Agentic JJ - Regras de Workflow para Agentes de IA
ulisten-ai/agentic-jj
Regras de workflow jj otimizadas para agentes de IA evitarem falhas comuns
Ver no GitHubPlugin que carrega regras sempre-ativas no contexto do Claude Code para guiar agentes de IA no uso do Jujutsu (jj). Evita armadilhas comuns quando LLMs dirigem jj autonomamente e estabelece um workflow consistente baseado em commits pequenos, snapshots frequentes e reescrita agressiva de histórico.
Instalação
npx convez-ai ulisten-ai/agentic-jj1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Guiar Claude Code ao usar Jujutsu para controle de versão
- Padronizar workflows de commit em sessões paralelas de agentes
- Evitar hanging em prompts de editor e falhas de squash/restore
- Ensinar LLMs a trabalhar com revsets e operações jj complexas
Veredito

Solução especializada e prática que resolve problemas reais de agentes LLM pilotando jj. Excelente para equipes usando Claude Code em repositórios jj, mas de escopo limitado e voltado para caso de uso específico.
README
A tradução para o português deste README ainda está na fila.
agentic-jj
Jujutsu (jj) workflow rules for AI coding agents. Currently ships as a Claude Code plugin; the rules themselves are tool-agnostic and future versions will package them for other agent harnesses (OpenAI Codex, etc.).
What it addresses
Most jj plugins for AI agents are on-demand skills, or read like jj documentation adapted for an agent — restating what models already know from training. This one loads as always-on rules into Claude's context, and addresses two specific things Claude needs help with:
1. Resistance to the failure modes that show up when an LLM drives jj
autonomously — editor prompts that hang the session, multi-line -m quoting
headaches, --from/--into squash surprises, jj op restore clobbering work
from concurrent jj workspaces (e.g. parallel Claude Code sessions in the same
repo), revset-syntax pitfalls, and so on.
2. A workflow shape to follow. Claude needs one to use jj effectively — without explicit guidance, jj's flexibility (multiple ways to checkpoint, clean up, recover) becomes a source of inconsistency. The shape this plugin teaches:
- Small, focused commits — default to separate, not combine.
- Snapshot-often, squash-later — many intermediate commits while working; consolidate when verified.
- Aggressive history rewriting — history is malleable until push.
- Describe early — set a description as soon as the change has a clear concept, not at the end.
If you have a different working style in mind, adapt the relevant rules into your own setup instead of installing wholesale.
PR-review workflow is out of scope — supplement separately if needed.
Install
/plugin marketplace add ulisten-ai/agentic-jj
/plugin install agentic-jj
The rules take effect in new Claude Code sessions after install — the
SessionStart hook can't fire retroactively in the session you ran the install
command from, so restart Claude Code (or open a fresh session) to see the rules
load.
Recommended one-time jj config
Run once after install:
jj config set --user ui.conflict-marker-style git
This changes the conflict markers jj writes into your working copy from its
default 3-way <<<<<<< / %%%%%%% / +++++++ / >>>>>>> format (which
encodes the base + each side's diff) to git's familiar 2-way
<<<<<<< / ======= / >>>>>>> format. LLMs are heavily trained on the
git format and read it more reliably; jj's 3-way format is information-richer
but unfamiliar to most models. The setting is local to your working copy —
it doesn't affect what's stored in the repo or what collaborators see.
How it works
On every SessionStart, a hook script syncs the rules file to
<project>/.claude/rules/ulisten/agentic-jj/jj-workflow.local.md. Claude Code auto-loads
every .md under <project>/.claude/rules/ recursively into the session's primary
context window — so the rules are always-on within the project, but scoped to it (no
global ~/.claude/rules/ pollution).
The destination uses the .local.md suffix to piggyback on the common *.local.md
gitignore convention; if your project's .gitignore doesn't cover it yet, the hook
appends *.local.md once. Idempotent on every run.
What's covered
The rules file (~285 lines) covers:
- Mental model — the working copy IS a change; change ID vs commit hash; the op log
- Non-interactive commands — alternatives to forms that open editors and hang the session
- Snapshot workflow — when to
jj new, when tojj describe, how often to checkpoint - Squash semantics —
--from/--intodirectionality, when diffs don't cancel out - Rebase flag semantics —
-svs-rvs-b, and what-ractually does to descendants - Splitting changes — file-based via
jj split -- <files>, restore-pivot for hunks - Revsets — correct
description(substring:"...")syntax (bare strings are glob, not substring); common patterns - Op log & recovery —
jj op revertoverjj op restore; concurrent-workspace caution - Multi-line commit messages —
printf '%s\n' ... | jj describe --stdinor$TMPDIR/jj-msg.tmp - Small-focused-commits philosophy + the
JJ_EDITOR=true jj commit <files>working-copy-split pattern - General tips — no stderr suppression, change ID conversational recognition,
(hidden)vsabandoned,--limitgotchas
Requirements
- Claude Code 2.x (plugin system)
- jj 0.25+ recommended (tested with 0.40)
Roadmap
v0 ships rules only. Planned for later versions:
- Skills for on-demand complex recipes:
jj-split,jj-rewrite-history,jj-recovery - Safety hooks: block-git (in jj repos), check-interactive (intercept editor-opening jj commands), auto-snapshot (preserve work on session stop), inject-state (show
jj statusat session start)