Markdown across every repo had accumulated a dozen ad-hoc names (refactor_claude_todo.md x23, DESIGN_AUDIT_TODO.md, NEXT_SESSION, HANDOFF, PLAN, SPEC) with nothing distinguishing a live task from a finished record. An audit found docs asserting work was pending that shipped months earlier, and one asserting completion while the work was open. Four namespaces now: README.md, CLAUDE*.md, DOCS*.md, TODO*.md. Task files carry the literal "REMOVE ME AFTER FINISH" and are deleted when the work lands. Enforced bidirectionally -- a TODO without the marker fails, AND any file carrying the marker that is not named TODO*.md fails. Without the reverse rule a task can be renamed into the permanent namespace and quietly live forever, which is the failure being fixed. md_naming/ mirrors the file_length topology and imports its tables (EXCLUDED_DIRS, THIRD_PARTY_REPOS, VENDORED_ANYWHERE, GENERATED_PATTERN) rather than restating them, so the two gates cannot disagree about the same file. 71 tests, 100% branch coverage. stale.py warns on TODOs past 90 days but never fails: no machine can prove a task is finished, so the honest mechanism is a nudge. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01A47aX5DJwmA3QFn3pkEWi7
10 KiB
Unified design tokens (frozen values)
Concrete instantiation of safe-design-rules' 28 rules. One identical
palette across every repo/stack (Flutter, web, Python/Tkinter) — not just
shared rules, the same hex values everywhere, including Flutter launcher-icon
glyphs (app-icon skill).
This table is checked, not trusted.
scripts/palette_check.pyparses all four token sources — this file,web_ui/src/tokens.css,design_system/lib/src/tokens.dartand gatelock'sLockConfig— and fails on any hex that disagrees across them, including a hex edited only here. It also fails on a colour token that no stack's map accounts for, so adding a token to one stack and forgetting the others is a CI failure rather than silent drift. Runs in.github/workflows/palette-drift.yml, deliberately with nopaths:filter: drift is a disagreement between packages, so a path-scoped trigger would skip exactly the edit that causes it. Editing any value below means editing it in all four places. Anchored on the shared icon charcoal#211D1Bso the UI palette and the icon family cohere instead of adding a second "near-black." Warm-tinted throughout (rule 10) — superseded from an earlier cool-tinted draft; every neutral is a byte-reversed (RRGGBB→BBGGRR) mirror of that draft's value, so relative lightness/contrast steps are unchanged, only the tint shifts from blue toward red/brown.
Neutrals
| Token | Hex | Notes |
|---|---|---|
ink |
#211D1B |
Dark bg. Same value as the icon charcoal. |
ink-raised-1 |
#2B2624 |
Dark elevated surface, step 1. |
ink-raised-2 |
#38312E |
Dark elevated surface, step 2 (≤12% luma vs bg). |
line-dark |
#463E3A |
Border on dark surfaces. |
text-on-dark |
#ECEAE9 |
Primary text on dark. Near-white, not pure. |
muted-on-dark |
#AAA09A |
Secondary/caption text on dark. |
paper |
#F6F4F3 |
Light bg. |
paper-raised |
#FCFBFB |
Light elevated surface — lighter than bg (rule 15). |
line-light |
#E0DAD7 |
Border on light surfaces (≥3:1 vs bg and card). |
text-on-light |
#211D1B |
Reuses ink — deliberate symmetry. |
muted-on-light |
#70625B |
Secondary/caption text on light. |
Accent + semantic roles
One set, used identically everywhere — UI, icons, everything. All four now cluster in the red-orange-yellow-olive wedge, distinguished mainly by lightness (danger 130 < accent 143 < success ~150 < warning 171 — a flagged, deliberate tradeoff, not an oversight): a colorblind/at-a-glance distinguishing risk the user accepted over keeping green/red as a semantic exception.
| Token | Hex |
|---|---|
accent |
#B8862E |
info |
#B8862E (= accent) |
success |
#8A9A3C |
warning |
#E0A63C |
danger |
#E2585F |
on-fill |
#211D1B (= ink) |
on-fill: text/icon color for anything drawn on top of a filled
accent/success/warning/danger surface (a primary button, a filled badge) —
never text-on-dark. All four fill colors sit in the same mid-light
lightness band (that's the accent/success/warning/danger clustering noted
above), so near-white text under-contrasts on every one of them (measured
2.2:1–3.6:1 against WCAG's 4.5:1 body-text floor); ink text passes on all
four (5.8:1–9.7:1). This is a closed rule, not a per-color judgment call —
never pick text color per fill, always on-fill.
Do not confuse on-fill with text on a scrim/overlay (a dark
semi-transparent backdrop behind a modal or a translucent control chip over
an image) — those stay text-on-dark (near-white), fixed regardless of
page theme, since the scrim itself is always dark. on-fill only applies to
the four opaque semantic fills above.
Categorical ramp
For charts, tags and category dots — anything that needs N mutually distinguishable colours with no inherent meaning. Never for semantic state: success/warning/danger still own that, and mixing the two vocabularies is what makes a "red" category read as an error.
| Token | Hex | L* |
|---|---|---|
cat-1 |
#C57293 |
58 |
cat-2 |
#398FC0 |
56 |
cat-3 |
#C85A32 |
52 |
cat-4 |
#228736 |
49 |
cat-5 |
#8D58BB |
47 |
cat-6 |
#686D2C |
44 |
This ramp deliberately leaves the warm wedge, and it is the only token group that does. Accent/success/warning/danger cluster in red-orange-yellow-olive and are separated mainly by lightness — a tradeoff that works when meaning carries the signal. A categorical ramp has no meaning to lean on, so hue separation is the requirement. Treating this as drift and "correcting" it back toward the wedge would destroy the only property it has.
Chosen by search under four hard constraints, not by eye: pairwise CIE ΔE ≥ 20 under normal vision and simulated deuteranopia/protanopia/tritanopia, ≥ 3:1 against both backgrounds, monotonically decreasing lightness, and muted saturation.
Do not copy the measured margins into this file. Run the check — it prints every one of them and fails if any regresses:
python3 unified-design-system/scripts/ramp_check.py
Two consequences worth keeping in prose, because the numbers alone do not say
them. The both-backgrounds rule is the binding constraint: colours legible on
a near-black and a near-white field only exist around L* 44–59, which is why
the ramp is mid-toned rather than spanning light to dark. And because lightness
falls monotonically, hue and lightness encode redundantly, so cat-1…6 still
reads as an ordinal scale in greyscale or under severe CVD.
Use them in order for ordered data, and by position for unordered data — never pick "the green one" because green means good, which is the semantic vocabulary leaking back in. Need a seventh category? That is a signal to group the tail into "other", not to extend the ramp: a 7th hue cannot clear ΔE 20 against all six under CVD within the L* 44–59 band.
Spacing scale (4px base)
xs 4 · sm 8 · md 16 · lg 24 · xl 32 · xxl 48
Every spacing/sizing value in a file should land on this scale. Round outliers to the nearest step rather than introducing a new one.
Corner radius scale
sm 8 (buttons, inputs, chips) · md 12 (cards) · lg 16 (sheets, dialogs).
Nesting (rule 24): an inner element's radius = outer radius − the gap between
inner and outer edges, computed per instance — not a fixed constant. E.g. an
lg (16px) container with 4px of padding before its child starts should give
that child a 12px radius, not reuse 16 or invent an unrelated value.
Typography scale
| Role | Size (px) | Tracking |
|---|---|---|
display |
32 | tight (e.g. −0.2px) |
title |
24 | tight (e.g. −0.1px) |
subtitle |
20 | default |
body |
16 | default — the floor (rule 20) |
label |
14 | loose (+0.05–0.1em), chrome only |
caption |
12 | loose (+0.1–0.2em), chrome only |
body (16px) is the minimum for anything a user actually reads.
label/caption are for UI chrome (timestamps, badges, tags) — never for
primary reading content.
Two typefaces max (rule 23): one system sans stack for all UI text
(-apple-system, "Segoe UI", Roboto, system-ui, sans-serif on web; platform
default on Flutter/Tk), one monospace stack reserved for code/data-dense
contexts only.
Line length (rule 21): prose/paragraph blocks capped at max-width: 40rem
(~640px, ~65–70 characters).
The sizes above are pixels. On Tk that means a negative number.
Tk encodes the unit in the sign of the font size: positive = points,
negative = pixels. So ("Arial", 16) is 16 points, not 16 pixels, and
renders ~37% larger than this scale specifies. Measured:
| Spec | size=16 (points) |
size=-16 (pixels) |
|---|---|---|
linespace |
26px | 19px |
ascent |
21px | 15px |
Always write the negative form in Tk — font=(family, -16) for body. Passing
the positive px value straight through inflates every string in the app by a
third, which is enough on its own to push a 768px-tall layout off-screen (it
did: it was the root cause of the diet-guard meal gate needing 974px in a 724px
pane). This is a silent failure — nothing errors, the text is just wrong-sized —
so it is checked by the gate rather than left to review.
Flutter fontSize and CSS px are already pixels; no sign convention there.
Buttons
Horizontal padding = 2× vertical (rule 22). Canonical: vertical 12px,
horizontal 24px (both on the spacing scale: 12 = 3×4, 24 = 6×4).
Shadow policy (rules 16, 26, 27)
Pick exactly one depth technique per surface — never a shadow and a lighter fill on the same element, never a border-only style next to a shadowed one at the same z-level.
- Dark surfaces: no shadows, ever. Elevation is
ink-raised-1/-2fill steps only. - Light surfaces: shadows allowed only on floating/overlay elements
(dropdown menus, modal dialogs) — never on inline cards or list rows.
Two tiers, blur = 2× offset, tinted from
ink(never pure black):- Level 1 (e.g. a dropdown):
0 2px 4px rgba(33, 29, 27, 0.12) - Level 2 (e.g. a modal):
0 4px 8px rgba(33, 29, 27, 0.16)
- Level 1 (e.g. a dropdown):
Icon-next-to-text contrast (rule 28)
Icons adjacent to text render at ~72% opacity of the text's color, or use the
muted role instead of the full-strength text role — never equal contrast.
Focus ring
| Token | Value | Notes |
|---|---|---|
focus-ring |
#B8862E (= accent) |
2px, offset 1px. |
Why the platform default is not good enough, and the Tk
highlightcolor/highlightbackground trap, are in
operability.md.
Operability and screen size
Keyboard-only reachability and the small-screen budget are requirements, not
suggestions: several of these apps are lockers that hold a global input grab,
so a control only a pointer can reach is a bricked machine. Both, plus the
focus-ring token, are in operability.md.