fix(usage,sessions): resolve token/subscription usage integration disparities

[junhoyeo]

Summary

Fixes the actionable disparities found in an audit of all token-usage + subscription-usage integrations (11 subscription providers in usage/*.rs, 34 session parsers in sessions/*.rs). The unifying theme: omit or fall back, never fabricate — several integrations emitted confident-looking empty/zero/100% data or silently dropped real records.

Subscription-usage providers (fix(usage))

• amp (high — confirmed live): fetch() returned Ok with no metrics when its $X/$Y display-text scrape didn't match, rendering a bare "Amp" header with no data. Now bails so fetch_all drops it. Also made the byte-offset slicing char-boundary-safe (it could panic on multibyte display_text).
• copilot, minimax_tokenplan: same empty-metrics guard / skip empty rows.
• warp: the "Spend" and no-limit "requests" metrics set 0% remaining → a false "exhausted" bar; now read full with the figure in the label. Dropped the placeholder "Aggregate API cache" plan string.
• zai: a missing percentage rendered as "100% left"; now skipped.
• kimi: dedup key now includes resets_at so distinct windows with equal numbers don't collapse.

Token-usage session parsers (fix(sessions))

• Provider canonicalization (opencode, micode, kilo, mux, antigravity, jcode): these stored a raw provider field verbatim, so aliases (fireworks, vertex, gemini, azure) split into separate aggregation buckets and undercounted per-provider totals. Now routed through provider_identity::canonical_provider, matching the other parsers.
• codex: provider defaulted to "openai" when session meta was absent; now infers from the model first, keeping "openai" only as the final fallback.
• gjc, pi (high — data loss): dropped messages that had valid tokens but no provider field; now recover via model inference (fallback "gjc"/"pi") instead of discarding the spend.
• droid: dropped a whole session when its timestamp resolved to 0; now falls back to file mtime.
• qwen, mux: had no dedup key (double-count risk on incremental re-parse); now keyed stably.
• utils / kiro / goose: seconds→ms conversion uses saturating_mul; timestamp paths hardened against overflow.
• kiro / goose / crush: documented source limitations (estimated tokens / heuristic reasoning / cost-only) so 0/estimated values aren't mistaken for bugs.

Deliberately NOT changed

• roocode / cline / kilocode intentionally preserve raw apiProtocol (has a test) — not canonicalized.
• grok input-only, codex API-specific credit fields, amp parser's Anthropic-first default — documented source behavior.

Testing

• cargo build --workspace clean; cargo test --workspace 1850 passed / 0 failed; cargo clippy --workspace --all-targets clean; cargo fmt --all -- --check passes.
• New unit tests for the amp (empty/multibyte) and warp (full-not-exhausted) behavior changes.
• Test updates were limited to ones encoding the old behavior: fireworks → fireworks_ai canonicalization and gjc now keeping a previously-dropped record.
• Live-validated on a real machine: the Amp ghost-row is gone (Amp drops cleanly instead of rendering an empty header); Claude/Codex/Copilot render full data.
• Also swept 3 pre-existing clippy useless-vec warnings in report.rs test helpers to keep the gate green (style(report) commit).

---

Summary by cubic

Fixes incorrect or missing usage data across subscription providers and session parsers by failing soft on empty responses, canonicalizing providers, and adding safe fallbacks. This removes empty headers, corrects Warp bars, prevents double counts, and improves per‑provider totals.

• Bug Fixes

  • Subscription usage: bail on empty metrics (Amp, Copilot, MiniMax); make Amp parsing char‑boundary safe; show Warp "Spend" and unlimited "Requests" as full with value in label; skip Zai limits without percentage; include resets_at in Kimi dedup.
  • Sessions: canonicalize providers (opencode, micode, kilo, mux, antigravity, jcode); Codex infers provider from model before "openai"; GJC/PI keep messages with missing provider via model inference (fallback to "gjc"/"pi"); Droid falls back to file mtime when timestamp is 0.
  • Data integrity: add stable dedup keys (Qwen, Mux, Zed); overflow‑safe seconds→ms conversions (utils, Kiro, Goose).

• Refactors

  • Documented source limits: Kiro/Goose use estimated tokens; Crush is cost‑only (zero tokens expected).
  • Dropped Warp placeholder plan; minor test updates and clippy cleanups.

^{Written for commit fbd2bd6. Summary will update on new commits.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
tokscale	Ignored	Preview	Jun 22, 2026 8:48am

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fbd2bd6070

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]

Drop Z.ai output when all quota rows are skipped

When Z.ai returns limits that omit percentage (the exact case handled here), every row can be skipped but fetch() still returns Ok(UsageOutput { metrics: vec![] }), so fetch_all() includes it and the renderer shows a bare Z.ai card instead of failing soft like the Amp/Copilot/MiniMax guards added in this commit. Add an empty-metrics bail after assembling metrics (or compute from other fields) so unparseable quota responses are dropped rather than rendered as an empty provider.

Useful? React with 👍 / 👎.