The interval floor was applied with max(), so the zero that --force
picked could never win: max(0, 15) is 15. Forcing a refresh within
fifteen seconds of the last probe silently served the cache instead,
though --force documents itself as ignoring them.
The window exists to absorb a panel opened and shut repeatedly, which
arrives as --limits-only. --force is a person pressing refresh, and it
should outrank a window meant for flicks.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Only the Claude Code CLI can refresh the OAuth token it saves; the
collector just reads it. A machine left alone long enough finds the
token lapsed, and that branch returned an empty limits list with no
status text at all, so the panel hid its whole limits section and
explained nothing.
Say what is wrong, and fall back to the cached limits already on disk
rather than discarding them. Cached windows are kept only until they
reset: a percentage from a window that has rolled over describes a
period that is over, and pinning a stale 78% on an allowance that is
now untouched would be worse than showing nothing. The probe-failure
path gets the same filtering for the same reason.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
A scoped window read as "Fable weekly" beside "Session" and "Weekly", so
the one row that names a model was also the one row in lowercase.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Show the per-model weekly limit Claude's usage endpoint reports
Model-scoped allowances arrive in the payload's limits array, not in the
seven_day_<model> buckets, which come back null. Read them so a window
like Fable's own weekly limit stops being spent against invisibly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* Read every model-scoped window, and title it for what it is
A model can hold more than one scoped window, so keying the dedupe on the
model alone dropped whichever came second — including a fuller one that
decides the headline. The model and the window kind together make the key,
and both make the title, so one model's two rows read apart.
The panel guesses a window out of the label, and that guess cannot survive a
model name: "Opus 5 (1M context)" parses as a one-minute window and renders
as a second "Session". The collector states the title outright now and the
panel takes it, eliding a long one rather than running it into the percentage.
Scoped percentages are read on whatever scale the payload speaks, the way the
flat buckets already are, rather than assuming percentages, and a model that
names only an id still names a window worth showing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: David Heinemeier Hansson <david@hey.com>
Pi/omp sessions kept falling into the Claude record when the api
field merely started with 'anthropic'. Kimi and other providers that
speak the anthropic-messages protocol (kimi-coding) were therefore
charged against Claude Code, showing k3 buckets under a Claude tab
even for a user without a Claude login.
Match the codex collector's provider-only attribution: only sessions
whose provider is exactly 'anthropic' count toward Claude usage.
Add a test proving a kimi-coding session sharing the
anthropic-messages api does not land in the Claude record.
Co-authored-by: Luca <luca@itwasarch>
Two Claude collectors running at once both wrote the cache through a temp
path derived from the target, so the second replace found the file already
moved away and crashed the update with a FileNotFoundError.
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Add a Fireworks balance collector and teach the agents panel prepaid ledgers
The omarchy-agent-usage-fireworks collector reads serverless token usage
from the Fireworks billing API, grouped by day and model for the last 30
days, and reshapes it into the shared record contract. Fireworks does not
expose its prepaid ledger through the documented API, so the record carries
an estimated balance instead of rate limits: credits configured in
~/.config/omarchy/agents/fireworks.json minus rated account costs since the
funding date. Credentials come from FIREWORKS_API_KEY/FIREWORKS_ACCOUNT_ID,
the auth.ini that firectl set-api-key writes, or — last, so an explicit
login wins — the key opencode stores for its fireworks-ai provider.
The panel gains two generic capabilities any agent record can use: a
balance object draws a BALANCE section — remaining credit, a fuel-gauge
meter that drains toward empty and lights the bar alarm below 10%, and
funded-versus-spent detail — and hasPromptStats: false keeps prompt and
session counts out of today's tooltip for agents whose billing API only
ever reports tokens, on this machine and through synced snapshots.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Feed Claude and Codex usage from pi, omp, and opencode sessions
A subscription burned entirely through another coding agent leaves no
native Claude Code transcripts and no Codex session files, so the panel
showed nothing for it. pi and omp write compatible JSONL sessions, and
opencode records per-message provider, model, and token usage in its
message database; the claude and codex collectors now scan all three —
filtered to Anthropic and OpenAI providers respectively — and merge those
numbers into their local stats. Fireworks stays out on purpose: its billing
API already sees that traffic server-side, and a local scan would count the
same tokens twice.
The collector tests pin XDG_DATA_HOME so a developer's real opencode
history cannot leak into fixture runs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* Add agent usage collectors that write display-ready data files
One omarchy-agent-usage-scan-<agent> collector per AI coding agent prints a
complete display-ready usage record — identity, tier, status, rate limits,
and today/week/all-time stats. omarchy-agent-usage-update runs every
collector it finds and writes the records atomically to
~/.local/state/omarchy/agents/usage/, so anything that displays usage only
ever reads JSON from there.
The Claude collector absorbs what the shell previously did in-process:
transcript scanning, the stats-cache/history fallback, credentials parsing,
and the OAuth limits probe, now with a probe throttle and last-good limits
kept across network failures. The Codex collector is the existing scanner
reshaped to the shared record contract.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Redo the model-usage plugin as omarchy.agents watching usage data files
The panel is now strictly a display. It discovers the JSON records that
omarchy-agent-usage-update maintains under
~/.local/state/omarchy/agents/usage/, watches them for changes, and draws
whatever appears — so adding an agent means shipping a collector, never
touching the panel. Marks resolve by convention (assets/<id>.svg with an
optional -light twin), the limits meters read a generic limits array, and
the per-provider QML adapters and in-plugin scanner scripts are gone.
Cross-device sync aggregation stays in the shell and keeps the snapshot
field names older versions wrote, so mixed-version fleets still merge in
both directions.
With the provider fan-out gone, the widget takes its real name: the plugin
id becomes omarchy.agents. A migration renames it wherever a user's config
mentions it — layout entries keep their settings and position, a disabled
widget stays disabled — then primes the data files once and drops the old
scanner cache. The migration test also drops a stale assertion that expected
migrations to restart the shell themselves, which c992cdff moved to
omarchy update.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Address Codex review: synced-only tabs, limits retry, history fallback
Three data-availability gaps from review. An agent whose records only exist
in synced snapshots — a collector installed on just one machine — now gets
its tab by unioning the synced aggregate into the provider list, with rate
limits blank since those never travel. A Claude limits probe that reaches no
server at all writes retryAdvised into its record, and the shell honors it
with one 30-second retry instead of waiting out the full refresh interval,
restoring the old boot-before-DHCP behavior. And a machine with only
history.jsonl — no transcripts, no stats-cache — still reports today's
prompt and session counts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Address second Codex pass: history-only visibility, targeted retries
Today's prompt and session counts now count toward an agent's presence in
the bar, so a machine whose only Claude source is history.jsonl shows up
without waiting for limits. And the 30-second limits retry passes the
advising agent ids to the updater, so an outage at one provider no longer
puts every other collector on a retry treadmill.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Drop omarchy-cmd-present jq guards from the agents migrations
jq ships in the default package set, which makes it a runtime invariant per
AGENTS.md — call it directly. The migration tests lose their now-unused
omarchy-cmd-present stubs with it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Drop the scan infix from the collector command names
Collectors are omarchy-agent-usage-<agent>; the updater skips its own name
when globbing them, and the update test proves it with a decoy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Keep the credential store out of the printed usage record
The Claude collector now reads .credentials.json once into three scalars —
the access token, its expiry, and the plan label — instead of passing the
parsed store around. The token reaches nothing but the Authorization header
of the limits probe, and only the plan label may travel into the record,
which is what CodeQL's clear-text-logging alert on the record print was
unable to see when the whole dict flowed through.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
