ART orchestrator — beads, agents, evolve loops. Skills for autonomous work management across all ART repos.
System-level technical vision — evaluates architecture trade-offs, writes ADRs, decomposes decisions into dispatchable work via BDR. Operates across the full ART ecosystem. Not an executor — produces plans that other agents implement.
Implementation quality perspective — finds complexity hotspots, dead code, TODO debt, hardcoded values, and copy-paste duplication at function level. Language-agnostic structural patterns.
Cross-file coherence perspective — finds circular dependencies, scattered functionality, duplicated data access patterns, inconsistent error handling, and API contract drift. Language-agnostic structural patterns.
You are a scheduled hygiene checker. You survey the codebase and bead backlog for structural drift, staleness, and complexity accumulation. You do NOT fix anything — you find problems and file beads or comment on existing ones.
Strategic perspective — examines cross-repo overlap, abandoned experiments, commit velocity, scope creep, and doc staleness. Low frequency filter in the survey graded filter bank. Can create and link beads across repos.
Continuous improvement loop — assesses open beads, classifies by effort, dispatches generator→evaluator pipeline with file-based handoff. No per-agent commits — evaluator gates all changes. Post-mortem refines agent prompts. Modes: --simplify (refactor only, no net new code), --security, --fe, --prune. Composable with /loop for cron scheduling.
File a bead from conversation context. Use when the user mentions something that should be tracked — a bug, idea, task, or observation. Synthesizes a bead title and description from the user's input and surrounding conversation. Dedup-checks against existing beads before creating.
Discover and document architecture seams — the interfaces, data flows, and integration points between projects or dependencies. Use when exploring how codebases connect, onboarding to a new ecosystem, or documenting service boundaries. Works on single projects or multi-project ecosystems.
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.
Experimental software. APIs, schemas, and behaviors change without notice. Use at your own risk. Contributions welcome — expect rough edges.
⚠️ Dispatch surface is currently broken. Recent Claude Code harness changes regressed the
claude -psubprocess invocation path used byrsry_dispatchand the reconciler's hub-to-spoke sync (see rosary-a53b91 + rosary-403a1a). The bead/task-tracking surface (CLI + MCP) is unaffected — see "Use rosary as a task tracker only" below if that's all you need. The architectural fix is a tmux-per-agent spawn substrate (rosary-75555d, pre-scoped into 6 children); contributions welcome.
Autonomous work orchestrator for AI agents across multiple code repos. Local-first, open source.
Rosary structures work as beads — small, trackable units stored in each repo via Dolt. A reconciliation loop scans for ready beads, dispatches AI agents (Claude, Gemini) to execute them in isolated workspaces, verifies the results, and syncs status to Linear for human review.
The human reviews 5-10 feature PRs a day. The agents handle the atoms.
graph LR
A[Scan repos] --> B[Triage & prioritize]
B --> C[Dispatch to agent]
C --> D[Agent works in<br/>isolated workspace]
D --> E{Verify}
E -->|pass| F[Close bead]
E -->|fail| G[Retry with backoff]
E -->|deadletter| H[Flag for human]
G --> C
stateDiagram-v2
[*] --> open
open --> dispatched: agent assigned
dispatched --> verifying: agent completes
verifying --> done: checks pass
verifying --> open: retry
verifying --> blocked: deadlettered
blocked --> open: human /resume
Work items live in each repo as beads — an AI-native issue tracker backed by Dolt (version-controlled SQL). Rosary reads and writes beads directly over MySQL, no CLI shelling.
Beads are organized into threads (ordered progressions of related work) and decades (ADR-level groupings) via the BDR harmony lattice.
# Beads are managed via rosary's MCP tools or CLI:
rsry bead create "Fix auth bug" --priority 1 --issue-type bug --files src/auth.rs
rsry bead list
rsry bead search "auth"
rsry bead close rsry-abc123 # requires a verifiable test command in the description
rsry bead close rsry-abc123 --force # override (legacy / non-impl beads)
# Transcript → BeadSpecs (Session provenance)
rsry capture --from-session sessions/2026-04-29.md --commit
# Source file → BeadSpecs (Code provenance, optionally scoped to a symbol)
rsry capture --from-code rosary src/bead.rs --symbol BeadSpec --commit
# Markdown ADR → beads + Rust stubs for design review
rsry decompose docs/adr/0009-feature.md --stub-output .
task build # requires Task (taskfile.dev)
task test
# Register repos to watch
rsry enable ~/code/my-app
rsry enable ~/code/my-lib
# See what's ready
rsry status
# Dry run — see what would be dispatched
rsry run --once --dry-run
# Real run — dispatch agents, verify, close
rsry run --once --concurrency 3
# Continuous loop
rsry run
Use
task build/task testinstead of rawcargo— the Taskfile setsPKG_CONFIG_PATHfor the fuse-t dependency via ley-line.
The dispatch + reconciler surface is currently broken (see banner above). The bead/task surface is fully functional today — rsry bead * CLI subcommands, the MCP server (rsry serve --transport stdio), and the Linear sync all work independently of the broken dispatch path.
If you just want issue tracking — backed by Dolt, queryable via MCP, optionally synced to Linear — you can use rosary today as a pure task tracker. Two install paths:
macOS (full install — codesigns binary + sets up an HTTP MCP launchd service):
task build
task install # macOS-specific: codesigns + launchd; binary lands at ~/.local/bin/rsry
Linux / Windows / any platform (build + copy; skips the macOS-only codesign + launchd steps):
cargo build --release
mkdir -p ~/.local/bin && cp target/release/rsry ~/.local/bin/rsry # or any directory on $PATH
Then on any platform:
# Register your repos for tracking (no dispatch loop will run)
rsry enable ~/code/my-app
npx claudepluginhub agentic-research/rosary --plugin rosaryAuto-invalidate ley-line-open's parse + LSP cache when Claude Code edits a file. Keeps lsp_hover/defs/refs/diagnostics fresh inside long sessions.
Repowire mesh usage skills for AI coding agents: cross-agent review and planning, delegate, usage patterns, and install/update. Backend-agnostic and parameterised on the agent you choose.
Task distribution, agent coordination, progress monitoring - executes plans via subagents. Requires AI Maestro for inter-agent messaging.
Enterprise AI agent orchestration plugin with 150+ commands, 74+ specialized agents, SPARC methodology, swarm coordination, GitHub integration, and neural training capabilities
Multi-agent coordination with agent-swarm MCP
AI-supervised issue tracker for coding workflows. Manage tasks, discover work, and maintain context with simple CLI commands.
Lanes skills and slash commands for Claude Code