agent's papermill run hung on a cell the agent never interrupted, agent stuck waiting

TL;DR

An agent that starts a papermill run and waits forever is missing two things: a per-run timeout and the ability to interrupt the kernel. Set papermill's execution timeout so hangs become named failures, and give the agent a kernel-interrupt action it can use when progress stalls. The agent's run loop should watch cell progress (execution counts advancing) and treat a stalled count as a hang, not as patience.

agent's papermill run hung on a cell the agent never interrupted, agent stuck waiting

Use this when

• An agent's papermill invocation hangs and the agent waits indefinitely.
• The agent has no interrupt capability in its toolset.
• Runs hang on cells with loops, sleeps, or blocking I/O.

Not for this skill when

• A human's papermill hangs. Set the timeout and interrupt it yourself.
• The run fails fast. That is a bug, not a hang.

Steps

1. Set an execution timeout on every agent-driven papermill run. Verify: a hung run now fails with the cell named.
2. Give the agent a kernel interrupt tool (interrupt_kernel via the client). Verify: the agent can interrupt a test hang.
3. Add progress watching: the agent checks execution counts or log output periodically. Verify: a stalled count is detected within minutes.
4. Define the hang policy: interrupt once, then fail the run with the cell source attached. Verify: the policy triggers on a synthetic hang.
5. Fix the underlying hanging cell (loop guard, timeout, event wait). Verify: the run completes without needing the policy.

Variant phrasings

agent papermill hang no interrupt

Short form. Timeout plus interrupt tool.

papermill run stuck agent waiting

The symptom phrasing. Watch progress, not just completion.

agent cannot interrupt hung cell

The capability phrasing. Add the interrupt action.

Compatibility: papermill 2.x, jupyter_client kernel management. The pattern fits any agent framework.

Why it happens

Papermill's default is to wait forever, and an agent without an interrupt tool inherits that default with no escape hatch. Humans interrupt hung cells reflexively; agents need the capability explicitly built and the policy for when to use it. Without both, the agent's run loop is a fancy way to wait forever.

Edge cases / pitfalls

• An interrupt that the agent never uses is decoration. The progress watcher is what triggers it.
• Interrupting and retrying the same hanging cell loops forever. Interrupt once, then fail and report.
• The timeout must be generous enough for legitimately slow cells or the agent will interrupt real work.
• Log the cell source on hang. The human fixing it needs to see the code, not just the cell number.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstuEM3a2AjFoYZUWvdYdVyA