Flow Diagrams
Three maps. The PUMP powers the factory. Identity, memory, scaffolds, boundaries.
What does success look like?
Five binary outcomes. Each is a pass/fail test.
OUTCOME MAP — Agent Platform ═══════════════════════════════════════════════════════════════ Outcome Measure Current ────────────────────────────── ─────────────────────────────────── ─────── Dashboard math correct UI totals = CLI totals FAIL Scaffold creates valid files drmg scaffold skill → passes lint FAIL Agent scope enforced Edit outside scope → warning FAIL Patterns extracted cross-run 3+ runs → grouped findings FAIL Memory recalled by similarity Write → query → relevant result FAIL
Outcomes name what we're measuring. The value stream shows current vs target.
How does value flow?
Current state (scattered) vs target state (unified).
VALUE STREAM MAP — Current vs Target
═══════════════════════════════════════════════════════════════
CURRENT STATE
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Plans UI │ │ Scaffold funcs │ │ Agent configs │
│ Math wrong │ │ No CLI surface │ │ No scope files │
│ No drill-down │ │ Manual copy/paste│ │ Implicit bounds │
└─────────────────┘ └─────────────────┘ └─────────────────┘
↓ ↓ ↓
Operator runs CLI Dev copies files Agent breaks things
to get real numbers 15 min per scaffold Discovered days later
TARGET STATE
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Plans Dashboard │ │ drmg scaffold │ │ scope.json/agent│
│ Math = CLI │────→│ skill|hook|rule │────→│ Pre-edit hook │
│ Group + drill │ │ Type registry │ │ Violation log │
└────────┬────────┘ └────────┬────────┘ └────────┬────────┘
│ │ │
▼ ▼ ▼
┌─────────────────────────────────────────────────────────────────┐
│ Pattern Extractor — cross-run findings → prevention proposals │
│ Memory Writer — structured recall → agents learn from history │
└─────────────────────────────────────────────────────────────────┘The value stream shows what changes. The build order shows the sequence.
What's the build order?
Three phases. Each unblocks the next.
BUILD ORDER — Three Phases
═══════════════════════════════════════════════════════════════
PHASE 1: SURFACE (2 sessions)
┌──────────────────┐ ┌──────────────────┐ ┌─────────────┐
│ Fix plans math │────→│ Project grouping │────→│ Drill-down │
│ #1: PUMP-001 │ │ #2: PUMP-001 │ │ #3: PUMP-001│
└──────────────────┘ └──────────────────┘ └──────┬──────┘
│
PHASE 2: PREVENTION (3 sessions) ▼
┌──────────────────┐ ┌──────────────────┐ ┌─────────────┐
│ Wire scaffold │────→│ Type registry │ │ Scope files │
│ #4: PUMP-002 │ │ #5: PUMP-002 │ │ #6: PUMP-003│
└──────────────────┘ └──────────────────┘ └──────┬──────┘
│
┌──────────────────┐ ┌─────┴───────┐
│ Scope audit │←────│ Boundary │
│ #8: PUMP-003 │ │ hook #7 │
└──────────────────┘ └──────┬──────┘
│
PHASE 3: LEARNING (3 sessions) ▼
┌──────────────────┐ ┌──────────────────┐ ┌─────────────┐
│ Pattern extractor│────→│ Prevention │────→│ Memory write│
│ #9: PUMP-004 │ │ proposer #10 │ │ #11: PUMP-005│
└──────────────────┘ └──────────────────┘ └─────────────┘
First target: Plans dashboard (PUMP-001, issues #26-30)Questions
Should the pattern extractor propose hooks or rules — and what's the difference in blast radius?
- If scaffold templates are too rigid, will agents bypass them and write files directly?
- How do you measure whether a boundary hook is too restrictive vs appropriately cautious?
- At what scale does per-agent scope become unmaintainable — should scope be per-role instead?