Multi-agent workflow skills for coding agents you already use.
Give your agent a high-level goal. Get back deep research, a tech-validated task DAG, parallel workers in isolated git worktrees, reviewed merges, verification, audit, tech-debt cleanup, and an honest final report.
/swarmgoal Add dark mode toggle to settings with tests
/swarmgoal Fix auth redirect loop --mode fast
/swarmgoal --resume .swarmgoal/add-dark-mode-Ab3Kx9 --from execute
Works on Grok, Claude Code, Codex, and Cursor (with graceful fallback on unknown harnesses).
Swarmgoal synthesizes the best of three workflows:
| Source | What Swarmgoal inherits |
|---|---|
| Supergoal | Deep planning, two human gates, falsifiable criteria, final audit, 3-strike recovery |
| GoalGraph | DAG decomposition, file ownership, parallel batches, worktree isolation |
| execute-plan patterns | Harness adapters, persona injection, concurrency control |
The difference: Supergoal runs phases sequentially. GoalGraph skips research depth. Swarmgoal does both — research and parallel execution and audit/debt closeout — with modes, resume, and mechanical validators.
flowchart TD
Start(["/swarmgoal goal"]) --> S0["Stage 0<br/>Bootstrap + harness detect"]
S0 --> G1{"Gate 1<br/>Intake (full mode)"}
G1 --> R["Stage 1<br/>Research (per mode)"]
R --> P["Stage 2<br/>DAG plan + self-critique"]
P --> G2{"Gate 2<br/>Plan review + pre-flight"]
G2 -->|Revise| P
G2 -->|Start| E["Stage 3<br/>Parallel workers<br/>review + merge"]
E --> C["Stage 4<br/>Verify → Audit → Debt"]
C --> Done(["SWARMGOAL_RUN_COMPLETE"])
classDef human fill:#fef3c7,stroke:#d97706,color:#000
classDef done fill:#d1fae5,stroke:#059669,color:#000
class Start,G1,G2 human
class Done done
Yellow = human gates in full mode (fast skips Gate 1). Resume with --resume <run-root> --from execute if a session drops.
| Mode | When | Research | Gates | Parallel |
|---|---|---|---|---|
full |
Greenfield, large features | Deep + tech audit | 2 | Yes |
fast |
Brownfield fixes, scoped work | Scripts only | 1 (plan) | Yes |
sequential |
High-risk or unknown harness | Scripts only | 1 (plan) | No |
git clone https://github.com/robzilla1738/workflowskill.git
cd workflowskill
./install.shinstall.sh checks git and jq, symlinks swarmgoal/ into agent skill dirs, and links the swarmgoal CLI to ~/.local/bin.
Manual install:
ln -sfn "$(pwd)/swarmgoal" ~/.agents/skills/swarmgoal
ln -sfn "$(pwd)/bin/swarmgoal" ~/.local/bin/swarmgoalOpen your coding agent in any project and run:
/swarmgoal Build a REST API for todos with SQLite, tests, and README
/swarmgoal Fix checkout bug --mode fast
CLI (from any repo):
swarmgoal init --check
swarmgoal validate plan .swarmgoal/<run>/plan.json
swarmgoal status .swarmgoal/<run> --json
swarmgoal resume .swarmgoal/<run> --from execute
swarmgoal test.swarmgoal/<goal-slug>-<random>/
├── plan.json # Machine-readable DAG (jq-mutated only)
├── ROADMAP.md # Human-readable plan
├── STATE.md # Live progress + resume command
├── THINKING.md # Goals, risks, constraints
├── tech-audit.md # Tech freshness validation (full mode)
├── harness.md # Detected harness + concurrency
├── PROTOCOL.md # Execution manual
├── contracts/ # Per-task worker contracts
├── reports/ # Worker reports
├── reviews/ # Batch review outputs
└── phases/ # Audit/debt fix specs
| Harness | Spawn API | Default concurrency |
|---|---|---|
| Grok | spawn_subagent + persona prepend |
4 (max 8) |
| Claude Code | Agent + isolation: worktree |
4 (max 5) |
| Codex | Task tool |
4 (max 8) |
| Cursor | Task tool + git worktree add |
4 (max 8) |
| Unknown | Sequential + manual worktrees | 1 |
Validators perform semantic overlap detection — src/** and src/pages/Settings.tsx in the same batch is rejected. Optional worker pre-commit hooks enforce allowed_files at commit time.
git— worktree isolationjq— plan.json mutationsbash4+ — scripts and validators
swarmgoal test
# or
bash swarmgoal/tests/run-all.shCI runs the full suite on every push. See CONTRIBUTING.md and docs/ARCHITECTURE.md.
workflowskill/
├── bin/swarmgoal # CLI entry point
├── install.sh
├── swarmgoal/
│ ├── SKILL.md # Main orchestrator entry
│ ├── references/ # modes, resume, harness adapters, DAG design
│ ├── personas/ # researcher, worker, reviewer, tech-validator
│ ├── templates/ # ROADMAP, STATE, PROTOCOL, contracts
│ ├── schemas/ # plan.json + task JSON schemas
│ ├── scripts/ # validators, CLI helpers, glob-overlap lib
│ ├── subskills/ # research + execute routing
│ ├── fixtures/ # Sample plans + overlap test repo
│ └── tests/ # Full test suite (7 files)
└── docs/
└── ARCHITECTURE.md
- Supergoal — sequential deep planning + single
/goaldispatch - GoalGraph — parallel DAG, hands-off, less research overhead
If Swarmgoal saves you time, you can buy me a coffee. ☕
Current: v1.1.0. See CHANGELOG.md.
MIT. See LICENSE.