Generate system instructions from chat history - #403
Conversation
🦋 Changeset detectedLatest commit: 037a81a The changes in this PR will be included in the next version bump. This PR includes changesets to release 2 packages
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
b3f5f6a to
5e08c13
Compare
c7fe4cd to
9f22d41
Compare
Users can promote a working conversation into an editable draft instead of writing the agent prompt from scratch, and nothing is saved until they apply or copy it. Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Walk a turn's user messages newest-first so the transcript reverse keeps them in order.
9f22d41 to
0672b2f
Compare
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
The only conflicts were the import lists in sessions.ts and sessionRoutes.ts. Keep generate-instructions next to the new get-or-create-by-external-id route.
Keep the generate-from-chat slots next to the new library/sessions pages, and keep both harness tests.
Keep GenerateInstructionsButton next to SaveAgent, hide Save when agent config is open, and keep the new SaveAgentForm slots. The harness server still exposes generateInstructionsFromChat on the typed AgentChatServer from main.
|
Merged current Conflicts were in the chat chrome and harness adapter. |
Main dropped TENANT_ID and the global modelProviderStore. The handler now resolves tenant, agent, and provider the same way as the other session routes. Transcript clipping reserves user turns first so a long latest assistant reply cannot drop earlier preferences.
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit d887691. Configure here.
| } | ||
| pageToken = next; | ||
| } | ||
| return events; |
There was a problem hiding this comment.
Paging can 422 established chats
Medium Severity
loadInstructionTranscriptEvents stops after four newest pages of 100 events, and the short-chat gate only counts user text from that window. Tool-heavy turns can fill the window with non-user events, so an earlier preference-setting message never loads and a long chat returns 422 as if it were empty.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit d887691. Configure here.
2 participants
Fixes #357
System instructions currently have to be written from scratch, even when the chat already made it obvious how the user wants the agent to behave.
This adds a From chat button on the session header. It drafts instructions from the transcript, you can edit them in a dialog, and nothing is saved until you apply. Inline chats can apply the draft to that session. Named/registry agents only get copy, so we don't mutate the saved agent.
Short chats return 422 before we call the model, so a "hey"/"hi" conversation doesn't turn into a generic helpful-assistant prompt.
New route:
POST /api/v1/sessions/{session_id}/generate-instructionsIt only returns a suggestion (
instructions,current_instructions,sources). Saving still goes through the existing session update.I didn't regenerate the SDK. The UI adapter hits the new path with
client.fetch.Note
Medium Risk
Introduces a session-scoped LLM call over full chat history and a path that updates inline session
agentSpec.instructions; access is creator-only like other session routes, but wrong drafts or apply could change agent behavior for that chat.Overview
Adds From chat in the session header so users can draft system instructions from the conversation, edit them in a modal, and only persist on explicit Apply (mutable inline sessions). Chats bound to a saved registry agent still get the draft and Copy, but not in-place apply, so the named agent is not mutated.
Backend: New
POST /api/v1/sessions/{session_id}/generate-instructionsreturns a suggestion (instructions,current_instructions,sources) and does not save. Transcript is built from session events with short-chat and generic-output guards so trivial threads return 422 before an LLM call. Drafting uses the session’s configured model via a one-shot non-stream completion.UI: Optional
generateInstructionsFromChaton the server (Harness adapter calls the new route withclient.fetch),GenerateInstructionsButtonslot wired into drawer/sidebar/stack layouts, and tests for the button, adapter, and runtime transcript logic.Reviewed by Cursor Bugbot for commit 037a81a. Bugbot is set up for automated code reviews on this repo. Configure here.