| name | workon |
| description | Safely start, resume, inspect, and finish isolated repo work when the operator says "work on <repo>", "start working on", "pick up <task>", "parallel session", "is anyone else working on", or "hand this off / I'm done with this task". |
/workon
Keep simultaneous agent sessions from stepping on each other. Every task gets its own private copy of the repo, based on the shared version, and finished work is handed off for review.
First use
On the first use in a conversation, briefly introduce the skill:
I can keep this repo work isolated from other sessions. I will put this task in its own private copy, leave the main copy alone, and hand it off for review when it is ready.
Engine path (resolve once, before anything else)
The engine lives in the vault, but most of this skill's work happens inside a private copy of some OTHER repo — where a relative config/scripts/... path does not exist. Resolve the absolute path ONCE from the vault root (the session's starting directory) and reuse it in every mode:
ENGINE="$(pwd)/config/scripts/repo-session.sh"
Every command below uses "$ENGINE", never a relative path.
Modes
/workon <repo> <task>
Also applies to conversational equivalents like "work on WorkDesk OS multi-session" or "pick up the transcript bug".
-
Run:
"$ENGINE" start <repo> <task>
-
Read the final WORKTREE=<path> line from the output.
-
Change into that path before reading or editing repo files.
-
Confirm in plain language:
I am in your private copy for <task>. The main copy is untouched.
-
Continue with the operator's actual task inside that private copy.
If the script warns about another session's unsaved or unshared work, relay the warning in plain words. Do not try to fix the other session's state.
/workon status
Run:
"$ENGINE" status
Relay the table conversationally. Say "private copy" for each workspace and "in use by another session" for held resources.
/workon done
Also applies when the operator says they are finished, asks to hand this off, or asks to wrap the task.
-
Check whether there are uncommitted changes.
-
If there are changes, review them, run the relevant tests, and commit with a concise, sensible message.
-
Run:
"$ENGINE" finish <task>
-
Tell the operator the work was handed off for review. If the script prints a review URL, include it.
/workon abandon
Also applies when the operator says to drop, scrap, or forget a task — they don't want it handed off.
-
Confirm with the operator: anything not yet committed in that private copy will be discarded.
-
Run:
"$ENGINE" cleanup --days 0
This removes idle private copies immediately. It refuses to remove a copy holding unsaved work — if the operator truly wants that work gone too, discard the changes inside the copy first, then re-run.
-
Tell the operator the private copy is gone and nothing was handed off. (Quietly: any branch already created stays local and harmless; the weekly sweep and git worktree prune keep things tidy.)
Operator voice
Use zero Git vocabulary with the operator:
| Technical term | Say this instead |
|---|
| worktree | private copy |
| branch | do not mention it |
| pull request / PR | hand-off for review / review request |
| lock | in use by another session |
| remote / origin | shared version |
What NOT to do
- Never work directly in
~/code/<repo> when a multi-session scenario is possible. Always run "$ENGINE" start first.
- Never commit to
main or master in any shared repo. The main copy only moves when finished work is reviewed and merged.
- Never base work on the shared checkout's local state. The script guarantees work starts from the shared version; do not bypass it.
- If the script warns about another session's unshared work, relay the warning in plain words and leave that session alone.
- Do not copy files from the main checkout to "pick up" someone else's local progress unless the operator explicitly asks for that exact recovery.
Source
config/scripts/repo-session.sh owns the engine behavior.config/rules/multi-session-repos.md owns the rule.- Incidents from 2026-06-10: one review carried another session's local work, and concurrent release checks corrupted the canary workspace.
