githubnext/ace2Public

Ace: a collaborative coding environment on the team's own machines

Add an optional pane that explains what the agent built and why

IA

Add an optional split pane that explains, in plain language, what the agent has built or done in the channel. The pane's purpose is to teach the participant: unfamiliar concepts, the decisions the agent made, and how the pieces fit together. Summarizing commands is not enough.

For example, building native computer use for Ace involved many mechanisms that Nate did not necessarily already know, such as Accessibility elements, snapshots, window targeting, background event delivery, and clipboard paste transactions. After that work, the pane should explain what changed, why, and how the mechanisms work together.

## Acceptance

- The pane can be opened beside a chat, and it can be closed and turned off. Its on/off state persists, and turning it off leaves the chat unchanged. - It explains what was built or changed, why, and how the parts interact, at the level of concepts and decisions rather than individual commands. - It introduces unfamiliar concepts when they first matter and links them to the concrete changes they relate to, such as files, diffs, commits, PRs, and issues. - It uses rich context where it helps: diagrams, short code excerpts, screenshots or renders, and links to docs. - It stays tied to the channel's actual work and its live account (#99). The pane must not claim anything that the run history does not support. - Dogfood: after a real multi-step change in an unfamiliar area, someone new to that area can explain from the pane alone what changed and why.

## Open questions

- When is the explanation produced: continuously during a run, at the end of each run, or on request? What does that cost in model time? - Should it follow one chat, a lane, or the whole channel (including subagent chats)? - Should it be a split beside the Chat tab, or a separate tab? - Can it be shared with other participants or exported into PR descriptions or docs?

Depends on #99, which supplies the plain-language account of outcomes this pane builds on. Personalization of explanations is tracked separately in #101.

Part of #98. Product feedback from Nate.

  1. iamnbutlerreferenced this in [Meta] Help participants follow and understand agent work #98last month on GitHub

  2. iamnbutlerreferenced this in Design: tailor explanations to concepts a participant marks as known #101last month on GitHub

  3. iamnbutlerparent-issue-addedlast month on GitHub