zugzbot 1.0.21 → 1.0.23

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/senior-dev/src/components/Pagination.tsx

@@ -1,4 +0,0 @@
-import { forwardRef } from 'react';
-export const Pagination = forwardRef<HTMLDivElement, { className?: string }>(function Pagination({ className = '' }, ref) {
-  return <div ref={ref} className={`px-3 py-2 rounded-md ${className}`} data-component="Pagination">Pagination</div>;
-});

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/senior-dev/src/components/Popover.tsx

@@ -1,4 +0,0 @@
-import { forwardRef } from 'react';
-export const Popover = forwardRef<HTMLDivElement, { className?: string }>(function Popover({ className = '' }, ref) {
-  return <div ref={ref} className={`px-3 py-2 rounded-md ${className}`} data-component="Popover">Popover</div>;
-});

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/senior-dev/src/components/Select.tsx

@@ -1,4 +0,0 @@
-import { forwardRef } from 'react';
-export const Select = forwardRef<HTMLDivElement, { className?: string }>(function Select({ className = '' }, ref) {
-  return <div ref={ref} className={`px-3 py-2 rounded-md ${className}`} data-component="Select">Select</div>;
-});

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/senior-dev/src/components/Sheet.tsx

@@ -1,4 +0,0 @@
-import { forwardRef } from 'react';
-export const Sheet = forwardRef<HTMLDivElement, { className?: string }>(function Sheet({ className = '' }, ref) {
-  return <div ref={ref} className={`px-3 py-2 rounded-md ${className}`} data-component="Sheet">Sheet</div>;
-});

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/senior-dev/src/components/Spinner.tsx

@@ -1,4 +0,0 @@
-import { forwardRef } from 'react';
-export const Spinner = forwardRef<HTMLDivElement, { className?: string }>(function Spinner({ className = '' }, ref) {
-  return <div ref={ref} className={`px-3 py-2 rounded-md ${className}`} data-component="Spinner">Spinner</div>;
-});

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/senior-dev/src/components/Tabs.tsx

@@ -1,4 +0,0 @@
-import { forwardRef } from 'react';
-export const Tabs = forwardRef<HTMLDivElement, { className?: string }>(function Tabs({ className = '' }, ref) {
-  return <div ref={ref} className={`px-3 py-2 rounded-md ${className}`} data-component="Tabs">Tabs</div>;
-});

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/senior-dev/src/components/Toast.tsx

@@ -1,4 +0,0 @@
-import { forwardRef } from 'react';
-export const Toast = forwardRef<HTMLDivElement, { className?: string }>(function Toast({ className = '' }, ref) {
-  return <div ref={ref} className={`px-3 py-2 rounded-md ${className}`} data-component="Toast">Toast</div>;
-});

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/senior-dev/src/components/Tooltip.tsx

@@ -1,4 +0,0 @@
-import { forwardRef } from 'react';
-export const Tooltip = forwardRef<HTMLDivElement, { className?: string }>(function Tooltip({ className = '' }, ref) {
-  return <div ref={ref} className={`px-3 py-2 rounded-md ${className}`} data-component="Tooltip">Tooltip</div>;
-});

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/senior-dev/src/components/VirtualList.tsx

@@ -1,4 +0,0 @@
-import { forwardRef } from 'react';
-export const VirtualList = forwardRef<HTMLDivElement, { className?: string }>(function VirtualList({ className = '' }, ref) {
-  return <div ref={ref} className={`px-3 py-2 rounded-md ${className}`} data-component="VirtualList">VirtualList</div>;
-});

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Alert.tsx

@@ -1,3 +0,0 @@
-export function Alert({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Alert</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Avatar.tsx

@@ -1,3 +0,0 @@
-export function Avatar({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Avatar</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Badge.tsx

@@ -1,3 +0,0 @@
-export function Badge({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Badge</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Button.tsx

@@ -1,3 +0,0 @@
-export function Button({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Button</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Card.tsx

@@ -1,3 +0,0 @@
-export function Card({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Card</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Input.tsx

@@ -1,3 +0,0 @@
-export function Input({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Input</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Modal.tsx

@@ -1,3 +0,0 @@
-export function Modal({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Modal</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Spinner.tsx

@@ -1,3 +0,0 @@
-export function Spinner({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Spinner</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Tabs.tsx

@@ -1,3 +0,0 @@
-export function Tabs({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Tabs</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/components/Toast.tsx

@@ -1,3 +0,0 @@
-export function Toast({ className = '' }: { className?: string }) {
-  return <div className={`bg-brand-500 text-white px-4 py-2 rounded ${className}`}>Toast</div>;
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/src/styles/globals.css

@@ -1,25 +0,0 @@
-@tailwind base;
-@tailwind components;
-@tailwind utilities;
-
-:root {
-  --bg: #ffffff;
-  --fg: #0f172a;
-  --muted: #64748b;
-  --border: #e2e8f0;
-  --primary: #4f46e5;
-  --primary-fg: #ffffff;
-  --space-1: 4px;
-  --space-2: 8px;
-  --space-3: 16px;
-  --space-4: 24px;
-  --space-5: 32px;
-}
-
-@media (prefers-color-scheme: dark) {
-  :root {
-    --bg: #0a0a0a;
-    --fg: #f5f5f5;
-    --border: #2a2a2a;
-  }
-}

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/vibe-coder/tailwind.config.ts

@@ -1,28 +0,0 @@
-import type { Config } from 'tailwindcss';
-
-const config: Config = {
-  content: ['./src/**/*.{ts,tsx}'],
-  theme: {
-    extend: {
-      colors: {
-        brand: {
-          50: '#eef4ff',
-          100: '#dde7ff',
-          500: '#4f46e5',
-          600: '#4338ca',
-          700: '#3730a3',
-          900: '#312e81',
-        },
-        accent: '#22c55e',
-      },
-      fontFamily: {
-        sans: ['Inter', 'system-ui', 'sans-serif'],
-      },
-      borderRadius: {
-        DEFAULT: '6px',
-      },
-    },
-  },
-};
-
-export default config;

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/iterations/i1-findings.md

@@ -1,51 +0,0 @@
-# i1-findings.md — Iteration 1 (unit-test driven)
-
-**Date:** 2026-04-29
-**Method:** Wrote 4 v4 test suites (context-detect, code-introspect, signal-classifier, budget-tracker, memory-foldin) that exercise actual modules against persona fixtures. Ran. Observed failures. Applied fixes. Re-tested.
-
-## Bugs found and fixed
-
-### CRITICAL
-
-1. **Trailing `</content></invoke>` in fixture/test files** (5+ files)
-   - **Symptom:** package.json invalid JSON; tests fail at parse time.
-   - **Root cause:** XML-style tag artifacts leaked into Write tool content (recurring issue across the project).
-   - **Fix:** Python script strips both forms (`</content>\n</invoke>` and bare `</content>`/`</invoke>`).
-   - **Lesson:** Need a *pre-commit hook* (or build-time guard) that rejects any source file containing these tokens. Queued.
-
-2. **`founder` fixture not actually empty** — empty subdirs (src/components, src/styles, src/app) auto-created with the other personas falsely triggered `is_frontend = true`.
-   - **Fix:** rmdir empty subdirs.
-   - **Lesson:** fixture creation should be persona-specific, not bulk mkdir.
-
-### MAJOR
-
-3. **`signal-classifier` Korean directive regex** — `\b` word boundary fails for Korean characters in default ASCII regex (e.g., `\b줘\b` won't match in "만들어줘").
-   - **Fix:** Drop `\b` for Korean tokens, keep for English tokens only.
-   - **Lesson:** Always test regex with Korean text in unit tests.
-
-4. **`signal-classifier` persona heuristic threshold** — `wordCount >= 25` was too aggressive for Korean (Korean text packs more meaning per word due to particles).
-   - **Fix:** `isLongAnswer = wordCount >= 18 || charCount >= 80`. Also expanded `designVocab` and `opinionVocab` with Korean equivalents.
-   - **Lesson:** Word count alone is locale-dependent; always include char count fallback.
-
-### MINOR
-
-5. **`vitest.config.ts` missed `test/v4/`** — new test directory not in include glob.
-   - **Fix:** Added `test/v4/**/*.test.{ts,mjs}`.
-
-## Test counts
-
-| Before iter 1 | After iter 1 |
-|---|---|
-| 505/505 (no v4 tests run) | 545/545 (40 v4 tests added, all green) |
-
-## What this iteration *cannot* catch
-
-These tests are *unit-level*. They don't catch:
-- Hook contracts in real Claude Code session (still untested)
-- Master prompt actually following 8 rules under real model
-- dembrandt CLI integration with real URL
-- OMD-PLAN.md flow end-to-end
-- Persona drift across multi-turn dialogue
-
-→ **Iteration 2** dispatches audit subagents to walk through scenarios prompt-by-prompt.
-

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/iterations/i2-audit-F.md

@@ -1,133 +0,0 @@
-# i2-audit-F.md — Persona F (Solo Founder / Indie Hacker)
-
-**Date:** 2026-04-27
-**Persona:** F — empty repo, "Stripe like" reference, zero design vocab, 5-minute patience cap, will smash `go` on every default, wants handoff zip ASAP for v0/Cursor.
-**Scenario:** `omd install-skills` → `claude` → `/omd-harness Stripe 같이 만들어줘 결제 SaaS 메인`
-**Fixture:** `/Users/kwakseongjae/Desktop/projects/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/fixtures/founder/README.md` (empty repo, billing-saas, references stripe.com/atlas)
-
----
-
-## Walk-through findings
-
-### Step 1 — Skill registration
-
-`omd install-skills` then `claude` in same session = **subagent NOT registered**. Skill SKILL.md §0 "Subagent registration 확인" tells user to run `/agents` to force re-scan, or restart Claude. Founder F sees this and patience drops *instantly*.
-
-- **MAJOR** — first turn for F is not "design", it's "your harness isn't loaded, restart Claude." That's a 30-60s detour for a persona with 5-minute total budget. Worth it? Probably yes (no false-success), but the message buries the lede. Reorder: "1-line action: run `/agents` now" then explanation underneath.
-- **MINOR** — message is bilingual-mixed Korean. F just typed Korean so OK, but no English fallback if F's terminal is English locale.
-
-### Step 2 — INTAKE / context-detect on empty folder
-
-Master spec (omd-master.md §1 INTAKE): `empty folder → SLOT_GATE (greenfield mode)`. Good. No code-introspection wasted on empty repo.
-
-But the task string `Stripe 같이 만들어줘 결제 SaaS 메인` contains a brand hint and a domain. SKILL.md §2.b says skill (launcher) does silent keyword extraction → catalog match → presents top-1 with `go` shortcut. **This is the right move for F.**
-
-Question: does master also do this, or only the skill? Reading SKILL.md again — yes, *the skill itself* does the matching (skill-side, no API key). Master is spawned only after `go`/ref id is chosen. So F sees:
-
-```
-DESIGN.md가 없어서 reference 한 개를 골라 부트스트랩할게요. … stripe가 가장 잘 맞을 것 같아요 …
-이대로 가시려면 go (또는 stripe).
-다른 후보: ramp · linear · vercel · …
-```
-
-F types `go`. **Wall time: ~2-5s for skill keyword match (in-head, deterministic)**. Within budget.
-
-- **CRITICAL question — catalog id validation**: SKILL.md §2.d explicitly requires `id ∈ items[].id` of `reference-fingerprints.json`. I verified `stripe` exists (grep confirms 1 hit). ✓ Validated. **No hallucination risk for this path.**
-- **MINOR** — top-5 list is shown even when top-1 score is overwhelmingly higher (brand hint = +5 pts, "stripe" → "stripe"). For a smash-`go` persona, lengthy alternatives are noise. Could collapse to "stripe (go) · 다른 후보 보기" expander.
-
-### Step 3 — URL extraction / dembrandt
-
-The task does NOT include a URL — it includes brand name "Stripe". The fixture README *mentions* `https://stripe.com/atlas` but harness skill does NOT scan README for URLs (verified — SKILL.md §0–§2 only reads task arg + catalog). So **dembrandt is never invoked for F's flow**. Catalog match supersedes.
-
-- **No CRITICAL** — dembrandt failure is irrelevant to F because catalog match wins first.
-- **However:** if F had typed a URL instead (`/omd-harness https://stripe.com 결제 SaaS`), master would hit `URL_EXTRACT` branch. Reading `url-token-extractor.ts`:
-  - `tryDembrandt` runs `npx -y dembrandt@latest extract "<url>" --format dtcg --output -` with **90s timeout**.
-  - **CRITICAL — 30s patience violation**: 90s is 3× F's patience cap. Cold `npx -y` install on first run pulls Playwright + Chromium = often >60s on cold cache. F abandons before output.
-  - **CRITICAL — fallback unimplemented**: `extractUrl` returns `dtcg: null` with warning `'Fallback extraction unimplemented — please install Playwright or use a different reference (catalog id)'`. The fallback path simply tells the user to go install Playwright. For F, that's a hard stop.
-  - **CRITICAL — failure handling not visible in master**: omd-master.md mentions `omd extract <url> --internal` as a future helper (§9 "Phase B helper, future"). Master spec does not specify what state transition happens when `tool: 'fallback'` and `dtcg: null` come back. Best-case: master falls back to catalog match on the URL hostname (`stripe.com` → `stripe`). Worst-case: master writes a placeholder reference and marches on. Neither is documented. **GAP.**
-- F's actual scenario (brand name, not URL) sidesteps all of this. But the audit brief explicitly asked, so flagging.
-
-### Step 4 — Catalog id verification
-
-Confirmed: `stripe` is one of 67 ids (`grep -c '"id"' = 67`, `"id": "stripe"` present). Skill spec mandates verification before recommendation (§2.d "검증 — hallucination 방지"). ✓
-
-The 999999-rule ("Never invent reference ids") in master also re-asserts this. Two-layer guard. Good.
-
-### Step 5 — Wall time to OMD-PLAN.md
-
-Best-case path for F (smashing `go` on every default):
-
-1. Skill pre-flight + catalog match: **2-5s**
-2. F types `go` for ref pick: user time, ignore
-3. Master spawn #1 (INTAKE → SLOT_GATE → ASK_TEST): cold spawn ~5-15s + opus reasoning ~10-20s = **15-35s**
-4. Phase 1 round 1 = 1 picker (audience/tone/actions/anti — 4 questions in one turn per RULE 2). F smashes defaults: user time
-5. Master spawn #2 (CLASSIFY_SIGNAL → SLOT_GATE → either ASK_TEST round 2 OR FAST_EXIT). Persona F (`detected_persona_signal=F`, budget cap=10) typically gets second round: **15-35s**
-6. F smashes again. Master spawn #3 → PROPOSE_PLAN writes OMD-PLAN.md: **10-25s**
-
-**Estimated wall time to OMD-PLAN.md: 60-120s of model time + 3 user smashes**. Each spawn could exceed 30s on opus-with-1M ctx for cold cache.
-
-- **MAJOR — per-turn 30s risk**: Opus cold spawn + reading `.omd/state.md` + `.omd/timeline.md` + `.omd/runs/INDEX.md` + `.omd/preferences.md` + `.omd/context.json` on every spawn = lots of Reads. For an empty `.omd/` (first run, all files missing or empty), that's just failed Reads (fast) → OK. But spec doesn't guard the missing-file case explicitly; if Read errors abort the master turn, F is dead. **Verify: does master gracefully no-op on missing `.omd/state.md`?** Spec says "auto-injected by SessionStart hook — already in your prompt" (omd-master.md §0.1). For first-ever run, the hook may not have populated it. **GAP — needs robustness check.**
-- **MAJOR — 8-16 turn average is too long for F**: omd-master.md frontmatter promises "8-16 turn 평균 (페르소나 적응)". RULE 5 budget says F=10 cap. Even 8 turns × 20s = 160s = 2.6min just on model latency. F's 5-min hard cap is at risk.
-- **OPPORTUNITY**: a *true* F-FAST path would skip SLOT_GATE entirely on detected F-signal — go straight from INTAKE → PROPOSE_PLAN with all defaults filled (`audience="[FILL IN]"`, `tone_seed=stripe`, `exit_scope=handoff-zip`, `a11y_floor=AA`, `asset_policy=all-auto`). That's 1 spawn to plan. Not implemented.
-
-### Step 6 — Defaults
-
-Reading `plan-emitter.ts` defaults (PlanInputs):
-- `exit_scope='wireframe-and-spec'` is the documented default. **F wants `handoff-zip`** (zip for v0/Cursor). Mismatch.
-- `a11y_floor='AA'` ✓
-- `asset_policy='all-auto'` ✓
-- `audience='[FILL IN]'` — placeholder. F may not edit, leaves it. DESIGN.md §11–13 will be `[FILL IN]` per 999-rule.
-
-- **CRITICAL — default exit_scope wrong for F**: F profile says "gives wireframe + DESIGN.md to v0/Cursor immediately." That's `handoff-zip`, not `wireframe-and-spec`. Smash-`go` F gets the smaller package and has to re-run with explicit scope. Two-roundtrip = patience burn.
-  - **Fix**: persona-detected F → default exit_scope='handoff-zip'. Or: make the smash-`go` choice include "handoff-zip" as the recommended default in the picker.
-- **MINOR — `[FILL IN]` audience leaks into DESIGN.md**: Spec says §11–13 placeholders OK (999-rule), but for F who'll hand to v0 directly, v0 doesn't know what to do with `[FILL IN]`. v0 will either ignore or hallucinate. Better default: synthesize a generic-but-plausible audience from the task ("결제 SaaS 메인" → "B2B SaaS 결제 운영자, finance-ops 담당, 30-40대"). Master could spawn omd-ux-researcher silently for 1 plausible persona. Trade speed for acceptance.
-
-### Step 7 — Final handoff zip
-
-Documented at `.omd/runs/run-…/handoff/v0.zip` (skill SKILL.md §"산출물 위치"). Master Phase 9 emits it (per plan-emitter §7). 
-
-Does v0 actually accept it? Spec doesn't say. Looking at `manifest.json` + `wireframes/*.md` + `DESIGN.md.patch` — that's a code-LM-friendly bundle but **v0 specifically expects either prompt text or a single tailwind/jsx component**. A zip with a patch file + ASCII wireframes is a Cursor/Claude Code shape, not a v0-paste shape.
-
-- **MAJOR — v0 acceptance unverified**: zip is named `v0.zip` but no test confirms v0 (vercel) actually consumes it. Likely founder must unzip → copy DESIGN.md.patch contents → paste into v0 prompt manually. Acceptable but spec oversells. Rename to `v0-bundle.zip` and document the paste procedure, OR generate a single `v0-prompt.md` (concatenated brief + DESIGN.md text) inside the zip optimized for paste.
-- **MINOR** — `cursor.zip`/`subframe.zip` also produced; for F who only uses one tool, that's 3× the artifact noise. Default to single zip matching detected persona (F → v0).
-
----
-
-## Findings table
-
-| Sev | # | Finding | Step |
-|---|---|---|---|
-| CRITICAL | 1 | dembrandt 90s timeout > F's 30s patience cap; URL path effectively unusable for F | 3 |
-| CRITICAL | 2 | dembrandt fallback unimplemented — returns null + "install Playwright" message | 3 |
-| CRITICAL | 3 | Master state transition on `tool='fallback'` + `dtcg=null` undocumented | 3 |
-| CRITICAL | 4 | Default `exit_scope='wireframe-and-spec'` mismatches F's need (`handoff-zip`); smash-`go` gets wrong package | 6 |
-| MAJOR | 5 | First turn after `omd install-skills` likely needs `/agents` re-scan — bilingual error message buries the 1-line action | 1 |
-| MAJOR | 6 | `.omd/state.md` etc. may be missing on first run; master spec assumes hook populated them — no missing-file robustness | 5 |
-| MAJOR | 7 | 8-16 turn avg + opus cold spawn ≈ 2-3min model time — flirts with F's 5-min hard cap | 5 |
-| MAJOR | 8 | F-FAST path (INTAKE → PROPOSE_PLAN, all defaults) not implemented; persona detection happens but doesn't shorten the loop | 5 |
-| MAJOR | 9 | `v0.zip` not verified to be paste-into-v0-shape; bundle is Cursor-flavored | 7 |
-| MINOR | 10 | Top-5 alternatives shown even when brand match is overwhelming — noise for smash-`go` user | 2 |
-| MINOR | 11 | `audience=[FILL IN]` propagates to DESIGN.md §13; v0 hallucinates — synthesize plausible default instead | 6 |
-| MINOR | 12 | 3 zips emitted (v0/cursor/subframe); F only needs one — match to detected persona | 7 |
-| GOOD | — | Catalog id `stripe` exists; two-layer hallucination guard (skill §2.d + master rule 999999) | 4 |
-| GOOD | — | Skill-side keyword/brand match avoids spawning master for ref-pick (saves 1 spawn ≈ 15-30s) | 2 |
-| GOOD | — | Empty-folder INTAKE branch skips code-introspection waste | 2 |
-
----
-
-## Verdict
-
-**FAIL for F as currently specced** — 4 CRITICAL items, 5 MAJOR.
-
-The dominant blockers are #4 (wrong default scope) and #8 (no F-FAST path). #4 is a 1-line fix in plan-emitter (or persona-conditional default). #8 is a real spec change: master needs an explicit `INTAKE → PROPOSE_PLAN_FAST` edge that fires when `detected_persona_signal='F'` AND task contains a brand hint. With those two fixes, F flow becomes:
-
-1. `/omd-harness Stripe 같이 만들어줘 결제 SaaS 메인` → skill matches `stripe` (3s)
-2. F: `go` → master single spawn → OMD-PLAN.md with all-defaults including `exit_scope=handoff-zip` (~25s)
-3. F: `go` → DESIGN_GENERATION pipeline runs (Phase 2-9) → handoff zip emitted (~3-5min model time, but F is fine because it's *one* wait, not 8 turns)
-4. F gets `v0-bundle.zip` and pastes into Cursor.
-
-That's 2 user actions, 1 plan-review checkpoint, 1 ship-gate checkpoint. Roughly 4-7min total — survivable.
-
-The dembrandt items (#1-3) don't block F's primary flow but represent latent danger if F ever supplies a URL. Defer to iteration 3+ unless URL-flow is on the v4 critical path.
-
-**Recommendation:** Block v4 GA on items #4, #5, #6, #8. Items #1-3 → backlog with a `requires-Playwright` env hint added now to prevent silent-fail. Items #9-12 → polish post-GA.

package/.opencode/oh-my-design/skills/omd-lab-02-design-harness/runs/v4-self-test/iterations/i2-audit-J.md

@@ -1,171 +0,0 @@
-# i2-audit-J.md — Persona J (Junior Designer pivoting to dev) walk-through
-
-**Date:** 2026-04-27
-**Auditor persona:** J — Figma-fluent, design-vocab sensitive, expects 10-15 turn depth, will reject fluff & vocabulary errors.
-**Scenario:** `/omd-harness https://figma.com/file/abc...` against a folder with minimal Next.js code + `@figma/code-connect` in devDependencies + no DESIGN.md.
-
----
-
-## Step 1 — `/omd-harness https://figma.com/file/abc...`
-
-**Path:** `skills/omd-harness/SKILL.md:17-26` extracts the trailing arg as `task`.
-
-The harness skill captures the Figma URL as the literal task string. Nothing in the SKILL.md treats `https://figma.com/...` as a *URL signal* vs free-text intent. There is **no URL-shape detection** in the launcher prior to handing to master.
-
-**Verdict:** Launcher passes URL through, but blind. Master will be the first thing that has to recognize a URL.
-
----
-
-## Step 2 — Does master DETECT the URL? Does it call url-token-extractor? Does the extractor handle figma.com (vs stripe.com)?
-
-### CRITICAL — Figma URL is a documented branch but has zero implementation
-
-`agents/omd-master.md:56` lists the INTAKE branch:
-
-> `URL given in task → URL_EXTRACT, then SLOT_GATE`
-
-This is the **only** mention of `URL_EXTRACT` in the master agent. There is:
-- no state definition for `URL_EXTRACT` in the §1 state machine table (`omd-master.md:53-67`)
-- no instruction telling master to call the extractor (the listed CLI on `omd-master.md:258` is `omd extract <url> --internal` annotated `(Phase B helper, future)` — i.e., not wired)
-- no mention of `extractUrl()` from `src/core/url-token-extractor.ts` in master's tool inventory
-
-### CRITICAL — `url-token-extractor.ts` cannot handle figma.com
-
-`src/core/url-token-extractor.ts:47-89` is hard-coded to `npx -y dembrandt@latest extract "<url>" --format dtcg`. dembrandt is a **Playwright-based DOM/computed-style scraper for live web pages**. It expects to render a page and read CSS custom properties. Figma file URLs (`figma.com/file/<key>/...`) are:
-
-1. Behind auth — dembrandt has no Figma OAuth/PAT
-2. A canvas-rendered editor, not a DOM with semantic CSS — there are no `--color-*` custom props on the relevant nodes
-3. Served as the Figma editor SPA, not the design surface
-
-The proper extractor for `figma.com/file/...` URLs is the **Figma REST API** (`/v1/files/<key>` + `/v1/files/<key>/styles` + `/v1/files/<key>/variables`) with a `FIGMA_TOKEN`. None of this exists in the codebase. dembrandt against a Figma file URL will return either an auth-wall page or empty DTCG → fall through `url-token-extractor.ts:192` to `'Fallback extraction unimplemented'`.
-
-So even *if* master calls `extractUrl()`, the result on a Figma URL is guaranteed to be `tool: 'fallback', dtcg: null, warnings: ['dembrandt failed: ...', 'Fallback extraction unimplemented...']`.
-
-### MAJOR — `@figma/code-connect` detected but never used as a token source
-
-`src/core/context-detect.ts:104` notices `@figma/code-connect` in devDependencies and uses it **only** as a heuristic for persona=J. There is no follow-up action like "scan `figma.config.json` / `*.figma.tsx` files for component bindings" — which is exactly what code-connect provides (component → Figma node mapping that is locally readable, no token needed). This is a missed extraction opportunity that is *strictly available offline*. As J I notice this immediately: "you knew I had code-connect, why didn't you parse the config?"
-
-**Verdict:** Master has no live Figma path. The architecture has a stub branch (`URL_EXTRACT`), a generic web-DOM extractor (dembrandt), and a persona heuristic — none of these compose into "process a Figma URL." This is the central missing functionality for persona J.
-
----
-
-## Step 3 — Fallback to catalog matching when extraction fails
-
-`skills/omd-harness/SKILL.md:60-103` describes the catalog fallback. It is reasonably robust (loads `reference-fingerprints.json`, scores keyword/brand/category, shows top-5). The 67-entry fingerprints file is intact (verified — `grep -c '"id"' = 67`, `airbnb`/`apple`/etc present with `tone_keywords`, `visual_theme`, `antipatterns`).
-
-But the fallback has **no Figma-specific branch**. The skill does not say "if URL is figma.com and extraction returns null, ask the user to paste a screenshot OR provide the Figma PAT OR point at the local `figma.config.json`." It just degrades to "pick a catalog reference" — which to J feels insulting (*"I gave you my actual design system; you handed me a list of other companies"*).
-
-There is also a perverse incentive: `data/reference-fingerprints.json:6` lists a `figma` reference id, and `web/references/figma/DESIGN.md` exists. The catalog scorer will likely surface `figma` as a top match given the Figma URL keyword — i.e., the system will recommend Figma's own brand DS *as a substitute for the user's Figma file*. That is wrong-shaped: J shared their own design source, not a request to imitate Figma.com.
-
-**Verdict:** Catalog fallback exists and works mechanically, but the *semantics* on a Figma URL is misleading. MAJOR.
-
----
-
-## Step 4 — Long thoughtful answer ("저는 spacing scale로 8pt grid 쓰고, type ramp는 modular 1.25 비율로...")
-
-### J classification — works (with caveats)
-
-`src/core/signal-classifier.ts:126-134`:
-- `isLongAnswer = wordCount >= 18 || charCount >= 80` — my Korean prose is ~30+ chars per clause; will trigger
-- `designVocab` regex (line 127) includes `spacing|hierarchy|호흡|여백|간격|위계` — "spacing scale", "type ramp", "modular 1.25" → **only `spacing` matches**. "type ramp" / "modular scale" / "8pt grid" are NOT in the regex. "ramp" / "grid" / "scale" / "modular" / "ratio" all absent.
-- One match is sufficient (regex.test returns true), so `personaSignal = 'J'` will fire — *if* `isLongAnswer` is true. With my example (~30 Korean chars first clause, ~80+ total) it should flip to J.
-
-But notice: the regex misses the **most distinctively J vocabulary** (8pt grid, modular scale, type ramp, line-height, leading, tracking, x-height, kerning, optical adjustment). A Junior who *only* uses these terms (without `spacing`, `padding`, `margin`, `호흡`, `여백`, `위계`) will be misclassified as `unclear` or even `S` (if they say `원칙` or `prefer`). The vocab list is partial.
-
-### Pacing adjustment — partially works
-
-`agents/omd-master.md:103-108` (RULE 4 Mirror & mode):
-- ≥ 20 words → respond 50-80 words, prose with options
-- Korean colloquial → respond Korean colloquial
-
-J signal does NOT change probe count — `omd-master.md:109-110` (RULE 5) sets J cap = 12 turns. So "wants 10-15 turn depth" is *partially* honored: 12 cap is in the J ballpark but trims the upper end. There's no instruction to *increase depth of each probe* (e.g., "if J: ask about modular ratio explicitly, ask about token naming convention, ask about elevation system"). The probes remain the same generic audience/tone/anti slot as for V.
-
-### MAJOR — no J-specific question library
-
-The slot definitions (`omd-master.md:69-89`) are persona-agnostic. A Junior wants questions like "8pt or 4pt base?", "type scale ratio?", "do you want me to honor your existing radius scale or propose one?". None of these exist. Master will ask `audience` / `tone_seed` — exactly what V/F get. As J I will read this as the system not knowing what to do with my depth.
-
-**Verdict:** Classification fires (luckily), but pacing/depth payoff is thin. MAJOR.
-
----
-
-## Step 5 — Plan emit: does OMD-PLAN.md respect my taste signals?
-
-`src/core/plan-emitter.ts:10-25` defines `PlanInputs`. Notable absences for J:
-
-- No `spacing_scale` field (my 8pt grid)
-- No `type_scale_ratio` field (my 1.25 modular)
-- No `radius_scale` field
-- No `elevation_system` field
-- No `extracted_tokens` field — even if `extractUrl()` had returned a `dtcg` blob, there is **nowhere in `PlanInputs` to put it**. The plan would not surface palette/typography/depth from the URL extraction. `plan-emitter.ts:56-65` (§3 Reference & Tone) only accepts `reference_id`, `reference_url` (URL string), `tone_seed`, `reference_urls[]`. The OMD-PLAN section emits the URL as a one-liner, not as a structured token table.
-
-`plan-emitter.ts:128-132` writes a hidden HTML comment with `persona_signal` — internal only, never surfaces to the user. Plan is identical for J vs V vs F vs S except for that comment. The comment promises adaptation that the file does not deliver.
-
-### MAJOR — fabrication risk
-
-If master *did* run dembrandt on the Figma URL and got `dtcg: null`, but then drafted a plan with "tone_seed: editorial-warm" by guessing from the URL slug, that violates `omd-master.md:303-304` rule "Never fabricate §11–13 facts." There is no guard in `plan-emitter.ts` ensuring `tone_seed` traces to either an actually-extracted token or a user-stated keyword. As J I will catch this on read — and reject.
-
-**Verdict:** Plan ignores extracted tokens (because there's no schema for them) and ignores design-vocab depth I shared. The plan is shaped for V, not J. MAJOR.
-
----
-
-## Step 6 — ui-junior generates wireframes — do they cite tokens correctly?
-
-`agents/omd-ui-junior.md:43-71` mandates `## Tokens cited` with `§<n>` references and rule on line 121: "Never introduce a token not in DESIGN.md."
-
-This is solid IF DESIGN.md has tokens. But in J's flow, DESIGN.md does not yet exist (Phase 5 generates it via `omd init prepare --ref <id>`). The only `<id>` that can be passed is a **catalog reference id** (because URL extraction failed, see Step 2). So the Phase 5 patch will populate §2/§3 with, e.g., Toss Blue `#3182f6` from the toss reference — *not* from my Figma file.
-
-`omd-ui-junior.md:39` example wireframe even hard-codes `Toss Blue (#3182f6)` as the example. Junior will read my Figma palette as irrelevant. The wireframes will cite tokens correctly *to a DESIGN.md that contains the wrong tokens*.
-
-There is no rule like "if user provided a Figma URL, before Phase 5 emit, confirm token table matches user's source." Phase 5 has a mandatory gate (`omd-master.md:219`) — but as J I'd have to spot the mismatch in the patch myself.
-
-**Verdict:** Citation discipline is fine. The upstream substitution of catalog tokens for Figma tokens is what bites. CRITICAL because it pretends to know my design system while citing someone else's.
-
----
-
-## Step 7 — microcopy — is §10 Voice respected from my Figma extraction?
-
-`agents/omd-microcopy.md:21-23` reads §10 Voice from DESIGN.md. §10 in the catalog-matched DESIGN.md will be the *catalog reference's* voice fingerprint (e.g., toss `voice_fingerprint.register = "trustworthy-neutral"` — see `data/reference-fingerprints.json`). Not mine.
-
-`agents/omd-microcopy.md:54-58` Korean rules ("Honorific level follows §10 — never mix") would apply that catalog register strictly. If I am J building a playful internal tool, but caught the toss reference fallback, I'll get formal honorific microcopy with "송금하기" archetypes. The microcopy agent has no path to read tone signals from a Figma file's text-style names (which often encode voice — e.g., `display/serious` vs `display/playful`).
-
-There is no §10 Voice extraction from URL in `url-token-extractor.ts` either — the extractor maps DTCG to §1/§2/§3/§6 only (`url-token-extractor.ts:139-149`), with `section_7_dos_donts: '[user fills]'` and no §10. So even on the happy path, voice never propagates from the URL.
-
-**Verdict:** Microcopy rigor is high *within* the §10 it sees, but §10 it sees is the catalog substitute, not mine. MAJOR.
-
----
-
-## Cross-cutting findings
-
-### CRITICAL
-1. **Figma URL handling is a stub.** `omd-master.md:56` lists `URL_EXTRACT` as a branch but nothing implements it; `url-token-extractor.ts` only knows dembrandt (Playwright DOM scraper) which cannot see Figma file canvases without the Figma REST API + token. (`agents/omd-master.md:56`, `src/core/url-token-extractor.ts:47-89`)
-2. **`@figma/code-connect` detected but unused as a token source.** Persona heuristic only. No parser for `figma.config.json` / `*.figma.tsx` bindings, which would be the offline win. (`src/core/context-detect.ts:104-107`)
-3. **Catalog fallback is semantically wrong on a Figma URL.** Will likely top-rank the `figma` catalog reference (Figma.com's own DS) as a substitute for the user's Figma file. (`skills/omd-harness/SKILL.md:76-83` + `data/reference-fingerprints.json` containing `figma` id)
-4. **`PlanInputs` schema has no slot for extracted tokens.** Even if extraction worked, the plan cannot surface palette/typography/depth structurally — only as a URL string. (`src/core/plan-emitter.ts:10-25`)
-
-### MAJOR
-5. **`signal-classifier` designVocab regex misses core J vocabulary.** `8pt grid`, `type ramp`, `modular scale`, `tracking`, `leading`, `line-height`, `x-height`, `radius scale`, `elevation` not present. (`src/core/signal-classifier.ts:127`)
-6. **No J-specific question library.** Slot table is persona-agnostic; J gets the same `audience`/`tone_seed` probes as V. (`agents/omd-master.md:69-89`)
-7. **`personaSignal` written to plan as HTML comment only** — promises adaptation that the rendered plan doesn't deliver. (`src/core/plan-emitter.ts:128-132`)
-8. **Phase 5 patch will cite catalog tokens as if they were user's.** No "user's source confirms token table" gate before patch. (`agents/omd-master.md:219`, `agents/omd-ui-junior.md:39`)
-9. **Microcopy §10 Voice never propagates from URL.** `dtcgToOmdPreview` maps §1/§2/§3/§6 only; §7/§8/§10 marked `[user fills]` / `[needs ...]`. (`src/core/url-token-extractor.ts:139-149`)
-10. **`reference-fingerprints.json` voice_fingerprint is sometimes null** (e.g., airtable, line 40) — so microcopy via catalog may have nothing to cite; agent-side handling unclear (`omd-microcopy.md:62-64` says halt if §10 missing, but doesn't address nullish fingerprint).
-
-### MINOR
-11. `agents/omd-master.md:283-296` trace example dated `2026-04-28` — fine, but cosmetic.
-12. `src/core/visual-anchor.ts:36-80` ASCII variant `C-airy` line 71 has trailing-whitespace asymmetry (`(water)     │` vs siblings) — purely cosmetic.
-13. `omd-microcopy.md:64` typo-ish: "Run omd init Phase 4.5" — there is no documented Phase 4.5 in `omd-master.md` (Phase 4 → Phase 5).
-14. `omd-master.md:307` "Never invent reference ids — only the 67 in `reference-fingerprints.json` are valid" — verified count is 67. Good. But no analogous rule for "never invent tokens from a URL extraction that returned null."
-
-### Design-vocab nits I'd call out as J
-- `agents/omd-microcopy.md:50` lists `"seamless", "powerful", "robust", "delightful"` as forbidden marketing fluff — agree, but `delightful` is contested (commonly used by mature DS docs e.g. Material). Minor; not a vocab error per se.
-- `data/reference-fingerprints.json:11` describes Airbnb as `"warm, photography-forward"` with `tone_keywords: ["warm", ...]` — consistent with #ff385c Rausch. No "warm-applied-to-cool-palette" error caught in the spot-checks I did. Good.
-
----
-
-## Verdict
-
-**REJECT.** The advertised entry point (`/omd-harness <figma-url>` against a Next.js + `@figma/code-connect` repo) is the highest-value scenario for persona J — and it is the worst-served path in the system. URL_EXTRACT is a stub; the only extractor (dembrandt) cannot see Figma files; code-connect bindings are detected but discarded; catalog fallback semantically substitutes someone else's brand DS for mine; PlanInputs has no schema for extracted tokens; J vocabulary is half-present in the classifier; downstream Phase 5/6/7 will cite catalog tokens as if they were mine. As J I would close the terminal at Phase 5 patch review.
-
-**To unblock J:** (1) implement a `figma:` URL branch using the Figma REST API + `FIGMA_TOKEN` env var, (2) implement a `figma.config.json` / code-connect parser as a no-token offline path, (3) extend `PlanInputs` with a `extracted_tokens` block, (4) extend `designVocab` regex with `8pt|4pt|grid|ramp|modular|leading|tracking|line.height|x.height|radius.scale|elevation`, (5) add J-tier slot questions (base unit, type ratio, radius scale, elevation system) gated on `inferred_persona_signal === 'J'`, (6) add a Phase 5 pre-patch gate that diffs proposed §2/§3 tokens against any user-provided source.
-
-The catalog/system below the URL layer (67 fingerprints, ui-junior token discipline, microcopy §10 enforcement, plan-as-file pattern) is genuinely good — it just doesn't reach J because Step 2 fails.