You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
feat(goal): remember the prompt template and hand unmet criteria back
Send with Goal and Generate Prompt keep the last Prompt template. The
default judge prompt now classifies instead of writing the follow-up.
Complete reasons that start with "need human-in-the-loop" or "max
iterations reached" stop the loop with the criterion still unmet.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Copy file name to clipboardExpand all lines: README.md
+1-1Lines changed: 1 addition & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -10,7 +10,7 @@ You shouldn't have to leave your editor to talk to an AI, run git, or open a she
10
10
11
11
NotepadAI speaks the Agent Client Protocol (ACP) over stdio. Claude Code and Codex are built in. From Settings you can add any other ACP-compatible agent, including Gemini, Auggie, or your own command. Agents read and write files, run terminal commands, and see your workspace context, so they work on the same code you do. Start a session from the AI menu or by right-clicking a folder in the file tree.
12
12
13
-
There is also a Goal Agent: pick an ACP judge (Claude Code, Codex, Opencode) or a Custom API endpoint that speaks Anthropic's Messages API. You set success criteria and it drives the working agent until those criteria hold, or attach a goal to an existing conversation. Send with Goal defaults to the agent's own `/goal` command instead: each criterion goes out on its own, with no separate judge, and a `/goal` typed in the composer can be sent while a turn is still running. When a goal is achieved, the session can compact automatically. AI chats group by project, with a session strip to switch conversations in the same dock. The composer can ask the agent to work in a new git worktree. When you commit, AI can write the commit message from your staged diff. That generator (and other LLM calls) can talk to an OpenAI-compatible endpoint or Anthropic's Messages API. Leave the commit box empty (when not amending) and Commit lets you pick an ACP agent to do the commit. Protocol details and how to wire up a custom ACP agent are in [doc/AcpAgents.md](doc/AcpAgents.md).
13
+
There is also a Goal Agent: pick an ACP judge (Claude Code, Codex, Opencode) or a Custom API endpoint that speaks Anthropic's Messages API. You set success criteria and it drives the working agent until those criteria hold, or attach a goal to an existing conversation. A judge that needs a human action, or that hits the iteration cap with the criterion still unmet, stops and hands the criterion back instead of marking the goal done. Send with Goal defaults to the agent's own `/goal` command instead: each criterion goes out on its own, with no separate judge, and a `/goal` typed in the composer can be sent while a turn is still running. When a goal is achieved, the session can compact automatically. AI chats group by project, with a session strip to switch conversations in the same dock. The composer can ask the agent to work in a new git worktree. When you commit, AI can write the commit message from your staged diff. That generator (and other LLM calls) can talk to an OpenAI-compatible endpoint or Anthropic's Messages API. Leave the commit box empty (when not amending) and Commit lets you pick an ACP agent to do the commit. Protocol details and how to wire up a custom ACP agent are in [doc/AcpAgents.md](doc/AcpAgents.md).
0 commit comments