Cursor
meshpoint connect cursorwrites entries into ~/.cursor/hooks.json (file format {"version":1,"hooks":{...}}) for sessionStart, beforeSubmitPrompt, afterAgentResponse, stop, preCompact and sessionEnd, each running meshpoint tap cursor, and the skill at ~/.cursor/skills/meshpoint/SKILL.md. With --project <repo> both land in <repo>/.cursor/. Cursor reloads hooks.json on save. Existing hooks are kept; --remove deletes only Meshpoint’s entries and the skill.
How it works
Section titled “How it works”- Recall and presence arrive as
additional_contextat session start. - Prompts and replies come from the
beforeSubmitPromptandafterAgentResponsepayloads. - A due checkpoint is delivered as a
followup_messagefrom thestophook, which Cursor submits as the next user message. Cursor caps auto follow-ups at five per conversation. preCompactandsessionEndsend the spool.
Cursor must be allowed to run meshpoint items put and meshpoint items retract in its shell approval settings, or the checkpoint command waits for a click.
Status
Section titled “Status”Ported from Cursor’s published hook documentation; not yet run end to end against the platform. Unconfirmed:
- the transcript file format (touched paths and the nudge counter depend on it),
- whether
afterAgentResponsefires once per turn, - whether the agent picks the skill up from the follow-up message on its own.
Cursor also loads .claude/settings.json hooks. A repository connected for both Claude Code and Cursor at project scope would run two taps; connect at user scope.
If you run Cursor with Meshpoint, tell us what you see; a meshpoint doctor output and the spool file after one session are enough to close these.
