Improve weekly pace projection using learned time-of-day profiles

[EthanLipnik]

Why

The old pace logic was purely linear, so it assumed people use Codex uniformly 24/7. In
practice that over-reported deficit/runout (especially overnight) and frequently showed
misleading “~2 days left” style projections.

What Changed

• Added a new weekly profile model (UsagePaceProfile) with 168 hourly bins and confidence
  thresholds.
• Extended UsagePace with:
  • model (linear vs timeOfDayProfile)
  • confidence (high / low)
  • isFallbackLinear
• Updated weekly pace calculation to:
  • use learned time-of-day profile when data is sufficient
  • fall back to linear when data is sparse, marked as low confidence
• Added persistent per-provider learning/store (UsagePaceProfileStore):
  • learns hourly usage rates from weekly snapshot deltas
  • saves to ~/Library/Application Support/com.steipete.codexbar/pace-profiles-v1.json
• Wired profile learning into refresh flow and profile-based pace into:
  • menu bar text
  • menu card details
  • menu descriptor summaries
  • provider preferences views
• Removed codex-only restriction for weekly pace display when a valid weekly window exists.
• Added/updated tests for:
  • profile-based pacing
  • low-confidence fallback behavior
  • provider-agnostic weekly pace display
  • updated status item/menu-bar expectations

User Impact

• Pace/deficit is now personalized to actual usage patterns instead of assumed constant usage.
• Overnight inactivity no longer skews projections as heavily.
• Early in learning, UI clearly signals fallback with low confidence.

Validation

• pnpm check
• swift test -q (passes: 756 tests / 122 suites)
• ./Scripts/compile_and_run.sh (build/package/relaunch verified)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 37ac40c964

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Require explicit weekly duration before showing pace

isWeeklyWindow currently treats window.windowMinutes == nil as a 7-day window, which makes non-weekly quotas eligible for weekly pace output. For example, Cursor emits windowMinutes: nil with a monthly billingCycleEnd (Sources/CodexBarCore/Providers/Cursor/CursorStatusProbe.swift:259-263), so this path will start showing a “weekly” pace during the last 7 days of a monthly cycle, producing misleading deficit/runway values.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Exclude unknown-duration windows from profile learning

The profile store uses the same weekly check that defaults nil minutes to 7 days, so it will learn “weekly” time-of-day rates from windows whose period is actually unknown or non-weekly. With providers like Cursor (monthly reset, no windowMinutes), this contaminates pace-profiles-v1.json with non-weekly behavior and then feeds those profiles back into pace projections when a reset enters the final week.

Useful? React with 👍 / 👎.