Handbook
Respecting resources and autonomy
Platform operators need a clear line between interactive Cursor planning (human in the loop each turn) and bounded autonomous execution (unattended or PDCA loops). This page is adoption-focused; canonical methodology…
Related in-repo: Cursor rules and skills (cost-aware planning), Agents and workcells, Governed workflows and rules.
Why Platform cares
Forge Platform is a governed delivery control plane with pluggable agent workcells — not a single opaque agent product. Unattended runs still consume tokens, local compute, and human review when they touch merge, approval, or release boundaries.
Respecting resources aligns with the operating model (three planes): cognition and policy stay inspectable; execution stays bounded; evidence feeds Assay and approval gates.
Interactive vs autonomous
| Mode | When | Read |
|---|---|---|
| Interactive | Human steers each Cursor turn; plans before code | Cost-aware planning and model tiering |
| Autonomous / bounded | Campaigns, patch loops, worker ladders with verify/repair | Respecting resources · Autonomy levels |
Cost-aware rules answer how to plan in the IDE. The autonomy ladder answers how much change an unattended run may attempt and where humans must gate.
Example at a glance (L1)
A real Dark Factory L1 run, from the worked examples:
- Goal: fix a broken markdown link in
README.md— classifiedcomplicated / S / high, routed to thelocaltier. - What ran: the local model stalled, so the worker ladder stepped to Cursor (
composer-2.5) — a within-loop recovery, not a human escalation. - Where you still gate: the change lands on a branch for approve/merge, and guarded promotion is skipped unless the live tree is clean (no auto-commit).
The Blueprints bounded execution examples page carries the full L1/L2 walkthroughs plus loop, routing, PDCA, and dual-wiki diagrams.
Autonomy ladder (L0–L3 summary)
Full L0–L8 table and enforcement detail: Autonomy levels (canonical).
| Level | Unit of delivery | Human gate (short) |
|---|---|---|
| L0 | Suggestions only | Continuous |
| L1 | One function / contract-bound change | Approve branch/merge |
| L2 | Multi-file change-set, no rearchitecture | Acceptance criteria + merge |
| L3 | End-to-end use-case slice in one app | Intent in; review out |
Levels L4+ (feature, subsystem, product increment, multi-platform) add ADR, go/no-go, and strategic checkpoints — see the handbook page.
Forge Dark Factory (PoC)
Forge Dark Factory is the current PoC reference implementation for a sequential local-first autonomous coding loop. It is not a Platform product submodule today.
| Aspect | PoC notes |
|---|---|
| Target | L1 today; foundation for L2–L3 |
| Stack | forge-lcdl patch units + verify/repair/proof; optional forge-workcells local worker |
| Loop | Classify → route → plan → draft → apply → verify → repair → proof → dual-wiki → escalate |
| Trace | Machine record + generated human report with freeze gate |
Use Dark Factory to validate ladder and resource rules in code; use this handbook and Blueprints pages for policy regardless of whether you run the PoC.
Platform touchpoints
- Agents and workcells — ForgeRun, AgentRun, workcell modes
- Governed workflows and rules — approval gates, campaigns, LCDL graphs
- Approval model — human/policy gates before write-capable runs
- Ecosystem reference — repos and integration topology