youmd 0.8.16 → 0.8.18

package/skills/remote-machine.md

@@ -0,0 +1,69 @@
+---
+name: remote-machine
+version: 1.0.0
+scope: shared
+identity_fields: [profile.about, preferences.agent]
+description: Check a synced machine's agent/work status and, if needed, trigger it to commit and push — so the user can resume on another computer.
+---
+
+# remote-machine
+
+Use this skill when the user asks about work happening on **another one of their
+machines** and wants to check on it or hand it off — e.g.:
+
+- "Is the work on my Mac mini done? Is it committed and pushed?"
+- "What was the last update on my office machine?"
+- "Have my Mac mini commit and push the youmd work so I can pull it down here."
+- "Check <machine> and if it's not pushed, push it."
+
+This runs entirely through the you.md synced connection (the agent bus relayed via
+Convex). The user's machines are not directly reachable; commands and results flow
+machine → Convex → machine. **Zero user steps beyond the sentence** — never ask the
+user to run anything by hand.
+
+## Identity Context
+
+- **About:** {{profile.about}}
+- **Agent preferences:** {{preferences.agent}}
+
+## Workflow
+
+1. **Identify the target machine.** If the user names it ("Mac mini", "office"),
+   match it against their machines. If unsure, call `remote_machine_status` with no
+   machine to list them, or `you remote list`, and pick/confirm the obvious one.
+
+2. **Check status (read-only first).** Call the MCP tool `remote_machine_status`
+   (or `you remote status <machine>`). Report concisely:
+   - last readiness proof (ready / warn / failed) + when
+   - git state if available: branch, dirty?, ahead/behind, last commit
+   - last agent-bus activity (what the remote agent was doing)
+
+3. **Decide if action is needed.** If the work is **clean and pushed** (not dirty,
+   ahead = 0), tell the user it's already safe to pull — done. If it's **dirty or
+   ahead of upstream** (uncommitted or unpushed commits), proceed to step 4.
+
+4. **Trigger commit + push (only if needed).** Call `remote_machine_run` with
+   `action: "git.commit_push"` and `args: { project: "<name>", message: "<short wip msg>" }`
+   (or `you remote run <machine> git.commit_push --project <name> --message "..."`).
+   This requires the opt-in `remote:command` scope on the key. The remote daemon
+   executes a **whitelisted** git action only (never arbitrary commands).
+
+5. **Confirm + report.** Read the returned result (`ok`, `gitState`, `output`).
+   Tell the user exactly what happened and that they can now `git pull` here to
+   resume. If it failed (e.g. auth, conflicts), report the redacted error and the
+   safe next step — do not retry destructively.
+
+## Allowed remote actions (the security boundary)
+
+Only these may ever be triggered remotely: `git.status`, `git.last_activity`,
+`git.commit_push`, `git.pull`, `agent.status`. Everything else is rejected by the
+daemon. There is **no** arbitrary shell, force-push, reset, or file write. If the
+user wants something outside this list, explain it's not a permitted remote action
+and offer the closest safe option.
+
+## Notes
+
+- Prefer the read-only status call before any mutating one.
+- All dispatches + results are audited (brain activity) and owner-scoped — a
+  command can only target the user's own machines.
+- See `project-context/CROSS-MACHINE-AGENTS.md` for the full architecture.