A structured, multi-agent software development workflow that separates cognitive concerns: creative agents write specs and code, critical agents review them, and adversarial agents break assumptions.
Your OpenClaw Instance
│
│ "start engineering workflow for project-x: add user auth"
│
├── Supervisor (Forge) persistent sub-agent, manages pipeline
│ ├── Analyst / Spec Writer drafts specs from requirements
│ ├── Spec Reviewer validates specs, produces test criteria
│ ├── Implementer (Codex) writes code + tests (task-spec driven)
│ ├── PR Reviewer reviews code against spec (diff-based)
│ ├── Code Reviewer holistic codebase quality review
│ └── Chaos Agent (Ralph) adversarial testing, breaks assumptions
│
├── CI/CD Pipeline Monitor polls CI, auto-diagnoses failures
├── GitHub Issue & PR Lifecycle triage, track, close, review iterations
└── Human approval gates at spec approval + merge
The supervisor runs as a persistent sub-agent session — your main OpenClaw agent stays general-purpose and spawns workflow supervisors on demand. Multiple workflows can run in parallel.
# Clone
git clone https://github.com/clawosiris/forge.git
cd forge
# Deploy (new standalone instance with container sandboxing)
./deploy.sh
# Or add to an existing instance
./deploy.sh --addon| File | Purpose |
|---|---|
deploy.sh |
Deployment script with container sandboxing setup |
config/openclaw-standalone.json5 |
Complete config for a fresh instance |
config/openclaw-addon.json5 |
Merge guide for existing instances |
workspace/AGENTS.md |
Main agent instructions (includes workflow routing) |
workspace/SOUL.md |
Main agent persona |
workspace/templates/forge-supervisor.md |
Supervisor spawn template |
workspace/templates/agents/ |
Specialist agent prompt templates |
workspace/knowledge/ |
Project knowledge directory (populate per-project) |
docs/deployment-plan.md |
Full architecture documentation |
Not every change needs the full pipeline:
| Tier | When | Agents Used |
|---|---|---|
| Small | Bug fix, config, <50 LOC | Implementer + PR Reviewer |
| Medium | Feature, refactor, 50-500 LOC | + Analyst + Spec Reviewer |
| Large | New system, architecture, >500 LOC | + Chaos Agent + Code Reviewer (parallel) |
The supervisor monitors CI after implementation, auto-diagnoses failures, and spawns targeted fixes (max 3 attempts before escalating). PR review is gated on CI green.
Dedicated holistic code reviewer (separate from PR reviewer) that analyzes the full codebase for quality, architecture, and security — not just the diff.
Implementer uses a task-spec-driven pattern: write TASK.md → Codex --full-auto → validate → iterate. This produces the most reliable output.
After implementation, the supervisor compares what was built against the spec and produces a spec-delta.md flagging scope creep and gaps.
Optional compliance stage with SPDX header enforcement, dependency auditing, and SBOM generation. Driven by knowledge/compliance.md.
Analyst produces test-infrastructure.md alongside the OpenSpec. Mock servers, fixtures, and test doubles get equal review rigor.
Optional release stage: version bump, tag, CHANGELOG, release workflow monitoring — all tracked in workflow-state.json.
Track downstream impacts across repos. Supervisor creates follow-up issues in dependent repos and coordinates sequential cross-repo work.
Implementer runs formatter → linter → tests → docs before declaring complete. Eliminates trivial CI failures.
Patterns for feature-gated tests, graceful PermissionDenied skips, and cwd-relative paths for container compatibility.
Full lifecycle management: issue triage, label-based state tracking, PR review iteration handling, and optional cron-based monitoring.
- OpenClaw installed
- Docker (for container sandboxing)
- API keys: Anthropic (primary), OpenAI (fallback)
- At least one messaging channel configured (Signal, Discord, Telegram, etc.)
See docs/deployment-plan.md for the full architecture, state machine, failure handling, and observability details.
MIT