tool result - ERROR]

⏹️ Operation cancelled: call_tool 'progress_task' on server 'progress_server'

─────────────────────────────────────────────────────────────────────────────────────────


▎◀ agent ────────────────────────────────────────────────────────────────────────────────

I apologize, but it seems there was an error executing the progress_task. The operation  
was cancelled. This could be due to various reasons such as:                             

 1 Temporary server unavailability                                                       
 2 Configuration issues                                                                  
 3 Unexpected interruption                                                               

Would you like me to try running the task again or help you with something else?         

─| progress_server |─────────────────────────────────────────────────────────────────────


Last turn: 573 Input, 75 Output, 1 tool calls (0.3%)

agent >

please see https://modelcontextprotocol.io/specification/2025-06-18/basic/utilities/cancellation#implementation-notes

if you want i can propose the watchdog in an a fresh issue.

[evalstate]

There's nothing explicitly supporting "ESC" or "CTRL+C" to cancel either inference or tool calling at the moment (but would absolutely love to have this - and would love help too).

I think the best use of time is going to be to start on the 0.3.0 branch; if you look at examples/new-api/simple_llm and tool_agent_sync.py.

(although the folder is called new-api that's not really the case, it's more a rearrangement of existing code). at the moment, i'm half-way through migrating the message display code for user/assistant, and then moving the stop reason handling in to the llm_agent. that should mean we have far skinnier llm classes, and make it easier to do implement these type of features without too much repetition. what are your thoughts on designing a cancel intercept/handler? (by the way, i'm out at a meetup this evening so won't be updating these classes in the immediate future).

[usamaJ17]

Hi @evalstate , I am also facing this issue,

/fast-agent/.venv/lib/python3.11/site-packages/httpx/_transports/default.py", line 118, in map_httpcore_exceptions
    raise mapped_exc(message) from exc
httpx.RemoteProtocolError: peer closed connection without sending complete message body (incomplete chunked read)