I think maybe we should disallow concurrent runs

@fresh-machine-37269 check if we have issues open on the automation repo about all of these comments and if not examine the code base together Fuller context about a possible solution and open an issue

🤖 On it! View progress here: oss-agent-canvas.ngrok.dev/conversations/…

But actually Robert I think we don't want to disable concurrent runs because that would make it impossible to have things like good PR reviewer?

ooh good point

✅ Done! Checked the OpenHands/automation repo — here's where each comment lands: 1. "Disallow concurrent runs" — already tracked: ◦ #218 "Add a minimal per-automation overlap guard (skip when a run is already in-flight)" (open) ◦ #72 "Implement hierarchical concurrency limits for automation runs" (open) 1. "Would be nice to have a cancel button" — already tracked: ◦ #154 "Add API and UI support to cancel automation runs" (closed — the API endpoint

POST /runs/{run_id}/cancel

exists; UI may already be covered by that work) 1. "Crashes report as RUNNING, not FAILED — import blew up before fire_callback, seven runs hung open, only one recorded failure" — not covered, so I opened a new issue. I traced the code: in

presets/prompt/sdk_main.py

and

presets/plugin/sdk_main.py

, the completion callback is only sent by the workspace context manager's

__exit__

. Everything before the

with workspace_ctx as workspace:

block — the

from openhands.sdk import ...

at line 132/128 and the

sys.exit(1)

calls for missing env vars / invalid

WORKSPACE_BASE

— exits with no callback. The run then sits

RUNNING

until the watchdog times it out (~10min + margin) before marking it FAILED. New issue opened: #406 — "Preset entrypoint crashes before the workspace context never fire the completion callback — runs hang in RUNNING instead of FAILED" github.com/OpenHands/automation/issues/406 It proposes: (a) POST a FAILED callback on early exits (mirroring

__exit__

), (b) defer the SDK imports into the workspace context, and (c) an optional service-side short probe to detect finished-but-uncall'd commands sooner. Related issues #203 (hard SIGTERM kills) and #337 (ephemeral bash_events) are cross-referenced. This reply was generated by an AI agent (OpenHands) on behalf of the user.