jig-ui 0.7.1 → 0.8.1

package/CHANGELOG.md

@@ -0,0 +1,822 @@
+# Changelog
+
+## 0.8.1
+
+Both fixes here are the same shape: 0.8.0 corrected the instance it was looking
+at and left the class alone, and in each case the commit message claimed a
+verification that had only checked the file it had just edited. Both were found
+by inspecting the published tarball rather than the working tree.
+
+### Fixed
+
+- **Four rule files still shipped author notes.** 0.7.x removed the section
+  headed "Notes for the author (not for the agent)" from `00-anti-patterns.md`
+  after an agent read it as "license to go tighter than the shared default" and
+  halved the radius scale on that authority. `01-modes.md`, `03-patterns.md`,
+  `04-principles.md` and `05-copy.md` carried a section under the same heading
+  and kept shipping it to every agent that installed Jig.
+
+  `01-modes.md` was the worst of them: it told the reader to **"Change them in
+  `tokens/mode.*.css`"** and closed with **"it is your call"** — an invitation
+  to edit the token layer, which is the one thing Jig's architecture reserves
+  for a human. `03-patterns.md` described navigation and cards as "deliberately
+  absent… add them once you have built enough". `05-copy.md` pointed at
+  `03-brand.md`, which does not exist. The content moves to
+  `docs/house-positions.md` unchanged; only its audience changes.
+
+- **`02-tokens.md` contradicted its own Tailwind fix 352 lines earlier.** The
+  compatibility table near the top still read *"Tailwind v4 | Wrap in
+  `@theme { }`"* — the exact instruction the section below it retracts with
+  *"Earlier versions of this file told you to, and Tailwind rejects it
+  outright."* The table is the part an agent reads first, so the correction
+  shipped underneath the error it corrected. A second, softer restatement
+  ("the same file can be wrapped in `@theme`") is gone too. `@theme` now first
+  appears in the section that explains it correctly.
+
+- **`jig explain` rendered a dangling `---` in 14 of the 15 pattern and mode
+  specs.** `parse.ts` has always dropped bare separator lines; `specs.ts` is a
+  separate code path and never did, so every `P-` and `M-` spec that is
+  followed by a separator in the source carried it into the rendered body,
+  between the last paragraph and the footer. Spotted on `P-12`, but it was
+  never about `P-12`. Table separators (`| --- |`) are untouched.
+
+- **`jig init`'s refusal without a terminal offered a way out that does not
+  work.** The message ended *"(To choose the mode without a terminal, write
+  jig.config.json first — init honours it.)"* The guard runs before any config
+  is read, so a config alone still exits 1. The sentence was true about mode
+  selection and false in a paragraph about not having a terminal, so it read as
+  a third alternative when it is a modifier on the first — a cold agent
+  followed it literally, hit the identical error, and allocated a
+  pseudo-terminal with Python's `pty` to get past it. It now says a config does
+  not replace `--yes`, and what the two do together. Nothing covered this path;
+  three tests now do.
+
+### Added
+
+- **`check-tokens` rule 13 — no shipped rule file addresses the author.** The
+  guard that should have existed for the 0.7.x fix. It reads `rules/` from disk
+  rather than a hardcoded list, so a rule file added later cannot escape it the
+  way those four did, and it checks the second-person tells ("your taste", "it
+  is your call", "my inclination is") as well as the heading, because a rename
+  would otherwise defeat it. Each tell is verified to fire on reintroduction.
+
+## 0.8.0
+
+Everything here was found by handing Jig to agents that had never seen it and
+asking them to build something real — its own documentation site. None of it
+was found by the test suite, which passed throughout.
+
+Minor rather than patch: `explain` prints lines it did not print before, and
+`init` can write a file it did not write before.
+
+### Fixed
+
+- **`jig explain` discarded most of the reasoning in the rules.** The parser
+  kept the first `❌` and the first `✅` and threw away everything else. **41 of
+  104 rules carry prose outside that pair — 103 lines.** `A-04` explains why
+  shadow-only definition cannot hold 3:1; `C-22` loses sixteen lines; `C-49`
+  leads with a three-row table distinguishing when a link needs colour, when it
+  needs an underline, and when neither is required — and its pair alone says
+  "keep the underline". All of it shipped in the tarball, installed into every
+  skill directory, and was unreachable through the command built to read it.
+
+- **`02-tokens.md` documented a Tailwind arrangement that fails.** It told every
+  reader to write `@theme { @import "./jig/theme.css"; }` and claimed it yields
+  `bg-surface`, `rounded-surface`, `p-card` and `text-body` as utilities.
+  Tailwind 4 rejects it: *"@theme blocks must only contain custom properties or
+  @keyframes."* The tokens cannot move into `@theme` regardless — it requires
+  them top-level and unnested, while they live in `:root` and are redeclared
+  under `[data-theme="dark"]` and a `prefers-color-scheme` query, which is what
+  makes dark mode work. The flat import is what works, and is what `init`
+  already wired without being told.
+
+- **`02-tokens.md` named a token location the tool had stopped using.** It said
+  tokens live at `.jig/tokens/`, that this was "the only location, in every
+  scope and every project", and that "nothing relocates them" — while `init`
+  printed `Token layer: src/styles/jig/` in the same run. `check-tokens` rule 4
+  hardcoded the same path in a comment anticipating and forbidding this exact
+  change, so the guard held the claim in place and would have failed the build
+  on anyone correcting it. Shipped in 0.7.1.
+
+- **A section headed "Notes for the author (not for the agent)" shipped to every
+  agent.** The heading was a label, not a mechanism: the file is in the tarball
+  and installs into every skill directory. A probe read it, quoted the
+  `A-07`/`A-08` line back as "the author's own notes… license to go tighter than
+  the shared default", and halved the radius scale on that authority. Moved to
+  `docs/house-positions.md`.
+
+- **The published package shipped no changelog**, despite the repo holding 32KB
+  of one. Staging and `files` now change together, which the tarball guard
+  already enforced.
+
+- **`A-01` had no stated scope for "then ask".** Two agents given the same brief
+  split on it — one shipped the unbranded default and deferred, the other
+  proposed a hue and argued the proposal *was* the ask — both citing the same
+  conflict-resolution clause. The rule now says a proposal is a question with a
+  suggested answer, not a decision.
+
+- **`B-11` capped line length and stated no floor**, so a column could be halved
+  indefinitely and still pass. Measured on the documentation site: two
+  side-by-side panels rendering prose at 41 characters, in a mode whose
+  `--measure-prose` selects 68. The 40–80 band existed in `02-tokens.md` but not
+  in the file an agent reads for a line-length decision.
+
+### Added
+
+- **Three detectors for the anti-slop rules, and a self-check that names them.**
+  Section A is Jig's answer to "what is AI slop" — fourteen rules for the
+  defaults a model reaches for when nothing was specified. **Three had
+  detectors.** The other eleven were enforced by one self-check question —
+  *"would this look different from a generic template?"* — which an agent that
+  has just produced a generic template answers yes to, and whose `A-01 → A-10`
+  range silently excluded `A-58`, `A-59`, `A-60` and `A-67`.
+
+  Jig's own documentation site shipped `<span aria-hidden="true">❌</span> Don't`
+  in the chrome of all 104 rule pages with `check --all` reporting nothing.
+
+  `A-05` (emoji as iconography) and `A-10` (placeholder content) were never
+  judgment calls — an emoji in a text node is a regex, and so is "lorem ipsum".
+  `A-09` (marketing voice) is the first **mode-gated** detector: it fires in
+  `product` and `operator`, stays silent in `editorial` where that register
+  belongs, and stays silent when no mode is declared rather than guessing.
+
+  The remaining eight stay judgment. Detecting "a container around every group"
+  needs to know what an element is and how heavy it looks, and six noisy checks
+  that fire on correct code is how a linter gets switched off. What reaches them
+  is the self-check, rewritten to name each tell the way every other item does.
+
+- **`check` reports what it examined.** `0 errors · 104 rules, 0 fired` was
+  byte-identical whether forty components were examined and found clean or
+  nothing was examined at all. It now reads `· 26 files, 4 with styles`, the
+  `JIG_CHECK:` record carries `files=` and `styled=`, and a run where nothing
+  carried a style region says **"Nothing inspected."** rather than "No
+  findings." Both counts, because `.ts` and `.tsx` are style-bearing by
+  extension and the file count includes parsers and configs that can never
+  produce a finding.
+
+  Worth knowing: `check` defaults to changed files, so on a clean tree it scans
+  almost nothing. That was always true and always looked like a pass.
+
+- **`init` offers a Tailwind v4 alias block.** The flat import gives you tokens,
+  not utility classes — Tailwind only generates those for names declared in
+  `@theme`. `init` can now generate that block, and **asks first**: it changes
+  how every component in a project is written, both styles are correct, and
+  silence is no. Under `--yes` it declines and says how to get it.
+
+  Only the 18 Tailwind namespaces are aliased. `--size-*`, `--measure-*`,
+  `--focus-ring-*`, `--border-width-*` and `--opacity-*` have none, and aliasing
+  one emits a declaration that generates nothing. 189 declared names filter to
+  87 — and the filter also excludes every primitive `02-tokens.md` says never to
+  consume directly.
+
+  The generated file explains why `--radius-surface: var(--radius-surface)`
+  appears beside Jig's own declaration in the compiled CSS: Tailwind's lands in
+  `@layer theme`, Jig's is unlayered, and unlayered beats layered regardless of
+  source order. Without that note, someone finds it and "fixes" it.
+
+- **The skill treats Tailwind setup as an ask.** Finding Tailwind in a project
+  is not permission to change how every component in it is written. The agent
+  reports what it found, shows both arrangements, and waits — and is told never
+  to write `@import "tailwindcss"` into a project that does not already have it.
+
+- **A guard that staged assets stay out of git.** Adding `CHANGELOG.md` to the
+  prepack staging list committed a build artifact, because
+  `packages/cli/.gitignore` is a hand-maintained list that nothing forced into
+  step. On its first run the new guard caught a second, pre-existing hole:
+  `references` had been staged and unignored since it was added, saved only by
+  the directory not existing yet.
+
+### Corrected
+
+- **0.7.0's notes said the non-TTY refusal "names the three ways out".** It
+  names three; two work. The guard runs before any config is read, so
+  `jig.config.json` alone still exits 1 — it selects the mode once you are past
+  the guard, it does not get you past it. The message is unchanged here and
+  recorded in `docs/known-follow-ups.md`, because the right wording depends on
+  whether the guard should consult the config first, which is a behaviour
+  question rather than a copy one.
+
+## 0.7.1
+
+A rule file that contradicted the tool, and the guard that kept it that way.
+
+Patch: the shipped rules change, but no command writes anything different.
+
+### Fixed
+
+- **`02-tokens.md` named a token location that 0.7.0 stopped using.** It said
+  tokens live at `.jig/tokens/`, that this was "the only location, in every
+  scope and every project", and that "nothing relocates them". Since 0.7.0 the
+  layer follows the project: `init` puts it beside the stylesheet it wires and
+  prints the path it chose. So a single `init` run told you `Token layer:
+  src/styles/jig/` while the rule file an agent is explicitly told to load
+  before writing token code insisted that directory could not exist.
+
+  The file also disagreed with itself — four absolute examples, and one
+  relative one added when the Tailwind section was written.
+
+  It now describes the real behaviour, says pre-0.7.0 projects keep their
+  layout, and points at the `theme.css` barrel so relocating the layer costs
+  one line instead of every stylesheet.
+
+- **`check-tokens` rule 4 was holding the claim in place.** It hardcoded the
+  same path, in a comment that anticipated and forbade this exact change:
+  "nothing — including a future `init` — relocates them". It was checking the
+  documentation against itself rather than against the CLI, so it could not
+  see the drift and would have failed the build on anyone correcting it.
+
+  It now asserts what stays true as the layer moves: token imports shown in
+  the rules must be relative. Mutation-tested in both directions.
+
+### Corrected
+
+- **0.7.0's notes said the non-TTY refusal "names the three ways out:
+  `--yes`, a real terminal, or writing `jig.config.json` first."** The third
+  is not a way out on its own. The guard runs before any config is read, so a
+  config alone changes nothing — it selects the *mode* once you are past the
+  guard. `jig.config.json` plus `--yes` works; `jig.config.json` alone exits 1
+  with the same message that suggested it.
+
+  The message itself is unchanged in this release and still reads as though
+  the config were a third alternative. It is recorded in
+  `docs/known-follow-ups.md` rather than reworded here, because the wording
+  that is actually right depends on whether the guard should consult the
+  config first — a behaviour question, not a copy question.
+
+### How these were found
+
+Not by the test suite, which passes 631 tests either way. By handing the rules
+to agents that had never seen this repository and asking them to build
+something real. Everyone who could have caught the token-location defect
+already knew where the tokens were.
+
+## 0.7.0
+
+A silent no-op, fixed, plus the missing half of 0.6.0's token-layer move.
+
+Minor rather than patch, and the call is arguable. The non-TTY fix is a
+correction; the relocation offer is new behaviour that can move files, though
+only interactively and only with consent. Taking the higher of the two is the
+reading that does not understate the change, and anyone who might be prompted
+to move their token layer should read these notes.
+
+### Fixed
+
+- **`jig init` without `--yes` exited 0 having written nothing when stdin was
+  not a terminal.** It printed the first prompt, read EOF, and stopped. Every CI
+  step, every piped invocation, and every agent shelling out without a TTY got a
+  silent no-op — and exit 0 having done nothing is the worst outcome available,
+  because the caller cannot tell it from success and then acts on a token layer
+  that was never created. Confirmed identical in 0.5.0, so this predates the
+  prompts added since. It now fails where the cause is still visible and names
+  the three ways out: `--yes`, a real terminal, or writing `jig.config.json`
+  first.
+
+### Added
+
+- **`init` offers to move a legacy `.jig/tokens/` layout** to the location your
+  project's own structure suggests. 0.6.0 deliberately left an upgraded project
+  where it was — a silent relocation could break an import you wrote that `init`
+  knows nothing about — but that meant the improvement only ever reached new
+  projects, and the way out was one line of output that is easy to miss.
+
+  Interactively it asks; under `--yes` it declines and says how. Moving is all
+  four halves or none: write at the new location, remove the old copies, strip
+  the stale import before wiring the new one, and update `jig.config.json` so
+  the next run does not go back for them. A file you have edited is never
+  removed — it is named and left for you to clear.
+
+  ```text
+  BEFORE  @import "../../.jig/tokens/brand.acme.css";
+          @import "../../.jig/tokens/mode.product.css";
+  AFTER   @import "./jig/theme.css";
+  ```
+
+## 0.6.0
+
+The token layer stops hiding in a dotfolder, and `check` starts reading it back.
+
+Both came from the same question: who should write the tokens. The answer turned
+out to depend on something that did not exist — **nothing validated the token
+layer's own declarations.** `init` checked a brand colour once, at write time,
+and no command ever looked again. A generated file edited afterwards, by a person
+or an agent, went unexamined: `--color-text-weak` dropped to 22% opacity and
+`check` reported "No findings". With that closed, where the files live and who
+writes them become ordinary decisions rather than load-bearing ones.
+
+**Upgrading:** run `npx jig-ui@latest update`, then `npx jig-ui@latest init`.
+Existing installs keep their `.jig/tokens/` layout — nothing moves unless you
+move it, because relocating files could break an import you wrote yourself.
+
+### Added
+
+- **`check` validates the token layer** (`check/token-audit.ts`). Text roles to
+  4.5:1, interface strokes to 3:1, **in both themes**, plus `--text-prose` at
+  18px and `--size-touch-target` at 48px. Alpha foregrounds are composited over
+  each surface first — Jig's foregrounds are alpha by design, so a ratio before
+  compositing means nothing. Only floors, never density: `--size-control` at 28px
+  is a deliberate `operator` choice, and reporting it would teach you to ignore
+  the ones that matter.
+- **`H-47` enforces the half of itself it only stated.** Its correction always
+  read "consume the semantic role, not the primitive"; the detector caught raw
+  hex and raw px and nothing else. `color: var(--brand-l)` was silent — and it is
+  the worse case, because it looks exactly like correct token usage while reading
+  a bare number and bypassing every theme override.
+- **`exempt` in `jig.config.json`.** Some surfaces render outside the cascade: an
+  OG card in an SVG `foreignObject` carries no stylesheet, a PDF renderer never
+  sees CSS. Those files were in permanent violation, which is an adoption blocker
+  — a check that cannot pass is a check people switch off.
+
+  Nothing is exempt by default and no filename is baked in; this list is the only
+  source. Every run reports the **pattern** and what it excused, because an
+  over-broad glob does not fail — `check` simply gets quieter, and quieter looks
+  like progress:
+
+  ```text
+  5 file(s) exempt via jig.config.json and not scanned:
+    src/*-card.css  (5 files — likely too broad, review it)
+    src/nope.css  (matches nothing — check the path)
+  ```
+
+  Prefer an exact path. An exemption is a claim about one file's rendering
+  context, and that is usually literally true of one file. `src/cv/pdf/**` is a
+  fact about a tree; `**/*-card.tsx` is a naming coincidence that would also
+  excuse every real card component in the project.
+- **One import per surface, through a barrel.** `jig/theme.css` imports the brand
+  file and one mode file; your stylesheet imports that single line and then never
+  changes again. Switching mode rewrites Jig's file, not yours.
+
+### Changed
+
+- **The token layer follows your project's layout.** `.jig/tokens/` was right
+  everywhere by being right nowhere — a tool dotdir holding product source,
+  reached from a Rails stylesheet by `@import "../../../.jig/tokens/…"`. The
+  default is now a `jig/` directory beside the stylesheet being wired, so the
+  import is `./jig/theme.css` in every ecosystem. `jig.config.json`'s `brand`
+  now decides placement, not just wiring — it previously honoured the path only
+  when the file already existed, which is the one case where it does not matter.
+- **The mode is chosen before `init` runs.** It is the most consequential thing
+  `init` writes and the thing it is worst at choosing: `--yes` took `'/' → product`
+  without reading the project, and that config then outranks every agent's later
+  inference. The command file now tells the agent to ask what the product is, map
+  each surface, write the config, and only then run. `init` also reports the
+  surfaces it **used** rather than the ones it defaulted to — with a config
+  present it had been announcing a default it was about to ignore.
+- A barrel holds exactly one mode, never a merge. Importing all three into one
+  document leaves only the last; verified in a browser, it yields `operator`
+  throughout. `01-modes.md` already said why that is not a loss: density switches
+  at the route boundary, never inside one view.
+
+### Fixed
+
+- `--text-prose`, `--size-touch-target` and every semantic colour are now held to
+  their floors after `init` as well as during it.
+- **`H-47`'s built-in token-layer skip was over-broad.** It matched
+  `(brand|mode).*.css` with the directory optional, so a project's own
+  `src/legacy/brand.colors.css` was silently exempt from the primitive check
+  wherever it sat. Scoped to the token layer's directory now — Jig owns `jig/`
+  outright, and a filename that merely looks like one of ours is a coincidence.
+- The token audit follows the token layer instead of reading a fixed directory.
+  It was written when `.jig/tokens/` was the only possible answer, so moving the
+  layer made it find nothing and report nothing — caught in a pre-release smoke
+  test, where a brand file edited to 22% opacity passed cleanly. A check that
+  goes quiet when its subject moves is worse than one that was never written,
+  because the silence reads as a pass.
+- Tailwind v4 does not scan workspace packages — documented, with the `@source`
+  directive it needs. Nothing errors when this bites: the class lands on the
+  element, no rule exists to match it, and the style simply does not apply.
+
+## 0.5.0
+
+Four things shipped in 0.4.0 were broken in ways that reported success. `/jig
+update` refreshed to the version already installed and said "Updated Jig →
+0.4.0". Dark mode could not be reached by choosing it. `check` skipped nearly
+every colour in a project that had not run `init`. And the reconciliation of
+every numeric default against an external reference is finished — 0 rows open —
+which is where most of the rest of this release came from.
+
+**If you are on 0.4.0, run `npx jig-ui@latest update` from a terminal.** The
+`/jig update` fix cannot deliver itself: your command file is the broken one, so
+the slash command will keep reporting a successful no-op until the CLI replaces
+it. Once, from the terminal, and the slash command works from then on.
+
+### Fixed
+
+- **`/jig update` could never upgrade anyone.** Every subcommand is invoked at
+  the installed pin so the CLI and the vendored rules agree; `update` is the one
+  exception, because its job is to move that pin. The skill file knew that and
+  the slash command did not, so it ran `npx jig-ui@<installed> update` — a no-op
+  that reports success, which is worse than an error.
+- **Dark mode was unreachable by explicit choice.** `brand.default.css` had one
+  dark block, inside `@media (prefers-color-scheme: dark)`. A selector inside a
+  media query cannot match when the query is false, so `data-theme="dark"` on a
+  light-mode system produced no dark tokens at all. A second, unmediated block
+  now carries the same declarations, and a check keeps the two identical.
+- **`check` did not resolve the project's own tokens.** The token map held only
+  Jig's vendored `.jig/tokens/*.css`, so any `var(--your-token)` was
+  unresolvable and skipped. On a project that had never run `jig init`,
+  `contrast-floor` and `violet-band-hue` evaluated very nearly nothing. A name
+  declared in two places with different values is still skipped — which value a
+  browser uses depends on import order, and a wrong guess reports a finding
+  against a value the page never renders.
+- **`operator` prose text was 16px** against the 18px floor `B-75` states and
+  `02-tokens.md` repeats. It is the mode most likely to be read for hours.
+- **Rules cited tokens that do not exist** — the `--color-danger` family,
+  `--color-surface`, `--leading-heading`, `--leading-display`,
+  `--font-weight-body`, `--spacing-unit`. An agent following those wrote a
+  `var()` resolving to nothing, with no error anywhere.
+- **`C-19` named the same token twice**, once "for large text only".
+  `--color-text-weak` clears 4.5:1 at any size, and there is no lighter grey.
+- **The error message in `P-03` moved above the input**, where autofill menus
+  and on-screen keyboards do not cover it.
+- **CSS nesting and line endings.** `install` and `update` had three near-copies
+  of one write helper and one had lost its line-ending handling, so `update`
+  flattened a CRLF token file to LF while leaving the rule files beside it
+  alone. All three now share `install/writer.ts`, and writes are atomic.
+
+### Added
+
+- **`P-13 · Ambient motion`** — a third category of motion the system lacked.
+  Interaction and transition motion are milliseconds; ambient motion is slow,
+  looping and decorative, and must never be noticed. `G-44` forbade all of it by
+  stating a 300ms ceiling it had never scoped, so an agent asked for a slow
+  decorative loop would have refused, citing a rule about something else.
+- **`--duration-ambient-fast/base/slow`** (3s / 4.7s / 7.1s), `editorial` only.
+  The values are mutually prime in tenths of a second so layered loops do not
+  re-align into one visible pulse.
+- **Fluid headings.** `--text-h1` and `--text-h2` are `clamp()` in `editorial`
+  and `product`, reaching their minimum at a 360px viewport and their maximum at
+  1024px. Every term is `rem`-based: a `px` or bare-`vw` bound ignores a
+  reader's font-size preference, which is a WCAG 1.4.4 failure.
+- **`explain` finds rules you cannot name.** A word searches every title and
+  body; one match prints in full. `--list` prints every id, or one section's.
+- **Easing direction** — `--ease-out` on entry, `--ease-in-out` on exit, never
+  linear for anything that moves. Both tokens shipped with no rule for choosing.
+
+### Changed
+
+- **`--text-h1` resolves to 32px on a phone**, not 48px. This is the change most
+  likely to be visible in an existing project, and it is the point: at a fixed
+  48px, a 45-character headline set as four lines and 211px of headline on a
+  360px screen.
+- **Line heights are documented per mode.** They always were per mode; the docs
+  quoted one mode's values as though they were everyone's, and were wrong for
+  two modes out of three.
+- **The README covers installing and using Jig, and nothing else.** How to
+  change a rule and how to test that a rule earns its context cost moved to
+  `AGENTS.md`, where an agent working on this repository will read them.
+
+### Reconciliation
+
+`RECONCILE.md` is at **0 open rows**. Every numeric default is either adopted
+from the reference or deliberately kept with its argument stated in one line.
+The motion rows are the exception worth knowing: that reference has no motion
+chapter, so those values were settled against separate sources, and where no
+source gave a number they are kept as ours on stated reasoning rather than
+adopted.
+
+`check-tokens` grew from 5 rules to 12, each mutation-tested. Two of the new
+ones exist because a guard had been passing vacuously: rule 6's regex was
+line-anchored and so covered 103 of 133 tokens while claiming to cover all of
+them, and nothing at all checked that a token cited in the rules exists.
+
+## 0.4.0
+
+Jig stops copying itself into your project. It is a skill an agent reads, and
+0.3.0 wrote 220KB across 17 files into every consuming repo to deliver it —
+roughly 200KB of that Jig's own property, read by an agent that already had it
+from the skill install. A single-mode project now gets **three** files, all of
+them its own.
+
+### Changed
+
+- **`install` writes one place: the harness's skill directory.** `SKILL.md`, the
+  rules, `rules.index.json` and attribution live beside each other at
+  `<harness>/skills/jig/`, project or global. Nothing lands in `.jig/` any more.
+
+  The rules are not a build input — nothing compiles them, and the agent reading
+  them already has them. The genuine exception is CSS: a stylesheet `@import` is
+  an edge in a build graph and must resolve inside the project, on every machine
+  that builds it. That is the one category `init` still copies.
+
+- **`.jig/` holds only what belongs to the project** — the brand file, the mode
+  files for the modes actually declared (not all three), and `state.json`.
+
+- **A project install refuses when a global one exists.** Two `jig` skills
+  registered with the same harness, whose rule paths point at different places,
+  is exactly the incoherence this layout exists to prevent.
+
+- **Every harness comes from one table.** Five adapters — Claude Code, Cursor,
+  opencode, Gemini CLI, and a generic `.agents/skills` — share the
+  `<harness>/skills/<name>/` convention, so adding one is a row rather than a
+  file. Codex keeps a bespoke implementation because `AGENTS.md` genuinely is a
+  different mechanism.
+
+### Fixed
+
+- **The skill told agents `check` was planned.** It shipped in 0.3.0. A baseline
+  run reported "the CLI is ahead of the doc" and worked by hand rather than
+  running it. The guard that should have caught this compared the metadata
+  against a hand-maintained list, so when `check` landed and neither was
+  updated, the two agreed and the test stayed green. The list is now read out of
+  the CLI source, and agreement is asserted in both directions.
+
+- **The skill sent agents to an unpinned CLI.** `npx jig-ui` resolves to whatever
+  is latest on npm, which need not be the version that wrote the bundle. Two
+  baselines hit this and fell back to working by hand. The skill now names the
+  version that built it, and `jig update` moves both together.
+
+- **`JIG_CHECK:` named two different records.** The CLI emitted `version=
+  mechanical= judgment=`; the skill told the agent to emit `version= mode=
+  self_check=` — disjoint fields under one label. There is now one field set,
+  `version mode mechanical judgment`, asserted against both sources. An emitter
+  that cannot determine a field says so in the value rather than dropping it.
+
+- **Every vendored file cited a licence path that no longer exists.** The header
+  hardcoded `.jig/LICENSE`, true only while install vendored into `.jig/`. In
+  the new layout the one line whose job is directing a reader to the licence
+  directed them nowhere. It is now computed from the file's depth in the bundle.
+
+- **`update` refreshed only one harness.** A project can hold several installs,
+  each with its own manifest; the rest stayed pinned at their install version
+  with nothing said about them. It now refreshes every one and reports each.
+
+- **`update` wrote files before checking the path.** `assertSafeRelPath` covered
+  adapter-rendered files but not `referenceDir`, which the harness table derives
+  just as directly. The shipped table is asserted at module load, so a bad entry
+  fails at import rather than at whichever command writes first.
+
+- **A project-scope Codex install wrote Jig's rules into your project's
+  `.jig/`.** Codex has no skill-directory convention, so its bundle went to a
+  bare `.jig` — the one directory this release reserves for the project's own
+  material. Install plus `init` left 13 files there, Jig's rules and manifest
+  interleaved with your tokens and state, and the harness that most needed the
+  0.4.0 separation was the only one that did not get it.
+
+  Two concrete harms beyond the untidiness: `init`'s legacy scanner looks in
+  `.jig/` for exactly those artifacts and offers to remove them, so it could
+  offer to delete a live install; and one directory holding two manifests is
+  what let a stale `.jig/manifest.json` hijack `update` in the first place.
+
+  Codex now uses `.codex/.jig/` at both scopes, mirroring the path its global
+  install already used. Verified against the real Codex CLI: it reads
+  `AGENTS.md`, resolves every rule path under the new location, and picks up the
+  declared mode. Upgrading leaves the old bundle in place — nothing is deleted —
+  and `update` now names it rather than reporting a bare "Jig is not installed"
+  to someone looking straight at the files.
+
+- **`update` could never move an install forward.** Pinning the CLI fixed version
+  skew, but applied to every command it trapped the install: `npx jig-ui@0.4.0
+  update` refreshes to 0.4.0, reports success, and changes nothing. `update` is
+  the one command whose job is moving the pin, so it is the one that does not
+  carry it — the skill renders `npx jig-ui@latest update`.
+
+- **`init --yes` chose a mode silently.** It takes `'/' → product` without
+  inferring anything, and `01-modes.md` rule 1 makes the config authoritative
+  over an agent's own inference — so the default is not a neutral placeholder,
+  it binds every agent that reads the project afterwards. Two baseline runs on
+  an `ops-console` project read every signal as `operator`, found `product`, and
+  correctly deferred to it; one noted the density difference is expensive to
+  reverse. The brand colour already stated its default and why. The mode now
+  does too, and says where to change it.
+
+- **A source build could stamp a pin to a version that was never published**,
+  silently, until an agent tried to run it. `install` and `update` now say so
+  when the running CLI is not an npm-installed package.
+
+- **Concurrent runs lost each other's manifest entries.** Two runs against one
+  install — two agents, or a script running `jig init` across a monorepo against
+  a shared global install — both read the manifest and both wrote it, and the
+  last writer's copy dropped the other's records of "Jig owns this file".
+  Losing one makes a later `update` treat that file as the user's and stop
+  refreshing it: the safe direction, but silent.
+
+  Fixed without a lock. The `files` map is additive and per-file — a run only
+  records entries for files it actually wrote — so merging against whatever is
+  on disk at write time is correct, and needs none of the stale-lock recovery a
+  mutex would after a process is killed mid-run. Writes are atomic
+  (temp + rename), and verified-and-retried to close the window between the read
+  and the rename. Ten parallel processes writing fifty entries now keep all
+  fifty; without the merge, one survives.
+
+- **An asset could be staged at prepack and left out of the tarball** — correct
+  code reading an asset that never shipped. Both lists must now agree, verified
+  against the real `npm pack` output.
+
+- **A legacy `.jig/manifest.json` hijacked `update`** and resurrected the whole
+  vendored layout.
+
+- **Agents invented token values when `init` had not been run.** A baseline run
+  with the skill installed but the project not initialised authored its own
+  `:root` block — "control height (32), row height (48), duration (120ms) and
+  the near-black brand default are my resolved values, not the system's" — and
+  wrote it into the project's stylesheet. Step 5 said "consume tokens by
+  semantic name only", and it complied to the letter by inventing the
+  definitions behind the names.
+
+  `SKILL.md` now opens with the precondition: no `jig.config.json` means no
+  token layer, so run `init` and stop, and do not author a `:root` block of your
+  own. Step 5 gained the counter — a token with no value is a finding to report,
+  not a number to supply. Re-run on the same fixture, the agent invented
+  nothing, and reported a real gap in the token contract instead.
+
+### Added
+
+- **`/jig` slash commands.** `install` writes a command file for every harness
+  that has a slash-command mechanism, so `/jig init`, `/jig check --all` and
+  `/jig update` work without leaving the session. One file named `jig`
+  dispatching on its arguments, which is what gives `/jig init` with a space
+  rather than a separate `/jig-init` per subcommand.
+
+  The command is not a thin wrapper: `/jig check` runs the CLI for the seven
+  mechanical rules, then applies the 97 judgment rules to the same files and
+  merges both halves into one report keyed by rule id — the half a CLI cannot
+  do, which is the reason the command exists at all.
+
+  Claude Code, Cursor, opencode and Gemini CLI (which gets its own TOML shape).
+  Not Codex: its custom prompts are not expanded by `codex exec` — probed
+  directly — so a file there could sit and never fire. Only subcommands the CLI
+  actually registers are offered, since a slash command that errors out is worse
+  than one that does not exist.
+
+- **`check` reads CSS wherever it lives.** It read `.css`/`.scss`/`.less` only,
+  which for most projects meant it examined almost nothing — and said nothing
+  about that. A Tailwind v4 fixture with `bg-[#6D28D9] p-[13px] rounded-[12px]
+  text-[22px] h-[32px]` reported "No findings · 0 errors · mechanical=pass:0". A
+  clean bill of health for a project where every value bypassed the token layer.
+
+  Now covered: `<style>` blocks (HTML, Astro, Vue, Svelte, PHP, ERB, Twig,
+  Handlebars, MDX, ASP and ASP.NET, Razor, JSP, Phoenix, EJS, Nunjucks, Liquid,
+  Jinja, Velocity, FreeMarker), the indentation-delimited templates (Pug's
+  `style.`, Haml's `:css`, Slim's `css:`, each with its own attribute syntax),
+  `style="…"` and `style={{ }}`, CSS-in-JS tagged templates
+  (`styled.button`, `styled(Link)`, `css`, `createGlobalStyle`, `keyframes`),
+  Tailwind arbitrary values, and Tailwind default-palette contrast pairs.
+  `@theme` counts as the token layer, so a v4 project that has adopted Jig is
+  recognised as having done so.
+
+  The mechanism is extraction rather than a detector per language: everything
+  that is not CSS is blanked, preserving character positions, and the existing
+  detectors run unchanged — so a finding in a `.vue` file still points at the
+  right line, and all seven rules gained host-language support at once.
+
+  Two deliberate limits. A bare `p-4` is not a finding: it resolves through a
+  scale, and flagging it would mean flagging correct Tailwind. A colour outside
+  the default palette is not resolved rather than guessed at.
+
+- **Reference files ship beside the skill.** `references/**` in the package
+  installs into the bundle with its subdirectory shape preserved, refreshed by
+  `update` under the same rule as the rules: replace when untouched, skip when
+  you have edited it. Adding one is a file drop, no code change.
+
+  **No reference ships in 0.4.0.** Four were planned — a procedure for `init`,
+  for `check`, for building a component, and one for resolving apparent rule
+  conflicts. Each was tested first by running an agent on the task with no
+  guidance, and in every case the agent already did the right thing: it inferred
+  the mode with signals stated, kept a 32px control inside a 48px target rather
+  than reading the two as contradictory, surfaced the brand question instead of
+  guessing, and handled loading, error and never-checked states unprompted.
+  `04-principles.md`'s tiebreakers were doing the work the references were meant
+  to do. Writing them anyway would have added words the rules already carry.
+
+### Migration
+
+`install` no longer writes rule files into `.jig/`, so a pre-0.4.0 project has
+rules in two places. The old copies are yours to remove; nothing deletes them
+for you, because you may have edited one and that edit is yours to keep. Run
+`jig install --agent <name>` to place the new bundle, then delete `.jig/rules/`
+and `.jig/rules.index.json` once you have checked them for your own changes.
+
+## 0.3.0
+
+Two new commands. `install` and `update` put the system in place; these two make
+it do something.
+
+### Added
+
+- **`jig check`** — verifies a consumer's code against the rules. Seven
+  detectors, exactly the ones `rules.index.json` already named: `gradient-text`
+  (A-02), `backdrop-blur` (A-04), `pure-black-white` (C-18), `contrast-floor`
+  (C-19), `focus-removed` (E-29), `hardcoded-value` (H-47), and
+  `violet-band-hue` (A-01, hybrid — it asks rather than fails). With `--all`,
+  `--ci` (mechanical bucket only, no model, deterministic) and `--json`.
+
+  H-47 runs only on files that reference a Jig token or import a vendored token
+  file. "Hard-coded *past* the token layer" means nothing for a file that never
+  adopted it, and without that scope it produced 13 of 13 findings on a 20-line
+  stylesheet. A project where nothing participates is told how to start.
+
+- **`jig init`** — sets a project up to use the system. Detects the CSS system,
+  derives a brand colour from existing code rather than interviewing cold,
+  validates it against the contract `brand.default.css` already states, writes
+  the brand file and `jig.config.json`, wires the imports, and runs `check` for
+  a baseline.
+
+  This is also what makes global installs coherent. The brand file and config
+  always live in the project, and for a global install the one selected mode
+  file is copied into the project's `.jig/tokens/` so the import is
+  project-relative. A `$HOME`-relative CSS import resolves only on the machine
+  that generated it; a stylesheet is committed and must build everywhere.
+
+- `oklch()` is parsed. Tailwind v4 and shadcn emit it by default, so skipping it
+  meant `init` derived nothing on a large share of new projects and `C-19`
+  computed no contrast against Jig's own `--color-bg-base`.
+
+### Fixed
+
+- **CSS nesting hid the parent's declarations.** A block containing braces was
+  discarded whole, so in `.card { color: red; .h { … } }` the `color: red`
+  belonged to no block and was invisible to every detector. On a Sass codebase
+  that was most declarations, reported as a clean result.
+
+- Comments produced findings, and a commented-out `:focus-visible` silenced
+  E-29 for a whole file — a dead detector reporting success.
+
+- `var(--x, fallback)` was resolved to the fallback and reported as fact.
+
+- The large-text contrast exemption was inert because only `px` font-sizes were
+  recognised, so `2rem` at 3.54:1 was flagged as failing a 4.5:1 floor that did
+  not apply to it.
+
+
+## 0.2.1
+
+### Fixed
+
+- **`--color-text-warning` and `--color-text-success` failed their contrast
+  floors.** Warning was 3.64:1 and success 4.48:1 against the surfaces they can
+  land on, where text requires 4.5:1 and strokes 3:1. The source is explicit:
+  system colours used for text need 4.5:1; used for interface elements and
+  icons, 3:1. `RECONCILE.md` also lists the contrast floors as the one category
+  not up for reconciliation, and `C-19` is a rule about this exact failure — so
+  the system was breaking its own hardest rule, in a shipped default that every
+  consumer inherits unless they override it.
+
+  Warning lightness 36% → 29%, success 26% → 23%. Hue and saturation unchanged,
+  so both are the same colour, darker. All four semantic colours now clear both
+  floors against `bg-base`, `bg-raised`, and `fill` on either.
+
+- The system colours now state what a replacement must satisfy. The brand colour
+  already carried that contract; these did not, so a user bringing their own
+  error or warning colour — the intended workflow — had no floor to hit.
+
+### Added
+
+- `scripts/check-tokens.mjs` gains two rules, both mutation-tested. Rule 5
+  computes contrast from the token values and fails the build; this is the only
+  defect class here that arithmetic can catch, and it shipped twice because
+  nobody was doing the arithmetic. Rule 6 fails the build when a token is
+  defined but never rendered by the preview.
+
+- `packages/preview` — a rendering harness. Every prior check verified the
+  system by arithmetic or grep; nothing had looked at it. Plain HTML and CSS,
+  no build, not published. It found two gaps on its first run: there is no
+  border-width token and no focus-ring geometry tokens.
+
+## 0.2.0
+
+The first release that actually works end to end. `0.1.0` shipped rules that
+cited tokens it never installed, and a token name that did not exist.
+
+### Fixed
+
+- **`--color-brand` was cited nine times and defined nowhere**, including
+  `03-patterns.md`'s primary-button spec. An agent following that rule wrote
+  `background: var(--color-brand)` and got an unset custom property.
+- **The design tokens were never installed.** The rules cite tokens 22 times
+  and `02-tokens.md` instructs the reader to import them, but `install` only
+  wrote the rule markdown. The CSS was in the published tarball the whole time
+  and never copied out.
+- **Rule `C-49` had no correction and `I-80` had no anti-pattern**, in a file
+  whose own contract states every rule pairs both. The totals hid it — the two
+  defects cancelled at 103/103.
+- **Ten values in the mode profile tables contradicted the tokens**, including
+  editorial body size, section rhythm, card padding, and a heading step that
+  did not exist.
+- **`A-07`'s correction had silently inverted from true to false.** A mechanical
+  rename of `--radius-base` to `--radius-sm` turned a correct statement about
+  radius derivation into an incorrect one, with nothing to catch it.
+- **`A-07`'s prohibition used a `16px+` threshold** that flagged `--radius-md`,
+  the value the rule exists to sanction.
+- **`C-68` depended on a radius token it never named.** "A badge is more rounded
+  than a button" now cites `--radius-full`.
+
+### Changed
+
+- Tokens vendor to **`.jig/tokens/`** on install, in every scope. That is the
+  only location; `update` refreshes them there and nothing relocates them.
+- Editing a vendored token file is expected: `install` and `update` both leave
+  files you have changed alone and report them skipped.
+- The mode profile tables cite token names instead of literals; the comparison
+  table is ordinal. A number in prose is a call site.
+- The semantic colours are defined once each as `--<name>-h/s/l`. Every variation
+  previously repeated the same literal four to five times, so changing a colour
+  meant editing every copy or the variations desynchronised.
+
+### Added
+
+- `scripts/check-tokens.mjs`, run by `npm test`. Four rules, each mutation-tested
+  against the drift it guards: the type table must match its tokens by name, no
+  unanchored literal in a prose table, no chosen colour literal repeated in a
+  token file, and every token import must use the canonical path.
+