Handbook
Product Boundaries
Each product must have one primary operational responsibility. Cross-product behavior must be expressed through contracts, not fuzzy integration.
Boundary principle
Each product must have one primary operational responsibility. Cross-product behavior must be expressed through contracts, not fuzzy integration.
| Product | Owns | Does not own |
|---|---|---|
| ForgeSDLC | methodology, vocabulary, operating model | runtime, UI, agent execution |
| Blueprints | canonical policy, recipes, Versonas, evidence expectations | live run state, execution, agent memory |
| Lenses | local control plane, run spine, approvals, evidence review | deep LLM reasoning, Docker execution, canonical policy |
| LCDL | governed LLM reasoning, contracts, verification, repair | UI, job execution, product policy |
| Fleet | controlled execution, templates, logs, telemetry | planning, approval, LLM reasoning |
| Hermes | adaptive operator, memory, skills, MCP orchestration | canonical governance, approval authority, evidence system of record |
| Factory Droid | implementation-heavy coding, review, patch proposal | Forge methodology, final approval, system of record |
| OpenClaw | messaging ingress, channel routing, status relay | approval authority, execution, canonical run state |
| forge-workcells (private repo) | workcell runner code (local_llm_worker, future runners), pack loader, run store helpers |
canonical schemas, product policy, control-plane UI, domain micro-packs |
| forgesdlc-kitchensink (design system + tools) | micro-packs, UX harness glue, shared CSS/JS | workcell runtime, LLM transport, run spine |
| forge-cdp-manager | Edge CDP surface leases, registry, optional HTTP control plane (:18770) |
Harvest orchestration, SQLite, LCDL task bodies |
Critical rule
Lenses is the system of record for runs, approvals, evidence, and human review.
Agent workcell rule
No agent workcell writes final state directly. Every workcell returns a WorkcellResult linked to a ForgeRun and reviewed through Lenses.
Fleet rule
Agents must not receive raw execution power by default. Fleet should expose approved templates first, raw job submission later if explicitly allowed.
LCDL rule
When a result affects a decision, route it through an LCDL contract or traceable governed reasoning flow.
Blueprints rule
Blueprints define what should be judged. They do not execute jobs or privately remember user preferences.
Forge Campaign rule
| Role | Owns | Does not own |
|---|---|---|
| Forge Campaign (ForgeRun + manifest) | Intent, repo list, branch naming, evidence expectations | Container runtime |
campaign_orchestrator workcell |
Git branch, pack/upload, Fleet API poll, integrate, PR open (when policy allows) | Final merge approval, canonical policy text |
fleet_ux_worker workcell |
In-container audit/remediation, worker_progress / worker_result |
Opening PRs, Lenses decision state |
| Fleet | docker_argv, workspace extract, job telemetry |
Campaign manifest, git integration |
| Lenses | Run visibility, approval UI, decision on frun_* |
Executing remediation inside Docker |
See Agents and workcells, Forge Campaign, and Forge Agent.