| Vision Claim | What We Built | Status |
|---|---|---|
| Fresh session per card (no context rot) | Ralph v3 stateless runner | Done |
| 2-Hat loop (Build + Audit) | System prompt + scorer | Done |
| Cards as brief AND audit trail | Forward + backward-looking fields | Done |
| Skills loaded at execution | Session-zero loads SKILL.md | Done |
| Typed errors for routing | 7 error types, auto-routing | Done |
| KFS write on completion | ralph-kfs-writer.sh | Done |
| Concurrent lanes | 4 systemd timers | Done |
| Vision Claim | Reality | Severity |
|---|---|---|
| VERIFY → LOOP closes automatically | Ralph doesn't trigger new perceive cycle. Cards independent. | HIGH |
| Forge as always-on PERCEIVE daemon | Nothing watches state changes. Jason queues everything manually. | HIGH |
| COORDINATOR for multi-agent work | No coordinator. No task decomposition. No merge agent. | HIGH |
| Expert clone as running loop agent | MasteryOS expert system is concept only. | HIGH |
| Governance lane perpendicular | No governance agents watch all lanes. | MEDIUM |
| Handoffs carry full history | Handoff protocol designed, not implemented. | MEDIUM |
| Brief Library (reusable briefs) | No brief library. Every brief from scratch. | MEDIUM |
| Pattern Library | No curated pattern entries. | MEDIUM |
| Git as loop state machine | Commits don't trigger PERCEIVE. | MEDIUM |
| Forge UI as loop controller | Dashboard shows nothing. No phase gates. | MEDIUM |
Vision: Forge is a PERCEIVE daemon that watches the world, detects what needs doing, decomposes into cards, executes with governance, closes VERIFY → LOOP, and learns.
Reality: Ralph executes cards that Jason manually queues. We built the HANDS (execution engine). We haven't built the EYES (perceive), the BRAIN (plan/coordinate), or the MEMORY (extraction/learning loop).
The real domino: Not "Ralph reliable" — that's done. It's something that generates cards without Jason. A PERCEIVE agent that watches state and auto-generates work.
When errors recur X times, Ralph spawns a /btw investigation card in parallel (researcher lane). Card produces diagnosis + fix + test plan. Safe fixes apply automatically. Risky fixes surface to Jason. Findings persist in KFS so the same investigation never repeats.
Persistent background agent capturing decisions, judgment patterns, corrections, and roadmap ideas in real-time. Like a court stenographer for strategic decisions.
V1: PostToolUse hook → Haiku classifies → KFS write
V2: Hooks + periodic Sonnet extraction for judgment/taste
Key filter: Can it be verified by reading docs/code? YES = structural (don't train). NO = judgment (capture).
| Card | Task | Score | Attempts |
|---|---|---|---|
| #445 | Review ralph-runner.sh error handling | 88/100 | 1 |
| #447 | Write CHANGELOG.md for Ralph v3 | 92/100 | 1 |
| #448 | Create ralph-v3-status.sh | 95/100 | 2 |
| #446 | Add /v1/context/card/:id endpoint | Re-queued, executing... | |
Updated: 2026-04-02 12:20 UTC