forgelisted

# The Forge — Core Protocol Plus Claude Code and Codex Drivers A structured, task-driven improvement protocol with KPI guardrails. Forge tracks coverage/speed/quality with baselines and targets, derives or records success criteria for the task itself, evaluates with fresh-context audits, rotates strategies when stagnating, and records lessons across iterations. Built on the Ralph Wiggum loop pattern (Geoff Huntley), informed by Karpathy's autoregressive philosophy, pi-autoresearch's measurement discipline, and SICA's compounding iteration approach. ## Activation Triggers - "forge it", "forge this", "forge loop" - "quality loop", "kpi loop", "improvement loop" - "codebase improvement loop" - When user wants structured, KPI-driven autonomous improvement ## Architecture ``` Forge Core ├── Protocol phases (A through H) ├── State format and KPI model ├── Strategy selection + stagnation logic └── Fresh-context evaluation expectations Claude Code driver ├── /forge command ├── /forge-cancel command ├── .claude/forge-state.SESSION.md ├── .claude/forge-loop.SESSION.local.md └── Stop hook re-injects prompt on session exit Each iteration (one OODA cycle): ├── A. ORIENT — Read forge-state, understand position + trends ├── B. MEASURE — Run tests, capture KPIs ├── C. EVALUATE — Every 3rd iteration: fresh-context reality-check subagent ├── D. DECIDE — Pick strategy + target from KPI gaps + findings ├── E. EXECUTE — Apply ONE focused transformation ├