Context

#102784 surfaces provider plan usage (5h/weekly/model windows, credits, plan label) in the chat context popover and gates per-token cost display on whether the session's provider has subscription windows. Two structural limitations remain because usage/billing attribution is provider-keyed, not credential-keyed:

1. Cost gating is provider-level. Session rows do not record which auth profile served a run. A provider configured with both an API key and a subscription profile (e.g. anthropic api_key + synced claude-cli OAuth, or openai api_key + Codex OAuth) always resolves to subscription display, hiding Est. cost / cost-by-type for sessions that actually ran on the API key.
2. Usage snapshots are one-per-provider. loadProviderUsageSummary fetches one snapshot per usage provider id. When a native anthropic OAuth account and a different Claude Code (claude-cli) account coexist, the native account's token wins and its quota/plan is attached to both auth rows (and to CLI sessions). Same conflation exists on the /usage page, openclaw status, and openclaw models today.

Proposed direction

• Record the serving auth profile (or at least credential type: subscription vs API key) on run/session projections so cost display can be gated per run.
• Key usage snapshots by auth profile (or provider+account) instead of provider id, and attach them to the matching auth rows; aggregate views can still group by provider.

Both need a data-contract decision (session row schema + usage summary shape), which is why they were split out of #102784 (flagged by autoreview there).