By rp4ri
Multi-agent orchestration toolkit for Claude Code — tmux director/worker patterns, PR bot-review triage, tag-driven multi-platform release trains, partitioned audit swarms, worker failure recovery, and recurring-bug family hunting. Distilled from real production orchestration sessions.
Triage one GitHub PR's automated bot reviews (Gemini Code Assist, Codex, etc.): list unanswered bot observations with severity, classify each as legit or rejectable with reasoning, and report a merge-readiness verdict. Read-only by default — it drafts replies and fixes but does not push or merge unless explicitly told the PR is its own.
Read-only monitor for tmux/rmux worker panes. Use to check what every worker in a multi-agent orchestration is doing: classifies each pane WORKING/IDLE (with truncation-safe detection), extracts final reports from scrollback, detects failure signatures (rate limit, usage-policy block, blocking questions), and verifies expected artifacts exist. Never sends keys.
Run a deep codebase audit by partitioning it into areas and delegating each to a parallel worker agent with a findings-only mandate, then consolidating the reports into one prioritized document. Use when the user asks for a thorough audit, wants to map every component for a class of bugs, or wants to verify code health across a large codebase.
When a bug recurs ("we fixed this last week!") or a root cause is confirmed, hunt the ENTIRE family: find every surface with the same pattern, fix the root once with a shared resolver, and add an anti-recurrence guard so the pattern cannot be reintroduced. Use for recurring bugs, post-fix sweeps, and "why does this keep coming back" investigations.
Merge N CLEAN PRs through one integration branch: combine them on an aux branch, verify the COMBINED build/typecheck, open a single PR to main, and sweep the issues that "Closes #N" did not auto-close. Use when a wave produced several ready PRs at once, when sequential direct-to-main merges risk semantic conflicts, or when consolidating before a release.
Run a tmux/rmux director-workers orchestration: spawn and command multiple Claude Code workers in panes, each in its own git worktree, monitor their state with background watchers, and coordinate their output into PRs. Use when the user wants to parallelize work across multiple agent sessions, set up a multi-agent workflow on a server, or asks to "orchestrate" tasks across tmux panes.
Triage automated bot reviews (Gemini Code Assist, Codex, etc.) on GitHub PRs before merging: wait out the review cycle, count unanswered bot observations, fix the legitimate ones with commits, justify rejections in-thread, and gate merges on zero unanswered comments. Use when handling PR reviews, deciding whether a PR is ready to merge, or coordinating merges across many PRs.
Executes bash commands
Hook triggers when Bash tool is used
Uses power tools
Uses Bash, Write, or Edit tools
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
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).
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.
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.
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/)| 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.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 |
npx claudepluginhub rp4ri/orchestration-studioA virtual design team for Claude Code, Cursor, Windsurf, Gemini CLI, and Copilot — powered by Design Studio. Assembles specialist roles — UI designer, UX researcher, content designer, Figma expert, data viz, email, social, motion, presentation, brand strategy, illustration, video, conversational, spatial, compliance, and more — for any design task. 26 roles, 60 commands, 13,800+ lines of expert design knowledge. Your agency's design brain, inside your terminal.
A virtual marketing team for Claude Code. Assembles specialist roles — SEO strategist, growth hacker, AI citation optimizer, content creator, social media strategist, and platform specialists — to plan and execute marketing for developer products, SaaS apps, and open-source projects. 8 roles, 10 commands.
v9.44.1 — Patch release for Gemini environment/version detection and qwen auth gating. Run /octo:setup.
Complete creative writing suite with 10 specialized agents covering the full writing process: research gathering, character development, story architecture, world-building, dialogue coaching, editing/review, outlining, content strategy, believability auditing, and prose style/voice analysis. Includes genre-specific guides, templates, and quality checklists.
Upstash Context7 MCP server for up-to-date documentation lookup. Pull version-specific documentation and code examples directly from source repositories into your LLM context.
Comprehensive startup business analysis with market sizing (TAM/SAM/SOM), financial modeling, team planning, and strategic research
Permanent coding companion for Claude Code — survives any update. MCP-based terminal pet with ASCII art, stats, reactions, and personality.
Comprehensive .NET development skills for modern C#, ASP.NET, MAUI, Blazor, Aspire, EF Core, Native AOT, testing, security, performance optimization, CI/CD, and cloud-native applications