Hello everyone, I’ve released the next automation version, v.1.15.1 Thank you very much! thankyou

I hit a an automation skill-routing issue. An existing scheduled run mentioned “automation,” which activated the creation-oriented skill. The agent then tried to inspect the scheduler instead of executing the actual task. And then everything went sideways. I fixed the prompt and that fixed it. It might make sense adding an run-guard to the existing automation skill to prevent this when creating / modifying automations.

I've been testing OAuth MCP connections in Agent Canvas (GitLab, Atlassian Rovo, and Linear), and they're working pretty reliably. One gap: MCP connections don't currently help us create event-driven automations. Most MCP servers don't expose tools for creating webhooks, so the agent needs to make raw API calls to the provider. It would be interesting if the agent could reuse the OAuth access token from the MCP connection for those API calls, enabling an end-to-end flow for creating Event-driven Automations. Potential flow: 1. User clicks on a "Linear Issue to PR" Automation Template 2. User connects a provider via OAuth MCP (e.g. Linear) 3. System stores the OAuth token as a Secret (e.g., LINEAR_ACCESS_TOKEN) 4. FastMCP refreshes the Secret with a new access token as needed 5. Agent uses the Secret to call the Linear API and create the webhook (<-- this is what storing the OAuth Access Token as a Secret solves) 6. OpenHands configures that webhook as the automation's event source Is something like this possible with our current MCP/OAuth architecture?

One piece of feedback would be to allow configuration of pre/post automation hooks E.g pr review via github its not obvious whether the webhook made it to the pending queue; some kind of hook that posts back in acknowledgement would be great

github.com/OpenHands/automation/pull/538 - bump sdk version cc @millions-crayon-32880

github.com/OpenHands/automation/issues/542 - @limited-student-10747 proposal for adding a more principled trace layer to automation service; imo this would be a good start for debugging issues with software factory automations

github.com/OpenHands/software-agent-sdk/pull/5462 github.com/OpenHands/automation/pull/544 adding otel to automation and some extra plumbing which allows us to build a trace and query it from event payload events e.g i can find the entire trace from webhook received to pr review conversation kicked off by querying for the repo + pr number

@quick-window-91705 Added a few screenshots from running this Events tab PR locally: github.com/OpenHands/OpenHands/pull/17873 . Mostly a poc/mvp based on some of the criteria we talked about last week. Events tab shows any registered webhooks, lets you add webhooks, generates a one time secret to copy. Doesn't have delete or update actions to keep it minimal. LMK what you think!

PR with a fix to the git sync when an encryption key changes, or is added: Fix git sync when encryption key changes

cc @limited-student-10747 Re: display LLM profile in the comments from automations Fix: show LLM provenance in automation output Please see below how this looks like, in live tests. I'd love to know if you have concerns about this PR. (Your agent

requested changes

asking for fresh evidence, calling October 5th evidence “historical”. 😅 My agent re-ran them just now, on the 9th, updated the PR description, and the last comment also describes in detail what it did.)

@quick-window-91705 I'm seeing the default python started show when attempting to edit an automation that already has files in it. Did a quick agent investigation here github.com/OpenHands/OpenHands/issues/18341. Before making AC and a PR I wanted to get your eyes on it to see if that looks correct

I’m thinking of a scenario where you could spawn an automation automatically after another automation completes. Sort of like an event-driven automation, but the event is “automation XYZ has completed” Scenario: A “QA” automation starts once an “Issue to PR” automation completes. Perhaps they both use the same sandbox to keep context and state.

Hey @modern-fireman-99755, now that we have the secrets on boarding, what are the next steps to a user with a blank canvas saying "build a software factory for me" and the agent helping them onboard? Maybe just a skill?

Quick note cc: @hundreds-twilight-32833 F28 got a run this morning: all specified behaviors were tested from user POV, by agents driving OpenHands. 🙏 They found one tiny blip found and fixed it. (again pesky contrast on some themes) It should be perfectly fine to just ask your agent to tell you the status or whatever, the behaviors (current, with or without bugs or gotchas) should be in the feature map in the repo.

github.com/OpenHands/automation/pull/580 supporting plugin spec based automation cc @strong-ice-84145 @curved-angle-80449

Flagging #557: prompt-based automations launch without the sub-agent

task

tool, even when sub-agents are enabled in the saved profile. I have demos with automations using subagents and people expect them to work in prompt based workflows. You shouldn't have to write a custom script to use subagents.

Is there a reason why OpenHands logs automations with each run its own directory? I mean, it does this, instead of maybe:

workspaces/automation-abc/run-1234/

, `workspaces/automation-abc/run-3456``

Is this acceptable behavior, WDYT? To let the user rotate the “session api key”:

AUTOMATION_KV_SECRET explicitly set? → use it
otherwise:
  first launch → pin current session key in automation-kv-secret.txt
  later launch → reuse pin, even after SESSION_API_KEY rotates

pin contract:
  regular non-symlink · private permissions · fail closed
  new file → exclusive hardlink; losing starters read the winner
  blank file → shared lock + reread + atomic replacement

@quick-window-91705 New automation UI requires setup script path for custom automations but doesn't appear to use it. c.f. github.com/OpenHands/OpenHands/issues/18652 The related PR removes it but wasn't sure if there was some future state reason to have it there. For the present, its confusing UX. Let me know if we should merge this PR or close it in favor of a different approach...

Hey @calm-artist-7206, you requested changes on github.com/OpenHands/automation/pull/420 and the OSS contributor followed up on it. Do you want to take a second look at it, or can I dismiss your review?

[Automations/Plugins and Agent Profile] 🧵

github.com/OpenHands/OpenHands/issues/18690 github.com/OpenHands/automation/issues/613 github.com/OpenHands/automation/issues/612 github.com/OpenHands/automation/issues/610 github.com/OpenHands/OpenHands/issues/18691 @hundreds-twilight-32833 some follow-on tickets for plugin specs (deprecations, missing UI pieces, bug fixes, etc)

I think this tiny detail makes a rather good experience… it’s convenient, non-intrusive, the app does the smart thing without being told (this is CC)

Hi team I'm looking for a review for automation service, thank you!: github.com/OpenHands/automation/pull/593: Automation runs no longer fail for good when their sandbox doesn’t start. A burst of webhook events, like a Jira bulk transition, can hit slow cold starts. Each run whose sandbox failed was marked FAILED on the first try, and Jira doesn’t redeliver, so that work was lost. Now the dispatcher retries up to 3 times with backoff, deletes the failed sandbox, and records why it failed. On a test install with slow cold starts, a burst of 7 runs went from 1 completed to 7.

Sharing my recent experience setting up webhooks in OpenHands -- I think the Automation UI is a step in the right direction, but we have some usability issues with setting up and creating event sources Automation event source registration feels incomplete

github.com/OpenHands/extensions/pull/853 github.com/OpenHands/OpenHands/pull/19034 Adding event source page to canvas dashboard for automations; extensions PR is a pre-req

I had someone mention the Ask Agent button wasn't working in Automations, I saw this PR: github.com/OpenHands/OpenHands/pull/19093 Anything I can do to help on this?

Looks like picking an agent profile got unsupported for automations in some cases. I was just noodling with some ui/ux and thought I'd broken it. Asked an agent to add agent profile selector back in:

I added the agent profile picker back for script (“Custom Python”) automations, but not for Prompt and Plugin ones. For those, it isn’t just a UI change.

Why Prompt and Plugin are left out: the picker was removed on purpose in OpenHands/OpenHands#18708. The backend endpoints for Prompt and Plugin automations reject agent_profile_id, so picking a profile made saves fail with a 422 (OpenHands/OpenHands#18569). The backend fix is still open (OpenHands/automation#600 and #615). Adding the picker back now would bring that bug back.

@hundreds-twilight-32833 @quick-window-91705 This is what I was speaking the other day. Actually, it is super easy to implement, because every automation has a description where we can write mermaid graphs, and then we can just use agent-canvas to format them in the right way

I am looking forward to the Events UI in Automations. I think these form fields will need serious documentation, and probably example implementation docs for GitHub and GitLab to help users configure this stuff