The indicator read alwaysThinkingEnabled from settings.json, which Option+T
never writes - it toggles thinking for the session only. The key was also
absent here, so `// false` pinned the segment to the hollow "off" diamond
while thinking was actually on: an inverted indicator, not just a stale one.
Claude Code pipes the live state in as .thinking.enabled, alongside
.effort.level. Read both, in one jq pass since the line re-renders on every
redraw. Absent thinking means enabled, matching the renderer's own
`thinking:{enabled: lt !== false}`.
The effort level now replaces the static "thinking" label, so the segment
says something that changes: "◆ high" rather than a word that was there
either way. Unrecognised levels fall back to the old label, since the wrap
branch sizes line one from this string.
Test widths are usable widths, not raw terminal widths: the tiers are
computed after padding is subtracted, so a raw width would change tier
whenever statusLine.padding did. The usage cache is seeded for the same
reason width_test.sh seeds it - an unseeded cache makes the first render
fetch live usage over the network.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A failed fetch left the cache mtime untouched, so with no network every
redraw — and a redraw happens on every keystroke — tried curl again and
could wait --max-time 10 for it. Cutting the TTL from an hour to five
minutes made that start 55 minutes sooner, so it is worth fixing now.
The marker is a separate file rather than a touch of the cached
response: a cold /tmp has no response to touch, and that is exactly the
case with no stale data to fall back on, where the timeout is paid in
full with nothing to show for it.
It goes down before the request and is cleared when that lands, not
written afterwards. A fetch is in flight for as long as curl takes to
give up, and this cache dir is shared by every session, so redraws
starting inside that window are real rather than hypothetical: they now
serve stale data instead of each launching their own doomed request.
Every attempt writes it, so a repeated failure restarts the backoff, and
only an attempt does, so the redraws it suppresses cannot keep
re-stamping it and it always lapses.
Missing credentials stays outside the backoff — it costs no timeout, and
the redraw after a login should show the bars rather than wait out a
retry window it had no part in earning.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The 5h/7d bars were served from a cache with a one-hour TTL, so a bar
could sit unchanged for most of a 5-hour window. The TTL is now a named
config setting next to GIT_CACHE_SECS at 300s: worst case 12 requests an
hour, shared across every session through /tmp/claude.
The section header claimed "cached 60s" — inherited from upstream and
never true here — which is what set the expectation the code did not
meet. It now names the setting instead of a number that can drift.
The refresh is gated on the same width-tier check as the bars, so a
render that shows them is a render that refreshes them: the new tests
pin that from full width down to WRAP_FLOOR, and pin that one column
below it, where the bars are dropped, no request is paid for either.
Those widths are given in usable columns, as in the sibling width test,
so the suite does not depend on the padding in the live settings.json.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
In wrap mode the branch is truncated to whatever the rest of line one leaves,
and "the rest" was a flat 56 columns. It is not flat: the token counts and the
cost figure vary with the session, and with a four-digit cost the real width is
57, so line one came out exactly one column over and the renderer clipped it.
Reachable at COLUMNS 73-81 with padding 2 — a narrow split pane.
Add up the segments that follow the branch instead, from the same values that
render them, and count the ahead/behind markers too, which the constant also
ignored. That needs the cost formatted and the token bar width chosen before
the git segment is built, so both move up; neither depends on anything in
between.
The budget is now exact rather than approximate, so the branch also gets the
columns the old constant was over-reserving when the cost was short.
Tests: the large-cost sweep only used the short branch 'main', where the
overflow cannot show. Repeat it with the long branch, which fails on the old
budget at usable 68, 69, 70 and 75 by exactly one column.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The measured ladder tried every single-line variant before considering a
second line, so whenever line one plus the full group did not fit — from
around 92 usable columns upwards, depending on how long the branch, cwd and
model names are — it emitted rl_bare (5h and 7d only) and silently dropped
the per-model (Fable) bar that is the main reason the group is worth
rendering. On a 110-column laptop the limit was invisible.
Reorder so a second line beats losing that bar: one line rich/mid/lean, then
wrap, and rl_bare only when WRAP_NARROW is false. Reset times and
extra-usage credits are still given up rather than wrapped for, which makes
the rendered content non-monotone in width; the comment on the ladder spells
that out.
Also stop the tests writing fixtures to the caches the live status line
reads. ~/.claude/statusline.sh is a symlink to the script, so a test run was
visibly rendering fabricated usage — a Fable bar at 10%, extra-usage credits
of $1234.56/$2000.00 — in whatever session happened to be open, and a run
killed before its trap fired would have left that in place for an hour. The
cache directory is now $STATUSLINE_CACHE_DIR (default /tmp/claude) and each
suite points it at a temporary directory, which also removes the
backup/restore dance and the deletion of the live git-status cache.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
/usage reports a weekly limit scoped to a single model (Fable) that the
status line did not surface. It is absent from both the status line's stdin
JSON and the legacy seven_day_opus/seven_day_sonnet fields, which are always
null now; it lives in the usage API's .limits[] under kind "weekly_scoped".
Render it labelled by scope.model.display_name so it follows whichever model
the limit applies to, inside the 7d segment since both are weekly limits
sharing one reset time.
Fitting it exposed a problem with sizing by width tier. Tiers only know the
terminal width, so content that varies with session state — branch name, cwd,
model display name — could push a line past the edge and be clipped by the
renderer. Measure the assembled line instead (vis_len) and emit the richest of
four rate-limit variants that fits, on one line or two: with reset times, with
extra-usage credits, bars only, or bars without the per-model segment. Every
branch is fit-checked, including with wrapping disabled. Two lines render
correctly in the status line.
Cap the cwd basename as well: nothing else shortened line one, so a long
project directory could overflow it on its own.
Correct the usable width. Claude Code exports COLUMNS but applies the `padding`
setting on top of it rather than deducting it first, so a line sized to COLUMNS
is clipped; deduct padding on both sides plus a column of margin.
Sanitise payload data before rendering. printf %b interprets backslash escapes,
which is how the colour variables work, so a cwd or model name containing a
literal \n — or a real newline, which jq decodes from the JSON — would split
the line and break both the width measurement and the two-line guarantee.
Parse the usage payload in one jq pass rather than ten. The status line runs on
every redraw and per-field parsing had become the dominant cost; this brings a
render back to roughly what it was before (~155ms vs ~115ms parent), the
remainder being the reset times the tiered version did not show at this width.
Fields are read one per line rather than via @tsv: tab is IFS whitespace, so
`read` collapses runs of it and an empty field — no scoped limit, which is the
common case — silently shifted every later field along by one.
Tests cover the API payload shapes including malformed and hostile input, the
width invariants (never exceed the usable width, never more than two lines,
never wrap unnecessarily), and a sweep over widths, branch lengths, model
names, cwd lengths and extra-usage. They restore the live usage cache on exit,
and remove the fixture outright when there was no cache to restore — otherwise
a test run would leave fabricated usage figures live for an hour.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
