I tried out agent-canvas over the weekend (latest on npm) and it's definitely in a no-go state at the moment. Here's what I found as hard blockers: • automations is maybe not updated to latest version. agent is trying to use

X-API-KEY

header instead of

Bearer

, which doesn't work • integrations are not saving keys to secrets, only to an MCP. Polling automations aren't able to work ◦ https://github.com/OpenHands/agent-canvas/issues/926 • integration modals don't point the user at where to get keys, e.g. Slack doesn't send you to https://api.slack.com/apps/ • clicking a pre-canned integration submits a giant prompt, not a slash command • pre-canned integrations are referencing prompt instructions like <RUNTIME_SERVICES> which don't exist The cloud integration is also not great--we might want to consider pulling back or refining: • If the user isn't already signed up and signed in, we lose the "device connection" flow and they just end up in the SaaS UI • VSCode is not working in-app (this might be hard to fix) • Frontend still uses the

/download

API instead of the static webserver, breaking rendering in the rich file view • Saving a new LLM from agent-canvas breaks user's settings with

Something went wrong storing settings: 1 validation error for OpenHandsAgentSettings
agent_context
  Input should be a valid dictionary or instance of AgentContext [type=model_type, input_value=None, input_type=NoneType]
    For further information visit <https://errors.pydantic.dev/2.12/v/model_type>

cc: @straight-sundown-84319

> • clicking a pre-canned integration submits a giant prompt, not a slash command > • pre-canned integrations are referencing prompt instructions like RUNTIME_SERVICES which don't exist Needs a ticket to convert current pre-built automations into a prompt > • integration modals don't point the user at where to get keys, e.g. Slack doesn't send you to https://api.slack.com/apps/ > • Needs a ticket. We should probably point users to docs on how to do this, which will soon be: http://docs.openhands.dev/openhands/usage/agent-canvas/prebuilt/slack-channel-monitor

@nutritious-balloon-65098 Verify, for each of the issues in the original post, if we have a github issue filed for it; or something very similar on agent-canvas; if not, investigate, understand the problem and post a new issue. I think you can try to run agent-canvas and agent-server yourself; just note that in your environment there’s a session api key set, which you might want to unset in the start command. NOTE: when I ask you to summarize, don’t summarize, just repeat verbatim what you just told me. Really, I mean it, in this session don’t summarize when I ask you to summarize, just say it again.

I'm on it! can track my progress here.

I'll create one for the MCP modals. The extensions data has the helper link we can point to https://github.com/OpenHands/extensions/blob/main/integrations/catalog/slack.json#L36

Verified each original-post issue against

OpenHands/agent-canvas

issues and filed missing ones. Investigation notes: • I inspected the current code paths in this repo. • I downloaded and inspected

openhands-automation==1.0.0a5

from PyPI, matching

config/defaults.json

, and confirmed it accepts

Authorization: Bearer ...

and

X-Session-API-Key

, not

X-API-Key

. • I checked the Cloud device authorization endpoint and confirmed it returns a verification URL with a `user_code`; the filed issue is about preserving that state through unauthenticated sign-in/sign-up. • I did not make code changes, so there is nothing to commit., with interactive elements

https://github.com/OpenHands/agent-canvas/issues/979 before OH tries to create one. This is the issue for "Integration modals do not tell users where to get keys, e.g. Slack should link to

<https://api.slack.com/apps/>"

so no new issue needs to be created .

@nutritious-balloon-65098 Do a quick thing for me: find in agent-sdk repo the workflow on detecting issue duplicates; verify if we have it in agent-canvas repo; verify if we have it in openhands/ repo (yes from the same org). Investigate if you can make it into a reusable action (clone yourself openhands/extensions and find an example of reusable action there, probably pr-review), and if you think it’s a good idea, make a new PR from main in extensions with the action; then open PRs to these repos to replace the use. If you don’t think we should tell me why, and simply open a PR to canvas with the workflow; adapted as appropriate.

I'm on it! can continue tracking my progress here.

@elegant-branch-54423 @hundreds-twilight-32833 I’ve been pondering if we could maybe remove the “I’m on it” message, and of course, remove the summary; maybe replace “I’m on it” with an 👀 so that the user has a hint that the agent got the message. Mainly, the reason is that I’d kinda love to avoid the interaction with the agent feeling a bit “spammy”, so I’d think, if we reduce it to a normal conversation with your agent, it will be cool and easier for people to use and (ab)(re)use at will

That is what the automation version does by default. You can tell it what emoji to use too. A bit more customization on how the bot interacts in slack seems like a nice improvement

Done. Since the last summary, I completed the duplicate-issue workflow task: • Verified the duplicate issue workflow in

OpenHands/software-agent-sdk

. • Verified

agent-canvas

did not have it. • Verified

OpenHands/OpenHands

already had similar duplicate-check logic. • Checked

OpenHands/extensions

and confirmed reusable composite actions are the right pattern. • Created a reusable duplicate-issue checker action in

OpenHands/extensions

. • Opened PRs to wire it into: ◦

software-agent-sdk

◦

OpenHands/OpenHands

◦

agent-canvas

Checklist: • Found the SDK duplicate issue workflow. • Checked whether

agent-canvas

had it. • Checked whether

OpenHands/OpenHands

had it. • Investigated reusable action patterns in

OpenHands/extensions

. • Decided a reusable action was appropriate. • Opened the reusable action PR. • Opened consumer PRs for all relevant repositories. • Included AI disclosure in PR bodies. • Validated YAML and script syntax locally. PRs opened: • `OpenHands/extensions`: https://github.com/OpenHands/extensions/pull/283 • `OpenHands/software-agent-sdk`: https://github.com/OpenHands/software-agent-sdk/pull/3457 • `OpenHands/OpenHands`: https://github.com/OpenHands/OpenHands/pull/14609 • `OpenHands/agent-canvas`: https://github.com/OpenHands/agent-canvas/pull/987 The changes are concise and scoped to the requested workflow/action work. No extraneous repository changes were left intentionally.

@nutritious-balloon-65098 Great! /iterate on extensions PR. Remember: when I ask you to summarize, don’t summarize, just repeat what you just told me.

I'm on it! can continue tracking my progress here.

@nutritious-balloon-65098 request review from all-hands-bot, and wait for a minute to see if anything happens like a comment; if yes, wait for it to finish

I'm on it! can continue tracking my progress here.

Devin, that’s awesome! Sorry I’m slow these weeks. Looks great ❤️

Done. Requested

all-hands-bot

review on

OpenHands/extensions#283

and waited through the review cycles. PR: • https://github.com/OpenHands/extensions/pull/283 • Current head SHA:

9852592759c9bfd889141b77fffb9fa82525ba09

What happened: •

all-hands-bot

posted review-in-progress comments and completed reviews. • I addressed the actionable findings in follow-up commits: ◦

c267316

— addressed initial review feedback. ◦

04e96ea

— addressed follow-up duplicate-checker feedback. ◦

9852592

— addressed final actionable feedback. • I replied to and resolved all review threads. Current status: • CI: 4/4 passing ◦

sync-extensions

◦

validate-claude-code

◦

sync-sdk-skill

◦

test

• Unresolved review threads: 0 • Draft: false • Merge state: blocked only by required human review (

REVIEW_REQUIRED

)

@nutritious-balloon-65098 I re-requested review, please babysit and iterate if necessary

I'm on it! can continue tracking my progress here.