Goldfinch — Lovable Prep: Plan v2 Proposal

Fresh-eyes reassessment (Opus 4.8) + integration of three new capabilities. Prepared for Brandon's review.

2026-06-04 Strategist · Goldfinch Tracer Bullet TB#2 Decision surface · scroll to the bottom to choose
BLUF. The v1 plan's bones are right — keep the mem-synthesis rule and the autopilot-vs-substantive discipline. But v1 is linear, single-chat, and renders its review checkpoints as dense markdown (the rubber-stamp risk). Recommend a v2 rework that: (1) renders the two real decision points as HTML dashboards; (2) restructures into an INDEX-spine + streams with boot-without-handoff; (3) reserves fan-out for one adversarial gap-check on the Lovable prompt (the cheap-but-load-bearing reliability move). The weekend tracer bullet is achievable and thin: a thin reader + the Affirmations surface + formation-first framing → a shareable Lovable URL for Andrew.
⏱ Timeline correction: the prior "this weekend" was anchored to 2026-05-28; this chat paused ~a week. Today is Thu 2026-06-04; the tracer-bullet weekend renews to ~June 6-8.

What v1 got right · and 7 gaps the rework fixes

Keep verbatim

  • The mem-synthesis discipline rule (synthesis layers ≠ ratified facts)
  • The substantive-vs-autopilot table
  • Default-route-to-backlog + elevation pass
  • Multi-session split; C2 deferred to a reground session

Seven gaps

1
Linear / single-chat; no fork-spine or concurrency→ INDEX-spine + streams (your TB#1 Entry 1)
2
Checkpoints are dense markdown → rubber-stamp risk→ HTML decision dashboards (Entry 2 / Thariq)
3
No crisp definition-of-success → over-build risk→ a pinned thinnest-real-shareable minimum
4
No quality gate on the Lovable prompt→ 3-lens adversarial gap-check (AIOPS-058)
5
Reader scope for prompt #1 ambiguous→ explicitly THIN; depth → Stream S1
6
"First prompt" conflated with full V1 scope→ prompt = a thin SLICE of the matrix
7
S1/S2 "TBD" reads as blockers→ explicitly non-blocking for the weekend

The three new capabilities · honest fit

High fit

HTML as a medium

Render the 2 substantive checkpoints (V1 matrix, Lovable prompt) as HTML dashboards. This proposal dogfoods it.

  • New move: HTML for the working plan + the micro-UI decision surface (this page).
  • Guardrail: markdown stays the builder contract; HTML = human surface only; watch token cost.
Structural fit

Fork-spine multi-chat

INDEX-spine + per-stream stubs + boot-without-handoff (lighter than a HANDOFF doc).

  • Guardrail: you're a single reviewer → cap at primary + 1 optional secondary, serialized review (Entry 1's "don't force parallelism").
  • Weekend: Stream P only; parallelism pays off the week after.
Selective

Dynamic Workflows

Your GC run was 1.37M tokens / 12 agents. Too heavy to review one prompt.

  • Use the primitive, not the tool: 3 blind reviewers vs. the canonical docs as ground truth — via the lighter Agent fan-out.
  • Not for routing: "right V1 scope" has no ground-truth oracle; keep it in-context.

Recommended v2 structure

Stream P Route → backlog→ V1 scope matrix→ ✦ CHECKPOINT 1 (HTML)→ Lovable prompt (thin slice)→ ⚔ 3-lens gap-check→ ✦ CHECKPOINT 2 (HTML)→ Send → prototype
Stream S1 Bible reader UX deep scope (non-blocking) Titan reader-research + highlighting doc + Track A → catalog for Lovable prompt #2+
Stream S2 Service Layer reground (the deferred C2) re-ground in raw Mem audio per the mem-synthesis rule
Stream X PRODUCT.md Pillars + Wedge major revision closeout, after P + S1/S2 inform it
✦ = Brandon HTML checkpoint ⚔ = adversarial gap-check Weekend runs Stream P only.

Definition of success · the weekend tracer bullet (thin on purpose)

A shareable Lovable URL Andrew can open that reads as a spiritual formation app, not a Bible reader. Thinnest real slice:

  • Thin Bible reader — pick a passage, read, switch translation, save a verse. No word-tap / Strong's / highlighting / audio yet.
  • Affirmations widget surface (C4-canonical) as the lead formation feature.
  • Home/dashboard IA presenting formation + reader as siblings — the framing signals breadth.
  • Optional if cheap: one more light formation stub (a daily-practice card or a capture button).

Explicitly OUT (later streams): deep reader, Service Layer, entity capture, voice-search backend, audio.

Your decisions · select, then compile & copy back to me

Each one is pre-set to my recommendation (★). Change any you want, add notes, then hit Compile → Copy and paste the block back in chat. Or just reply in prose.

D1Rework scope
How much of v1 do we change?
Supersede v1 with the v2 spinev1 marked superseded; v2 becomes active. ~90% of v1 content is reused.★ REC
Light patch to v1Keep v1 as the file; bolt on the HTML-checkpoint + gap-check notes only.
Keep v1 as-isNo structural change; proceed on the original 4-phase shape.
D2Tracer-bullet definition of success
Thin reader + Affirmations surface + formation-first home IA → shareable URL. (See the card above.)
Approve the thin-slice minimumExactly as stated above; reader stays thin; depth defers.★ REC
Tighten itDrop something (e.g., the optional second formation stub).
Expand itAdd scope (note what — risks blowing the weekend).
D3Plan structure
How is the work organized across chats?
INDEX-spine + streams + boot-without-handoffThe spine + WORKING-CONTEXT pin is the handoff; each stream boots just-in-time.★ REC
Single linear handoff (v1 style)One HANDOFF doc, one fresh chat, sequential.
D4Weekend concurrency posture
How many streams run this weekend?
Stream P only (focus)One critical path; least review load; recommended for a 2-3 day window.★ REC
Stream P + boot S1 in parallelFill the model-thinking lulls with the reader-scope stream; more juggling.
D5Fan-out on the Lovable prompt
How do we quality-gate the highest-leverage artifact before you send it?
3-lens gap-check via Agent fan-out (light)Blind reviewers: conviction / mem-synthesis / formation-framing. The AIOPS-058 primitive, cheap.★ REC
Full dynamic WorkflowDeterministic JS spine + gates. Powerful but heavy (~1M+ tokens). Overkill for one prompt.
NoneI draft, you review directly, no independent check.
D6HTML review surfaces
Which checkpoints get an HTML dashboard like this one?
Both checkpoints (V1 matrix + Lovable prompt)Full dogfood of the pattern at the two decision points.★ REC
Lovable prompt onlyHTML for the prompt; V1 matrix stays markdown.
Markdown onlySkip HTML; this proposal was a one-off.

If you approve · Stream P boots from this prompt

Goldfinch Lovable prep — Stream P (primary critical path). Boot from the v2 spine.

Read in order:
1. apps/goldfinch/planning/_plans/PLAN_lovable-prep_INDEX_v2_2026-06-04.md (the spine — Stream P is your scope)
2. apps/goldfinch/WORKING-CONTEXT.md
3. apps/goldfinch/.rules/00-conviction-primacy.md AND .rules/01-mem-synthesis-discipline.md (apply explicitly)
4. apps/goldfinch/knowledge/analyses/INTEGRATION_ANALYSIS_mem-delta-since-2026-04-17_2026-05-28.md (SYNTHESIS LAYER 2 — candidates, not facts; cite source layer)
5. apps/goldfinch/planning/handoffs/TRACK_D_RETURN_bible-data-architecture_2026-05-28.md (Lovable-needs-to-know-now split)
6. apps/goldfinch/knowledge/research/competitive/lit-bible-v0-capture-2026-05-27/LIT_BIBLE_V0_CAPTURE_2026-05-27.md (§12 binding decisions)

Definition of success (weekend tracer bullet — do NOT over-scope): a shareable Lovable URL that reads as a SPIRITUAL FORMATION app, not a Bible reader. (a) thin reader: pick passage, read, switch translation, save a verse — no word-tap/Strong's/highlighting; (b) Affirmations widget surface (C4) as lead formation feature; (c) home/dashboard IA presenting formation + reader as siblings.

Sequence: (1) light routing pass, default to backlog; (2) V1 scope matrix → HTML dashboard (CHECKPOINT 1); (3) first Lovable prompt (thin slice) → 3-lens adversarial gap-check (blind Agent reviewers: conviction-primacy / mem-synthesis / formation-framing+no-Capacitor-leak+not-over-scoped vs the canonical docs) → HTML dashboard (CHECKPOINT 2) → Brandon sends.

Discipline: Strategist boundary holds; markdown is the builder contract, HTML is the human surface; cite source layers; don't autopilot routing into the matrix.
Copied