Claude Code Plugins for jj (Jujutsu)

Claude Code plugins for jj (Jujutsu) workflows — project setup, workspace isolation, commit management, peer review, and autonomous permission gating.

Who this is for

These plugins are for people who have chosen jj (Jujutsu) and want their coding agents to stay on it. They are opinionated and jj-only by design: the jj plugins install hard walls, not gentle reminders. If you're git-first, or want something neutral between the two, this isn't it.

Why hard walls instead of instructions

LLM agents have a strong reflex toward git. It dominates their training data, so git add / git commit / git status are what they reach for automatically — often mid-task, even when the project explicitly specifies jj, and even when the agent agrees jj is the better choice. A written rule like "use jj, not git" is a suggestion the model can rationalize past a moment later.

So enforcement sits below the level the model can argue with. A PreToolUse hook (block-raw-git.sh) in the jj plugins (project-setup, workspace, commit-commands, peer-review) intercepts every Bash call, blocks raw git, and hands back the jj equivalent — turning a reflexive git commit into a redirect the agent recovers from. The two deliberate exceptions are jj git subcommands (e.g. jj git push) and the gh CLI, the legitimate git-interop seams. Permission-gateway is standalone and works in any repo, jj or git.

All jj output commands (jj log, jj diff, jj bookmark list, jj op log, jj workspace list, jj show, jj evolog, jj op show, jj config list, jj tag list) use JSON templates (-T 'json(self)') by default, giving Claude Code structured, machine-parseable output instead of human-readable text. Requires jj >= 0.31.0.

Plugins

project-setup-jj

Bootstrap jj workflow enforcement for any project with a single command. Sets up a SessionStart hook (shows jj context each session), a PreToolUse guard hook (prompts jj new before editing non-empty changes), permissions (allow jj/gh, deny git), and a CLAUDE.md policy directive.

Setup:

# 1. Install from plugin manager
/plugin install project-setup-jj@muloka-claude-plugins

# 2. Run setup in your jj project
/project-setup

# 3. Restart Claude Code for SessionStart hook

Requires: jj and jq

workspace-jj

Enables Claude Code's --worktree flag and subagent isolation: "worktree" in jj repositories. Claude Code uses git worktrees by default for isolated parallel sessions — this plugin replaces that with jj workspaces via WorktreeCreate and WorktreeRemove hooks, so --worktree works natively in jj repos.

Setup:

# 1. Install from plugin manager
/plugin install workspace-jj@muloka-claude-plugins

# 2. Run setup in your jj project (copies hook scripts, configures settings)
/workspace-setup

# 3. Restart Claude Code, then use worktrees
claude --worktree feature-auth

Claude Code doesn't pick up WorktreeCreate/WorktreeRemove hooks from plugins — they must be in project settings. The /workspace-setup command handles this by copying scripts to .claude/scripts/ and configuring .claude/settings.local.json.

Requires: jj and jq

commit-commands-jj

Streamline your jj commit workflow with simple slash commands.

Commands: /commit, /commit-push-pr, /new, /edit, /describe, /squash, /abandon, /sync, /undo, /finish, /clean_stale, /show, /evolog, /op-show, /tag-list

peer-review-jj

Unified change review for jj repos. Two-phase pipeline (requesting → receiving) with generalist-first architecture and emergent specialists.

Command: /peer-review

/peer-review                          # review current change (@)
/peer-review <revision>               # review specific change
/peer-review --deep errors types      # generalist + specialist dispatch
/peer-review --track                  # enable progress tracking (duplicate+squash)
/peer-review --post                   # post findings to GitHub PR
/peer-review --json                   # raw structured output

Agent: change-reviewer — generalist reviewer that scales with change size (1 per ~300 lines). Returns structured JSON findings with severity tiers and confidence scoring (>= 80 threshold). Recommends specialists for deeper analysis when needed.

Specialist emergence: After 3+ reviews flag distinct patterns for a concern type, the plugin prompts to create a project-specific specialist at .claude/peer-review/specialists/.

Replaces the deprecated code-review-jj, pr-review-toolkit-jj, and feature-dev-jj plugins. See design doc for full details.

permission-gateway

Tiered permission gateway for autonomous subagent workflows. When running multiple subagents in parallel, each making dozens of tool calls, you either pre-approve everything (dangerous) or get 60+ confirmation prompts (kills parallelism). Permission gateway is the middle ground.

Evaluation order: Gate-the-Gate → Deny (immutable floor) → .local.md rules → Confirm → Approve → Tier 2 (LLM eval)

Tool call fires
     │
     ▼
 Gate → Deny → .local.md → Confirm → Approve → Tier 2 LLM
  │      │         │          │          │          │
  ▼      ▼         ▼          ▼          ▼          ▼
PROMPT BLOCK    per rule    PROMPT    SILENT     LLM+PROMPT

Security: One-way ratchet — hardcoded deny is an immutable floor that .local.md cannot override. Writes to permission-gateway config files require human confirmation (gate-the-gate). Dangerous patterns are scanned in the full command string to prevent bypass via find -exec, xargs, or redirect clobbers.

Self-tuning: All decisions logged to .claude/permission-gateway.log. Review the log to promote frequently-confirmed commands to .local.md approve rules.

Commands: /tune — scan decision log and propose .local.md rule promotions

Requires: jq. Tier 2 uses Claude Code's built-in prompt hook evaluation — no separate API key or CLI needed.

Installation

Add the marketplace and install plugins via the plugin manager:

/plugin marketplace add muloka/claude-plugins
/plugin install peer-review-jj@muloka-claude-plugins

Or browse available plugins:

/plugin

Note: After installing workspace-jj, run /workspace-setup in your jj project and restart Claude Code.

Relationship to claude-plugins-official

This repo started as a fork of Anthropic's claude-plugins-official and has evolved through two phases:

Phase 1: jj translations — Replaced Anthropic's git-based plugins (commit-commands, code-review, feature-dev, pr-review-toolkit) with jj-native equivalents. Same capabilities, different VCS.

Phase 2: jj-native capabilities — Built features that leverage jj's model in ways git can't easily support. Lightweight workspaces for parallel subagent isolation (fan-flames), first-class conflicts for multi-workspace merging, automatic working-copy snapshots eliminating the commit/stage ceremony, and operation-log-based undo for safe experimentation. These aren't ports of git workflows — they're new patterns that emerge from jj's architecture.

Removed from original: code-review, commit-commands, feature-dev, pr-review-toolkit — replaced by jj-native equivalents above.

Net new (no upstream equivalent): permission-gateway, workspace-jj, project-setup-jj, fan-flames skill.

License

See each plugin directory for the relevant LICENSE file.