jujutsu-colocated（柔术并置）

Jujutsu Colocated Workflow

Use this skill to execute jj-first workflows safely in repositories where git and jj coexist.

Detect VCS First

Run this before any VCS command:

if jj root &>/dev/null; then echo "jj"
elif git rev-parse --show-toplevel &>/dev/null; then echo "git"
else echo "none"
fi

Interpretation:

• jj: continue with this skill (also covers colocated repos).
• git: stop and report that this is a Git-only repository. Ask the user to colocate before continuing with this skill (for example, run jj git init --git-repo. from the repo root, then verify with jj git colocation status).
• none: stop and report no repository detected.

Default to jj help <command> for command syntax to keep context lean. Use Context7 selectively for nuanced or best-practice questions (for example colocation caveats, safety guidance, and version-migration differences).

In colocated repositories, prefer read-only git commands and use jj for mutating operations.

Follow Non-Negotiable Safety Rules

• Treat working copy @ as a real commit, not a staging area.
• Use one workspace per agent for parallel work; never share a workspace path across active agents.
• Refer to work using change IDs when coordinating; commit IDs can change after rewrites.
• Use jj op log as the recovery source of truth.
• Treat bookmarks as publication pointers, not a current-branch model.
• Never run destructive history/state operations without explicit written approval in-thread.
• Never edit .env or environment variable files.
• Never revert or delete other agents' work without coordination.

Workspace and Worktree Naming Convention

Use one canonical token for JJ workspace names and Git worktree directory names:

<repo>-<agent>-<task>

Rules:

• repo: short repository slug (for example billing-api).
• agent: stable agent identifier (for example hephaestus, reviewer-2).
• task: short kebab-case task slug (for example token-refresh).
• Keep names lowercase and ASCII; avoid spaces and punctuation beyond -.
• If the same agent repeats the same task, append a short disambiguator (for example -v2, -2, -20260210).

Example:

billing-api-hephaestus-token-refresh

Cross-tool policy:

• JJ: pass --name <repo>-<agent>-<task> explicitly so workspace identity does not depend on path basename inference.
• Git worktree: pass -b <branch> explicitly; do not rely on git worktree add <path> default branch-from-path behavior.
• Recommended Git branch shape for agent worktrees: wip/<agent>/<task>.

Forbidden without explicit written approval:

• jj op restore...
• jj op abandon...
• jj util gc
• jj bookmark set --allow-backwards...
• jj abandon <shared-change>
• jj bookmark delete <shared-bookmark>
• destructive Git fallbacks (git reset --hard, git checkout --, broad git restore, rollback rm)

Use Conventional Commit Descriptions

In jj, the -m value is the change description (commit message equivalent). Require Conventional Commit format:

<type>(<scope>): <description>

Allowed types:

• feat, fix, docs, style, refactor, perf, test, build, ci, chore, revert

Examples:

• feat(auth): add token refresh flow
• fix(sync): avoid duplicate bookmark tracking
• chore(integration): merge parallel agent changes

Core Workflow

1. Sync first:

jj git fetch --remote origin

1. Create a dedicated workspace for this agent from mainline, with the first scoped change already described:

jj workspace add --name <repo>-<agent>-<task> --revision main@origin \
  -m "<type(scope): description>" ../<repo>-<agent>-<task>
cd ../<repo>-<agent>-<task>
jj workspace list

1. Create additional scoped changes in that workspace only when needed:

jj new -m "<type(scope): description>"

1. Keep one logical concern per change:

jj split -i
jj squash --into <target-change>

1. Inspect and pick publish targets (plain English):

jj status
jj diff
jj log -r "main@origin..@" --reversed --no-graph \
  -T 'change_id.short() ++ " | " ++ description.first_line() ++ "\n"'

1. Publish selected change(s) explicitly (single PR or stack):

# Single PR in this workspace
jj bookmark set pr/<agent>/<task> -r <change-id>
jj git push --bookmark pr/<agent>/<task> --remote origin

# Stacked PRs in this workspace
jj bookmark set pr/<agent>/<task>/01 -r <change-id-1>
jj bookmark set pr/<agent>/<task>/02 -r <change-id-2>
jj git push --bookmark pr/<agent>/<task>/01 --bookmark pr/<agent>/<task>/02 --remote origin

CI guardrails:

• Push only selected bookmarks. Never use broad push patterns for publish.
• Reuse stable bookmark names with jj bookmark set to avoid creating extra CI-triggering branches.

1. Resolve conflicts intentionally (never by deleting peers' edits):

jj resolve --list
jj resolve

1. Retire temporary workspaces after integration:

jj workspace list
jj workspace forget <workspace-name>

Recovery Ladder

1. Use jj undo for immediate local mistakes.
2. Use jj op log and inspect prior states.
3. Use operation-level restore only with explicit approval.

References

• Colocated playbooks: references/colocated-workflows.md
• Git-to-JJ mental model: references/git-to-jj-map.md
• Safety and recovery rules: references/safety-and-recovery.md
• Parallel multi-agent contract: references/parallel-agent-contract.md