[LucasOl1337]

Summary

Stock Omarchy's Agents panel charts Claude, Codex, and Fireworks over the last seven days. This PR is the rest of that panel: every harness that actually runs on an Omarchy machine, filters that match the calendar instead of a quota window, project and live inspection, and totals that do not lie.

It is the same work that ships today as lol.agents v1.4.1. The plugin id stays omarchy.agents. Refresh still goes through omarchy-agent-usage-update. Extra collectors are new bin/omarchy-agent-usage-* commands, so the stock updater picks them up.

What stock does today

• One tab per record, no All
• Last seven days only
• Collectors: Claude, Codex, Fireworks
• No project or live view

What this PR adds

Views

• All — sums every enabled harness. Rate limits stay per-account and are never merged.
• Day / Week / Month / Total — week is the default. Total is all-time (or the collector's window), not a quota reset.
• Projetos / Tempo real — local sessions, compact rows, on-demand previews, no network. 25 records per page. 9Router stays in the filter but is excluded from the combined total so harness requests are not counted twice.

Collectors (stock updater discovers them)

Collector	Limits	Local stats
Antigravity	agy session + weekly pools	~/.gemini/antigravity-cli
Hermes	none	~/.hermes/state.db and profiles
Grok	SuperGrok weekly pool	~/.grok/sessions
Cursor	Cursor Models / Other Models from the same display sentences Settings shows	billing-cycle tokens by model
OpenCode	none	completed assistant turns in opencode.db, dated by the message clock
Devin	none	local Devin CLI sessions.db

Accuracy (the part that bit us on a real desktop)

• Stale today* is not today. A leftover OpenCode file from 6 Sep kept todayTotalTokens: 585147 with updatedAt: null. Day painted that as calendar today. The panel now trusts today* only when updatedAt or the history itself is from today.
• Day / Week / Month never fall back to all-time modelUsage. That fallback is how a billion Grok Bot tokens from Cursor's billing-cycle map showed up under Day. Addresses the review on the earlier revision: bounded periods use per-day history only; Total still shows the cycle / all-time breakdown. Regression: tests/test_period_model_map.py.
• Cursor meters match Settings. Plan & Usage on Ultra showed Cursor Models 1% / Other Models 0% / no on-demand pot. Reading autoPercentUsed and Ultra's includedAmountCents produced a fake 45% and a prepaid $400 wallet. Meters now parse the display sentences; a balance appears only when spendLimitUsage has a real remaining + limit.

Test plan

• bash test/shell.d/agents-panel-test.sh
• bash test/shell.d/agents-usage-accuracy-test.sh
• bash test/shell.d/agents-usage-collectors-test.sh
• bash test/shell.d/agents-tracking-test.sh
• bash test/shell.d/agent-usage-update-test.sh
• bash test/shell.d/agent-usage-hermes-scanner-test.sh
• git diff --check

On a live Omarchy desktop the same collectors already run as lol.agents 1.4.1: OpenCode today is 0 when the last turn was last week; Cursor chips show 1% / 0% next to Settings.

Notes

• Tracking is not a system-wide request monitor or a billing ledger. It reads default local paths only.
• 9Router is listed separately so the same request is not counted once by the harness and again by the router.
• Rebased onto current quattro.

[Brams-s]

Sharing related work for coordination around the shared Agents panel: #9204 agrees a focused Codex follow-up after #8862, emitting optional UTC recentModelUsageByHour buckets (hourStart / tokensByModel) from the existing native, pi/omp, and OpenCode scan paths.

Today's local prototype audit found a concrete coverage failure: the daily total included native Codex while the hourly view only included OpenCode, leaving a false zero for an hour with about 15.0M native tokens. The useful invariant for both efforts is that totals for the same source scope, window, and token definition reconcile, or the UI makes incomplete attribution explicit.

I have submitted only the independent native repeated-notification correction so far (#10531). The hourly UI remains a separate follow-up. Linking the discussions here so the optional field/coverage semantics can stay compatible as your aggregate and live views evolve.

[LucasOl1337]

Rebased onto current quattro (was 641 commits behind). Agent panel/tracking/Antigravity/Hermes tests still pass.

[Brams-s]

Thanks for the rebase update. I checked the period aggregation at d9902b0 and found a remaining time-window mismatch relevant to the coverage point above: periodModelMap() falls back to all-time modelUsage when the selected window lacks enough per-day model attribution. The All view then includes those all-time tokens while its footer labels them “this week”.

A small synthetic reproduction using the current helper functions gives 100 tokens in the weekly daily totals but 1,000 in All/week's model total when that provider has 1,000 all-time tokens and no per-day model breakdown. Could we keep the fallback explicitly labeled as all-time/incomplete, or show the selected-window model breakdown as unavailable rather than presenting it as a weekly total? A behavioral regression test for this case would also help keep the later optional hourly buckets consistent with these views.

[Brams-s]

Thanks for removing the all-time fallback. I checked 6ad89bd and found two remaining cases in the All/week footer:

• A provider with 100 weekly tokens but no per-day model breakdown shows “0 tokens this week · every harness”.
• Thirteen models with 10 tokens each show 120 rather than 130, because the footer sums the capped model rows.

Could the footer use the uncapped selected-period totals and label missing model attribution as incomplete? I reproduced both with the production Panel.qml helpers and synthetic records. The five Python tests pass, but they exercise a separate implementation; a regression against the QML helpers would catch these cases.

[dhh]

Thanks for all the work here. Thanks for this! The agents panel was redesigned in #13770 into a single page with every agent and account at once, so the All tab and period filters don't apply anymore. Grok landed in #13992, and Cursor, OpenCode, Copilot, Hermes, and Antigravity collectors are being brought in from their dedicated PRs. Closing.