luxe 0.1.4: remembered models, wheel-driven lists, and a steadier phone app

- luxe starts on the model you last chose in this project, then the last one anywhere. Signing in
  opens the model picker instead of choosing for you.
- Every list in a modal takes the mouse wheel and a draggable scrollbar.
- The phone app shows reasoning in full, stays put while you read back through a long session, and
  shows a question's answer on its card.
- On Windows, luxe update replaces the running luxe, and a signed-in install no longer sees setup
  again.
- The file tree no longer shows "engine busy" in a large project, loop output is easier to follow,
  and the name is written lowercase.

The CHANGELOG has the full list.
This commit is contained in:
Evan Whittaker
2026-10-20 13:51:34 -05:00
parent 6184aa21fb
commit a0666d6c81
591 changed files with 4926 additions and 2942 deletions

View File

@@ -4,7 +4,7 @@
Some surfaces do not survive being watched. Guided Review puts a chapter's argument beside the code
it argues about, and a reader meeting it for the first time needs to *stop* on it — played at demo
speed it registers as motion rather than as information. So the same capture that makes a demo also
makes the stills, and there is still exactly one source of truth for what Luxe looks like.
makes the stills, and there is still exactly one source of truth for what luxe looks like.
**The drawing is not done here.** This replays the recording with `pyte` to work out what is on the
screen, then hands that grid to `fixture_svg --grid`, which is the exporter the generated

View File

@@ -25,7 +25,7 @@ ROOT = pathlib.Path(__file__).resolve().parent.parent
GUIDE = ROOT / "docs" / "guide"
# Sidebar groups, in order. `web/src/content.config.ts` holds the same list.
GROUPS = ["Get started", "Using Luxe", "Customize", "Extend", "Reference"]
GROUPS = ["Get started", "Using luxe", "Customize", "Extend", "Reference"]
# Pages exempt from the voice rules while a rewrite is in progress. Empty: every page follows them.
# Links and frontmatter apply to every page regardless.

View File

@@ -1,7 +1,7 @@
# Functions allowed over clippy's too_many_lines threshold. See scripts/fn-length-check.py.
# <lines[,lines…]> <path> <fn> # reason
406 crates/luxe-tui/src/commands/built_in.rs built_in # the command catalog: one `Command` literal per entry, no logic to pull out, so a new command raises this by its own entry
369 crates/luxe-tui/src/app/messages.rs reduce_inner # the `Msg` reducer, one arm per message and nearly all of them one call; the fat arms already live in `app::messages::*`
366 crates/luxe-tui/src/app/messages.rs reduce_inner # the `Msg` reducer, one arm per message and nearly all of them one call; the fat arms already live in `app::messages::*`
351 crates/luxe-tui/src/app/commands.rs dispatch_action # maps each `UiAction` to the one call that handles it; arms are a line or two and a handler that grows moves out
237 crates/luxe-tui/src/app/events.rs reduce # the `Event` reducer, one arm per engine event; the long arms delegate to `app::events::*`
218 crates/luxe-core/src/engine/run.rs run # the engine action loop, a dispatch over `Action` where each arm is one handler call; the arms that did real work are methods

View File

@@ -15,13 +15,13 @@ that is wrong in the one direction that matters.
scripts/glyph-survey.py # report coverage for the current vocabulary
scripts/glyph-survey.py --candidates # which of a wider palette every face has
scripts/glyph-survey.py --check # portable.txt still holds (CI)
scripts/glyph-survey.py --face FILE... # does a face draw everything Luxe does, cleanly
scripts/glyph-survey.py --face FILE... # does a face draw everything luxe does, cleanly
U+2500-U+259F is not surveyed: terminals synthesize that range against the cell box, so the font is
not in the path. Everything else has to be in every face on the panel.
`--face` asks the opposite question, of a font you might *recommend* rather than one a reader
already has: does it carry every character Luxe can put on screen at the nerd tier, and does each
already has: does it carry every character luxe can put on screen at the nerd tier, and does each
one stay inside its cell? The second half is the one a coverage count misses. The non-`Mono` cut of
any Nerd Font has every icon and draws most of them twice the cell wide, and plain Iosevka has every
character while `→ … —` print over their neighbor. The character set is read out of the TUI's
@@ -45,7 +45,7 @@ import unicodedata
ROOT = pathlib.Path(__file__).resolve().parent.parent
PORTABLE = ROOT / "crates" / "luxe-tui" / "src" / "glyphs" / "portable.txt"
# Where `--face` reads what Luxe draws from.
# Where `--face` reads what luxe draws from.
DRAWING_SOURCES = (ROOT / "crates" / "luxe-tui" / "src", ROOT / "crates" / "luxe-cli" / "src")
DRAWING_ASSETS = (
ROOT / "web" / "src" / "assets" / "generated",
@@ -151,7 +151,7 @@ def parse_portable() -> list[int]:
return listed
# === What Luxe draws, for --face ===
# === What luxe draws, for --face ===
def rust_literals(src: str) -> tuple[list[tuple[int, int]], str]:
@@ -307,7 +307,7 @@ def braille(cp: int) -> bool:
def survey_faces(paths: list[str]) -> int:
"""Report each face against everything Luxe draws. Non-zero if any misses or spills anything."""
"""Report each face against everything luxe draws. Non-zero if any misses or spills anything."""
try:
from fontTools.pens.boundsPen import BoundsPen
from fontTools.ttLib import TTFont
@@ -322,7 +322,7 @@ def survey_faces(paths: list[str]) -> int:
}
classes["marks"] = [cp for cp in drawn if not any(cp in group for group in classes.values())]
print(
f"Luxe draws {len(drawn)} non-ASCII characters: "
f"luxe draws {len(drawn)} non-ASCII characters: "
+ ", ".join(f"{len(group)} {name}" for name, group in classes.items())
+ "\n",
file=sys.stderr,
@@ -375,7 +375,7 @@ def survey_faces(paths: list[str]) -> int:
if absent - len(missing["braille"]) or spilled:
failed = True
else:
print(" draws everything Luxe does, each inside its cell")
print(" draws everything luxe does, each inside its cell")
font.close()
return 1 if failed else 0
@@ -394,7 +394,7 @@ def main() -> int:
parser.add_argument("--check", action="store_true", help="verify portable.txt against the panel")
parser.add_argument("--candidates", action="store_true", help="survey the wider palette")
parser.add_argument(
"--face", nargs="+", metavar="FILE", help="measure font files against everything Luxe draws"
"--face", nargs="+", metavar="FILE", help="measure font files against everything luxe draws"
)
args = parser.parse_args()

View File

@@ -1,10 +1,10 @@
#!/usr/bin/env bash
# Materialize a clean, publishable copy of Luxe — without this repo's development
# Materialize a clean, publishable copy of luxe — without this repo's development
# artifacts (PROJECT.md, PROGRESS.md, docs/plan/, .pi/, …) or its git history.
#
# The exclusions are declared as `export-ignore` in .gitattributes, so `git archive`
# does the filtering for us. The result is a fresh tree suitable for `git init` in a
# new public repository. Luxe never requires the excluded files at runtime.
# new public repository. luxe never requires the excluded files at runtime.
#
# Usage:
# scripts/publish-clean.sh <target-dir> [ref]

View File

@@ -1,8 +1,8 @@
#!/usr/bin/env python3
"""Ask real models to fix review comments and see whether Luxe can read what they report.
"""Ask real models to fix review comments and see whether luxe can read what they report.
A fix request ends with a remediation report, and the review learns what the agent did only from
it. Whether a model writes one Luxe can read is a property of the model and of the prompt, so it is
it. Whether a model writes one luxe can read is a property of the model and of the prompt, so it is
measured against live models rather than asserted in a unit test. Each run is a scratch repository
with two bugs, a working-tree review, a comment on each, and a fix request over `luxe --rpc`; the
turn's assistant text is then read back from the session log and sorted by how it wrapped its

View File

@@ -5,7 +5,7 @@
# too): JetBrains Mono Nerd Font Mono, and Adwaita Mono for the marks it lacks. The subset differs.
# The site draws a known set of characters; the app draws whatever a model writes. So it keeps the
# ranges text is written in (Latin, Greek, Cyrillic, punctuation, arrows, maths, box drawing) plus
# every glyph Luxe draws: the TUI's glyph table at every tier (the glyph_codepoints example), and,
# every glyph luxe draws: the TUI's glyph table at every tier (the glyph_codepoints example), and,
# when the website has been built, its coverage.txt, which catches what captures show outside the
# table. Anything else falls through the CSS stack to the phone's own font, one glyph at a time.
#
@@ -69,7 +69,7 @@ for style in Regular Bold Italic; do
"${out_dir}/jetbrains-mono-${style,,}.woff2" "${unicodes}"
done
# Only Luxe's own marks come from Adwaita: filling every gap in the text ranges would add its
# Only luxe's own marks come from Adwaita: filling every gap in the text ranges would add its
# drawing of characters JetBrains Mono chose to leave out, for little a phone can't draw itself
gaps="$(comm -23 <(printf '%s' "${luxe}" | tr ',' '\n' | sort -u) \
<(cmap "${out_dir}/jetbrains-mono-regular.woff2" | sort -u) | paste -sd, -)"

View File

@@ -6,7 +6,7 @@
# fixture runner. That is the point — a screenshot taken by hand is a claim about a build nobody can
# reproduce, and it starts lying the first time a border changes.
#
# Run it after anything that changes how Luxe looks, and commit the result. CI runs it with
# Run it after anything that changes how luxe looks, and commit the result. CI runs it with
# `--check`, which regenerates into a scratch directory and fails if the committed assets differ.
#
# The *demos* are not here and cannot be. They are recordings of real sessions against a real
@@ -116,7 +116,7 @@ guide_dir="${repo_root}/docs/guide/images"
# The phone app's wordmark, from the same art. The app has no build step, so it ships the file
app_dir="${repo_root}/crates/luxe-remote/app/generated"
# Luxe's own theme, the one the site is set in. The colors are baked into each file rather than left
# luxe's own theme, the one the site is set in. The colors are baked into each file rather than left
# as custom properties, so a capture stays correct wherever it is embedded — a README, someone's
# blog. The theme preview is the one picture drawn in every theme, below.
THEMES=("luxe")

View File

@@ -1,5 +1,5 @@
#!/usr/bin/env python3
"""Drive a real Luxe session from a script, and record what the terminal saw.
"""Drive a real luxe session from a script, and record what the terminal saw.
Run *inside* a terminal, with the app under a pseudo-terminal this process owns:
@@ -12,7 +12,7 @@ desktop is to synthesize key events with `wtype`, and it works until focus moves
every remaining keystroke, `Ctrl+C` included, is typed into whatever window took it. Writing to the
pty cannot reach another window at all, and it does not need the recording to hold focus.
**The relay is transparent both ways.** Luxe asks the terminal what it can do — cursor position,
**The relay is transparent both ways.** luxe asks the terminal what it can do — cursor position,
device attributes, the glyph probe — and a driver that answers none of those hangs on the first
question. So the app's output goes to our stdout (the real terminal, which answers) and the real
terminal's replies go back into the pty. Only the outbound half is recorded; the terminal's answers
@@ -121,7 +121,7 @@ LEGACY = {
**dict(zip(("f1", "f2", "f3", "f4"), ("\x1bOP", "\x1bOQ", "\x1bOR", "\x1bOS"))),
}
# The same keys as Unicode codepoints, for the CSI-u form. Luxe negotiates the Kitty keyboard
# The same keys as Unicode codepoints, for the CSI-u form. luxe negotiates the Kitty keyboard
# protocol, so a chord the legacy encoding cannot spell — `Ctrl+Esc`, which is how a turn is
# interrupted — has a representation here rather than being unreachable.
CSI_U_CODE = {
@@ -267,7 +267,7 @@ class Session:
self._screen = pyte.Screen(cols, rows)
self._stream = pyte.Stream(self._screen)
# Incremental, and that matters. A read returns whatever bytes happened to be available, so
# a multi-byte character — every box-drawing rule and every Nerd Font glyph Luxe draws — can
# a multi-byte character — every box-drawing rule and every Nerd Font glyph luxe draws — can
# be split across two of them. Decoding each read on its own turns both halves into U+FFFD,
# which is how stray `<60>` diamonds ended up scattered through otherwise good takes. This
# holds the incomplete tail until the rest of it arrives.

View File

@@ -29,7 +29,7 @@
# on the web through the CSS stack and in a terminal through fontconfig.
#
# **Why two mono faces.** JetBrains Mono is the one people already run, so it is what the site and
# every capture are set in. It lacks six marks Luxe draws (`ℹ ↳ ↺ ⏸ ✔ ✖`), and a missing glyph on
# every capture are set in. It lacks six marks luxe draws (`ℹ ↳ ↺ ⏸ ✔ ✖`), and a missing glyph on
# the web is whatever the reader's system substitutes, at whatever width. Adwaita Mono has all six
# on the same 0.6em cell, so those codepoints, and only those, come from it; the CSS stack names it
# second and the browser falls through per glyph.
@@ -50,7 +50,7 @@ assets_dir="${repo_root}/web/src/assets/generated"
# face installed. Both directories feed the subset.
captures_dir="${repo_root}/web/src/assets/captures"
# The guide's screenshots are drawn on the site too, and so is the guide's own text, which names
# the symbols Luxe draws: set in a face that lacks them, a browser reaches for color emoji.
# the symbols luxe draws: set in a face that lacks them, a browser reaches for color emoji.
guide_images_dir="${repo_root}/docs/guide/images"
guide_dir="${repo_root}/docs/guide"
coverage_file="${out_dir}/coverage.txt"
@@ -191,7 +191,7 @@ subset() {
printf ' %-34s %6s\n' "$(basename "${dst}")" "$(stat -c%s "${dst}")"
}
# The terminal face. Regular and Bold are what Luxe's cells ask for; Italic is there because reasoning
# The terminal face. Regular and Bold are what luxe's cells ask for; Italic is there because reasoning
# and a few labels are set in it, and a synthesized slant of an upright mono reads as a rendering bug.
pyftsubset_python="$(head -1 "$(command -v pyftsubset)" | sed 's/^#!//')"
cmap() {

View File

@@ -1,5 +1,5 @@
#!/usr/bin/env bash
# Record a demo for the website by driving the real Luxe in a real terminal.
# Record a demo for the website by driving the real luxe in a real terminal.
#
# These used to be scripted through the reducer with no terminal involved, which made them exactly
# reproducible and exactly as much work as writing the app twice. This runs the shipped binary
@@ -19,8 +19,8 @@
# --cols N terminal columns (default 92 — see below)
# --rows N terminal rows (default 28)
# --font-size N Ghostty font size (default 13)
# --theme NAME Luxe theme (default luxe)
# --project DIR where to run Luxe (default: a scratch copy of web/demo)
# --theme NAME luxe theme (default luxe)
# --project DIR where to run luxe (default: a scratch copy of web/demo)
# --project-src DIR the project template to copy
# --project-name N what the scratch copy is called, which is the name every surface shows
# --max-idle SECS cap on a gap between frames after editing (default 1.2)
@@ -36,7 +36,7 @@
# --model P/ID which model answers (default: LUXE_RECORD_MODEL, else whatever is remembered)
# --thinking LEVEL reasoning level for the take
# --glyphs STYLE glyph fidelity: auto, nerd, nerd_wide, unicode, ascii
# --bin DIR put DIR first on Luxe's PATH (relative to web/demo-scripts), for stand-ins
# --bin DIR put DIR first on luxe's PATH (relative to web/demo-scripts), for stand-ins
#
# Both are how a demo gets a starting state without spending a turn on it: a changes-and-commit
# demo needs a dirty tree, and paying a provider to manufacture one so the demo can then talk about
@@ -52,7 +52,7 @@
# character on a phone, which is the same as not being there. 92 carries a two-pane review and
# still reads at 390px wide.
#
# The keystrokes go into Luxe's pty, never to the compositor. `wtype` was the obvious way to drive
# The keystrokes go into luxe's pty, never to the compositor. `wtype` was the obvious way to drive
# a TUI and it works right up until focus moves, after which every remaining key — `Ctrl+C`
# included — is typed into whatever window took it. `scripts/site-drive.py` writes to the pty
# instead, which cannot reach another window and does not need the recording to hold focus.
@@ -190,7 +190,7 @@ drive_log="${work}/drive.log"
# === The terminal ===
# Ghostty wears the Luxe palette. The app paints nearly every cell itself, but not the ones it has
# Ghostty wears the luxe palette. The app paints nearly every cell itself, but not the ones it has
# not reached yet — and a demo whose first frame is somebody else's background color announces that
# it was staged.
theme_json="${repo_root}/crates/luxe-core/src/themes/${theme}.json"
@@ -242,7 +242,7 @@ if [[ -z "$project" ]]; then
# cannot proceed, so an agent asked to commit correctly stops and asks who it is — which is good
# behavior and a poor demo, and it is the recorder's doing rather than something to talk the
# model out of in the prompt.
git -C "$project" config user.name "Luxe"
git -C "$project" config user.name "luxe"
git -C "$project" config user.email "demo@luxe.dev"
git -C "$project" add -A
git -C "$project" commit --quiet -m "Initial commit"
@@ -256,7 +256,7 @@ if [[ -n "$setup_cmd" ]]; then
(cd "$project" && DEMO_SCRIPTS="${repo_root}/web/demo-scripts" eval "$setup_cmd")
fi
# === Luxe's own config ===
# === luxe's own config ===
# Copied from the real one so the credentials and the model are the ones being demonstrated, then
# overridden where a demo needs a fixed answer. Never the real directory: a take must not be able
@@ -270,7 +270,7 @@ if [[ -d "$real_config" ]]; then
find "$real_config" -maxdepth 1 -name '*.json' ! -name 'auth.json' ! -name 'mcp.json' \
-exec cp {} "${config_home}/luxe/" \;
# Linked, not copied. An OAuth credential rotates: the provider hands back a new refresh token
# and invalidates the one just used, and Luxe writes it to whichever auth file it was given. A
# and invalidates the one just used, and luxe writes it to whichever auth file it was given. A
# copy in a scratch directory therefore takes the rotation with it when the directory is
# deleted, and every later take authenticates with a token the provider has already retired —
# which is a signed-out account after one recording, not a recording problem you notice at the
@@ -325,7 +325,7 @@ luxe_args=()
[[ -n "$thinking" ]] && luxe_args+=(--thinking "$thinking")
# A stand-in for a program the demo project has no real counterpart for, like `gh` for a repository
# that doesn't exist. First on PATH, so Luxe and every command the model runs find it.
# that doesn't exist. First on PATH, so luxe and every command the model runs find it.
record_path="$PATH"
if [[ -n "$bin_dir" ]]; then
[[ "$bin_dir" == /* ]] || bin_dir="${repo_root}/web/demo-scripts/${bin_dir}"
@@ -357,7 +357,7 @@ trap cleanup EXIT
# rather than a convenience.
#
# A take does not need to be seen: the recording comes from the pty the driver owns, and a hidden
# GTK window still answers the device queries Luxe asks. What an on-screen window *does* do is sit
# GTK window still answers the device queries luxe asks. What an on-screen window *does* do is sit
# in the way — it takes focus, it lands on whichever monitor the pointer last touched, and the
# operator cannot use their machine without wondering whether they just spoiled a minute of
# recording. `silent` keeps the workspace from being switched to.

View File

@@ -51,7 +51,7 @@ const PAGES = [
* Pages only applies `_headers` and a wrong charset turns every em dash into mojibake for a model.
*/
const TEXTS = [
{ path: "/llms.txt", type: "text/plain", expect: ["# Luxe", "/docs/install.md"] },
{ path: "/llms.txt", type: "text/plain", expect: ["# luxe", "/docs/install.md"] },
{ path: "/llms-full.txt", type: "text/plain", expect: ["# Install"] },
{ path: "/docs/install.md", type: "text/markdown", expect: ["# Install"] },
];
@@ -398,7 +398,7 @@ const CAPTURE_ADVANCE = 0.6;
* An SVG loaded through `<img src>` is a document of its own in secure static mode: it cannot see
* this page's `@font-face`, so the subset webfont never reached a single screenshot and every Nerd
* Font icon in them drew as tofu. None of that is visible to anyone with the face installed
* locally, which is anyone who works on Luxe — so the check has to be a measurement taken in the
* locally, which is anyone who works on luxe — so the check has to be a measurement taken in the
* page.
*
* `foreign` is the structural half: a capture that is not an `<svg>` holding real `<text>` in *this*

View File

@@ -4,7 +4,7 @@
# scripts/tool-bench.sh NAME=REPO:CASES ...
#
# Each argument is a checkout to measure in and the cases to run there (scripts/tool-bench/);
# Luxe's own checkout is always measured last, with the cases built into the example. The first
# luxe's own checkout is always measured last, with the cases built into the example. The first
# repository is the one the page draws its bars from.
#
# A recording, not a generated asset: the numbers depend on the trees they run in, build output
@@ -25,8 +25,8 @@ for spec in "$@"; do
(cd "$repo" && TOOL_BENCH_CASES="$root/$cases" "$bench" "$out/$name.json")
names+=("$name")
done
"$bench" "$out/Luxe.json"
names+=("Luxe")
"$bench" "$out/luxe.json"
names+=("luxe")
python3 - "$out" "${names[@]}" <<'PY' > web/src/data/tool-bench.json
import json, sys
out, names = sys.argv[1], sys.argv[2:]

View File

@@ -151,7 +151,7 @@ cleanup() {
}
trap cleanup EXIT
# The throwaway XDG_CONFIG_HOME is for Luxe alone. gh finds its hosts.yml under the same variable,
# The throwaway XDG_CONFIG_HOME is for luxe alone. gh finds its hosts.yml under the same variable,
# so without this every GitHub surface in the drive reads "not signed in" while gh is fine
export GH_CONFIG_DIR="${GH_CONFIG_DIR:-${XDG_CONFIG_HOME:-$HOME/.config}/gh}"

View File

@@ -1,7 +1,7 @@
#!/usr/bin/env bash
# Regenerate the embedded models.dev snapshot.
#
# Luxe resolves model metadata from four layers (defaults -> snapshot -> cache -> config). The
# luxe resolves model metadata from four layers (defaults -> snapshot -> cache -> config). The
# snapshot is the offline/first-run floor: it's what a fresh install with no network knows. The
# runtime refresh keeps a live install current, so this only needs to run at release cadence —
# CI does it weekly and opens a PR when the bytes change.