tqdm progress bar BrokenPipe silently discards completed generations in background/MCP mode

[Chaos2-lgtm]

BrokenPipe — tqdm progress bar kills valid generations in background + MCP/API mode

Suggested issue for: Comfy-Org/ComfyUI-Manager
(bug lives in prestartup_script.py of the Manager custom node, not ComfyUI core)

Summary

When ComfyUI is launched in background / non-interactive mode (which is how the
official MCP + API / agent-driven usage is meant to run) and you submit a prompt that
contains a KSampler, the generation actually completes, but the tqdm progress-bar
renderer writes to a severed stderr pipe during sampling, raises BrokenPipeError, and
that exception propagates up the execution chain — flagging the whole prompt as
status_str: "error". The finished image is never saved. A fully valid generation is
silently discarded because the display layer crashed.

Environment

• ComfyUI: 0.34.0 (WSL2 / Linux)
• ComfyUI-Manager: latest (d47c9346 "update DB") — bug present at HEAD
• Launch mode: detached background
  python main.py --listen 127.0.0.1 --port 8188 (stderr redirected to a closed pipe)
• Client: REST /prompt (same occurs via comfy MCP; both go through the same API)

Reproduce

1. Start ComfyUI in background so stderr has no live terminal:

   nohup .venv/bin/python main.py --listen 127.0.0.1 --port 8188 > comfyui.log 2>&1 &

2. POST any prompt graph containing a KSampler to http://127.0.0.1:8188/prompt.
3. Poll GET /history/<prompt_id>. Observe:

   status_str: "error"
   exception_type: BrokenPipeError
   exception_message: "[Errno 32] Broken pipe"
   executed: ["70", "86", "22", "45", "17", "1", "42"]   # KSampler ran, all deps ran
   

   and no output image in /history/<id>/outputs.
4. Note the server log actually reports the sampling finished:
   Prompt executed in 5.83 seconds

Root cause

The traceback funnels through:

k_diffusion/sampling.py  res_multistep()
  → tqdm progress bar (trange)
    → tqdm/std.py display() → print_status() → fp.write()
      → custom_nodes/ComfyUI-Manager/prestartup_script.py
          write_stderr(message)
          original_stderr.flush()      # ← BrokenPipeError here

ComfyUI-Manager's prestartup_script.py installs a stream wrapper that re-emits tqdm
progress to original_stderr. In background/detached mode the stderr pipe's read end has
already been closed, so each progress-bar refresh raises BrokenPipeError. Because this
happens inside the sampler loop, the exception is not contained — it bubbles all the way
up and marks the entire job failed, even though sampling succeeded.

Why this matters (official direction)

Background + MCP/API is the officially promoted way to drive ComfyUI (agents, automation,
headless servers). In that mode stderr is commonly redirected, so every KSampler job is
at risk of this. It is not an exotic one-off.

Suggested fix (minimal, display-only)

# custom_nodes/ComfyUI-Manager/prestartup_script.py  (tqdm branch)
if '100%' in message:
    self.sync_write(message)
else:
    try:
        write_stderr(message)
        original_stderr.flush()
    except (BrokenPipeError, OSError):
        pass   # progress-bar output is cosmetic; never fail the job because of it

• Touches only the display path. Does not alter sampling, output, or error handling.
• The progress bar is cosmetic; a broken pipe on it should never abort a completed generation.

Workaround (no code change)

Launch with stderr pointed at a real file/terminal, or ignore the error server-side —
but the honest fix is the two-line guard above, so completed generations are not dropped.

---

Filed from a background-launched, MCP-driven ComfyUI instance; patch verified locally
(ad-hoc py_compile + severed-pipe behavioural probe pass; generation returns success and
saves PNG after applying the guard).

以上內容是由Agent代為總結。