Runtime fix in live-browser.js: the 2s static-server fallback in handleAccept
now swaps the outer wrapper with the `[data-impeccable-variant="N"]` div itself
(+ display:contents), matching what live-accept.mjs writes to disk. Scope rules
anchored on the variant attribute keep matching on the non-HMR path, so the
accepted design no longer flashes unstyled until reload. Propagated to all
harness script copies.
/designing:
- §03 Polish redesigned as drenched magenta masthead: commands live in the band,
three title/description columns beneath on cream.
- §04 Maintain redesigned as architectural poster diptych: extract + document
vizzes become the hero element, caption below.
- §05 Interop section removed.
- §05 (was §06) "Pick a register" renamed to "Brand, or product." with a
two-lane hairline-divided layout and an auto-selected framing in the sub.
Live mode status: BETA → ALPHA across the periodic table, magazine spread,
and docs callout, reflecting real-world-testing readiness.
Skill bootstrap: removed the `<post-update-cleanup>` block from source/SKILL.md
(the source repo is the origin; running cleanup-deprecated here would touch
legitimate source). CLAUDE.md and AGENTS.md now document the skip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Section 7 no longer reads as default-zero: composition-sized targets,
freeform bias toward 1–2 dials on non-tiny surfaces, hard cap of four.
Cross-link freeform to §7 in the action loader; sync all harness copies.
Made-with: Cursor
Unify the design-system panel's data shape around DESIGN.md frontmatter
as the primary source of truth; the sidecar carries only what Stitch's
frontmatter schema can't (extensions + live component HTML + narrative).
Also fix a long-standing build bug that destroyed per-project config.
Shape changes:
- Server /design-system.json now returns { parsed, sidecar, hasMd,
hasSidecar, mdNewerThanJson, parseError?, sidecarError? }. No more
mode switching; both layers ship when present and the panel merges.
- Panel consolidates renderSidecarVisual + renderParsedMdVisual into a
single renderDesignVisual that merges frontmatter primitives with
sidecar extensions.colorMeta / typographyMeta. Helpers for color,
typography, radii model-building. Parsed-md narrative synthesis
survives as a fallback when no sidecar.
- DESIGN.json rewritten at schemaVersion 2: extensions.{colorMeta,
typographyMeta, shadows, motion, breakpoints} + components (with
refersTo pointing back to frontmatter component keys) + narrative.
Token primitives no longer duplicated in the sidecar.
Build fix:
- scripts/build.js:634 wiped .claude/skills/ (and every other harness
dir) on each rebuild, then recopied from dist. After commit b0feed0
unbundled per-project config.json from dist, the sync destroyed the
user's live-mode config on every build without replacing it.
- Added stashPerProjectArtifacts / restorePerProjectArtifacts in
scripts/lib/utils.js. Hoisted PER_PROJECT_SCRIPT_ARTIFACTS to a
module-level export so build.js and readSourceFiles share one
source of truth. Build now preserves config.json across the sync.
Verified in browser: panel renders 10 colors, 9 typography roles, 3
shadows, 6 grouped components, 9 rules, 25 do/don't items, all merged
correctly from frontmatter + v2 sidecar with zero console errors.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two loopholes in the critique procedure let the model shortcut past its own rigor: "SHOULD delegate" with a broad "if sub-agents are not available" escape, and "Browser visualization (when available)" framing that made the [Human] detector-overlay tab read as optional color. Both get rationalized away under context pressure even though the isolation is what makes the combined score honest and the overlay is the user-facing deliverable.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adopt the Stitch google-labs-code/design.md format's two-layer shape:
YAML frontmatter carries machine-readable tokens, prose body covers the
six canonical sections. The sidecar DESIGN.json now extends frontmatter
rather than duplicating it.
- document.md: new frontmatter section, Step 2b staging, sidecar scoped
to extensions, Do's/Don'ts now cite PRODUCT.md anti-references by name,
OKLCH vs hex presented as project posture not mandate.
- design-parser.mjs: tiny YAML-subset reader, exposes model.frontmatter,
schemaVersion bumped to 2, prose-scraping fallback intact.
- live-browser.js: "basic view" CTA copy reflects frontmatter-first model.
- DESIGN.md: add frontmatter with 10 colors, 9 type roles, 7 components;
OKLCH values direct per The OKLCH-Only Rule.
- tests/design-parser.test.mjs: coverage for no-frontmatter, Stitch-shape,
missing-terminator, comment handling.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Config drift was a real tripwire for projects with static generators: new
HTML files get added, never make it into config.files, silently skip
injection. Two additions.
config.files entries now accept glob patterns (**, *, ?) expanded via
fs.globSync in live-inject. Multi-page projects can write
["public/**/*.html"] once and never maintain the list again. New optional
exclude field filters out matched files (email templates, demo fixtures).
HARD_EXCLUDES of node_modules and .git are enforced regardless of user
config so vendor trees can never receive a tracking script.
live.mjs now runs a drift scan after inject: walks common page-source
roots (public/, src/, app/, pages/) and reports HTML files not covered
by the resolved inject targets. Respects user excludes so intentional
omissions aren't flagged. Output JSON carries configDrift: { orphans,
orphanCount, hint } or null. live.md documents the agent flow for
surfacing drift to users without auto-mutating the config.
Unbundle config.json from the distributable skill: it's a per-project
artifact, not skill code. readSourceFiles now skips any PER_PROJECT_ARTIFACTS
during source scan so build output to .claude/ .cursor/ etc never ships
one project's inject targets to another's install. The per-harness
copies stay gitignored via the existing **/skills/impeccable/scripts/config.json
rule; each consuming project writes its own on first /impeccable live.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Without comments or strokes the screenshot is pure visual anchoring,
biasing the model toward the existing rendering and fighting the
three-distinct-directions brief. Local blob still drives the shader
overlay; upload and screenshotPath are gated on annotation presence.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a coarse-controls ("Tune") popover that slides out from behind the
contextual bar via clip-path, showing 2-4 per-variant knobs (range / steps /
toggle) driven by a `data-impeccable-params` manifest. Range/toggle drive
CSS custom properties on the variant wrapper; steps toggle a data
attribute. Values reset per variant; on accept, current values are passed
through live-poll to live-accept as an `impeccable-param-values` comment
for the carbonize cleanup step to bake in.
Other live-UI work in this change:
- Theme-aware palette (barPaletteForTheme) now drives the contextual bar,
action picker, and tune popover. Dark sand on light pages, paper on
dark. Detection has a localStorage dev override for QA.
- Action picker chips get inline SVG icons (wand / bars / funnel / sparkle /
type ramp / circles / grid / devices / curve / star / bolt) stacked
above the label; selection state recolors via currentColor.
- Accept button switched to saturated site magenta with paper text.
- Cycle dots reworked: solid accent for active, neutral for arrived,
hairline ring for pending. No more magenta-on-gray noise.
- Tune chip sits in the cycling row with a count pill badge; open state
uses accentSoft bg + accent text (no ad-hoc white border).
- Popover uses the bar's palette with a deeper surface (surfaceDeep),
lives behind the bar via z-index so a 6px overlap reads as tucked under
it, and animates with clip-path inset() for reliable slide behavior.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Default browser focus-visible ring was a heavy blue outline that
clashed with the dark capsule. Inject a palette-aware inner box-shadow
ring (accentSoft + accent) scoped to the global bar buttons via a
one-time <style> tag. Keeps keyboard focus visible without the
visual noise.
Previously Escape while picking just hid the highlight and set state
to IDLE, leaving the global bar's Pick button visually active. Next
pick attempt fell into a broken state where the button looked on but
no picker ran. Route through togglePick() so the flag, the button,
the UI, and the state all flip together.
Small copy fix on Why panel 04: Figma stamp now reads "last touched
Q3 2025" (was 2024; should reflect closer to the current calendar).
Previous hardening emitted a 7-line todo array and a 10-line ASCII-bar
stderr banner on every accept event, both printed to the agent's
transcript. Per-event overhead added up fast on multi-variant sessions.
Keep the three-layer defence but shrink the per-event noise:
- todo is now a single string: "REQUIRED before next poll: carbonize
cleanup in FILE. See reference/live.md ..."
- stderr is one line with the same pointer.
- reference/live.md keeps the full five-step checklist (loaded once
per session, so its verbosity is a fixed cost — no repetition tax).
Attention signal still triple-redundant: stdout todo, stderr line,
reference section.
After an LLM-triggered session where the carbonize cleanup got skipped
entirely (the instruction was buried as a single bullet among four
cases in live.md, and `_acceptResult.handled: true` felt like a
"done" signal), add three redundant reinforcements:
1. live-accept.mjs now emits a `todo` array on the event payload when
carbonize is true, listing all five cleanup steps plus a pointer to
reference/live.md. The agent reads this as part of the event JSON.
2. live-poll.mjs prints a loud multi-line stderr banner on every
carbonize=true event. Even agents that parse only stdout JSON see
stderr output and can't treat the event as handled without action.
3. reference/live.md pulls the carbonize branch out of the "Handle
accept" bullet list into a dedicated "Required after accept
(carbonize)" section with a numbered five-step checklist, a
rationale paragraph for why skipping is bad, and an explicit
"do not poll again until the file is clean" instruction.
The three layers are deliberately redundant: a future LLM that ignores
the reference text should still be caught by the stderr banner or the
in-event todo, and vice versa.
The inline pre-restore wasn't actually fixing a timing issue — the
fix was the fonts.ready + load retries. Since live.js's own
top-level block runs before DOMContentLoaded and we can do the same
retries there, we don't need an inline script injected into every
user page. Simpler HTML, single source of truth.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
scrollTo(y) clamps to the current document.scrollHeight, which is
several hundred pixels short of the final value until async-loaded
fonts swap in (Cormorant Garamond italic grew consulting-section
layout by ~585px in the logs). The initial synchronous scroll was
clamping to ~6165 even though the Go-time target was 6749.5.
Retry on document.fonts.ready and on the window load event, both of
which fire once the document reaches its final height.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
external live.js is fetched, so by the time it runs the browser has
already queued its reload-scroll animation and history.scrollRestoration
='manual' has no effect. Inject a tiny inline synchronous <script> into
the same block live-inject writes, BEFORE the external live.js tag. It
sets scrollRestoration='manual' and does window.scrollTo(0, savedY)
during HTML parse — before the browser can animate anywhere.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
startScrollLock calls stopScrollLock at the top as a reset. I had
clearScrollY() inside stopScrollLock, so every Go sequence was:
writeScrollY(6749.5) → startScrollLock → stopScrollLock → clearScrollY
— the persisted value was wiped right after being written, so resume
after reload read null and locked to 0.
Move clearScrollY to the three genuine session-end sites (hideBar
error path, confirmed/accept, cleanup/discard). stopScrollLock no
longer touches persistent storage.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three concrete bugs from the diagnostic logs:
1. saveSession was writing scrollY alongside state, so every call during
resumeSession clobbered the Go-time value with whatever the browser
had left us at (typically 0). Move scrollY to its own localStorage
key, touched only at Go and on user-scroll reanchor.
2. history.scrollRestoration='manual' was being set inside init() at
DOMContentLoaded — by then the browser has already started animating
its restore, especially with scroll-behavior: smooth on html. Apply
it at script parse time, and apply the saved scrollY immediately
there too, before the browser's animation starts.
3. Corrections only fired on MutationObserver. A programmatic smooth
scroll (browser restore animation, or another script calling
scrollIntoView) produces zero DOM mutations — so we never caught it
walking scrollY from 0 up to 4800+ in the recorded session. Snap
back on every scroll event, gated by a 250ms user-gesture window so
we don't fight real user scrolls.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Log target-Y at Go, every mutation that triggers a correction (with the
mutation type + added nodes), every correct-or-noop (with from/to/delta),
every reanchor, and every external scroll event >5px. Lets us see which
step is actually moving the page during wrap / variant insert.
Element-based scroll tracking broke every time: Bun's HMR destroys the
target element, the browser's scroll anchoring picks a different nearby
element (e.g. the #downloads CTA) as its new anchor, and the page jumps
to wherever that surviving element is. My element-based correction then
computes against a replaced DOM node with stale / wrong geometry.
The primitive the user actually cares about is window.scrollY — they
want the page to stay where it is, regardless of which element survives
the patch. Pin scrollY directly: capture it at session start, restore it
on every mutation inside the wrapper, re-anchor on user scroll, store it
in saveSession for reload-resume.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two things were wrong. First, I capped large corrections — which was
backwards: a huge delta is exactly when we most need to restore (it
means the browser's own scroll anchoring drifted, which is what makes
the page 'jump to Get Started' when Bun's HMR destroys and re-inserts
our target). Remove the cap so any delta is corrected.
Second, the browser's built-in scroll anchoring was competing with us:
when Bun destroys our target element, the browser picks the nearest
surviving element (like a CTA anchor in another section) as its new
scroll anchor and scrolls to keep THAT stable. Disable overflow-anchor
on html and body for the duration of the session so we own scroll
entirely; restore the original values on stopScrollLock.
Kept the user-scroll grace window (400ms): wheel / touch / arrow keys
re-anchor and suppress corrections, so momentum scrolls don't get fought.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Watching document.body caught every mutation on the page — shader
animations, Bun HMR indicators, tooltips, anything — and fired a
correction on each one, which fought the user when they tried to scroll
mid-session. Now the observer only responds to mutations inside the
session's wrapper. On user scroll intent (wheel / touchstart / touchmove
/ arrow & page keys), cancel any pending rAF correction and re-anchor
to the element's new position, so momentum scrolls don't get yanked
back by a stale correction.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The reload-only fix missed the primary case: Bun's HTML loader hot-patches
the DOM in place rather than doing a full page load, so the resume
codepath never ran and the browser's scroll drifted wherever Bun's patch
left it. Likewise variant cycling (taller → shorter) and agent-driven
variant inserts both mutate layout without a reload.
Add a scroll lock: on Go (and on resume after a true reload), capture the
selected element's viewport-top and install a MutationObserver on body
that re-measures the target and corrects scroll on every batch of DOM
mutations. The target is re-resolved each pass via sessionId + visible
variant, so it survives DOM swaps that invalidate `selectedElement`.
Scroll intent events (wheel, touchstart, arrow/page keys) re-anchor to
the new position so we never fight a user who scrolls during a session.
Also set `scrollRestoration = 'manual'` at init so true reloads don't
land the user somewhere odd before our correction runs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When HMR misses and we fall back to window.location.reload(), the native
scroll restoration landed the page somewhere near the right region but
not on the selected element, because layout had shifted between the
save and the reload. Capture the element's getBoundingClientRect().top
into the session snapshot, disable native scroll restoration on resume,
and manually scroll the element back to that exact viewport-relative
position. Run a second correction pass after fonts and images settle to
absorb late layout shifts without animating the fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Screenshotting a transparent container rendered black because we were no
longer passing `backgroundColor` to modern-screenshot at all (to avoid
its `background-color !important` override on elements with their own
bg, like the teal card). That fix left elements without their own bg
rendering on a transparent canvas, which reads as black wherever the
PNG is previewed.
Now we resolve per-element: if the element has an opaque
background-color or a background-image, omit the option (element's own
bg renders, no override). If it's transparent, walk up ancestors to the
first opaque background (falling back to body/html) and pass that as
the canvas fill.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1. cleanup-deprecated: strip `i-` prefix before fingerprint lookup so
`i-harden` / `i-optimize` classify correctly (regression from the
prefixed-naming migration).
2. build: substitute `{{scripts_path}}` in reference/*.md the same way
it's substituted in SKILL.md. Previously the placeholder survived
unresolved in built reference files, so any reference that told the
agent to run a scripts path emitted a literal `{{scripts_path}}` to
the shell.
3. live-poll: drop the `undici` import. Node's built-in fetch enforces a
300s headers timeout that can't be lowered per-request, so we now cap
each poll slice at 270s and loop internally until a real event or the
caller's total timeout. Removes the hard `ERR_MODULE_NOT_FOUND`
failure when undici isn't transitively hoisted.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
modern-screenshot force-sets `background-color: X !important` on the root
clone's inline style when backgroundColor is passed, clobbering the
element's real background and rendering every captured card with the page
body color. Omit the option so the canvas stays transparent and the
element's own background renders into the foreignObject.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Surfaced during hands-on testing against a real Next 16 + Turbopack app
(EACManagement). All four compound to produce unusable live iteration
for React users; fixed bottom-up because each one blocked testing the
next.
## 1. Picker bar snaps to (0,0) on first variant arrival
In startVariantObserver, `showVariantInDOM(sessionId, 1)` hides the
original via display:none but we never re-pointed selectedElement.
Next frame, getBoundingClientRect() on the hidden original returns a
zero rect and the bar positions at (0,0). Clicking Next masked the
bug because cycleVariant already calls updateSelectedElement.
Fix: after showVariantInDOM, re-point selectedElement via
pickVariantContent(wrapper, visibleVariant) — same call the no-HMR
fallback and updateSelectedElement already use.
## 2. React NotFoundError on accept/discard (Next 16 / Turbopack)
handleAccept and cleanup both called
`wrapper.parentElement.replaceChild(...)` eagerly, before the agent's
source rewrite had propagated through HMR. That yanks children out
from under React's reconciler; when React later tries to remove/replace
the wrapper, its fiber tree no longer matches the DOM and it throws.
Fix, both paths:
- cleanup (discard): `wrapper.style.display = 'none'` so variants
disappear immediately, no structural DOM mutation.
- handleAccept: skip the eager replaceChild entirely. The accepted
variant is already the only visible child of the wrapper thanks to
the display: contents pattern; HMR cleans up the wrapper itself.
- Both paths schedule a 2s fallback replaceChild that runs only if
HMR hasn't cleaned up — keeps static-server / no-HMR flows working.
- Capture sessionId + visibleVariant in closure variables before the
1800ms cleanup timer zeros them, so the fallback still has context.
## 3. Server serves stale live.js forever
loadBrowserScripts() read live-browser.js once at startup into a
liveScript string. The /live.js handler served that cached string
with no cache headers. Every edit to the browser script was invisible
until a full server restart — silently broke the iteration loop on
fixes#1 and #2 for the user.
Fix:
- loadBrowserScripts returns { detectScript, livePath } — existence
check only, no caching.
- /live.js handler re-reads livePath on every request and prepends
__IMPECCABLE_TOKEN__ / __IMPECCABLE_PORT__ each time.
- Response headers: Cache-Control: no-store, no-cache, must-revalidate,
max-age=0 + Pragma: no-cache.
detect.js stays cached — it rarely changes during a session.
## 4. Picker stuck in GENERATING when HMR doesn't fire
The only 'done' fallback fired when arrivedVariants === 0 and called
injectVariantsFromSource, which parses raw source via DOMParser. That
can't work for TSX/JSX/Vue/Svelte — JSX expressions aren't valid HTML.
If HMR flaked or was slow (500+ line inserts on Next 16 are prone to
this), state stayed in GENERATING and the spinner ran forever.
Fix: give HMR a 2s grace window, then `window.location.reload()`.
resumeSession already counts variants off the rendered DOM on load
and transitions straight to CYCLING — reload is the universal
recovery path that works for any framework, HTML, static server,
anything.
injectVariantsFromSource is now dead code on the 'done' path. Kept
for potential pure-HTML-no-HMR future use.
## Credit
Precise repro + root-cause diagnosis from the other agent in the
EACManagement session. #2 and #4 are the high-impact ones for Next 16
/ Turbopack; #3 is the meta-fix that made iterating on #1 and #2
possible at all.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pre-prod — no need to carry forward the backwards-compat line in
SKILL.md or the historical note in CLAUDE.md. Existing PRODUCT.md
files with `register: editorial` will hit the "missing field" branch
and get re-inferred from content, which is fine for the tiny number
of projects that touched it during iteration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
"editorial" was doing semantic double duty — naming the strategic
distinction (design IS the product) AND a specific visual aesthetic
(editorial magazines, broadsheets, serif display, italic drop caps).
Models pattern-matched the aesthetic and defaulted to it on every
brand brief, producing magazine-shaped landing pages for hiking
brands, tech tools, restaurants.
The register name now describes the SURFACE KIND, not an aesthetic.
Brand covers every visual lane — tech-minimal, luxury, editorial-
magazine, consumer-warm, brutalist-grid, hand-drawn — each with
legitimate voice within the register.
## Changes
- `reference/editorial.md` → `reference/brand.md`. Content rewritten:
broadened typography guidance (pairing shapes per brand genre,
single-family commitment is valid), broadened color references
(Stripe, Vercel, Liquid Death alongside Klim, Condé Nast), added
a second slop test ("name your aesthetic lane") to prevent drift
into editorial-magazine defaults, added brand ban against the
drift itself.
- SKILL.md: register names brand/product; load brand.md.
- teach.md: register values brand/product; signals renamed; example
principles no longer use "editorial over marketing" phrasing.
- Six sub-commands (animate/bolder/colorize/delight/layout/quieter):
per-register subsections flipped Editorial: → Brand:.
- product.md: cross-references updated.
- live.md: register reference updated; density axis no longer uses
"editorial" as a synonym for "dense".
- typeset.md: per-register paragraph generalised beyond serif+sans
pairing.
- CLAUDE.md: architecture section rewritten; kept "editorial
wrapper" content-authoring term as-is (different meaning).
## Legacy handling
- `editorial` is accepted as an alias for `brand` on PRODUCT.md's
register field — agents treat it as `brand` without asking.
- Documented in SKILL.md setup section and CLAUDE.md.
## What's unchanged
- Register identification priority (task cue → surface → PRODUCT.md).
- Permission structure (brand can go big, product stays restrained).
- Shared design laws, absolute bans, color strategy vocabulary.
- Framework fixtures and tests.
Full build clean, test suite passes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Gemini 3 Flash baseline showed the Unsplash bullet wasn't directive
enough — the model still dropped imagery entirely on italian-
restaurant and vintage-moto-forum niches when the brief clearly
implied photography.
Changes:
- Added a MUST-ship-imagery lead paragraph listing the niches that
require photography (restaurant, hotel, magazine, etc.).
- Gave a literal Unsplash URL shape (`images.unsplash.com/photo-{id}?
auto=format&fit=crop&w=1600&q=80`) with real photo-id examples, so
weaker models have a concrete pattern to copy rather than inferring
the URL format.
- Promoted "zero imagery on an imagery brief" into the Editorial bans
list so it lands as a hard rule, not a nudge.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three related extraction bugs surfaced in the EAC session all rooted
in the line-based state machine:
1. `<style ... />` (JSX self-closing) had no separate `</style>` for
the "skip until close" mode to exit on, so the state machine stuck
and every `data-impeccable-variant` marker after it got missed.
Accept reported `handled: false, error: "Variant N not found"`.
2. A variant whose entire `<div ...>...</div>` sits on one line had
its body silently discarded — the marker line was `continue`d past,
and the extractor started capturing from the next line, which
usually belonged to a different variant or the wrapper close.
3. `extractCss` kept scanning for `</style>` after a self-closing
opener, greedily swallowing every subsequent variant div as "CSS".
Result: a mangled carbonize block stuffed with HTML and a duplicate
variant rendered below.
## Fix
Replaced the line-based state machine with a string-based flow:
- `stripStyleAndJoin(lines, block)` returns the wrapper text with
`<style>` elements fully removed. Handles self-closing, same-line
open+close, and multi-line open/close. Markers inside CSS strings
(e.g. `@scope ([data-impeccable-variant="1"])`) are gone by the
time extraction runs — no false positives.
- `extractInnerByAttr(text, attrMatch)` is a balanced-tag matcher that
walks the joined text finding `<TAG ...attrMatch...>…</TAG>` with
proper depth tracking for nested same-tag elements. Handles
single-line, multi-line, and deeply nested variants.
- `extractOriginal` and `extractVariant` are thin wrappers over the
above.
- `extractCss` gets explicit same-line handling: returns null for
self-closing (nothing to carbonize), extracts inner content via
regex for same-line `<style>…</style>`, falls through to the
existing multi-line path otherwise.
## Tests
New tests/live-accept.test.mjs with four cases — all failing before,
all passing after:
- Self-closing `<style />` with dangerouslySetInnerHTML
- Single-line `<style>…</style>`
- Multi-line `<style>...</style>` (regression baseline)
- Discard restores the original element after self-closing style
Wired into `bun run test`. Full suite passes.
Credit: precise repro + root-cause trace from the other agent in the
EAC session.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five related bugs that surfaced in a real Next.js App Router project
(EAC) all rooted in live-wrap.mjs treating source as line-anchored HTML:
1. findElement matched on raw substring anywhere, so it landed on a
className continuation line of a multi-line JSX tag whose class
happened to collide with a later target. The wrong tag got wrapped
(really, its attribute line got wrapped, producing broken JSX).
2. findClosingLine's opener regex required whitespace or `>` after the
tag name, so a bare `<section\n className="..."\n>` opener was
unrecognised; it returned `start` silently, capturing only one line.
3. buildSearchQueries only emitted `class="..."`, missing React's
`className="..."`. The full-combo query never fired in JSX, so
search silently degraded to single-class substring matching.
4. Wrapper output used `style="display: contents"` unconditionally,
which is invalid JSX (type error in strict setups, parser hazard
in production transforms).
5. --tag was ignored during the primary class search. Ambiguous class
hits inside the wrong element type weren't filtered out.
## Fixes
- New OPENER_RE `/<([A-Za-z][A-Za-z0-9]*)(?=[\s/>]|$)/` recognises
tag openers at end-of-line too.
- New findOpenerLine(lines, matchLine, tag): walks up to 10 lines
backward to the enclosing opener when the match lands on a
continuation line. Aborts the walk if it hits a different tag.
- findElement now iterates all matches (not just the first), takes
a tag parameter, and routes through findOpenerLine; wrapCli passes
--tag through.
- buildSearchQueries emits both `class="..."` and `className="..."`
for multi-class queries, and both `<tag class="..."` /
`<tag className="..."` for tag+class combos.
- Wrapper builder emits `style={{ display: "contents" }}` when
commentSyntax is JSX and `style="display: contents"` otherwise.
- findClosingLine uses the same OPENER_RE so its tag-name extraction
works on multi-line openers too.
## Tests
Five new regression tests in tests/live-wrap.test.mjs, all failing
before the fix, all passing after:
- wraps the correct <section> when a class collides with a multi-line
tag elsewhere
- emits JSX-safe style attribute ({{ }}) in .tsx files
- finds elements via className= (React) when the exact class combo is
unique there
- respects --tag to reject matches inside the wrong element type
- findClosingLine recognises an opener line where the tag sits at
end-of-line (multi-line JSX)
31/31 in tests/live-wrap.test.mjs and 54/54 in
tests/framework-fixtures.test.mjs pass.
Credit: precise bug report from the other agent in the EAC session
made diagnosis and test design straightforward.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The content heuristic for deciding whether a deprecated skill dir belongs
to us returned false for harden and optimize (their v2.x SKILL.md never
said "impeccable"), while lock-entry cleanup used the authoritative
source field. Result: lock entries purged, dirs orphaned.
Layer three signals now: lock source (authoritative), word heuristic,
then per-skill description fingerprints for the two stock v2.x skills
that predate the self-identification convention.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Shape names renamed to describe the patch mechanism (what the agent does)
rather than the origin (where the CSP lives). One template now covers
multiple frameworks.
## Shape rename
- shared-helper → append-arrays
- inline-headers → append-string
append-arrays applies wherever CSP is a structured directive array.
append-string applies wherever CSP is a literal value string.
## New detection coverage
- SvelteKit kit.csp.directives in svelte.config.js → append-arrays
- Nuxt routeRules / nitro.routeRules CSP header → append-string
- Nuxt-security module's contentSecurityPolicy → append-arrays
## New fixtures
- sveltekit-csp/: SvelteKit config with kit.csp.directives. Includes
expected-after-patch.js showing the array spread.
- nuxt-csp/: Nuxt 3 config with routeRules CSP. Includes
expected-after-patch.ts showing the string splice.
## Skill docs
Single append-arrays template covers Next monorepo, SvelteKit, and
Nuxt-security. Single append-string template covers inline Next
headers() and Nuxt routeRules. Per-framework specifics listed as
sub-bullets under each shape.
54 tests across 9 fixtures, all passing. Clean fixtures (plain vite,
nextjs-app, astro, sveltekit, multipage-with-generator) still classify
as shape: null.
Astro and Vue (non-Nuxt) left unhandled by design: Astro has no
first-party CSP mechanism; Vue without Nuxt is covered by the existing
Vite fixture. Plain Svelte has no framework CSP primitive and inherits
from its bundler (Vite/Rollup).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Real-world tests (EAC Next turborepo) confirmed that CSP is the common
blocker for live mode. Adds setup-time detection with a one-time user
consent flow — the patch becomes a permanent, dev-guarded entry in the
user's own config, not a transient add/remove.
## Changes
- New detect-csp.mjs helper: grep-based classifier returning
{ shape, signals }. Shape is one of:
- "shared-helper" (monorepo CSP helper with additional*Src arrays)
- "inline-headers" (literal CSP string in headers())
- "middleware" (response.headers.set in middleware.ts; detect-only v1)
- "meta-tag" (<meta http-equiv>; detect-only v1)
- null (no CSP)
Max depth 6, skips node_modules / build / cache dirs, 64KB per file.
- cspChecked boolean on config.json. First-run setup runs detection;
subsequent runs skip. Users re-trigger by deleting the flag.
Validator accepts it.
- Skill live.md gains:
- CSP detection step in first-time setup (gated by cspChecked)
- Consent-prompt template (so every agent phrases it the same way)
- Shape 1 patch template: append `...__impeccableLiveDev` to
additionalScriptSrc/additionalConnectSrc in the app's config
- Shape 2 patch template: two-point edit — declare a dev-only
variable, interpolate into script-src and connect-src in the
CSP literal string
- Troubleshooting note for "said no but now live doesn't work"
## Fixtures
- nextjs-turborepo/: Turborepo shape (shared CSP helper with
additionalScriptSrc options). Sanitized from a real monorepo so the
patch mechanics get tested against realistic layering. Includes
expected-after-patch.ts for human/agent review.
- nextjs-inline-csp/: app-level next.config.js with a literal CSP
string. Includes expected-after-patch.js showing the Shape 2 edit.
## Tests
Framework-fixture harness extended with a detect-csp shape-classification
assertion per fixture. 42 tests across 7 fixtures pass. Clean fixtures
(vite-react, nextjs-app, astro, sveltekit, multipage-with-generator)
correctly return shape: null.
## Deliberately not doing
- No patches[] array, no marker-based rollback, no add/remove lifecycle.
The patch is a permanent dev-guarded config line — the same kind of
edit a user would make themselves.
- No base URL rewriting or proxy mechanism. Script tag still points at
localhost:8400; CSP permits it once patched. No browser-side changes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five representative project shapes under tests/framework-fixtures/ that
stage into fresh tmp git repos and drive the live scripts against each:
- vite-react: tracked index.html shell + src/App.jsx
- nextjs-app: app/layout.tsx as JSX inject target
- astro: src/layouts/Layout.astro
- sveltekit: src/app.html shell + src/routes/+page.svelte
- multipage-with-generator: src/ tracked, dist/ gitignored (our own
repo's shape); exercises the is-generated guard and
element_not_in_source fallback
Each fixture declares its config, expected source/generated paths, and
wrap cases in fixture.json. The harness copies into tmpdir, applies
gitignore, commits, then asserts:
- inject --port lands the script tag at the correct anchor across all
configured files
- inject --remove strips it cleanly
- is-generated classifies source vs generated paths correctly
- wrap routes to the expected source file or emits the expected
fallback error
Plumbing + bug caught while building out the matrix:
- IMPECCABLE_LIVE_CONFIG env var so tests can point live-inject at a
fixture-specific config.json without clobbering the harness copy.
Backwards-compatible.
- live-wrap.mjs no longer hardcodes dist/build in its directory skip
list. Only node_modules and .git remain universal skips; the
isGeneratedFile check is now the sole guard for generated paths. This
lets the includeGenerated second pass find elements in dist/ and
report generatedMatch, which is what the multipage-with-generator
fixture needs to exercise.
Wired into bun run test. 25 tests, 5 suites.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses every issue surfaced during hands-on live-mode testing.
## Injection across multi-page sites
- Config schema: `file` → `files: string[]` so multi-page static sites can
opt into script-tag injection across every HTML entry the browser loads.
- `live-inject.mjs` loops the array, reports per-file results, and
refuses silently with `config_invalid` if the schema is stale.
- `insertBefore` switched from first-match to last-match (lastIndexOf)
so the anchor lands at the true close of `</body>`, not the first one
embedded inside a `<pre><code>` documentation sample.
## Source-vs-generated detection
- New `is-generated.mjs` helper: gitignore check + generated-header
markers. Edge-case `generatedFiles` config dropped — the two real
signals cover every project shape we tested.
- `live-wrap.mjs` excludes generated files from auto-search and returns
clear fallback errors: `file_is_generated`, `element_not_in_source`
(with `generatedMatch` path), and `element_not_found`.
- `live-accept.mjs` refuses to persist into generated files; returns
`mode: "fallback"` so the agent takes over via the Handle fallback
flow.
## Accept correctness
- `extractVariant` / `extractOriginal` now skip `<style>` regions when
matching markers. Previous regex substring match treated
`@scope ([data-impeccable-variant="N"])` in CSS as the target HTML
div, capturing garbage and producing orphan CSS that rendered as
prose on the page.
- On accept, the chosen variant's content is wrapped in
`<div data-impeccable-variant="N" style="display: contents">` so the
carbonize block's `@scope` selectors keep matching. Users see the
accepted design immediately; no pre-carbonize dead state.
## Browser-side UI
- `positionBar` gains a third case: when the selected element is taller
than the viewport, pin the bar to a stable viewport anchor instead of
teleporting between top and bottom as the user scrolls.
- No-HMR source-fetch path (`injectVariantsFromSource`) now calls
`hideShaderOverlay()` on state transition to CYCLING. Previously the
shader kept running after variants arrived via the fetch fallback.
- `pickVariantContent` helper replaces fragile `> :first-child`
selection for outline positioning. Skips non-visual tags (style,
script, link, meta, template) and falls back to the variant div
itself when a variant contains multiple visual children.
- `resumeSession` re-captures and restarts the shader overlay when
the page reloads mid-generation (Bun HTML HMR does a full reload
and destroys the canvas).
- MutationObserver re-anchors `selectedElement` when the original
element is detached by HMR, preventing zero-rect highlight drift.
## Skill docs
- `live.md` reframes `config.files` as "the HTML files the browser
actually loads" and documents the regen-wipes-inject caveat for
multi-page generator projects.
- New Handle fallback section covers the three wrap error shapes and
how the agent should manually wrap for preview and commit to real
source on accept.
- Handle accept documents the new `data-impeccable-variant` wrapper
and the carbonize agent's duty to strip it.
## Prefetch feature (landed but disabled)
A `prefetch` event fires from the browser on first CONFIGURING per
route so the agent can pre-Read the source file before Go. Real latency
win in the linger-before-Go case but costs a harness round trip when
Go fires quickly. Disabled via a `PREFETCH_ENABLED = false` flag in
`live-browser.js`; server validator and skill dispatch stay so re-
enabling (with a browser-side debounce) is a one-line change.
## Harness guidance
Earlier skill rewrite compressed two load-bearing instructions:
- Restored prescriptive wording for "open the tab via Chrome MCP
before the first poll" and the Claude Code background-poll policy.
- Flag-mapping for `live-wrap` rewritten as explicit bullets so models
don't collapse `--element-id`/`--classes`/`--tag` into a single
`--query` argument.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bun's HTML HMR does a full page reload when the live-wrap.mjs edit
lands, so the shader canvas is destroyed and in-memory capture blob is
lost. resumeSession rehydrated state from localStorage but never
restarted the overlay, so the wait went dead.
resumeSession now re-captures the original's content (still in the DOM
inside the variant wrapper) and restarts showShaderOverlay when we
reload mid-generation. Also swaps the two remaining :first-child
selectors in resumeSession for pickVariantContent so the earlier
loose-children robustness fix carries across reloads.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two regressions surfaced in smoke testing, both traceable to state
drift when live-wrap.mjs rewrites the source file and HMR swaps the
DOM.
1. Variant outline on the wrong element. The skill rewrite lost the
explicit "each variant must be a complete element replacement"
rule and dropped the "full element replacement" comments from
variants 2 and 3. Models started producing variants with loose
sibling children, so live-browser's :first-child selector framed
only the first sibling. Restored the rule, made all three comments
consistent, and replaced :first-child with pickVariantContent —
which skips non-visual tags (style/script/link/meta/template) and
falls back to the variant div itself when a model still ships
multiple visual children.
2. Loading shader freezes after wrap. The MutationObserver only woke
up when new non-original variants arrived, so when the wrapper
first appeared via HMR with just the original inside, selectedElement
was left dangling on the now-detached pre-wrap node. Scroll-tracking
read a zero rect on every frame and collapsed the shader canvas to
0x0. The observer now re-anchors selectedElement to the original's
content the moment the wrapper shows up, keeping overlays positioned
until real variants land.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The skill rewrite compressed two load-bearing instructions into ambient
context:
- "Navigate the browser to the URL" lost the signal that models with
Chrome MCP should proactively open the tab before the first poll.
Restored the forcing phrasing and the "before the first poll" anchor.
- "Claude Code can background the poll" read as permission rather than
prescription. Models fell back to foreground blocking by default.
Restructured harness guidance as a bulleted policy, prescriptive per
harness, with the reason attached (harness notifies on completion so
the conversation stays free).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Splits the skill into two register references (editorial, product),
replaces category-based theme selection with a forced physical-scene
inference, and introduces a four-step color strategy axis (Restrained /
Committed / Full palette / Drenched) with editorial permission for the
bold three.
Adds a seed mode to /impeccable document for pre-implementation
projects, updates /impeccable teach Step 5 to offer the seed path, and
grows /impeccable shape with Design Direction + Scope intake
(fidelity, breadth, interactivity, time). Extends live-mode variant
distinctness to forbid three variants sharing theme and dominant hue.
Also drops the anti-pattern validator coupling, consolidates a11y into
audit.md, and updates CLAUDE.md with the register architecture.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a full annotation pipeline to /impeccable live. On Go, the browser
captures the selected element as a PNG (with annotations composed in),
uploads it to the live helper, and sends the generate event with the
screenshot path so the agent reads user intent visually instead of from
HTML alone.
Annotation tools (while an element is picked):
- Click inside the outline to drop a magenta comment pin with a text input
- Drag to paint a magenta SVG stroke (5 px click-vs-drag threshold)
- Click a pin to edit; double-click to delete; drag a pin to reposition
- Click a stroke to delete it (wider invisible hit path)
- Clear chip top-right wipes everything; hidden when no annotations
Capture pipeline:
- modern-screenshot vendored as an IIFE (scripts/modern-screenshot.umd.js)
and lazy-loaded from the live helper
- Font fix: cross-origin @font-face rules are fetched and fonts are inlined
as base64 data URIs before being handed to modern-screenshot via
font.cssText, since SVGs rasterized via canvas can't fetch external
resources (fix for "Impeccable" rendering bold-serif and items wrapping
wrong in the capture)
- Annotations are temporarily attached to the live element (not only the
clone) so computed styles resolve during the embed pass
- Session screenshots live in .impeccable-live/annotations/session-*/ in
the project root (gitignored) so the agent's Read tool doesn't trip a
per-path permission prompt
Loading shader (activates during GENERATING):
- WebGL overlay rendering the captured PNG as a halftone — cells with
luma-driven dot radius, rendered on paper-cream underneath a magenta
roller that sweeps top-to-bottom with a 3.4s cycle and clean overshoot
- Fixed asymmetric bandAt() using one-sided smoothsteps (previous reversed
smoothstep was undefined on d>0, giving "trail=1 everywhere below")
- Graceful <img> fallback when WebGL is unavailable; prefers-reduced-motion
freezes the band at t=0
Server:
- POST /annotation endpoint (raw image/png body, token + eventId query),
session-scoped tmpdir cleaned up on shutdown
- GET /modern-screenshot.js serves the vendored UMD with aggressive caching
- Optional screenshotPath / comments / strokes fields on generate events
- Fixed pre-existing /source crash on ENOENT (writeHead called twice)
Agent side:
- reference/live.md step 0 tells the agent to Read the screenshot first,
with four rules for interpreting annotations: comments are position-
anchored and scoped to the sub-element under their {x,y}; strokes are
gestures (loop=focus, arrow=direction, cross=delete); comments and
strokes are independent unless adjacent; don't silently guess on
ambiguous strokes
Also:
- Generating bar no longer claims "Generating 1 of 3..." (variants arrive
atomically) — now says "Generating N variants..."
- tests/live-server.test.mjs fixed to read the PID file from project root,
matching the server; adds coverage for the new endpoints and validator
fields
- .impeccable-live/ added to .gitignore
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
live-server.mjs stop now runs live-inject.mjs --remove after the HTTP
server shuts down, so HTML entries do not keep loading a dead localhost
live.js URL. Add stop --keep-inject to stop only the helper.
Update reference/live.md cleanup steps and sync all provider skill copies.
Made-with: Cursor
Document that Cursor Composer should run live-poll blocking in the same turn
(with a link to Cursor subagent foreground vs background docs). Claude Code may
use a background poll when the harness surfaces completion. Sync reference to
all provider skill bundles.
Made-with: Cursor
typography.md had a parallel 4-step font-selection procedure and a smaller
banned-fonts list (5 fonts: Inter, Roboto, Open Sans, Lato, Montserrat)
that duplicated SKILL.md's authoritative <font_selection_procedure> with
its 23-font list. Removed the duplicate procedure and deferred to SKILL.md
for the banned list. Kept the unique material: anti-reflex corrections,
system-font note, pairing principles, web font loading, OpenType, fluid
type guidance, accessibility — and all of the scale/rhythm/measure
content that SKILL.md doesn't cover.
pin.md removed for the same reason as context.md: SKILL.md's inlined
pin section already covers what an agent needs (what pin does, usage,
valid commands, how to report back). No value in the indirection.
SKILL.md: 386 → 388 lines (slight growth from inlining pin details)
typography.md: 142 → 132 lines
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The main skill now has enough detail to stand alone: two files, load
command, no-truncate rule, never-infer warning, session cache, teach
fallback, DESIGN.md nudge, and the live-mode "already warmed"
exception. context.md was indirection without added value.
Inlined the two bits from context.md worth keeping:
- Content validity: treat empty / <200 chars / [TODO]-placeholder
PRODUCT.md as missing
- live.mjs auto-warms, don't double-load with load-context.mjs
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Before: 422 lines. After: 386 lines. The cut is conservative because
evals/AGENT.md revealed that most of what looked like bloat is actually
load-bearing: the font_selection_procedure with its 23-font ban list
(lesson 6), the theme_selection with audience examples (lesson 9), and
the absolute_bans with literal CSS patterns (lesson 7) all drive
measurable eval improvements and must stay inline.
What moved out of SKILL.md:
- Context Gathering Protocol (52 → ~18 lines). The full protocol — cache
semantics, dispatch tree, teach/document/live exceptions, why-it-matters
— moved to reference/context.md. SKILL.md keeps only the compact hook:
load command, "never infer from codebase" warning, and pointer.
- Pin/Unpin (14 → 6 lines). Details moved to reference/pin.md.
- Spatial principles: dropped 4pt-vs-8pt rationale, gap-vs-margins CSS
technicality, and container-queries-vs-viewport explanation (not
load-bearing in the main skill). Kept all load-bearing rules
including the 80-char body-text line (detector-backed).
What did NOT move (load-bearing per evals):
- <font_selection_procedure> with the 23-font ban list
- <theme_selection> with the 8 audience examples
- <absolute_bans> with literal CSS patterns
- All XML tag structure (lesson 8: XML works better than markdown
for reasoning models, especially OpenAI)
Also added:
- reference/context.md (new) — full context protocol
- reference/pin.md (new) — full pin/unpin docs
- "Never infer brand, audience, or tone from the codebase" warning
restored to SKILL.md (was dropped in an earlier refactor)
Fixed:
- reference/colorize.md had "Accent borders: Add colored left/top
borders to cards or sections" which directly contradicted
SKILL.md's absolute_ban on border-left/right > 1px. Rewrote the
accent-border advice to use hairline borders, surface tints, or
leading glyphs instead, with an explicit reference to the ban.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two observed failure modes:
1. Smoke-test style truncation (`| head -N`) in bash commands defeats the
whole point of load-context.mjs — Claude needs the FULL file contents,
not the first few lines of JSON.
2. The old protocol didn't clearly explain session caching, leading to
repeated load-context.mjs calls across commands in the same session
(thousands of wasted tokens on 3-5KB files re-fetched 3-5 times).
Context Gathering Protocol rewrite:
- PRODUCT.md required (blocker), DESIGN.md optional (one-line nudge if
missing). Greenfield projects can't yet have a DESIGN.md to document.
- Explicit session cache: if content is in conversation history, do not
re-fetch. Exceptions listed (after teach/document/manual edit).
- Explicit "never truncate" rule: consume the full load-context.mjs
output, never pipe through head/tail/grep/jq with field filters.
- Content validity check: hasProduct=true but content <200 chars or
full of [TODO] markers = treat as missing, run teach.
- Missing-PRODUCT.md flow spells out task resumption: user asked for
/impeccable polish ButtonGroup, we must run teach, then RESUME polish
of ButtonGroup with fresh context — not silently abandon intent.
- Three explicit exceptions to the protocol:
- /impeccable teach skips it (teach creates PRODUCT.md)
- /impeccable document loads PRODUCT.md only (creates DESIGN.md)
- /impeccable live already warms context via live.mjs — don't also
run load-context.mjs
teach.md Step 6 and document.md Step 5 now re-run load-context.mjs at
the end so the freshly-written files surface in conversation history
and subsequent commands use the new version, not a stale earlier read.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
