Integrate models.dev pricing flow

[iam-brain]

Summary

• Prefer cached models.dev pricing for Codex and Claude cost calculations before bundled fallback tables.
• Thread provider-scoped models.dev catalog/cache context through Codex, Claude, and Pi session cost scanners.
• Recompute report costs from the current catalog before falling back to cached packed cost rows, with focused regression coverage.

Reasoning

• PRs Feat: Add models.dev pricing metadata pipeline #863 and Fix models.dev pricing refresh continuity #881 added the models.dev metadata pipeline; Fix: Align models.dev refresh guard with lookup #883 aligned refresh continuity. This moves the Codex and Claude pricing layer into that flow while preserving bundled fallback behavior.
• Provider-scoped lookup keeps same-name or overlapping model ids from crossing provider boundaries.

Scope

• Codex and Claude pricing layer.
• Cost usage scanner report construction.
• Pi session merged cost reports.
• Focused pricing tests for models.dev precedence, threshold pricing, research-preview override, and isolated bundled fallback behavior.

Screenshots

N/A

GIFs

N/A

References

• Builds on Feat: Add models.dev pricing metadata pipeline #863, Fix models.dev pricing refresh continuity #881, and Fix: Align models.dev refresh guard with lookup #883.

Validation

• swift test --filter CostUsagePricingTests --filter PiSessionCostScannerTests --filter CostUsageScannerTests
• make check
• ./Scripts/compile_and_run.sh

Audit

• GPT-5.5 high subagent audit completed. Initial findings were fixed, and the final pass found no material issues.

[steipete]

Pushed a small test-only CI fix in 59a1ad7: precompute the pricing expectations so Swift does not time out type-checking #expect arithmetic expressions.\n\nValidation:\n- swift test --filter 'CostUsagePricingTests|PiSessionCostScannerTests|CostUsageScannerTests'\n- make check\n- CI is green on the pushed commit.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 680561a4a3

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".

[chatgpt-codex-connector]

Preserve per-request threshold costs for Claude rows

This now prioritizes currentPricingCost over cachedCost, but currentPricingCost is computed from day/model aggregates instead of original per-request usage. For Claude models with long-context threshold pricing, aggregating many sub-200k requests can incorrectly cross the 200k tier and overstate cost (or understate in other mixes). Before this change, cached nanos preserved per-request pricing behavior, so this introduces report inaccuracies whenever threshold-priced models are used multiple times in a day.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Avoid re-rating aggregated Pi usage with tiered pricing

computedCostUSD is being re-run on the aggregated packed totals for each model/day, which loses per-message boundaries needed for correct tiered Claude pricing. If multiple requests are each below the long-context threshold, recomputing on the sum can apply above-threshold rates that never actually occurred per request. The prior behavior used accumulated packed.costNanos from parse-time samples and did not introduce this aggregation error.

Useful? React with 👍 / 👎.