Fix token cost calculation discrepancies

[validatedev]

TL;DR

Corrects local token-cost estimates for large and priority-tier Codex requests, preserves whether cache discounts were actually published, and counts nested Claude advisor usage under the advisor model. It also normalizes Codex -fast aliases before provider pricing and verifies that persisted claude -p usage is included.

What was happening

• Codex requests above 272K input tokens were always priced at base rates, even when OpenAI applies higher rates to the whole request.
• Codex priority pricing fell back to one generic multiplier, which underpriced GPT-5.5 and could apply Cursor-specific fast multipliers to GPT-5.6.
• Codex model aliases ending in -fast could retain Cursor-scaled rates and bypass the Codex long-context and priority switches.
• Pricing catalogs synthesized a 10% cache-read rate without retaining whether the provider published that discount; models with no cache discount could therefore be underpriced.
• Claude advisor iterations were ignored, leaving their tokens and locally imputed costs out of the spend totals.
• Persisted Claude print-mode usage was not explicitly covered or documented, making claude -p and --no-session-persistence behavior unclear.

What this changes

• Adds a configurable long-context threshold to the shared pricing engine and applies Codex's 272K threshold with model-specific GPT-5.4, GPT-5.5, and GPT-5.6 rates.
• Applies provider-specific Codex priority multipliers: 2.5x for GPT-5.5 and 2x for supported GPT-5.4/GPT-5.6 models.
• Resolves Codex -fast aliases through the unscaled base model, then applies the Codex multiplier exactly once for both short and long-context requests.
• Carries cache-read provenance through live feeds, the supplement, compact caches, and the snapshot generator; Codex uses full input rates when no discount is published.
• Keeps explicit no-discount handling for GPT-5.4 Pro and GPT-5.5 Pro compatible with older bundled snapshots.
• Expands Claude advisor_message iterations into deduplicated entries priced under their own model while leaving ordinary iterations in the parent totals.
• Adds regression coverage for persisted claude -p (sdk-cli) records and documents that --no-session-persistence writes no log for OpenUsage to read.
• Documents the updated Codex and Claude spend behavior.

Heads-up

• Claude log-reported usage.speed remains authoritative. No date-derived fast-mode override was added because post-retirement Opus 4.6 responses already report standard; overriding that field could misprice real events.
• claude -p counts when Claude persists the session. Calls using --no-session-persistence cannot be recovered from Claude's local history because that flag intentionally prevents the usage record from being written.
• Bundled pricing snapshots were not refreshed, avoiding unrelated catalog churn. The updated generator will persist cache-rate provenance on the next scheduled snapshot refresh, while current live refreshes persist it immediately.

Tests

• swift test --filter 'CodexLogUsageScannerTests|ClaudeLogUsageScannerTests|ModelPricingTests'
• swift test — 1,144 XCTest cases passed, 3 opt-in tests skipped; 3 Swift Testing cases passed.
• ./script/build_and_run.sh verify — release build, signing, relaunch, and process verification succeeded. The existing icon compiler fallback used assets/AppIcon.prebuilt; the local build had no matching iCloud provisioning profile.

[validatedev]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bfb325c802

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[robinebers]

Pushed cfdfdd0, which closes the remaining gap in how the fast/priority tier is detected.

Why

The multiplier corrections in this PR are right (verified against OpenAI's published priority table: 2.5× for GPT-5.5, 2× for GPT-5.4/5.6), but the detection still read service_tier from the current config.toml and applied it to the entire 30-day history. That caused a real incident: priority was enabled for one task, and every past Codex day repriced ~2× — confirmed by comparing iCloud history snapshots of the same token days ($0.77/MTok vs $1.55/MTok, an exact 2.0×). The inflated history then synced to every Mac until the config was flipped back.

What changed

• The scanner now tracks the tier per session from thread_settings_applied lines (Codex CLI records these since ~Jul 2026), exactly like it already tracks the model from turn_context. Each event is stamped isFast and only turns that actually ran fast/priority price at fast rates — including mid-session tier switches.
• config.toml is no longer consulted at all. usesFastServiceTier is deleted and the global fastTier parameter is gone from aggregate.
• Sessions with no recorded tier (pre-Jul-2026 logs, older CLIs) price at standard rates. Deliberate trade-off: the worst case becomes slightly under-pricing an old fast turn, instead of retroactively inflating 30 days of history and the synced iCloud documents.
• Existing inflated history self-heals: the next refresh reprices from the logs and overwrites the device's iCloud history file.

Tests

• New parse tests: tier tracked across thread_settings_applied transitions (default → priority → default), fast spelling, and no-metadata sessions defaulting to standard.
• New end-to-end regression: a home whose config.toml says service_tier = "priority" still prices at standard rates.
• Full suite: 1,144 tests pass. Parity harness against real local logs matches the expected standard-rate figures.