Claude Code
meshpoint connect claude-codewrites hook entries into ~/.claude/settings.json and the skill at ~/.claude/skills/meshpoint/SKILL.md. With --project <repo> the same lands in <repo>/.claude/. Start a new session to load them.
Every hook runs meshpoint tap claude-code with Claude Code’s event payload on stdin.
| Hook | What happens |
|---|---|
SessionStart (startup, resume, clear, compact) |
Recall for the repository’s origin remote and branch, teammates fetched in the background (waits up to 1.5 s), both rendered into additionalContext. |
UserPromptSubmit |
The prompt is spooled as a [user] turn. |
Stop |
The reply since the last prompt is spooled as an [assistant] turn with the files it touched through tools. When a checkpoint is due, the hook blocks the stop once and asks the agent to run the meshpoint skill. |
PreCompact |
The spool is sent now. |
SessionEnd |
The spool is sent now. |
The Stop hook never nudges twice in a row: when Claude Code reports stop_hook_active, the stop is already a continuation and goes through.
The skill
Section titled “The skill”The skill is allowed-tools: Bash(meshpoint items put:*) and Bash(meshpoint items retract:*), so the checkpoint command runs without a permission prompt. It asks for the plain argument form, --item kind 'title' 'body', because Claude Code’s command guard asks for approval on heredocs and quoted JSON even when the command is allowlisted.
What you see
Section titled “What you see”At session start the context contains a “Team knowledge” block: items with their id prefixes, the time of the last checkpoint, and who else is live on the branch. After twenty replies Claude Code shows a stop-hook message, “Meshpoint checkpoint due (20 replies since the last one). Run the meshpoint skill now, then continue.”, runs the skill, writes items, and keeps going.
- Claude Code loads project hooks from
<repo>/.claude/settings.jsonand user hooks from~/.claude/settings.json; connect once at one scope. - Cursor also reads
.claude/settings.json. A repository connected for both would run two taps; connect Cursor at user scope instead. - Headless runs (
claude -p) run the hooks too, including the nudge. Pass--session-idto keep one session across prompts.
