Paul's decision on the tenth sweep's design question: hard vetoes (missing region, contradicted plate or text, SVG illustration, clipped plate, invented ink) stay unconditional; at overall >= HERO_MIN the numeric readings (ink colour, letter-spacing, line pitch, strip heights, box positions) print as advisories with the pass and belong to the polish pass. Every sweep-10 sample closes its hero under this condition, which settles 07 without another paid round.
Ninth-sweep defects: codeDrawn / container / bleed now persist into spec.json with WARN lines (an overridden refusal used to vanish from the record); font-match probes os.tmpdir() and points TMPDIR at .impeccable/tmp when the sandbox /tmp is unwritable (every ninth-sweep rank silently fell back to the catalog).
AI-assisted (Claude Code).
A reference, not the page: --r-<id>-x/y/w/h in % of the comp (plus cap height, font-size, family, weight where measured) to bind to any markup, and hero-reference.html with every region at its box and every plate placed with object-fit: contain, as a check on positions. Attacks the most common execution failure of weaker builders (badly positioned, overflowing, pushed below the fold) without dictating structure to strong ones; overlapping boxes are overlapping boxes and the gate reads pixels regardless.
AI-assisted (Claude Code).
From the human review's most repeated pin ('terrible svg instead of asset', on every model) and from sessions that wrote 'ship' with the hero open. Icons, arrows, chevrons, and runtime data charts stay code; diagrams, notation, and leader lines are plates.
AI-assisted (Claude Code).
From the final review batch: 'bad asset crop (crops are never allowed)' twice, 'letter spacing way too wide'. A crop resampled to the region scores 99.8% structure against the raw region; a produced plate scores 30-60.
AI-assisted (Claude Code).
A session's own note said its hero sat at 67 because the 10x10 grid boxes straddled two elements each, and it was right: every downstream measurement (cap height, line count, structure) inherited the slop. Text and control regions now snap to the dominant connected ink inside the span (page-ground threshold, dilated cells, masses touching the span's sides lose to inside masses), keep the span on the record for coverage, and can opt out with snap: false.
AI-assisted (Claude Code).
From the second review batch: the best build of the fifth sweep passed the hero at 87% with the cover arch cut flat on the left (object-fit: cover on a box narrower than the shape), and legends, badges, and extra controls one or two cells wide slipped under the invented-ink floor. comp-spec measures the artwork's contiguous contact with each box edge against the page ground; the spec gate refuses such a box unless bleed is set. Human pass line landed at comp-diff 72-73; HERO_MIN stays 0.72.
AI-assisted (Claude Code).
From the first human review of sweep-3 builds (pins on 12 samples): fonts at the wrong size, weight, colour, or place; nav bars too tall; kickers and dividers the comp does not have; a footer strip pushed off the frame that read as drift on ground colour alone; a drawing filed as chrome with no note to catch it.
- lib/hero-checks.mjs: textRegionCheck (cap height, line count, ink density, ink colour, first-line offset vs the comp crop, measured fresh), chromeStripCheck (first rule row), inventedInk (build energy over a calm comp cell and neighbourhood). Wired into gateHero as reasons; invented ink vetoes at 4% of cells.
- comp-diff verdictFor: detailRaw < 0.15 is missing whatever the palette.
- comp-spec: every region carries a note.
- font-fingerprint: the tall-line filter takes its median over lines with real mass, so two display lines above a small line are not 'tall'.
AI-assisted (Claude Code).
A session named seven regions for a page with three plates, a table, a note, callouts and a spine, so the hero gate could name nothing and the score sat at 70. A code region is one element; a column is a container of several.
AI-assisted (Claude Code).
font-match / catalog index
- Index schema 2 adds a third render, 48c (48px cap, ALL CAPS text). Caps headline crops have no x-height band and ranked against mixed-case renders as barcode faces; they now route to the caps render.
- Non-text families (barcodes, redacted, flow, dingbats, effect faces) are excluded from candidates.
- The distance adds a gross width and weight gap (log ratio of advance and ink density) so a face 50% wider or 35% lighter cannot rank first on run-length detail alone; the index stores those readings.
- Multi-line all-caps crops vote on x-height across lines: one line's crossbars no longer give the crop a spurious x band.
- With no browser, --rank records the catalog's nearest face (source catalog, size estimated) so the spec gate can close; the NEXT line and new-work.md say not to install a browser or hand-write a choice.
- font-match stamps the choice it writes; the spec gate refuses a chosen face it did not write (sessions typed Arial Narrow into spec.json to pass).
- IMPECCABLE_NODE_MODULES lets a harness lend a playwright.
comp-spec
- A region note that describes painted material (diagram, drawing, photo, texture...) under a code kind is refused at the spec unless codeDrawn is set: the exploded carburetor filed as chrome is a plate about to be redrawn in SVG.
build-phase
- The control ink-box veto applies only when the comp's ink is a discrete element and the build's box is too; a full-width bar told one session six times that 1376x87 was 1382x102 with no edit able to move it.
AI-assisted (Claude Code).
lib/font-fingerprint.mjs replaces the three-number fingerprint with
size-invariant shape features (x-height ratio, stroke contrast, stem width,
run-length quantiles, roundness, serif signal, width spread) and a
noise-normalized distance; family recall on a held-out self-test rose from
13% to 72% top-5. data/font-index.json carries the whole Google Fonts
catalog (3,092 faces at two cap sizes, 707 KB); font-match --rank fingerprints
the comp crop, takes the 25 nearest faces from the index (plus the model's
own names), renders them at the comp's cap height, ranks by the same
distance, and prints a proof sheet and the CSS to use. scripts/build-font-
index.mjs rebuilds the index at release time.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
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>
Round-1 simulated builds passed the hero at 1536 and shipped a page whose
first viewport collapsed to one column at 1440 (comp-diff 50% on the
final capture, 82% on the hero). The responsive phase now requires
desktop.png and mobile.png and diffs desktop.png against the comp at 65%
with no region missing.
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>
concept-seed's choice ping now prints the next mandatory step from the
recorded build path (comp-led: build-phase.mjs start --direction <key>),
because every run that skipped the comp round did so right after that
ping. build-phase gains a comps phase ahead of spec (three sidecar'd comps
under .impeccable/mocks/, one approved) and records the approved comp on
close. The hero gate lists the worst region crops first with the fix class
per verdict, and refuses a third value-only attempt on the same stuck
region. Hero instruction is plates first, then the semantic layer. The
finish reviewer treats a comp-led build with no closed comps phase as a
material finding.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
The first live run produced a faithful carburetor plate, then drew the
region in SVG and left the plate on disk. Before diffing, the hero gate
now walks the artifact (or a bounded source tree) for every plate's file
name or a data URI named for it.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
The first live run forced past the plates gate with 'single-file HTML
delivery requires embedded CSS/SVG'. That is not a reason the comp's
authority moves for; the script now refuses a --force whose reason does
not quote the user, and new-work.md says a single-file deliverable inlines
the plate as a data URI. Texture plates are judged on palette and grain,
not size or structure, since they tile.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
new-work.md section 6 becomes the spec / plates / hero / sections / motion /
responsive phase list, each closed by build-phase.mjs advance; the
reproduction and hero-checkpoint prose that asked the model to compare
from memory is gone. visualize.md's inventory, medium gate, and produce
sections collapse into the spec and plate mechanism. The finish reviewer
takes the state file and comp-diff reports as inputs and starts its
fidelity matrix from the measured verdicts. docs/COMP-FIDELITY.md records
the design.
AI-assisted (Claude).
Co-Authored-By: Claude <noreply@anthropic.com>
Grok was classified as GitHub Copilot, so the design hook skipped every
edit with no-file-path and never ran Stop. Normalize toolInput/sessionId
and treat Stop additionalContext as the Grok product.
Prepared with AI assistance.
Co-authored-by: Cursor <cursoragent@cursor.com>
Claude Code now folds multi-edit behavior into Edit, so keep generated and repaired hook manifests aligned with the current Edit and Write tools. Grok keeps its compatibility matcher unchanged.
AI assistance was used to implement and validate this change.
* Fix: stop the direction page hanging forever after a re-roll (#469)
The re-roll leg of the decision-page protocol was documented only in
serve-question.mjs's own header, so agents never ran --update and the
open tab polled a round that could never arrive. Compounding failure
modes: the page poll swallowed every error, the daemon's --timeout was
an absolute guillotine that killed the server under a still-open tab,
a choice posted to a dead server confirmed nothing, and refresh or
Reload on an unresolved round resurrected heartbeats that held the
daemon alive indefinitely.
- new-work.md documents the re-roll leg: rerun concept-seed with
--from/--reroll, deliver with --update on the same key, never --start
a second server.
- The page poll terminates and says why: eight consecutive fetch
failures means the server is gone; the delivery deadline (the
server's own --idle-grace, inlined into the page) passing means the
hand never arrived. Both stop heartbeating.
- The daemon's --timeout bounds only the wait for a page to open; once
the page heartbeats, the server lives while the page does and exits
after --idle-grace (default 600s) without a beat, including under
--timeout 0.
- Build this and Re-roll against a dead server fail loudly instead of
silently swallowing the click.
- The server tracks the window between a collected re-roll answer and
the --update that replaces the round, and serves the page in waiting
mode there, so a native refresh re-enters the same bounded wait
instead of resurrecting dead cards; the in-page Reload button only
revives a delivered hand.
- --update is exempt from the headless gate and its liveness probe
trusts a fresh heartbeat over a failed kill probe (sandbox EPERM is
not death).
Squash of the six review-round commits on this branch, rebased onto
main after the decision-page revamp.
AI assistance: prepared with an AI agent operating under maintainer
instruction (abdulwahabone).
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review findings: persist the replacement deadline, refuse unloadable hands
A browser-native refresh of the waiting page re-entered the bounded wait
with a fresh delivery deadline and an immediate heartbeat, so refreshing
before each deadline expired could hold the daemon alive and keep --wait
on WAITING indefinitely. The server now records when the re-roll or
followup answer was collected, each served waiting page inherits only
what remains of that one allowance, and a page served after the deadline
renders stalled immediately and never starts its heartbeat.
And a next hand the round could not load used to reload-loop the tab:
GET /'s catch kept the file on disk, so /next-status stayed ready:true
forever. --update now refuses a payload without a non-empty options
array at the sender, and GET / discards an unloadable next file so the
bounded wait resumes.
AI-assisted (Cursor agent) under maintainer instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review finding: a stalled page recovers a late hand without a click
The stall silenced heartbeats so the idle grace could reclaim the
daemon, but that silence read as a closed tab: after a late --update,
--wait saw the stale beat and reported PAGE CLOSED while the user sat
on the Reload screen, so the agent abandoned the browser path the
recovery UI exists for. The stall screen now keeps a beat-free
/next-status watch that reloads into a delivered hand on its own
(GET never beats, so an abandoned flow is still reclaimed), and --wait
no longer concludes closure from a stale beat while an undelivered
next hand sits on disk.
AI-assisted (Cursor agent) under maintainer instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review finding: a delivered hand must not mask a closed page
The mid-delivery suppression keyed on the next file existing, but a
closed tab never claims that file, so an unconsumed delivery held
--wait on WAITING indefinitely instead of reporting the closed flow.
The suppression is now age-bound: a stalled page's watch reclaims a
delivered hand within seconds, so a file still unclaimed after a 10s
grace means no page is coming back and the stale beat reads as the
closed page it is.
AI-assisted (Cursor agent) under maintainer instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review finding: stamp the delivery clock at --update, not the copy
--wait's mid-delivery grace reads the next file's mtime, but
copyFileSync's timestamp behavior is the platform's business: a copy
that preserves the source payload's older mtime would start the grace
already spent and report PAGE CLOSED under a live stalled tab. --update
now touches the delivered file itself, so delivery time is delivery
time everywhere.
AI-assisted (Cursor agent) under maintainer instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review findings: disable canon during the wait, validate --timeout
The waiting and stall screens disabled only the re-roll buttons; the
footer canon action stayed clickable, and a canon pick posted after
--wait had consumed the re-roll could never be collected: it overwrote
the answer, marked the table closed, and exited the daemon under the
agent. Both disable sites now take the canon exit down with the re-roll
buttons; a delivered hand reloads the page and serves it live again.
And --timeout reached the lifetime timer unvalidated: NaN or a negative
value disarmed the no-page exit and the daemon leaked. It now takes the
default unless the value is a finite non-negative number, keeping 0 as
the explicit wait-forever.
AI-assisted (Cursor agent) under maintainer instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review finding: a second click must not renew the delivery deadline
dealAgain left the re-roll and canon controls live through the answer
POST and the 700ms fly-out, so a second click posted another re-roll
and the server restamped awaitingNextSince, renewing the deadline this
PR made non-renewable on refresh and on the stall screen. The controls
now go quiet at the click itself, in dealAgain and in answer(), and the
server stamps the allowance only on the transition into the wait, so a
duplicate answer racing the disable keeps the first stamp.
Regression coverage on both sides: the unit deadline test posts a
duplicate re-roll mid-allowance and asserts the budget shrank instead
of resetting, and the e2e stall test asserts both controls are disabled
immediately after the click, before the fly-out.
AI-assisted (Cursor agent) under maintainer instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review finding: a late delivery must survive its claim window
--update could land a replacement hand after the stalled page went
silent but moments before the daemon's idle deadline: the daemon exited
before the page's 1.5s watch could claim the hand, orphaning a delivery
--update had confirmed, and the next --wait reported a server failure.
The idle exit now defers while an unclaimed next hand is younger than
the claim grace --wait already reads (extracted as one shared
constant), so the page's watch deals it and heartbeats resume; a file
unclaimed past the grace still ends the daemon, bounded as before.
Regression test: deliver at idle-deadline-minus-a-beat, assert the
daemon survives past the deadline and serves the late hand.
AI-assisted (Cursor agent) under maintainer instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review finding: the claim itself must hold the daemon
The idle-exit hold read only the next file's freshness, but GET /
deletes that file when it serves the claimed round, before the
reloading page can post its first heartbeat: a lifetime tick in that
gap saw no pending hand and a stale beat, and exited under the hand
just claimed. GET / now stamps the claim when it consumes a pending
hand, and the idle exit honors the same bounded grace from that stamp,
so the reloading page gets its seconds to beat while an abandoned claim
still ends the daemon at the grace.
The claim-window regression test now also fetches after the claim, past
another lifetime tick, and asserts the daemon survived the gap;
verified it fails on the previous commit.
AI-assisted (Cursor agent) under maintainer instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix review finding: --wait must ride out the claim gap too
The claim deletes the next file --wait's mid-delivery grace watches,
and the reloading page has not beat yet, so --wait in that gap read the
stale beat as PAGE CLOSED while the daemon was alive serving the dealt
round, and the agent abandoned a browser session that had just
recovered. GET / now persists the claim stamp into the per-key state
file, and --wait's suppression honors it under the same bounded grace:
a fresh claim stays WAITING, a claim nobody followed with a beat still
reads as the closed page it is.
Regression test drives --wait through the gap (claim with a stale beat:
WAITING, not exit 4) and past it (backdated claim stamp: exit 4);
verified it fails on the previous commit.
AI-assisted (Cursor agent) under maintainer instruction.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Paul Bakaus <paul.bakaus@gmail.com>
The quality bar leaves the color-authority chain (it arrives as card
image paths and never governs composition). With no comp, a color
OWN-WORLD names is the target; when it names none, the review states
there is no GROUND authority instead of inventing a target. The build
side of the numeric comparison now samples the same way each record
was taken: patch average against patch average, gradient ends against
gradient ends.
AI-assisted change (Cursor), prepared under maintainer direction.
Co-authored-by: Cursor <cursoragent@cursor.com>
GROUND no longer lapses silently on code-led builds: with no comp to
sample, the authority is the colors OWN-WORLD and the quality bar name,
and no invented target beyond them. Non-uniform fields get sampling
rules (interior pixel, patch average for texture, both ends of a
gradient, never an edge), and the numeric comparison gets tolerance
semantics so render noise never fails a faithful build. The hunt hint
names the dark-ground prior beside the light one.
AI-assisted change (Cursor), prepared under maintainer direction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix: keep raster provenance through the finish-review fix loop
Three runs (two harnesses) showed the parent generating production
rasters after the producer returned: no exact embedded prompt, no
inventory row, orphan files. The asset contract in visualize.md was
phase-scoped to the build while new-work.md's fix loop licensed
"produce the named assets" with no rules attached.
- visualize.md: name the provenance contract, require the exact tool
payload, and scope it to the run, fix rounds and rebuilds included.
- new-work.md: bind fix/rebuild rasters to the contract, add an
embed-prompt --scan step before the verdict round, and extend the
FINISH line to carry the condition through long builds.
- embed-prompt.mjs: add --scan mode listing rasters missing a prompt
(exit 3 when any), reusing the existing read path.
AI-assisted change, prepared with Cursor under maintainer direction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add cursor-control-8 comp vs final screenshots for PR evidence
AI-assisted change (Cursor), prepared under maintainer direction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add cursor-control-9 comp vs final screenshots for PR evidence
AI-assisted change (Cursor), prepared under maintainer direction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Address review findings on the provenance gate
- Hoist the provenance rule out of the fix disposition into its own
paragraph binding rebuild and fix alike, gated before either round's
result goes back for review or verdict (Bugbot: rebuild skipped the
scan when its fresh review shipped).
- A scan-flagged raster gets the record it is missing embedded, exact
prompt for produced, origin for sourced/stock/pre-existing; deletion
is reserved for abandoned rasters, never scan hits (Bugbot: gate hit
non-generated assets on extensions).
- Document the scan command with its required directory argument
(Greptile: literal command exited before scanning).
- Align the FINISH line on the provenance token.
AI-assisted change (Cursor), prepared under maintainer direction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Remove evidence images from the diff; they live on the pr-evidence branch
AI-assisted change (Cursor), prepared under maintainer direction.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Accents join the sampled record alongside ground and dominant fields,
every recorded color (not only the ground) is compared by number
during the build, and the light-ground-only rationale clauses become
value-neutral so dark and saturated comps get the same protection.
Rule anchor renamed to skill-color-by-number to match its scope.
AI-assisted change (Cursor), prepared under maintainer direction.
Co-authored-by: Cursor <cursoragent@cursor.com>
Sample the approved comp's ground and dominant-field hexes into the
brief (visualize.md), judge the built page's ground by number against
that record including the net value under textures (new-work.md), and
make GROUND a mandatory fidelity-matrix row beside TYPE and MATERIAL
(finish reviewer). Pre-comp palette chips are retired at approval.
AI-assisted change (Cursor), prepared under maintainer direction.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Comp-fidelity review discipline + conciseness pass on core references
Process fixes derived from a real Codex session (Hanasaku landing page)
where a build drifted wholesale from the approved comp and still shipped
under a reviewer pass:
- finish reviewer: new Evidence check (check 0) with a fourth
disposition, recapture, for malformed screenshots; a review on invalid
evidence binds nothing and owes a full re-review, not a verdict pass
- finish reviewer: verdict passes exit scoring mode when recaptures fail
check 0 or when the packet carries user-supplied screenshots that
contradict a prior verdict (those force a fresh full review); a ship
earned in a verdict pass covers the scored fixes, not the whole surface
- new-work: capture-validity rules (settle entrance motion, capture from
document top, comp comparison at comp dimensions, open every file once
before sending); user's actual viewport joins the inspected sizes
- new-work: hero checkpoint now writes .impeccable/review/hero-repro.png
and the reviewer verifies it exists under Persistence
- new-work: comp authority is explicit (only the user can downgrade it);
handoff reports the verdict at its actual scope; user evidence reopens
a full review; documenter re-runs when fixes land after documentation
- craft-floor: Refuse entry for geometric masks approximating organic
photographic contours (the circular-cutout failure)
- editorial conciseness pass over new-work.md, visualize.md, and both
agent files: tighter sentences, no dropped rules, all rule markers and
mechanical tokens preserved
Assisted-by: Claude Code
* fix: define the ship disposition in new-work's action paragraph
Copilot review finding: the paragraph claimed exactly four disposition
words but defined only recapture, rebuild, and fix.
Assisted-by: Claude Code
* fix: rebuild returns get a full review; recapture return shape in preamble
Cursor Bugbot findings:
- a return following a rebuild directive is now a fresh full review on
both sides of the contract, never a verdict pass, so a wholesale
rebuild cannot earn a scoped ship on the directive alone
- the turn-ceiling preamble now names the recapture return shape instead
of contradicting it with "the five sections"
Assisted-by: Claude Code
* fix: absent required captures fail the evidence check
Greptile finding: a packet with no desktop.png/mobile.png (or missing
native device-class captures) routed to the missing-input notice and
could still reach ship. A required capture that is absent now fails
check 0 exactly like a malformed one and forces recapture; the
missing-input allowance in the preamble excludes captures.
Assisted-by: Claude Code
* fix: user-viewport capture is a required, named input to the review
Greptile finding: the evidence gate hard-coded web requirements to
desktop.png and mobile.png, so a reported user viewport could join the
inspected set and still ship uncaptured. The parent now saves it as
user-<width>.png and names every inspected viewport required in the
packet; check 0's required set includes every brief-named capture.
Assisted-by: Claude Code
Two review findings on #583.
"context.mjs reports it" undersold what counts: that directive only fires on an
OPENAI_API_KEY, so a harness with a native image tool and no key generates
images while the boot output says nothing. Read literally, the step would skip
the question exactly where the toggle belongs. It now names both sources and
says a silent boot is not evidence.
The unanswered branch said to state which path the session takes without saying
which one it is. Left implicit, an agent picks its own, which is the failure
this step exists to stop. It now names comp-first, the default new-work applies
when nothing is recorded.
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A live Codex session folded the build path into the stack question as a
trailing recommendation ("I recommend static and code-first"), never said what
either name means, treated its own recommendation as the user's answer, and
wrote `buildPath: code` as a standing default. The user disagreed and flipped
the board to comp-first, but a flip binds one session, so the unasked default
stayed on disk to steer every later round.
Step 5 said to "ask once ... stated as the trade it is", which the run
violated, but the step left the shortcut open: it sits after the interview
ends, it never says a recommendation is not an answer, and unlike the stack
question it offers no way to end without a value, so an agent holding no answer
writes one anyway.
Now: it is its own question, never a clause inside another; the trade is stated
in the question the user reads, because the two names mean nothing on first
contact; only the user's own choice is written; and an unanswered question
records nothing and says so. Unset is a working state, since the page toggle
governs the session and new-work's one-time offer still captures the answer at
the first flip.
The same run also wrote "Code-first build path" into PRODUCT.md's `## Stack`,
so the step now says the config is the only place this lives: a copy in product
truth outlives the setting and steers rounds nobody can trace back to it.
Written with AI assistance (Claude Code).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Round two of review findings, all four valid.
`doctor` builds its own finding list and never called `checkBuildPathUnset`,
so `config-build-path-unset` could not appear in the report even though
doctor.md documents it. That is also the only path left once stalenessCheck is
off, which is exactly when someone is looking for it.
The lookup chain included `process.cwd()`, which lets an ambient invoking
directory decide another project's workflow: run from workspace A with
--target resolving onto workspace B, and B inherited A's buildPath ahead of
the repository default. The chain is now the resolved project then the repo
root, matching `checkBuildPathUnset` exactly; cwd stands in only when no
project resolved at all.
Two prose contradictions, both mine. new-work said to write the value "when
the user says yes" and then to "record the answer either way", which reads as
persist-on-yes-only and leaves the decline to be asked again next session. It
now says the write always happens and the answer picks the value. The README
still pointed existing projects at re-running init, which is the problem this
PR exists to solve; it now names the toggle as the migration path.
The workspace-isolation test earned a correction of its own: the first version
passed a relative --target, which resolves against the caller's cwd and puts
projectRoot back on the calling workspace, so it asserted nothing.
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>
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>
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>
Paul Bakaus
