Count Codex cached input once, and read archived rollouts in the report

[tsouth89]

Fixes SOU-295 and SOU-296. Both are Codex accounting bugs with the same shape: a second implementation that did not know what the first one knew.

SOU-295 — per-model cache read showed 49% instead of ~97%

Codex reports cached input inside input_tokens. model_breakdown used the provider-blind ModelTokenCounts::processed() (input + output + cache_read + cache_write), so cached tokens landed in the denominator twice and the ratio collapsed toward one half:

displayed: 97 / (100 + 97) = 49.2%
correct:   97 / (  3 + 97) = 97%      (fresh input = 100 - 97)

The top card was already right, because token_breakdown subtracts cached input first. Two copies of one rule, one of them wrong.

Both paths now go through a single NormalizedTokens type in the scanner that owns the convention (provider_folds_cache_into_input), exposes processed() and cache_read_percent(), and is reachable from both CostSummary and ModelTokenCounts. The provider-blind processed() is gone, so there is no longer a way to build the wrong ratio.

Scope note: the original report suggested fixing "model rows, effort rows, projects, CSV exports." Verified that is broader than reality — .processed() had exactly one production call site, and effort/project/CSV compute no cache percentage. One field, one surface.

SOU-296 — archiving a Codex task shrank every total

scan_codex_report walked only sessions/, never archived_sessions/. It is the path behind the Charts cards, the reset-aligned token and dollar windows shipped in 1.3.1, and the estimated API value card — while the older summary scanner did read archives. The two disagreed for the same account on the same machine.

It also had no dedup at all (unlike the Claude report path), so one rollout present under two homes or WSL roots could be counted twice.

Rather than copy the archive walk (a second implementation is how these drifted in the first place), the per-rollout accounting is extracted into CodexReportRollups::ingest with a shared seen set. Both the date-nested tree and the flat archive dir now run through it, the archive gated by the existing archived_rollout_day_in_range filename check.

Tests

• codex_cache_percent_does_not_double_count_cached_input — 338.2M input / 329M cached reads ~97%, not 49%.
• claude_cache_percent_leaves_input_untouched — the correction is Codex-only.
• normalized_removes_cached_input_for_codex_so_cache_rate_is_honest / normalized_sums_every_bucket_for_a_provider_that_separates_cache / cache_read_percent_is_none_without_activity.
• codex_report_counts_archived_rollouts_exactly_once — archived-only rollout counted, shared rollout counted once. Confirmed to fail without the fix (1 session instead of 2), so it is a real regression test.

Shared lib 658 passed, desktop 395 passed, frontend 278 passed. Clippy -D warnings clean on both crates, cargo fmt --check clean, tsc --noEmit clean.

Summary by CodeRabbit

• Bug Fixes
  • Corrected token and cache-read percentage calculations across Codex and Claude usage reports.
  • Prevented cached input from being counted twice in Codex statistics.
  • Ensured Claude input metrics remain accurate.
  • Improved usage summaries and charts to consistently include archived sessions without duplicate counting.
• Tests
  • Added coverage validating provider-specific token breakdowns, cache percentages, and archived-session reporting.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	ceiling	0e6d498	Commit Preview URL Branch Preview URL	Jul 21 2026, 06:53 PM

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 2433c3f3-76c6-4ac3-8d2b-edb34f50d06e

📥 Commits

Reviewing files that changed from the base of the PR and between 8ea0222 and 0e6d498.

📒 Files selected for processing (2)

• apps/desktop-tauri/src-tauri/src/commands/chart.rs
• rust/src/cost_scanner.rs

---

📝 Walkthrough

Walkthrough

Changes

Provider token normalization

Layer / File(s)	Summary
Provider-aware token normalization rust/src/cost_scanner.rs	Adds normalized token buckets and provider-specific cache-folding behavior, with tests for Codex, Claude, and empty activity.
Chart breakdown integration apps/desktop-tauri/src-tauri/src/commands/chart.rs	Uses normalized tokens for token and model breakdowns, updates call sites, and validates provider-specific cache-read percentages.

Codex archived report rollups

Layer / File(s)	Summary
Shared Codex rollout accumulation rust/src/cost_scanner.rs	Consolidates nested and archived session ingestion through CodexReportRollups, including rollout deduplication and report finalization.
Archived rollout regression coverage rust/src/cost_scanner.rs	Tests archived-only inclusion and prevents duplicate counting for rollouts present in both session trees.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant Sessions as sessions/
  participant Archives as archived_sessions/
  participant Rollups as CodexReportRollups
  participant Report as scan_codex_report
  Sessions->>Rollups: ingest nested rollout JSONL
  Archives->>Rollups: ingest in-range archived rollout
  Rollups->>Rollups: deduplicate rollout filenames
  Rollups->>Report: finish(days)

Loading

Possibly related PRs

• tsouth89/ceiling#53: Modifies Codex report aggregation and archived rollout deduplication.
• tsouth89/ceiling#87: Overlaps with cache-read percentage and token-count processing changes.

Suggested reviewers: finesssee, darknight105

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes both main fixes: Codex cached input accounting and archived rollout ingestion.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/codex-cache-accounting-and-archives

---

_{Comment @coderabbitai help to get the list of available commands.}