| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
GNU ls follows a symlink to a directory with or without a trailing slash, so
both forms count the packages correctly. Only -d lists the link itself.
The original claim came from a count taken where the path did not resolve at
all, and it reached the spec, the plan and two commit messages. The count was
always right; the reason given for it was not, and a comment that teaches the
next reader something false is worse than no comment.
Records the real trap in its place: ls aliased to a long format counts its
"total" header as a package.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Fifteen tasks, TDD throughout: parsers and primitives first, then the
samplers, then the cards, then the edits to existing widgets.
Verified two assumptions the plan rests on rather than assuming them: Lua
5.4 provides utf8, so card.truncate can cut on a character boundary, and
conky's Cairo bindings export cairo_get_line_width, without which the ring
line-state fix has nothing to restore.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Eleven tasks. The palette's warning role first, since everything else
colours against it; then an offscreen render harness, so each later task
can look at its work without an install-restart-toggle cycle; then
parsers, samplers, primitives, the four cards, and the live pass.
Glyphs are taken from the old conky config and re-rendered rather than
guessed. That matters: a guess at the Slackware mark drew a stomach, and
the config had the real codepoint (U+F318) all along, along with the
disk and network-share icons. The icon is the label now, so no mount
name is drawn.
The two NFS shares are told apart by their server's address, read from
the mount's device field at runtime. The repo is public, so the address
belongs on screen and not in a committed file; deriving it also means
the card survives a renumbered LAN. df_parse keeps the device field for
this, and `host` is non-nil exactly for a network mount.
Two traps recorded because both would fail silently: df is aliased to
df -h in the user's shell, so the sampler calls /usr/bin/df -P -B1 by
absolute path or the parser reads '1.6G' where it expects bytes; and the
per-core counters must outlive a frame, since load is a delta and a
counter rebuilt each draw reports nil forever.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Placement is column 1 row 2, directly beneath the clock, which drops to
a single-row cell to make room.
The card's type now scales off the cell WIDTH rather than its height.
Scaling off height meant the text shrank every time the cell got
shorter, and beside the clock's numerals the card read as a whisper.
Width is the stable dimension for this layout and is what constrains the
text anyway, since the stat rows run edge to edge.
Verified against the live dashboard, not only the offscreen render: the
two disagreed about type size precisely because the offscreen harness
was cropped to the card and never showed it next to the clock.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The first draft placed every band at a fixed pixel offset and let the
arc absorb whatever height was left. In a 1x2 cell on this monitor that
is 794px of arc: a cathedral arch under a fifth of a card of content,
nothing like the mockup it was shaped after.
Three changes, all visible in an offscreen render of the real widget:
- Type sizes and band offsets scale with the rect, within clamps, so the
proportions hold on the 1920 monitor as well as this one.
- The arc gets a bounded band that follows the stats instead of being
pinned to the bottom, so no dead strip opens between them.
- The layout gives the widget a 1x1 cell. Its content is about 350px
tall and a full-height cell cannot be filled without inventing data.
Found before restarting conky by loading conky's own cairo bindings
standalone and rendering the widget to a PNG, which is a far tighter
loop than install, restart, switch workspace, screenshot.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
`set -a; . "$ENV_FILE"; set +a` on one line meant the
`# shellcheck source=/dev/null` above it attached to the `set`, not the
`.`, so SC1090 fired anyway. Split across lines, the directive binds and
the script lints clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The comment claimed a bad key returns HTTP 200 with an error body, so
the status code alone proves nothing. Measured against the live API,
that is wrong: a bad key returns 401 and an unknown city 404, and with
curl -f both fail at curl without the .cod check ever running.
The claim came from the polybar script, which ran curl WITHOUT -f and so
genuinely did receive error bodies with a success exit. The check was
ported but not the reason for it. It still earns its place behind -f,
for a 200 whose body is not usable weather, and the comment now says
that instead of something false. The spec carried the same wrong
sentence and is corrected too.
Separately, the cache landed at 644 in a directory where every other
generated file is 600, and the response carries the configured city and
its coordinates. A umask before the write puts it at 600 without a
window where it is briefly readable.
The .cod discrepancy was found by the Task 6 implementer, which noticed
the bad-key test produced a different message than the plan predicted
and said so rather than reporting the expected text as achieved.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
is_stale and age_str were the only functions in the module doing
arithmetic on an unchecked argument: a nil `now` raised, and a raise in
this project is a blank dashboard with no message. Every caller passes
os.time(), which cannot be nil, so this was unreachable, but it is the
same bug already fixed once in icon() and the module header promises
nothing raises.
age_str also printed a negative age when the observation timestamp was
in the future, which a clock skew can produce. "stale -17m" on the card
reads as a broken widget, so the age now floors at zero and such a
reading counts as fresh: weather from the future is not old.
With this the whole module is raise-free, verified across every function
against nil, string, table, boolean, NaN, negative and zero arguments in
all three positions.
Both cases were found by the Task 5 implementer, which was asked to
probe beyond its given tests and reported them without changing the code.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The numeric patterns stopped at the 'e', so a temp of 1.8e1 parsed as
1.8 and the card would have drawn 2 degrees instead of 18. Wrong in the
worst way available: silently, and still looking like weather.
OWM has not been seen to emit exponent form, and sampling several cities
near and below zero returned plain decimals throughout, so this was not
reachable in practice. Fixed regardless, because two characters in a
character class is cheaper than the reasoning required to be sure it
stays unreachable, and unlike the NaN case noted in sun_t this one fails
invisibly rather than leaving a mark.
Tests now cover both a negative temperature, which is ordinary here for
half the year, and the exponent form. Reverting the pattern fails them.
Found by adversarial probing of the committed parser, not by the plan's
own tests, which only exercise the fixture.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
NaN is a number, so the type guard admits it, and every comparison
against NaN is false, so neither clamp catches it and the function
returns NaN.
Left unguarded on purpose. No caller can produce it: parse() matches
%d+ for the sun times and `now` is os.time(). If one ever did, Cairo
ignores a NaN coordinate, so the arc would lose its dot rather than the
dashboard going blank. A ponytail: comment names the ceiling and where
the guard would go.
Found while probing the committed code, not by the tests, which do not
cover NaN.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Two comments, no behaviour change.
`force - 1` reads as an off-by-one until you know the scale starts at 0
while ipairs starts at 1, so the loop index is already force + 1. Said
plainly now, along with why the fall-through returns 12.
compass() has no caller: the card draws arrow(). It is kept because it
shares arrow()'s binning exactly, so testing both pins that shared logic
from two angles, and switching the card to "S 9 km/h" becomes a one-word
change. Without this note the next reader finds an unused function and
has to reconstruct the reasoning.
Raised by code review as the only two points worth acting on; no
Critical or Important issues were found.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
icon(800, nil, 1500, 1900) compared nil against a number and raised,
which in this project is a blank dashboard with no message. The module's
own header promises nothing raises, and is_day() already guards the same
case, so the inconsistency was the bug.
Not reachable from the widget as written: it passes os.time(), and
parse() refuses a response without sunrise and sunset, so no caller can
currently produce the failing combination. Fixed anyway, because the
contract is what future callers will rely on and the fix is one word.
The test file also now says why the codepoint VALUES carry no assertion:
comparing M.ICON.snow against a literal would only prove the escape was
typed twice. Whether E31A draws a snow cloud is what the glyph sheet is
for, and mutation testing confirms the assertions cannot see it.
Found by code review. The same fix is applied to the plan, so Tasks 2-5
inherit the guarded version.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
DAY was 1000 against a sunrise of 1500, so every timestamp in the test
evaluated as night and the first day assertion could never pass. The
implementation was right; the test constants were not.
DAY is now 1700, inside the 1500..1900 window, with a comment saying why
the value has to be checked against the window rather than just named.
Caught by the Task 1 implementer, which stopped rather than quietly
inverting the day/night comparison to make the assertions pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Nine tasks, TDD throughout: the domain tables and parser get tests, the
card gets a screenshot.
The icon codepoints in the plan were each rendered and inspected rather
than derived. That caught U+E30A, which is present in the font and draws
a sun-with-rain, standing in for snow; the correct glyph is U+E31A. A
wrong-but-present codepoint fails this way, plausibly, which is why the
plan tells the implementer to look at the rendered sheet instead of
trusting a fontconfig query.
Task 6 and Task 8 both state which outcomes are acceptable when the OWM
key is still activating, so an inactive key blocks neither.
Paths are relative to the repo root rather than absolute: the first
version of this file hardcoded a home directory in eighteen places and
the pre-commit hook rejected it, correctly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Hyprland's fullscreen rule resizes the window but not conky's drawing
surface, which stays at minimum_width/minimum_height. The dashboard
therefore painted its backdrop over roughly the top-left third of the
screen and left bare wallpaper beside it, visible as soon as the window
rules put it fullscreen on special:dash.
Set to the primary monitor, 2560x1080. The grid derives its cells from
the surface, so the layout still fills whatever size it is given; only
the surface needed to grow.
Task 4's development config keeps 1200x700 deliberately: a smaller
window is easier to screenshot and compare during that task, and it
never runs fullscreen.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Four defects, all found while running Task 8.
The documented toggle command does not work on this machine. Under the
Lua config `hyprctl dispatch` evaluates its argument as Lua, so the
plain `hyprctl dispatch togglespecialworkspace dash` fails with "')'
expected near 'dash'"; Hyprland even says so in its error. The working
form is `hyprctl dispatch 'hl.dsp.workspace.toggle_special("dash")'`,
verified toggling special:dash on and off. Also read the result from
`monitors[0].specialWorkspace.name`, since activeworkspace keeps
reporting the ordinary workspace underneath.
`fullscreen` and `no_focus` appear in no existing rule, so they were
checked against the runtime rather than left as caveats:
hl.window_rule rejects an unknown field with "unknown field '<name>'",
and neither produces that error, so both are accepted. A deliberate
bogus key confirmed the check itself works.
hl.mainMod is nil in any section other than keybindings.lua, which
declares `local mainMod = "SUPER"` and never exports it. The plan
hedged about this; it now states the literal to use.
Task 8's class check ran `conky -c ./conky.conf`, which is gitignored
generated output and absent on a fresh clone, so it now renders the
config first. Its `pkill -f` also matched the running script's own
command line and killed the shell's process group, returning 144 and
swallowing the log it had just been told to read; `pkill -x conky`
instead, fixed in Tasks 4 and 6 too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The out-of-grid check iterated `G.layout or {}`. With layout unexported
that walks an empty table, so the block passed vacuously and the test
could never detect the very condition it existed to check. Confirmed by
unexporting layout and watching the suite still pass.
Drops the fallback and asserts the export first. Verified by mutation:
unexporting layout now fails with "dashboard.lua must export layout",
and moving the clock to col=5 fails with "layout entry out of columns",
so both paths have teeth.
The plan's Step 3 expected a failure that could not happen; it now says
what the failure looks like and warns that a pass there means the test
is broken rather than the code.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The date rendered as "16 /SEP /2026", each slash jammed against the
next glyph. measure() returns the ink width, which ignores leading and
trailing spaces because a space carries no ink, and clock.lua was using
it to advance a cursor. Measured: " / " covers 6px of ink but advances
14px, so every separator lost 8px.
Splits the two uses that were conflated in one function. measure() is
for centring and right-aligning, where ink width is what you want;
advance() returns x_advance and is for laying out runs of text left to
right. The clock's date now uses advance() and reads "16 / SEP / 2026"
with even gaps, verified by screenshot.
Worth the split rather than changing measure() in place: text_right is
a correct caller of the ink width, so one function cannot serve both
and silently returning the advance would have broken alignment
instead.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
An earlier patch left "c8 = 200/255 alpha." on its own line directly
above the sentence that already says it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Two defects found while running Task 4, both confirmed against the
conky build rather than assumed.
Task 4's development config defined color1..color7 and default_color
but not default_shade_color, which palette() reads for the card fill,
so following the plan verbatim drew every card in the magenta sentinel.
The config now carries it, with a note saying why it is required.
own_window_argb_visual has been REMOVED in conky 1.24 (ARGB is always
on when available) and own_window_argb_value is deprecated; both warn
on this build. Replaced with own_window_colour '#AARRGGBB', which emits
no warnings. In the template that value follows the palette through a
new @BODY_SHADE_RAW@ placeholder, the shade colour without its leading
'#', so a scheme switch recolours the window backdrop too rather than
pinning it to one hardcoded hex.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Two defects, both found by review of the Cairo work and both confirmed
by experiment rather than reasoning.
The palette mechanism could not have worked. conky_parse('${color3}')
returns an empty string: conky's colour variables emit renderer escape
codes into conky.text, they never evaluate to a hex string, and
${default_shade_color} is not a variable at all, so conky_parse hands
the literal text back. conky.config is not exposed to Lua either
(_G.conky is nil) and conky_info carries only cpu_count and
update_interval. What Lua does get is conky_config, the config file's
path, so the colours are now parsed out of that file, where UDT has
already substituted real hex. The pattern needs [%w_]+ rather than %w+
or it misses default_color and default_shade_color.
The card fill also duplicated the border colour, both being color3, so
filling at 0.55 alpha and stroking the hairline at 0.9 in one hue made
the border invisible and the cards read as blobs rather than the
mockup's thin outlines. The fill is default_shade_color now, a role
every scheme already defines.
card.measure allocated a cairo_text_extents_t per call, which leaks:
5000 allocations grow Lua's heap by 182KB that GC never reclaims, and
:destroy() on each does not help, measured at the same 182KB. One
reused module-level struct costs 0KB. A draw hook running every 2s
would have bled memory for as long as conky stayed up, which no
screenshot would ever reveal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The test comment said "cpu_percent", but the API is new_cpu_counter()
plus :sample(). Prose only, the code was always consistent, but a
comment naming a nonexistent function sends the next reader looking
for it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The live config is hyprland.lua requiring sections/, with rules
declared through an hl helper API, so the planned .conf fragment with
source = could not have worked. Task 8 now writes a Lua section using
hl.window_rule and hl.bind, matching the existing rules.
Three facts checked against the running config:
- keybindings.lua already binds SUPER+D to the unnamed special
workspace, so the dashboard takes SUPER+S.
- autostart.lua already runs a bare conky. The dashboard replaces it,
so that line changes rather than a new exec being added, which would
run two instances.
- The unnamed special workspace hosts the btop scratchpad, so the
dashboard gets its own special:dash.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Ten tasks, TDD where there is logic to test. Ordered so nothing
half-built touches the working desktop: pure Lua and its tests first,
then pixels, then the palette and the widget, then window placement,
and only last does UDT switch over to this repo's template.
Leads with the platform facts verified during design, because each one
invalidates the approach an implementer would otherwise reach for:
conky_surface() instead of the Xlib idiom on Wayland, own_window_type
normal rather than desktop, a Lua fault presenting as a blank screen,
grim capturing screen coordinates so it needs the dashboard's workspace
active, and Cairo having no Black font weight.
The pcall error overlay is built before any widget and verified by
injecting a deliberate fault, since it is the only channel that makes a
later mistake visible.
Fixture arithmetic in the parser tests is checked rather than asserted
from memory: a wrong expected value sends the implementer chasing a bug
that is not there.
Paths are ~-relative throughout, so the plan does not pin to one home
directory.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
