orchestration-studio
rp4ri/orchestration-studio
Multi-agent orchestration toolkit for Claude Code — tmux director/worker patterns, PR bot-triage, multi-platform release trains, audit swarms, worker recovery, recurring-bug family hunting. Skills + agents + hooks, distilled from production sessions.
Instalação
npx convez-ai rp4ri/orchestration-studio1 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
README
A tradução para o português deste README ainda está na fila.
Orchestration Studio
Multi-agent orchestration toolkit for Claude Code — distilled from real production sessions running a director Claude Code instance commanding multiple worker instances over tmux/rmux, shipping multi-platform releases, triaging bot reviews, and auditing a large monorepo in parallel.
Everything in here was learned the hard way: each rule traces back to a real incident (a stale tag that shipped a release without its fix, a worker whose uncommitted files were wiped by a sibling, a "recurring" bug that was actually 7 copies of the same logic).
Install
claude plugin install https://github.com/rp4ri/orchestration-studio
# or load for one session:
claude --plugin-dir ./orchestration-studio
Skills are namespaced: /orchestration-studio:director, etc.
Start here
WORKFLOW.md — the canonical end-to-end agent flow (setup → dispatch → work → bot wait → triage → merge → release → monitor), including the worktree-per-worker isolation procedure and the exact rules that keep workers conflict-free.
skills/director/rmux-reference.md — the rmux/tmux cheat sheet: every key combination and command (send-keys sequence, capture-pane flags, the truncation-safe state grep, the watcher loop), with the reason behind each flag.
What's inside
Skills (skills/)
| Skill | What it teaches Claude |
|---|---|
| director | The tmux/rmux director-workers pattern: worktree-per-worker, the exact send-keys protocol (Escape → C-u → literal → Enter), ghost-text hazards, truncation-safe WORKING/IDLE detection (grep 'esc to' + debounce), background watchers as the event loop, dispatch mandate templates |
| pr-triage | Bot-review discipline: the unanswered-comment jq query, fix-legit-or-justify, reply-with-SHA, merge gating on zero unanswered, merge ordering across many PRs |
| release-train | Tag-driven multi-platform releases with independent cadences (per-platform tag namespaces), safe re-tagging (verify tag→tip BEFORE pushing), gh run watch background watchers with built-in post-release verification, first-platform-build failure catalog (Windows spawn/cross-drive, Android Tauri ACL), auto-updater manifest safety |
| audit-swarm | Partitioned multi-agent audits: findings-only mandates, enumerable-inventory framing, uniform severity contract, consolidation into cross-cutting themes, rate-limit/filter recovery during swarms |
| worker-recovery | The failure taxonomy: rate-limit cuts, usage-policy false positives (and the defensive re-framing that passes), false-idle detection, the verify-artifact protocol (idle + missing artifact = incident), context-preservation rules |
| bug-family-hunt | Killing recurring bugs: git archaeology of previous fixes, instance enumeration, one shared resolver, the two-test rule (literal regression + anti-recurrence source-scan guard), sibling sweeps from a confirmed root cause |
| consolidate | Merging N CLEAN PRs through one integration branch: combined-state CI as the real gate, conflict→exclude-and-report, and the "Closes #N doesn't fire via an aux branch" sweep |
| spec-discipline | Diagnosis discipline: falsifiable-hypothesis specs (never conclusions), mandatory phase-0 data audit for state bugs, the two-shipped-fixes stop-rule with canonical-repro release gates, and the premise-challenger role against "CLEAN theater" |
Agents (agents/)
| Agent | Role |
|---|---|
| worker-monitor | Read-only pane inspector: WORKING/IDLE/INCIDENT per pane, failure-signature detection, artifact verification. Never sends keys |
| pr-triager | Single-PR bot-review triage: unanswered observations table, LEGIT/REJECT classification with reasoning, merge-readiness verdict |
Hooks (hooks/hooks.json + scripts/)
| Hook | Event | What it does |
|---|---|---|
tag-push-guard.sh | PreToolUse(Bash) | When a git push carries a tag, warns if the tag's commit isn't contained in any remote-tracking branch — catches the stale-tag-ships-old-code incident before it happens |
leak-guard.sh | PreToolUse(Bash) | Before a git push/release upload, scans the OUTGOING commits' added lines for secrets and infra identifiers — including bare public IPs (not just ip:port), instance ids, internal hostnames, private-key blocks, .env/.pem additions. Secret scanners miss identifiers; a too-narrow IP regex once let the real ones through |
symlink-guard.sh | PreToolUse(Bash) | Before a git commit, flags staged symlinks (mode 120000) named node_modules or with absolute-path targets — a worker's broad git add once shipped them and broke Windows CI while macOS tolerated it |
detect-orchestration.sh | SessionStart | Injects the live tmux/rmux session/pane map into context, with the "pane titles are stale" reminder |
The five laws (if you read nothing else)
- One worker, one worktree. Shared checkouts end with one worker's git surgery destroying another's uncommitted state.
- Never blind-Enter a pane. Ghost text is not input.
C-ufirst, always. - Idle + missing artifact = incident. A dead worker (rate limit, filter block) is indistinguishable from a finished one until you check what it was supposed to produce.
- Verify a tag points at the intended tip before pushing it. A re-tag
whose
git tag -dsilently failed will ship a release without its fix. - A recurring bug is a pattern with N instances. Fix the resolver once, then add the guard test that fails when instance N+1 is written.
Provenance
Built from a live orchestration session over a production monorepo (Electron desktop + SvelteKit UI + Node relay + Tauri mobile): ~20 PRs authored/triaged/merged, 5 releases across 3 platforms, one prod outage diagnosed and reverted, one 80+-finding parallel audit — all coordinated by a director session driving 3 workers. The skills are the playbooks that session converged on.
License
MIT