[Bug]: Agent tends to use browser instead of GitHub API to open pull requests

[zchn]

Is there an existing issue for the same bug?

• I have checked the existing issues.

Describe the bug and reproduction steps

Despite having the github microagent in https://github.com/All-Hands-AI/OpenHands/blob/main/microagents/knowledge/github.md, I frequently run into cases where the agent tries to use the browser to open a new PR (and fails) instead of the preferred way of using GitHub API. Here is the most recent example:

https://www.all-hands.dev/share?share_id=883b5c047dcbb2df36e1b1c65f7676f25c282bf28a1bc6c896e5f69b2cfc1f56 where

1. I did mention GitHub in my prompt (screenshot 1),
2. but openhands still tries the browser first (screenshot 2),
3. and I have to tell openhands to use the GitHub API after it fails to use the browser to submit a PR (screenshot 3).

OpenHands Installation

Development workflow

OpenHands Version

main

Operating System

MacOS

Logs, Errors, Screenshots, and Additional Context

https://www.all-hands.dev/share?share_id=883b5c047dcbb2df36e1b1c65f7676f25c282bf28a1bc6c896e5f69b2cfc1f56

[enyst]

Thank you for the report! Was this LLM Claude Sonnet?

[zchn]

Yes Claude sonnet 3.5

[zchn]

(realized I did not upload the screenshots properly so just added them to the original post)

[enyst]

I wonder, the id of the action was 95, that might be relatively much for Sonnet to remember every detail from the beginning? Idk, in any case appreciate the report.

[mamoodi]

There were two other issues related to this bug I'm having a hard time finding them. It was also not using the github API when it should be.

[enyst]

There's a note here, but that was not really the same, because it was with the resolver: some initial prompts are different. The github microagent in particular is not enabled with the resolver, and instead it has prompting of its own.

[mamoodi]

Found it. Here is it:
#5634

[h55nick]

Can this be resolved by expanding the "triggers" in the microagent and then cleaning up the github prompt? This thought process doesn't seem baked into the system itself but more a prompting update.