Agent-Skills.md

Agent Skills: jj Workflow

Use this skill for ANY task involving jj or jujutsu version control. ALWAYS trigger when the user mentions jj, jujutsu, revsets, change IDs, bookmarks, or oplog. Also trigger when the user wants to squash, split, or reorder commits in a stack, write a revset query, absorb fixup changes, undo or restore a previous operation, resolve conflicts after rebasing, recover from force-pushes, rewrite protected/immutable commits, view change evolution (evolog), or try parallel approaches. Trigger even if "jj" is not explicitly said — "changes" instead of "commits", "stack" instead of "branch", "absorb", "squash into the right commit", "undo my last operation", "conflict after rebase", or "compare approaches in parallel" are strong jj signals. This skill contains critical non-obvious rules (like always using -m flags) that prevent broken workflows.

UncategorizedID: trevors/dot-claude/using-jj

Repository

trevors/dot-claude
TrevorS
5

Install this agent skill to your local

pnpm dlx add-skill https://github.com/TrevorS/dot-claude/tree/HEAD/skills/using-jj

Skill Files

Browse the full folder contents for using-jj.

Download Skill

Loading file tree…

skills/using-jj/SKILL.md

Skill Metadata

Name
using-jj
Description
Jujutsu (jj) version control reference. MUST load this skill whenever a query mentions jj, jujutsu, revsets, absorb, evolog, oplog, immutable_heads, divergent changes, or jj-specific concepts — jj differs from git in non-obvious ways and the skill contains critical constraints (e.g., always pass -m, never use split/diffedit) that prevent broken workflows. Skip only for trivial jj commands you're certain about (describe, new, commit, push).

jj Workflow

For daily-command tables, git equivalents, troubleshooting, parallel-experiment patterns, immutable-heads disable/restore commands, recommended config, and the full revset cheatsheet, see REFERENCE.md.

Philosophy

  1. Commits are cheap, descriptions are mandatory. The working copy is always a commit. Never leave it as "(no description set)".
  2. Experiment freely, the oplog is your safety net. jj undo and jj op restore make anything reversible.
  3. Conflicts are state, not emergencies. jj records conflicts in commits as structured data; rebase still succeeds.
  4. Change IDs are your handle on work. Commit hashes change on rewrite; change IDs don't.
  5. Bookmarks exist for GitHub, not for you. Work with anonymous changes; add bookmarks only when pushing.
  6. Keep the stack shallow. Squash early.
  7. Use absorb over manual squash routing. Let jj distribute hunks to the right ancestor.
  8. Colocated = invisible to the team. Teammates see standard git.

CRITICAL: AI-specific rules

Always pass -m — unset opens an editor and blocks the agent:

jj new -m "msg"
jj describe -m "msg"
jj commit -m "msg"
jj squash -m "msg"

These open an editor and hang the agent (jj 0.39). Two guards enforce this — hooks/jj_interactive_guard.sh blocks them pre-run, and $JJ_EDITOR (hooks/jj-reject-editor.sh) fail-fasts any editor jj still opens — but get them right the first time:

  • jj describe / jj commit with no -m → add -m "msg"
  • jj squash with no message → add -m "msg", or -u to reuse the destination's
  • jj commit / jj squash with -i / --interactive / --tool → opens the hunk picker even with -m; drop the flag, edit files, then jj squash -m
  • jj split, jj diffedit → no non-interactive mode; restructure with jj squash -m / jj new -m
  • jj resolve → edit the conflict markers in the files, then jj squash -m; or pass --tool
  • jj config edit → use jj config set <name> <value>

Renamed: untrack a file with jj file untrack <path> — jj forget is gone.

Core concepts

  • Working copy = commit. Every file edit is tracked in @. No staging area, no git add.
  • @ = current change, @- = parent, @-- = grandparent.
  • Change IDs (e.g. kpqxywon) are stable across rewrites. Use these, not commit hashes.
  • Conflicts are state, not emergencies — jj records them in commits and rebase still succeeds.
  • Previous versions: <change-id>/0 (latest), /1 (previous). jj restore --from xyz/1 --to xyz reverts to a prior state.

Workflows

Squash (recommended)

jj describe -m "feat: what I'm building"
jj new -m "wip"
# ... make changes ...
jj squash -m "feat: done"

Commit (simpler)

jj commit -m "feat: what I did"   # = describe + new

Edit (mid-stack fix)

jj edit <change-id>
# ... fix ...
jj new -m "back to work"   # descendants auto-rebased

Absorb

From @, jj absorb routes each hunk to the ancestor where those lines were last modified. Use instead of manual squash routing when fixing across a stack.

Bookmarks & pushing

Bookmarks don't auto-advance — move them explicitly. @ is typically empty; target @-.

jj bookmark set <name> -r @-
jj git push

Don't rewrite reviewed PR history

If a PR has review comments, do NOT squash or rewrite the original commits — review threads detach from line anchors. Add new commits on top instead, and only rewrite after Teej confirms. See rules/version-control.md.

Default (safe — preserves comment anchors):

jj new feature-x- -m "address feedback"
# ... make changes ...
jj describe -m "fixup: address feedback"
jj bookmark set feature-x -r @-
jj git push

Creating PRs (jj + gh)

In jj-colocated repos, git HEAD is detached — gh pr create fails with "not on any branch". Always pass --head <bookmark> — never rely on git auto-detection:

jj bookmark set feature-x -r @-
jj git push
gh pr create --head feature-x --title "..." --body "..."

Recovery

jj op log
jj undo
jj op restore <id>
jj evolog [-r <change-id>]

Immutable commits

Pushed commits are protected by immutable_heads(). Always ask Teej before disabling protection — rewriting remote bookmarks means force-pushing shared history. See REFERENCE.md for the disable/restore commands.

Revsets

jj log -r 'trunk()..@'              # everything between main and here
jj log -r '::@ & ~::trunk()'         # my branch only
jj log -r 'author("trevor")'         # my commits

Full cheatsheet in REFERENCE.md.