URL.pathname leaves spaces percent-encoded, so the embed helper never ran and the generator still claimed success. Convert with fileURLToPath, check the spawn result, and warn instead of lying.
Prepared with AI assistance.
Co-authored-by: Cursor <cursoragent@cursor.com>
- keyChroma re-encodes with the PNG's tEXt chunks intact (the embedded prompt survived generation but not keying)
- organic-clip-path counts relative curve commands too (path data letters are only commands, so the match is case-insensitive)
- buried-raster normalizes percentage alphas (parseFloat('80%') read as 80) and reads 4- and 8-digit hex alpha instead of treating #rrggbbaa as opaque
- the extension-injected-node skip in checkQuality runs before any finding is pushed (a low-opacity injected raster was recorded, then returned by the skip)
- fake-mode plates carry impeccable:fake tEXt and the plates gate's crop-identity refusal skips them (fake mode IS the crop by design; the refusal is for models shipping the comp's pixels as artwork)
Findings by cursor[bot] and Copilot on PR #599; detector engines rebuilt (build:browser, build:extension).
AI-assisted (Claude Code).
comp-spec, comp-diff, build-phase, font-match, and generate-image now decode any comp raster via loadRaster(), converting non-PNG input to <file>.png next to the source. Sessions used to hit 'not a PNG' and rewrite the .webp in place with PNG bytes, which broke transcript replay (a later step rewrites the comp beyond the cut) and left a mislabeled file.
AI-assisted (Claude Code).
The first paid confirmation sweep showed the failure: models rendered the
three comps first and ran build-phase.mjs start after, so a session cut at
the composition pick carried no state.json and the resumed model followed
the conversation ('translate the comp into HTML now') instead of the
phases. Decision comps (.impeccable/mocks/decision/) are unaffected;
--force-mock overrides.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
font-match.mjs fingerprints a text region's lettering in the comp (cap
height, glyph advance per cap = width class, ink fraction = weight class,
tracking), renders candidate faces at that cap height in a headless
browser (yours plus a shortlist for the width class), and ranks them by
distance with a proof sheet; the spec gate refuses to close until the lead
text region is measured and ranked. generate-image --plate keys ink-on-
ground plates to alpha (chroma) so the drawing sits on the page's own
ground; the plates gate scores keyed plates composited over the region's
ground. comp-diff reports each region's ink box; the hero gate names a
control whose box height or width differs from the comp. The hero
instruction copies the comp's words verbatim; rewording is a stated
decision after the hero passes.
Driven by a human review of the r3 side-by-sides: face width and weight,
plate ground, control row height, and content substitution.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
plateVerdict() is shared by the plates gate and generate-image's PLATE-WARN
so they cannot disagree; --plate picks a frame that clears the 1.5x width
floor (a square region wider than 682px takes the 1536 landscape frame);
'missing' on text/chrome/control regions requires the build region to be
near-empty, so a 12px rule a few pixels off reads as contradicted or drift,
not missing; the responsive gate does not re-litigate a plate that passed
the hero; record hero after close does not inflate the attempt count.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
From a forensics pass over twelve runs and two adversarial passes over
the metrics:
- comp-diff: detail is signed and penalizes invented energy; regions with
structure under 0.3 (or painted regions under 0.45 / added detail over
0.4) are contradicted whatever the mean says; palette ramp tightened;
region crops inherit the whole-image best translation so a shifted page
is not eight contradicted regions.
- hero gate: fails on any contradicted plate/image/text region (chrome and
controls keep the one-third allowance), on a capture that is not the
comp's frame, on a palette that is not the comp's, and on an organic
clip-path drawn inside a raster region's box.
- plates gate: scored against the comp crop with overlapping text/chrome
painted out (comp-spec plateReference; generate-image uses it too), with
a structure floor and an added-detail veto; the real plate passes, noise,
mirrors, mosaics, and other regions do not.
- comp-spec refuses a regions file that leaves comp ink unnamed.
- The direction-choice ping is folded into build-phase.mjs start
--direction --kind; the roll writes .impeccable/build/pending.json and
context.mjs / detect.mjs report COMP_ROUND_OPEN until the hero gate
passes. A code-led config makes start print the contract step and stop.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
generate-image.mjs --plate produces one raster region of the measured spec
from the comp crop, scores it against the crop, and refuses under --min.
The asset producer's job becomes producing the spec's plates. Detector
gains organic-clip-path (many-vertex polygon / curved path() clips) and
buried-raster (raster under a near-opaque wash or at near-zero opacity),
wired into both engines with fixtures.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
- concept-seed --scope surface deals three grounded-list indices (dice-picked,
primary leads) instead of one: a single card is not a choice, and the
no-lineup rule stays direction-only, where it was written for worlds
- serve-question renders a new per-card wireframe field as a layout schematic
in the media slot: the code-led channel's visualization, no image
generation needed, no card back, no salience weight
- generate-image gains --ref (repeatable): routes through the edits endpoint
with input images, so an established world's comp inherits identity from a
captured screenshot of a real page instead of a prose paraphrase; tested
against impeccable.style, where the reference-anchored comp reproduced the
live site's chrome and the prose-only comp drifted
- new-work rung two rewritten around the dealt hand: lock-in is the
approval, a locked comp builds comp-led and discharges the visualize.md
three-option round, a locked wireframe builds code-led; visualize.md
records the exemption and the reference-image discipline, including the
reference-leak caveat (chrome carries, the reference page's content
does not)
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The prompt behind a generated image was recorded three different ways,
a sidecar in the eval harness, nothing in the skill's API tool, nothing
for native tools, so intent survived or vanished depending on where you
ran. One dependency-free script now embeds the prompt inside the image
itself, PNG tEXt or JPEG COM with a sidecar fallback for other formats,
idempotent, and reads it back from any impeccable-generated file. The
API tool embeds automatically; the prose directs every native-tool
generation through it; copies between machines and harnesses keep their
intent. Comps meanwhile are declared the build thread's own work, never
delegated, and the comp-skeleton guidance now asks for the surface's
actual regions instead of prescribing navs onto pages that have none.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two findings from the first human-validated probe. The comps rendered
as scene vignettes because the generation prompts led with the world's
atmosphere; the model painted the fish market instead of the fish
market's website. The comp guidance now demands the page's literal
skeleton in the prompt, nav and its items, headline block, sections in
order, footer, with a self-check: a render that could hang as a poster
is not a comp. And generation context is part of the asset: the thread
that wrote a prompt knows what the image contains and why, so build-
critical imagery prefers the build thread, and subagent-produced assets
must carry their prompts, via the tool's new sidecar or the manifest,
read by the builder before composing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A cheap, LLM-free E2E tier for the interactive parts of new-work, mirroring
the two-layer live-e2e pattern (deterministic now, opt-in LLM tier later).
- generate-image.mjs: IMPECCABLE_IMAGE_GEN_FAKE=1 writes a deterministic
offline image (SVG with wrapped prompt + SYNTHETIC COMP label, or a valid
palette-stripe PNG carrying the prompt/marker in a tEXt chunk). Same CLI
contract, no key, no network, $0.00 cost line.
- tests/new-work-e2e/user-bot.mjs: scripted user bot (module + CLI) that
resolves the serve-question daemon from the workspace and drives the real
page via Playwright (pick, re-roll + steer, canon, tab close).
- tests/new-work-e2e.test.mjs: node --test coverage of the serve-question
cycles (pick + CHOSEN CARD, re-roll + --update re-deal, canon + CANON
CHOSEN, tab-close exit-4, text-only card) plus fake image determinism.
- Registered as the opt-in new-work-e2e suite; added test:new-work-e2e.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Paul's design, three pieces:
serve-question.mjs: the world decision presented as a themed page instead
of a text prompt. The script serves an impeccable-styled option board
(assigned direction leading with THE ROLL badge, dealt challengers as
alternates carrying their QUALITY BAR cards, re-roll and steer built in),
prints the URL, opens the browser, and blocks until the user chooses;
the answer lands on stdout as ANSWER JSON, so the shell call itself is
the wait and no harness machinery is needed. Local images are served by
the ephemeral server; nothing leaves the machine.
generate-image.mjs + context.mjs IMAGE_GEN_AVAILABLE: when an OpenAI key
is in the environment, context reports that image generation works even
without a harness-native tool (gpt-image-2, billed to the user's key,
stated before first use; Google skipped by decision). Harness-native
tools always win when present.
new-work.md: visualize-before-build is now the default whenever any
image generation exists, not a codex.md special case; the attended
presentation prefers the visual decision page and falls back to the
structured question tool. Evals keep the unattended path untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Abdul WahabandCursor
Paul Bakaus
