`initialized natural build` looked like a third defect on main: sonnet began
implementation before the attended concept checkpoint, three runs in a row.
It is flaky, not broken, and the measurement setup was the larger problem.
A run that stops to put the concept to the user before building takes about
579s on sonnet. A run that skips the checkpoint and fails the assertion
finishes in 130-200s. The suite capped each test at 300s, so the thorough path
was killed as a timeout and the hasty path was graded as a result: the cap was
selecting for the behavior the scenario exists to forbid. Raised to 900s, with
the reasoning recorded next to the number so it is not trimmed back as a
mystery constant.
The baseline is corrected accordingly: the scenario is flaky (1 of 4), not
failing, and readers are told to check a duration against the cap before
calling a slow failure a behavioral one.
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two defects the skill-behavior baseline had recorded as failing on main.
`UPDATE_AVAILABLE` told the agent to ask once, then said "If they agree, run
`npx impeccable update`", then said to continue without waiting. Nothing gated
the run on an answer, and the same sentence removed the wait that could have
produced one, so the command read as the next step and sonnet took it. The
offer stays; the command leaves the turn. Running it mid-session rewrites the
files the session is reading and only takes effect next session, so there is
nothing to gain by running it now, and the directive says that rather than
relying on the model to infer it. Failed 3 of 3 before, passes 3 of 3 after.
Scenario 15 was a broken fixture, not a routing defect. The iOS workspace held
PRODUCT.md and nothing else, so `audit the app in this workspace` named an app
that was not there: sonnet spent its step budget hunting for it, including a
`find /` across the filesystem, and read no reference file at all. The
assertion reported "loaded audit.md instead of the variant" when the truth was
"loaded neither". One SwiftUI screen makes the request answerable, and the
scenario then passes on unmodified main, which is the evidence that the skill
text was never at fault. This is the convention MINIMAL_LANDING_HTML already
established for the web scenarios; the native fixture never received it.
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The build-path preference shipped as a question only `init` asks, written to
a file only `init` writes. Nothing routes an initialized project back through
init, so every existing project took the comp-first default without anyone
choosing it, and the only recourse was a footer toggle that binds one session.
Neither the setting nor the round that preceded it ever reached a release
(skill-v4.0.4 has no `buildPath`, no `comp-led`, no `.impeccable/settings.json`),
so the PRODUCT.md standing-commitment fallback describes an era that never
existed publicly. It is deleted rather than honored: told a field might exist,
models go hunting for it and preserve it.
- `buildPath` moves from `.impeccable/settings.json` into the unified
`.impeccable/config.json`, which already has a known-keys registry, doctor
coverage, and a gitignored `config.local.json` override. Whether a machine
has an image tool is a property of that machine, so the local file wins.
- new-work captures the answer from behavior instead of an interview: a toggle
flip on a project recording nothing asks once, after the round closes,
whether to keep it. The answer is written either way, because a declined
offer nothing writes down is an offer the next session makes again.
- Two findings: `config-invalid-build-path` (an unread value rides the default
rather than the opposite path) and `config-build-path-unset`, gated on a
product record plus evidence of direction work so polish-and-audit projects
never hear about a setting they do not use.
- init treats a recorded value as a confirmed answer, resolving its conflict
with Step 1's "do not reopen confirmed fields".
- The setting was undocumented in the README and doctor.md. Both now cover it.
Also records a measured skill-behavior baseline. Three cells fail on unmodified
main (scenarios 9 and 15, `initialized natural build`), verified against a clean
worktree; the suite README now says so, so the next person does not spend the
hour attributing them to their own branch.
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two failures the trace test found were structural, not model quirks.
critique.md described the report's format and then went straight to writing a
temp file, with no step saying to output the report. gemini-3.6-flash and luna
both responded by bundling heredoc, snapshot write, trend read, and cleanup into
one bash call and stopping, leaving a perfect archive nobody had read. A
"Deliver the Report" step now precedes persistence, and persistence describes
itself as a copy of what was already sent. gemini-3.6-flash failed three
consecutive runs before this and its failures afterward all show the report
reaching chat.
The close is also step 6 of the persistence list rather than a section after it,
since the same shape is what fixed delivery.
Lineup: gpt-5.6-luna and deepseek-v4-flash out, gpt-5.6-terra in at
reasoningEffort high (IMPECCABLE_SKILL_BEHAVIOR_EFFORT overrides), gemini 3.5 to
3.6. Provider options resolve from the model object inside the harness so the 21
runTurn call sites are untouched. Verified the effort actually reaches the API
rather than being silently dropped.
The Gemini bump was not cosmetic: 3.5-flash passed critique closes twice and
3.6-flash then failed three times against identical text. A version bump inside
one family changed the outcome, so the README now treats cross-version carryover
as unmeasured.
Known floor, recorded: critique closes is flaky on gemini-3.6-flash, 1 run in 3.
Two structural attempts moved it from consistently failing to intermittently
passing and then stopped paying. claude-sonnet-5 and gpt-5.6-terra are clean.
Prepared with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The earlier fix in this branch was built on a wrong diagnosis. It assumed a
structured question hides any prose sharing its message, so it split report and
question across two turns. A controlled check showed prose before a question
renders fine; what hides a report is emitting it AFTER the question. The split
therefore fixed nothing and introduced a worse failure: a turn that ends on the
report is a turn that ends, and the questions never arrived at all.
Persistence returns to main's ordering, byte for byte, and the boundary prose is
gone. What replaces it is a position rule: the question is the last thing in the
response.
The trace test added here found two failures beyond the reported one. Critique
can fail to land in three ways, and they are now all asserted:
1. Question emitted before the report, hiding it behind the picker.
2. No close at all: no questions and no skip line, so polish inherits nothing.
3. Report authored into the persistence heredoc and never written to chat,
leaving a perfect snapshot and a user who sees nothing.
Mode 3 predates this branch entirely. Persistence step 1 now says the temp file
is an archive copy, not delivery.
The Codex final-question gate is promoted out of its <codex> fence, where it was
stripped for three of four providers, and the skip branch is now a countable
threshold (fewer than 3 Priority Issues) rather than a judgment call.
Known floor, recorded in the suite README: gpt-5.6-luna passes 1 run in 6 and
deepseek-v4-flash is flaky. claude-sonnet-5 and gemini-3.5-flash are consistent.
Prepared with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Review on #576 caught document.md:71 splicing {{ask_instruction}} after
"then", which is the same defect this branch set out to fix. Rendered for
Codex it produced "Show the user the existing file, then STOP and use Codex's
structured user-input/question tool...". The line now starts a new sentence.
The comment added to PROVIDER_PLACEHOLDERS asserted the contract without
enforcing it, which is exactly how four reference files shipped the splice in
the first place. validateAskInstructionSites() in build.js now checks every
call site and fails the build on a mid-sentence interpolation, and the comment
points at the gate instead of asking authors to remember.
Prepared with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The critique report and the AskUserQuestion call shipped in one assistant
message, so the report stayed hidden until the user answered the picker and
the command read as if it had never run.
Reorder critique's persistence steps so the temp-file cleanup runs after the
report and trend line are sent. That cleanup now ends the message carrying the
report, leaving the questions to open a fresh one. Both critique.md and
overdrive.md state the constraint and why it exists, so the ordering is not an
unexplained sequence a model can optimize away. Overdrive additionally moves
its direction descriptions inside the question options, where the user is
actually reading them.
Also fix the ask_instruction splices. The placeholder is a complete sentence,
but five call sites spliced it mid-sentence and shipped text like "stop and
STOP and call the AskUserQuestion tool to clarify. before expanding it". Every
call site is now sentence-initial and the twelve lowercase provider values are
capitalized to match, with a comment in utils.js pinning the contract.
Record a workflow-contract baseline for the current model lineup. The two
failures seen while validating this change are pre-existing: bolder refinement
fails on deepseek-v4-flash identically with bolder.md reverted to HEAD, and
redesign replaces DESIGN is flaky on assertions driven by new-work.md, which
this change does not touch.
Prepared with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Same line as the round's title rather than the brand mark: the control
reads as part of the round it configures, and the brand row stays clean.
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Top-left is the brand and reading-entry corner and mode controls belong top
right; on the brand row the toggle also costs no vertical space, so the
headline keeps its position.
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The toggle sits top-left under the brand instead of in the footer bar, and
flipping to comp-first now opens a confirm dialog before anything renders,
since the flip starts billed, minutes-long generation; flipping back stays
free and immediate. The dialog lives at the document root so it never loses
the stacking fight with the deck. The schema blob also states harder that
toggle: true may only be offered when image generation exists.
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The retired execution-contract round and the buildPath payload keep
main's text; the decision-page fallback keeps this branch's gate.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- serve-question: payload buildPath { value, toggle } renders a footer
segmented control (comp first / code first) with the trade stated in one
line; the default comes from settings, a flip binds that session only.
Code-led rounds treat declared comp paths as flip reserves: wireframes
render, a flip to comp shimmers the slots and surfaces once through
--wait as BUILD PATH FLIPPED so the agent starts generating mid-round;
the flip back is free and a landed comp stays. The ANSWER carries
buildPath and buildPathFlipped with a session-only directive.
- init Step 5 asks the preference once (only when image generation exists)
and writes .impeccable/settings.json; context.mjs surfaces the recorded
default every session; PRODUCT.md standing commitments stay honored as
the fallback.
- new-work retires the two-card execution-contract round: no round asks a
workflow preference. followup stays as the generic same-table mechanism.
- e2e: new toggle test (14/14 with the wireframe test).
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Bugbot's finding: exit 2 is overloaded, and at --wait it means the
question server died, so the unscoped rule would drop a live visual
round onto the text channel after a transient daemon loss. The gate now
names the serving invocation, which also settles Copilot's exit-code
ambiguity, and the display clause reads grammatically.
AI-assisted (Claude Fable 5).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Across every recorded gpt-5.6-sol session, serve-question.mjs was never
invoked once: the rule's prose list of fallback environments (headless,
CI, an eval worker, a remote shell) let the model match itself against
the list and take the structured question tool without running the
script, while claude-opus-5 on the identical harness runs the script
every time and the page works. The direction is then chosen with no
imagery on the table, the catalog challengers are weighed without their
art, and the session's own safest candidate wins: measured end to end
on the eval harness, this is where bland output enters.
The environment list is deleted; the script's own exit 2 is now the
only key to the fallback, and the script already prints the rationale
and the override at runtime to exactly the sessions that hit it. Same
gate on the comp round's approval point: inline image rendering earns
the in-harness path, and a text-only surface is not display.
One adversarial review pass; its two word-level findings are applied.
AI-assisted (Claude Fable 5), prepared for maintainer review.
Co-Authored-By: Claude Fable 5 <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 raise block drops the side-tab left border for a quiet patina panel,
and drops the poker jargon: "Improved by Impeccable's worlds" with
per-line "From <world>" donors, on single raises too; tooltip, aria, and
screen-reader copy follow
- The pick-card kicker convention renames MY PICK to IMPECCABLE'S PICK at
every definition site, so users stop reading "my" as themselves
- The footer (steer, registers, canon exit) is a sticky full-bleed bar on
wide viewports, sharing one --page-inset with the content column; portrait
keeps it in flow where the deck scrolls internally
- Short landscape viewports compact the headline and narrow the cards so a
full round fits 1440x800
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Cursor's finding was real: gating the density check on a named focal
moment let a busy comp pass whenever the direction named none, which is
the common case on the lane that produced the busy comps. The second leg
reuses the bullet's own distinction: several regions performing the
concept at once is the same shout; regions doing their jobs are not.
AI-assisted (Claude Fable 5).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Unify Svelte component accept and discard around one operation dispatch, source lock, error path, and result emission while preserving their existing CLI contracts. Add direct CLI characterization coverage for both operations.\n\nAI-assisted implementation under pbakaus's scheduled-refactor authorization.
Two independent skeptic passes over the added prose, one hunting
oversteer and example bias, one hunting mode and platform damage. What
they killed, and why:
- The absolute 'never a medium' rule contradicted the file's own
imagery-stance fixity two paragraphs up and stripped legitimate guards
(an illustration-committed world, a native app screen warding off
stock-photo drift). A medium ban now belongs to the committed imagery
stance, never to caution, and the rule appears once per reader context
instead of five times corpus-wide.
- The quoted incident string and the four-example subject list taught
the model the exact framings they existed to prevent. Gone; the
abstract rule plus the point-at-the-subject check carry it.
- 'A first-time visitor learns what this is, why it matters, and what
to do' was Persuade anatomy imposed on all four modes. The guard is
now mode-neutral: a quieted region keeps its information and stops
performing.
- 'Calm is what Operate and Read surfaces are for' contradicted
operate.md's density affordance. Deleted; modes stay defined in one
place.
- The focal-moment count now presupposes nothing: it fires only where
the direction names a focal moment, and only on same-scale rivalry,
so an even, calm field stops reading as a failure.
- The decision-comp clause and the mode bullet no longer restate what
they can reference.
Net: the prose additions drop from roughly 480 words to under 200, with
no quoted strings and no example lists.
AI-assisted (Claude Fable 5), prepared for maintainer review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Proven necessary by its own demo: the first re-render of the declined
moto-forum comp satisfied subject-present and one-dominant-move by
deleting the value proposition, leaving a members' index that told a
first-time visitor nothing about what this is or why to care. Paul
caught it. Quieting a region means it stops performing, not that it
leaves; empty is quieter, not calmer.
AI-assisted (Claude Fable 5), prepared for maintainer review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The declined moto-forum comp's prompt read 'no gradients, no rounded SaaS
cards, no photography, no fake member counts, no badges, no testimonials':
the reflex that rightly bans invented claims swallowed the one medium the
subject lives in, and that is exactly how a motorcycle forum got comped
with no motorcycles. The lektor prompt's 'no AI imagery', written by an
image model, is the same fingerprint.
One counterweight, phrased once per authoring surface: the comp
discipline's subject-presence check (which the decision comps already
bind), the asset producer's own prompt rules (a standalone agent that
never reads visualize.md), and new-work's author-assets law (the path a
code-led build takes without the comp round). Truth binds claims, not
demonstrations; a photo of the subject doing its job is a demonstration.
AI-assisted (Claude Fable 5), prepared for maintainer review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The deliverable died in #545; the word survived as the decision-page
payload's field name, annotated everywhere it appeared with the same
compatibility apology. The page and the skill text ship together and
payloads are per-session, so the compatibility burden is one input alias,
not a frozen name.
serve-question.mjs: the card field, the answer key, the schema docs, the
--schema example, the help text, and every internal identifier (compSrc,
data-comp, .media.comp-pending, img.comp, comp-note) now say comp; a
payload declaring the legacy sketch key still renders and answers
identically. new-work.md and the asset producer drop their wire-name
parentheticals. The unit suite covers the canonical answer key coming
back from a legacy-key payload; the new-work e2e's declined-card stray
comp stays declared as sketch, which doubles as alias coverage.
AI-assisted (Claude Fable 5), prepared for maintainer review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Factory review evidence (two batches, both lanes): generated comps drift
poster-ward. They render the world's atmosphere at high density, drop the
surface's subject (a motorcycle forum comped with no motorcycles), and stop
reading as screens a product would ship. The existing anti-vignette
self-check catches the fully collapsed case but says nothing about density
or subject presence, and new-work's "committed all the way" reads as a
coverage instruction.
Three sibling self-checks in visualize.md's comp discipline, each phrased
per mode (Persuade/Operate/Read/Experience) and platform-neutral: the
subject appears as the content the regions hold; the mode must be readable
from the image alone; commitment is depth, not coverage, with one dominant
move per viewport. new-work.md's decision-comp rule gains a clause binding
the same checks so the direction round inherits them explicitly.
AI-assisted (Claude Fable 5), prepared for maintainer review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Stop assuming white when a background cannot be read
Dark themes came back from a scan buried in low-contrast findings that
all claimed the light text sat on #ffffff. Two live runs against
impeccable.style produced 102 and 95 of them.
Two causes, both fixed here.
Parsing. Browsers keep the authored color space in getComputedStyle
output: oklch() stayed oklch, but color-mix results come back as
color(srgb 1.04 0.72 -0.21), wide-gamut authors get color(display-p3
...), and lch()/lab() survive verbatim. The parser read none of those, so
those surfaces registered as unset. parseGradientColors was worse: it
matched only rgba() and #hex, so a ground painted as
linear-gradient(oklch(...), oklch(...)) counted as a gradient with no
stops at all.
Guessing. When the ancestor walk ran out of readable color it returned
white, and on a body-level gradient it returned white without even
looking. Light copy on a lacquer-black page then measured 1.3:1 against a
canvas the visitor never sees.
resolveBackgroundInfo now separates three outcomes: a resolved surface, a
gradient the caller should fall back to stops for, and an unreadable
layer. The last one makes both color adapters skip their contrast checks
entirely. White survives in exactly one case, the one that earns it:
every layer up to the document root was genuinely transparent.
Color conversions moved to cli/engine/shared/color.mjs and gained lab,
lch, and color() for srgb, srgb-linear, and display-p3. Spaces outside
that set return null, which now routes to abstention rather than to a
color nobody painted. Every conversion is pinned against what Chrome
itself paints for the same string.
Rescanning impeccable.style: 102 low-contrast findings down to 30, none
of them on an invented white ground.
Assisted-by: Claude Code
* fix: address PR review bot findings on background resolution
- Treat a url() image layer stacked above a gradient as an occluding,
unreadable surface: resolveBackgroundInfo now returns unresolved so the
gradient-stop fallback never measures stops the image hides
(greptile-apps finding, reproduced in Chrome).
- Route the glow and AI-palette DOM adapters through resolveBackgroundInfo
so an unresolved surface makes them abstain instead of hunting gradient
ancestors past an unreadable layer (Cursor Bugbot finding).
- Resolve background-color keywords jsdom hands through verbatim:
inherit now reads as no-paint (the ancestor walk IS its resolution) and
currentcolor substitutes the element's own computed text color instead
of forcing an abstention (Copilot finding).
- Regression coverage in the dark-theme fixture for all three, asserted in
both the jsdom and real-Chrome suites; browser detector regenerated.
AI-assisted: prepared with Claude Code at the maintainer's direction.
Co-Authored-By: Claude <noreply@anthropic.com>
* fix: keep zero-offset glow findings when the surface is unreadable
The browser glow adapter abstained from the whole element when
resolveBackgroundInfo reported an unreadable surface, which also dropped
zero-offset chromatic halo findings that do not depend on the background
at all. It now skips only the gradient hunt past the unreadable layer and
scores the halo tell against a null surface, matching what the static
loop already did. Fixture cases pin both sides: the halo over a url()
image ancestor flags in both engines, and an offset chromatic shadow on
the same unknown surface stays abstained.
Also hardens the currentcolor background substitution with the
parseColorResolved fallback used by the text-color path, and adds fixture
coverage proving tokenized currentcolor surfaces already resolve through
the static cascade (flag when knowable, abstain when the token is
undefined).
Addresses Cursor Bugbot review findings on PR #541.
AI-assisted-by: Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
* fix: abstain on translucent gradients over images, drop phantom color-mix stops
Two follow-up review findings on the merge with main.
A gradient leading a url() layer was treated as a resolvable surface even
when its stops are translucent, so the glow and AI-palette hunts averaged
wash stops (a 20% black wash reads as pure black) while the real surface
blends with image pixels the engine cannot read. resolveBackgroundInfo now
marks gradient-over-image unresolved unless every readable stop of the
leading gradient is opaque, in which case the gradient provably covers the
image and remains the scorable surface.
parseGradientColorsModern predated this branch's parseGradientColors
rewrite: its second regex pass re-extracted color tokens nested inside
color-mix() stops that the shared parser already captures whole via
balanced-paren tokens, appending ingredient colors that are never painted.
The worst-case stop ratio then invented low-contrast findings against a
color nobody sees. The helper is removed; all callers use the shared
parser, which covers the modern syntaxes it existed for.
Fixture coverage pins both: the translucent-wash-over-image glow abstains
in both engines, an opaque gradient over an image still flags in the
browser, and the color-mix wash case stays clean in the static engine.
Each new assertion was verified to fail against the previous engine.
Addresses Greptile and Cursor Bugbot review findings on PR #541.
AI-assisted-by: Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
Paul BakausandClaude Opus 5
github-actions[bot]
Paul BakausandGitHub
