[omartelo]

Forty percent of a week gone on day two and forty percent gone on day six draw
exactly the same bar today, and only one of them runs out early. This adds the
one thing the gauge was missing to tell them apart.

What lands

• internal/quota/pace.go — markAhead(Window, time.Time) Window, pure, no I/O
  and no clock of its own. It places the current cycle by rolling ResetsAt —
  the next reset, the only timestamp either API gives — back whole windows
  until it lands at or before now, which stays right however many cycles that
  reset sits ahead. Ahead is true when percent - elapsed/period*100 >= 15.
• Window.Ahead bool \json:"ahead,omitempty"`on the struct, mirrored by hand infrontend/src/lib/api-types.ts(no codegen — rootCLAUDE.md`).
• QuotaGauge.tsx draws it as a lucide trending-up at size-3 after the
  window's name, with a tooltip. It sits in the label column so the two
  fixed-width numeric columns stay untouched and three stacked gauges keep
  ending in the same place, and it carries no colour of its own — the row
  already wears usageColor. The owner picked this over the two alternatives
  from a two-theme mockup.
• docs/ceilings.md and a CHANGELOG.md [Unreleased] entry.

Only weekly, and silent for a day

The five-hour window is out on purpose: it turns over too fast for a rate to
mean anything — the first spend in a fresh window is "ahead" almost by
definition — and the package's five-minute cache is a large fraction of five
hours while it is nothing against a week. Model-scoped caps are weekly and are
paced like the account-wide one. Codex's monthly free-tier window is not weekly
and is never marked.

Nothing is marked in the first 24 hours after a reset: the elapsed share is near
zero there and almost any use at all reads as far ahead. An elapsed == 0 guard
does not cover that — a reading taken an hour in has a small elapsed, not a zero
one.

The three constants are borrowed, and the diff says so

15 points, 24 hours, weekly-only: all three are calibration measured by the
claude-swap project (src/claude_swap/pace.py, its issue #125). That
calibration is the only thing taken from there; the code is ours. Each constant
carries the why in its comment, and docs/ceilings.md records that nothing
here has been tuned against a lich user's own accounts — plus the trap it sets:
a window 90% spent on day one is deliberately unmarked.

Invariant 5 — every provider

This is pure derivation over a Window that already exists, applied once in
Service.Plans rather than per provider, so it holds for anything reporting a
weekly window without knowing this code exists. That is Claude Code (weekly_all
and weekly_scoped, both weeklyWindow) and Codex (secondary_window, which
reports limit_window_seconds: 604800 — confirmed against the fixture in
quota_test.go, so no ceiling had to be written for a shape that didn't match).
The other six providers have no plan gauge at all, for the reason already in
docs/ceilings.md: they run on the user's own API keys and there is no plan to
report. No gap to record.

Tests

internal/quota/pace_test.go pins each boundary as a literal and mutates
N±1 around it — a test that reads aheadPoints keeps passing after someone
edits aheadPoints, which is the exact change it exists to catch. Covered:
weekly ahead (64/65/66 against a flat 50% expected), weekly on pace, one second
short of the grace vs. the grace exactly elapsed, the five-hour window, the
monthly window (the case that actually pins the weekly-only guard — the grace
alone already excludes the 5h one), a model-scoped cap, a reset three cycles
further out, a reset in the past, an absent reset, an unparseable one, and a
window with no reported length. Each case also asserts markAhead changed
nothing but the mark. All three constants were mutation-checked: each one moved
turns the suite red.

Gate

gofmt -l . clean · go vet ./... clean · go test ./... green (31 packages) ·
biome check . clean (17 pre-existing a11y warnings, none new) ·
vitest run green (114 files / 1368 tests, cache cleared first) · tsc -b and
vite build clean · GOOS=linux|darwin|windows build + vet clean.

The render path is not covered by the node-environment suite; AheadMark is the
same tooltip-inside-tooltip composition LockedReading already ships in this
file.

A second commit, and it is not this feature

os-tests (windows-latest) went red on this branch in internal/relay, a
package this feature does not touch (git diff main -- internal/relay was empty
at that point). It is a real product bug, reproduced and fixed here rather than
re-run:

errandOfLocked picks "the oldest message delivered to this session" by
comparing time.Time. Windows' clock only moves every ~15.6ms, so two hand-offs
inside one tick carry the same timestamp — Before is false both ways and the
winner falls to Go's randomised map iteration. turnErrand decides which errand
a finished turn belongs to by the same comparison. So on Windows, with two
errands open, lich reply "<answer>" could send the answer home to the wrong
sender and both senders would read a confident wrong report.

Hand-offs are now counted; deliverySeq is the order itself rather than a
reading of it. Reproduced on Linux by truncating the service's clock to 16ms:
TestReplyWithoutATicketAnswersTheOldestErrandDelivered failed 22/200
before and 0/200 after. The permanent guard drops the clock question
entirely — two tickets sharing a delivered timestamp exactly, resolved 50
times over — which is red on the timestamp comparison and green on the counter.
Gate re-run whole after it, cross-compile loop included.

Happy to split this into its own PR off main if you would rather review it
apart from the feature — say the word and I will move it.

https://claude.ai/code/session_01EtriuLefbWGY9wRt4HbkQc