when working with codex via ACP, I notice several weird things: • the agent writes files using python scripts, rather than using built-in tools for file editing (expected?) • we show very long commands rather than high-level summaries • all of the agent's commentary about its work comes in in a rush after its done working, instead of being interleaved with commands

the only things we can control is point 2, and should be just an UI change. But I think the problem here is how we are visualizing the Terminal Tool in the UI. for the other two points. This is on the model. OpenAI models are strongly trained on using the terminal (as I also noticed), and the fact the model gives commands and the a comment at the end is also because of the training.

🤔 there's an API parameter we may want to play with, on responses/ , verbosity; and GPTs are very prompt sensitive still, I think it’d listen if we tell it how to behave with the user

the agent writes files using python scripts, rather than using built-in tools for file editing

I see this on native Codex CLI too

the fact the model gives commands and the a comment at the end is also because of the training

I'm not sure this is actually the problem. What I'm seeing is a trajectory like • agent looks at a file • agent runs command

foo

• agent says "let me look at this file" • agent says "let me try running `foo`" • agent says "I ran

foo

and it worked" Based on the way the messages are phrased they should be appearing before the commands

@bland-bear-604 let's repro this out of order issue: note that it’s canvas UI, using codex cli via ACP, so an agent profile on codex cli. Keep codex on sub, so reusing our codex cli auth. Use control-openhands and verify-openhands as you know. Also test everything ACP (local), fix the map

Reproduced in Canvas with Codex via ACP 🐾 The cause is SDK-side: the ACP bridge buffers all of the agent's text for the whole turn and emits it once, as the turn's final message, while tool calls stream live. So narration like "let me run foo" always lands after every command. Fix: software-agent-sdk#6452 emits the text said before each tool call right before that call. Verified live (also stays put after reload, which fixes #15606). Still draft.

reviewed and merged! Thanks!

Thank you! Please do tell if what

smolpaws

is doing, at any time, is annoying slop or something 😅 It has full control of OpenHands and the local machine, and it tests live as a user, so personally, I find it easier to have confidence…

ACP feature map OpenHands#19314 is bot-approved, CI green: running ACP conversations are now mapped with live Codex recipes, incl. this ordering bug. 🐾