Provider-reported subscription quota telemetry for z.AI and GPT / Qtile #83

New issue

Open

opened 2026-09-14 14:15:57 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-14 14:15:57 +00:00

Owner

Requested behavior

Expose one llm-log quota interface for two Qtile widgets labelled exactly z.AI and GPT. Each cycles through reported five-hour, weekly, and other applicable quota windows. Display percentage USED, green then yellow then red as consumption increases. Support any provider-reported plan, including GPT Pro, without hard-coded entitlement tables.

Contract

Keep subscription quota separate from captured token statistics and local scheduler admission budgets.
Use the authenticated provider's reported usage, limits, window durations, reset times, plan metadata and limit IDs. Missing, stale, malformed, unavailable, and explicitly unlimited states must remain distinct; never fabricate 0% or infinity.
GPT adapter should use the supported Codex App Server account/read and account/rateLimits/read interface, including rateLimitsByLimitId. Retain meter scope; Codex quota must not be labelled as every ChatGPT web/voice quota.
z.AI adapter must use a source-verified usage endpoint/schema with credentials supplied locally. No browser credential scraping, entitlement guessing, token-to-credit guesses, or secrets in Qtile/cache/API output.
Read-only localhost API and sanitized atomic snapshot for desktop consumers; bounded fetches, cached refresh, no network calls or subprocess execution in the Qtile render loop.
Green below 70%, yellow 70–89%, red from 90%, configurable; stale/unknown displayed explicitly in neutral color. Rotation every five seconds with manual cycle/pause controls and reset details.
Keep packaged Common Lisp forwarding and Python capture/analytics runtime boundaries explicit. Do not imply a merge updates the live host.

Validation

Regression tests for plan changes, 5h/week mapping, multiple GPT buckets, missing/invalid usage, reset expiry, provider errors, stale cache, secret redaction, percent bounds, widget color thresholds/rotation, and Org/generated parity. Use existing transport blocker #74 and rate work #75 separately; telemetry must not claim those gates are green.

## Requested behavior Expose one llm-log quota interface for two Qtile widgets labelled exactly `z.AI` and `GPT`. Each cycles through reported five-hour, weekly, and other applicable quota windows. Display percentage USED, green then yellow then red as consumption increases. Support any provider-reported plan, including GPT Pro, without hard-coded entitlement tables. ## Contract - Keep subscription quota separate from captured token statistics and local scheduler admission budgets. - Use the authenticated provider's reported usage, limits, window durations, reset times, plan metadata and limit IDs. Missing, stale, malformed, unavailable, and explicitly unlimited states must remain distinct; never fabricate 0% or infinity. - GPT adapter should use the supported Codex App Server account/read and account/rateLimits/read interface, including rateLimitsByLimitId. Retain meter scope; Codex quota must not be labelled as every ChatGPT web/voice quota. - z.AI adapter must use a source-verified usage endpoint/schema with credentials supplied locally. No browser credential scraping, entitlement guessing, token-to-credit guesses, or secrets in Qtile/cache/API output. - Read-only localhost API and sanitized atomic snapshot for desktop consumers; bounded fetches, cached refresh, no network calls or subprocess execution in the Qtile render loop. - Green below 70%, yellow 70–89%, red from 90%, configurable; stale/unknown displayed explicitly in neutral color. Rotation every five seconds with manual cycle/pause controls and reset details. - Keep packaged Common Lisp forwarding and Python capture/analytics runtime boundaries explicit. Do not imply a merge updates the live host. ## Validation Regression tests for plan changes, 5h/week mapping, multiple GPT buckets, missing/invalid usage, reset expiry, provider errors, stale cache, secret redaction, percent bounds, widget color thresholds/rotation, and Org/generated parity. Use existing transport blocker #74 and rate work #75 separately; telemetry must not claim those gates are green.