/* ==UserStyle==
@name           DenseGPT
@namespace      github.com/yunmin311/DenseGPT
@version        2.5.1
@description    Verified compact reading layout for ChatGPT, two neutral tokens, four width presets plus a free width slider. No density algorithm, no theme detection, no page background or text colour changes.
@author         yunmin311
@license        MIT
@homepageURL    https://github.com/yunmin311/DenseGPT
@supportURL     https://github.com/yunmin311/DenseGPT/issues
@updateURL      https://raw.githubusercontent.com/yunmin311/DenseGPT/main/densegpt.user.css
@preprocessor   default

@var select dg-width-preset "Width preset" {
  "reading:Reading - long prose, narrowest column": "44rem",
  "balanced:Balanced - everyday reading*": "54rem",
  "wide:Wide - code blocks and tables": "66rem",
  "ultra:Ultra - large screens, reference material": "78rem",
  "custom:Custom - set it with the slider below": "initial"
}

@var range dg-width-custom "Custom width (used when preset is Custom)" [54, 36, 90, 1, "rem"]
==/UserStyle== */

@-moz-document domain("chatgpt.com") {

/* =====================================================================
   DenseGPT

   Scope for everything except content width:
       [data-message-author-role="assistant"] .markdown

   Nothing here reads or writes page background, body text colour or
   color-scheme. Light and dark are inherited from ChatGPT as shipped.
   ===================================================================== */


/* ========================  0. neutral tokens  =======================

   Two tokens, both plain currentColor at low alpha, for the two things
   that should stay neutral: the inline code fill and the horizontal
   rule. Lightness comes from the page's own text colour, so light and
   dark adapt with no theme detection.

   The blockquote does NOT come from here. Its colours are the values
   that were accepted by eye in the 1.1 baseline, written literally in
   the rule itself. Three rounds of deriving them from a shared cite
   ink - 15/6, 35/7, 48/7 with the rule at 35 then 25 percent - each
   measured correctly and none of them looked like the accepted
   version. A colour that was approved as a literal is stored as a
   literal.

   There is no cite ink token any more. It fed nothing but the
   blockquote, and an abstraction kept only for the symmetry of having
   one is dead weight.
   ==================================================================== */

[data-message-author-role="assistant"] .markdown {
  --dg-surface-soft: color-mix(in srgb, currentColor 5%, transparent);
  --dg-edge-soft: color-mix(in srgb, currentColor 18%, transparent);
}


/* =====================  1. verified spacing  ========================
   Unchanged since 1.2.0. Do not retune.
   ==================================================================== */

[data-message-author-role="assistant"] .markdown p {
  margin: 0.45em 0;
  line-height: 1.65;
}

[data-message-author-role="assistant"] .markdown h2 {
  margin: 1.15em 0 0.45em;
  font-size: 1.15em;
}

[data-message-author-role="assistant"] .markdown h3 {
  margin: 0.9em 0 0.35em;
  font-size: 1.05em;
}

[data-message-author-role="assistant"] .markdown :is(ul, ol) {
  margin-top: 0.35em;
  margin-bottom: 0.45em;
}

[data-message-author-role="assistant"] .markdown li {
  margin: 0.12em 0;
}


/* ========================  2. inline code  ==========================
   `:not(pre) > code` keeps this off fenced blocks, which stay exactly
   as ChatGPT ships them: background, colours, font size, highlighting.
   ==================================================================== */

[data-message-author-role="assistant"] .markdown :not(pre) > code {
  background: var(--dg-surface-soft);
  padding: 0.08em 0.28em;
  border-radius: 4px;
  border: none;
  box-shadow: none;
}


/* =========================  3. blockquote  ==========================

   Reset first, then one structure. !important is used here because the
   symptom is ChatGPT's own declarations winning the cascade; the reset
   has to land regardless of what specificity they use.

   Covers every source that can draw a second line on the blockquote
   itself: any border side (physical or logical), an inset box-shadow,
   a background image, and generated content. A bar drawn by a *child*
   element is the one case this cannot reach - see docs/selectors.md.
   ==================================================================== */

[data-message-author-role="assistant"] .markdown blockquote {
  border: 0 !important;
  border-inline: 0 !important;
  border-block: 0 !important;
  box-shadow: none !important;
  background-image: none !important;
  text-indent: 0 !important;

  border-left: 3px solid rgba(100, 116, 139, 0.85) !important;
  border-radius: 6px !important;
  background-color: rgba(100, 116, 139, 0.09) !important;
  padding: 0.55em 0.85em !important;
  margin: 0.65em 0 !important;
  color: inherit !important;
}

/* generated quote marks and pseudo-element bars */
[data-message-author-role="assistant"] .markdown blockquote::before,
[data-message-author-role="assistant"] .markdown blockquote::after,
[data-message-author-role="assistant"] .markdown blockquote p::before,
[data-message-author-role="assistant"] .markdown blockquote p::after {
  content: none !important;
}

/* horizontal offset: no inherited indent on the text inside */
[data-message-author-role="assistant"] .markdown blockquote p {
  margin-inline: 0 !important;
}

/* vertical offset: the block's own padding is the only top/bottom gap */
[data-message-author-role="assistant"] .markdown blockquote > :first-child {
  margin-top: 0 !important;
}

[data-message-author-role="assistant"] .markdown blockquote > :last-child {
  margin-bottom: 0 !important;
}


/* ==========================  4. emphasis  ===========================
   Italic alone is too weak to separate at this line-height. 500 is a
   half step, not a second bold. Bold-italic must stay bold.
   ==================================================================== */

[data-message-author-role="assistant"] .markdown :is(em, i) {
  font-weight: 500;
}

[data-message-author-role="assistant"] .markdown :is(strong, b) :is(em, i) {
  font-weight: inherit;
}


/* =======================  5. horizontal rule  =======================
   A hairline at the low end of the edge range, not a chapter break.
   ==================================================================== */

[data-message-author-role="assistant"] .markdown hr {
  border: 0;
  border-top: 1px solid var(--dg-edge-soft);
  margin: 1.1em 0;
}


/* =======================  6. content width  =========================

   Two rules, both needed, neither a blanket selector.

   <main> is the inheritance root: a semantic landmark element, not a
   utility class, so it is not a Tailwind-churn hook. Everything that
   merely inherits the variable - the composer among them - is covered
   from there.

   But <main> is NOT the only declaration site. ChatGPT sets the
   variable again further down, on the turn wrapper, through a Tailwind
   arbitrary property:

     <div data-conversation-screenshot-content=""
          class="[--thread-content-max-width:40rem]
                 @w-lg/main:[--thread-content-max-width:48rem] mx-auto ...">

   A declaration on the element itself beats an inherited value, so
   that 40rem shadowed everything above it and the column measured
   560px regardless of the preset. The second rule overrides the
   variable exactly where it is redeclared, matched by the *variable's
   own name* inside the class attribute rather than by any styling
   class, so it survives Tailwind churn and it also catches the
   container-query variant.

   This is deliberately not `main *`. Only the nodes that redeclare the
   variable need overriding; everything else inherits.

   Four recommended stops plus a free slider:

     Reading   44rem   prose, narrowest column
     Balanced  54rem   default, everyday reading
     Wide      66rem   code blocks and tables
     Ultra     78rem   large screens, reference material
     Custom    36-90rem in 1rem steps, from the slider

   Wiring. @preprocessor default compiles every @var into a CSS custom
   property in a :root block prepended to this section. It does NOT
   substitute USO-style square-bracket placeholders - those belong to
   @preprocessor uso. So the variables are read here with plain var(),
   under the exact names declared above: --dg-width-preset and
   --dg-width-custom.

   A USO placeholder in this file is not an error, which is what made
   it dangerous: the CSS parser treats it as a comment, the
   declaration ends up empty, an empty custom property is NOT the same
   as an unset one, so var(name, fallback) does not fall back, min()
   receives an empty argument, and the whole width feature silently
   does nothing. That was the state of every release up to 2.0.0.

   How the two settings combine, with no conditionals available in
   plain CSS: a preset compiles to its length, and the Custom option
   compiles to the keyword `initial`. A custom property set to
   `initial` is the guaranteed-invalid value, so
   var(--dg-width-preset, var(--dg-width-custom)) falls through to the
   slider exactly when Custom is selected, and uses the preset
   otherwise. The slider is ignored by the four presets.

   The second term is the responsive retreat: a narrow window shrinks
   the column instead of pinning text to the edges, keeping a 1.5rem
   gutter.

   It measures the container, not the window. Up to 2.5.0 it was
   calc(100vw - 3rem), and 100vw is the whole window including the
   sidebar - so the clamp allowed a column wider than the space <main>
   actually has. Measured in a mock of this layout: a 1440px window
   with a 260px sidebar leaves main 1143px, while the Ultra clamp
   still permitted 1248px. That is only inert while every consumer
   uses the variable as a max-width; anything using it as a width
   overflows by that difference, and an overflowing thread is exactly
   the pressure that can squeeze a shrinkable sibling - the sidebar.

   100cqi is the inline size of the nearest query container. ChatGPT
   declares one named `main` (the @w-lg/main: variant in the turn
   wrapper's class list is proof of it), so the clamp now tracks the
   real column. Container inline size is definite by containment, so
   unlike a percentage it can never fall into cyclic sizing inside an
   auto-sized grid track or a shrink-to-fit box. With no container
   ancestor at all, cqi falls back to the small viewport - the old
   behaviour, not a broken one.

   Option labels deliberately carry no rem numbers. Stylus stores a
   select choice by its label, so a label that names its own value
   would reset every user's saved choice the moment the value is
   tuned.
   ==================================================================== */

main,
main [class*="--thread-content-max-width:"] {
  --thread-content-max-width:
    min(var(--dg-width-preset, var(--dg-width-custom)), calc(100cqi - 3rem)) !important;
}

}