feat(agents): add aggregate, project and live usage views

[LucasOl1337]

The Agents panel gains an All view, Day / Week / Month / Total filters, Antigravity and Hermes collectors, and project and live activity inspection. Users can select a project, browse its usage records, and open a record for the associated user-message preview and token breakdown.

• Keep quota meters per account while aggregating token charts across enabled providers. Merge model totals per provider so a partial daily breakdown from one provider does not hide another provider's usage of the same model.
• Add compact Projetos and Tempo real tabs with search, provider filters, 25-record pages, pause and record details. Inspection views fit within the monitor and expand to at most 900 × 600 theme units.
• Index local Codex, Claude, Grok, Hermes, OpenCode and 9Router records incrementally. Refresh every 5 seconds in the live view and every 30 seconds in projects, only while the corresponding view is open. Read previews and details on demand; store message offsets instead of conversation text. The tracking collector makes no network or model calls.
• Add Antigravity quota collection through agy -p /usage --output-format json and Hermes model usage from its local database. Label Antigravity as AGY. Keep the stock plugin ID and usage updater.

Record granularity depends on the source: individual responses where available, Grok turns, or Hermes session/model totals. These types are labeled in the UI. Project attribution uses recorded directories, with existing Git worktrees grouped by their common repository. Unknown projects remain unassigned. 9Router is excluded from the combined tracking total to avoid counting tool requests twice. Tracking currently reads default source paths and is not a system-wide request monitor or billing ledger.

Validation passed:

• bash test/shell.d/agents-panel-test.sh
• bash test/shell.d/agents-tracking-test.sh, including seven tests for token deduplication, cache accounting, worktrees, filters, pagination, incremental appends, partial records, rewrites and message lookup
• bash test/shell.d/agent-usage-antigravity-scanner-test.sh
• bash test/shell.d/agent-usage-hermes-scanner-test.sh
• bash test/shell.d/qml-text-format-test.sh
• git diff --check

The panel was exercised locally in Omarchy, with compact project/live lists and detail layouts inspected through QML renders. bin/benchmark-tracking.py --cold provides a repeatable local measurement; the included metrics show a 1.94-second initial index and a 56 ms median warm refresh on this machine. No session logs or conversation screenshots are included.

[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.