Files
meadows/DESIGN.md
2026-06-20 02:09:09 +02:00

13 KiB
Raw Blame History

meadows — Design (QFD)

Goal-driven design for the meadows editor, phase 1 (diagram; numeric simulation is a later phase). Scope: data model, canvas/rendering, interaction model, and persistence for building stock-and-flow Models. It relies on CONTEXT.md for vocabulary and ADR-0001 (loops detected from a polarity-annotated graph, not stored). "Simulation-ready" is a constraint here, not a goal: nothing in the data model should need a rewrite when simulation lands.

Strength weights used in matrices: 9 strong, 3 medium, 1 weak, blank none.


1. Goals — the WHATs

ID Goal Weight Source
G1 Build a model with near-zero friction 9 original request; CONTEXT.md
G2 Feedback structure surfaces as insight (live) 8 original request; ADR-0001
G3 Models are safe and portable 7 original request

2. Functions — the HOWs

ID Function Dir Target (now) Target (future)
F1 Place a node in one gesture (palette → canvas, auto-named) 1 click/drag, 0 dialogs
F2 Connect by direct manipulation (auto source/sink cloud on open end) 1 drag, 0 menus
F3 Render & manipulate the canvas smoothly 60 fps pan/zoom/drag @ ≤150 elements ≥300 elements
F4 Enforce valid structure inline 0 invalid models persistable
F5 Detect & classify loops live (R/B) on every structural change ≤16 ms recompute @ ≤150 elements ≤16 ms @ ≥300
F6 Capture link polarity frictionlessly default on create, 1-click toggle
F7 Autosave locally, continuously debounced ≤500 ms; survive reload
F8 Round-trip Models as versioned JSON export→import identity; version schema migrations
F9 Undo/redo every model change ≥50 steps, all structural ops

3. Cascade — Goals → Functions → How → Components

  • G1 Build a model with near-zero friction W:9
    • F1 Place a node in one gesture ↓ 1 click/drag
      • How: Vue Flow drag-from-palette → store action creates the node
        • Component: C1 Vue Flow canvas · C2 custom node components · C6 projection
    • F2 Connect by direct manipulation (auto cloud on open end) ↓ 1 drag
      • How: drag from a node handle; open end auto-spawns a Cloud
        • Component: C2 node handles · C3 edge components · C5 store actions · C7 structure guard
    • F3 Render & manipulate smoothly ↑ 60 fps @ ≤150
      • How: Vue Flow viewport (GPU transforms); Model→VueFlow projection memoised
        • Component: C1 Vue Flow canvas · C6 projection
    • F9 Undo/redo ↑ ≥50 steps
      • How: ring buffer of deep-cloned Model snapshots in the store
        • Component: C5 model store (history)
  • G2 Feedback structure surfaces as insight (live) W:8
    • F5 Detect & classify loops live ↓ ≤16 ms @ ≤150
      • How: derive directed graph → Tarjan SCC → capped Johnson → polarity product (R/B)
      • How: SCC-only — rejected (loses per-loop insight), see T3
        • Component: C8 loop engine · C11 loop overlay
    • F6 Capture link polarity frictionlessly ↓ default + toggle
      • How: InfoLink defaults to + on create; 1-click +/ toggle on the edge badge
        • Component: C3 edge components · C5 store actions
    • F4 Enforce valid structure inline → 0 invalid persistable
      • How: guard guides (disallow invalid handle targets) rather than erroring after
        • Component: C7 structure guard · C4 domain types
  • G3 Models are safe and portable W:7
    • F7 Autosave locally, continuously → ≤500 ms debounce
      • How: store subscribes to Model changes → debounced repository.save()
        • Component: C5 store · C9 ModelRepository (IndexedDB/idb)
    • F8 Round-trip Models as versioned JSON → lossless
      • How: serialise Model (version field) → download; import → validate → load
      • How: PouchDB/cloud sync — deferred (needs a backend), see Tensions
        • Component: C10 import/export · C4 domain types
    • F9 (also serves G3 as a safety net) → C5 model store

4. House — Goals × Functions

Cells: strength 9/3/1. Importance Σ = Σ(weight × strength).

F1 F2 F3 F4 F5 F6 F7 F8 F9
G1 (9) 9 9 9 3 1 3 1 3
G2 (8) 3 1 3 9 9 1
G3 (7) 1 9 9 9
Σ 81 105 89 58 81 99 72 63 98

Top engineering priorities (from importance): the matrix puts F2 (connect, 105), F6 (polarity, 99) and F9 (undo, 98) above the headline detector F5 (81). The lesson: loop insight is only as good as its inputs and ergonomics — getting wiring (F2) and polarity capture (F6) effortless matters as much as the detection itself, and a reliable safety net (F9) underwrites fearless building. F4 (58) is real but lowest-leverage — keep it lightweight (guide, don't nag).

5. Roof — Function × Function tradeoffs

strong reinforce · mild reinforce · × mild conflict · strong conflict.

F2 F3 F4 F5 F6 F9
F2 ×
F3 ×
F4
F5
F6
F9

Conflicts that actually shape the design:

  • F5 ⊗ F3 — live loop enumeration on every change vs the 60 fps frame budget. The design-shaping conflict. Mitigated by the cap (≤200 cycles) + running detection off the render frame (debounce / idle callback / worker). Owned by ADR-0001 + the §7 budget.
  • F2 × F4 — guardrails can add friction to connecting. Mitigated by guiding the gesture (invalid handles refuse the drop) instead of erroring afterwards.
  • F3 × F9 — snapshot cloning on every change costs memory/time. Bounded by small models + a fixed-size ring buffer; revisit with structural sharing only if profiling demands.
  • F2 ◎ F6, F5 ◎ F6 — polarity is captured in the connect gesture and is the detector's input; these reinforce, so build them together.

6. Components & Function → Component map

ID Component ADR
C1 Vue Flow canvas (@vue-flow/core + addons) ADR-0002
C2 Custom node components (Stock/Flow/Converter/Cloud) ADR-0002
C3 Custom edge components (InfoLink badge, Flow pipe) ADR-0002
C4 Domain Model + types (model/types.ts) ADR-0003
C5 Model store (Pinia) + snapshot history
C6 Model ↔ Vue Flow projection (projection.ts) ADR-0002
C7 Structure guard (validation.ts) ADR-0003
C8 Loop engine (loops.ts: SCC + capped Johnson) ADR-0001
C9 ModelRepository (IndexedDB via idb)
C10 Import/Export (versioned JSON io.ts)
C11 Loop overlay UI (R/B badges, hover highlight) ADR-0001
C12 Palette + chrome (DaisyUI toolbar, model list)
C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12
F1 9 9 3 9 9
F2 9 9 9 9 3 9
F3 9 3 3 9
F4 1 1 3 3 9
F5 3 1 9 9
F6 9 3 3 1
F7 9 9
F8 3 1 9
F9 9

7. Critical performance budget

Rank Function Target Watched on If we miss it
1 F5 ≤16 ms recompute @ ≤150 bench on synthetic dense graphs enforce cycle cap; move detection off-frame (idle/worker); debounce
2 F3 60 fps pan/zoom/drag @ ≤150 frame profiling in DevTools memoise projection; simplify node DOM; virtualise off-screen nodes
3 F2 1 drag, 0 menus manual UX walkthrough simplify handle model; reduce connect steps
4 F7 ≤500 ms debounce, 0 loss reload-after-edit integration test flush on visibilitychange/beforeunload
5 F8 export→import identity property test (round-trip) pin schema; add migration on version bump

8. Tradeoffs — Got / Paid / ADR

ID Tradeoff Got Paid ADR
T1 Vue Flow over hand-rolled SVG fast F1F3, batteries-included editor mechanics a dependency owning the view layer + a Model↔VueFlow projection to maintain ADR-0002
T2 Flow-as-node + materialised Clouds uniform graph for loops/render; info-links can target flows; sim-clean Cloud lifecycle management; a node that renders as a pipe ADR-0003
T3 Enumerate loops with a cap per-loop R/B insight (G2) within the frame budget capped completeness on pathological graphs ("+N more") ADR-0001
T4 IndexedDB(idb) behind ModelRepository size headroom + swappable engine (sync later) async wrapper + a small dependency; no sync today
T5 Snapshot undo over command pattern simplicity and correctness memory per snapshot (bounded by ring buffer + small models)

Tensions being watched (unresolved by design)

  • Cross-device sync / sharing. Not built; G3 is met by local persistence + JSON export. Trigger to revisit: a real multi-device or collaboration need → add a PouchdbModelRepository or API-backed repository behind the existing interface.
  • Auto-layout. Manual placement only in phase 1. Trigger to revisit: users report large models becoming unreadable → add an optional layout pass (e.g. ELK/dagre) over the projection.

9. Inconsistencies spotted and fixed

  • Brief vocabulary vs glossary. The original request said "functions," "arrows," "faucets"; DESIGN uses the resolved CONTEXT.md terms verbatim — Converter, Flow/Information Link, faucet dropped.
  • G3 ambiguity. "Safe and portable" initially read as possible cloud sync; clarified — G3 is satisfied by local persistence + JSON export, not sync (PouchDB deferred, see Tensions).
  • Storage default drift. Initial recommendation was localStorage; resolved to IndexedDB via idb, behind a ModelRepository interface so the engine choice is no longer hard-to-reverse.

How to keep this honest

  • When a new ADR lands → add its components to §6 and re-score affected rows.
  • When a spike / measurement returns numbers → update §7 Target / Watched on.
  • WHATs change rarely; HOWs change with each release; matrices are recomputed when either side changes.
  • If a section becomes empty after edits, delete it — empty sections lie.