Dashboard feedback repairs and uncommitted-code review, 2026-10-03
Reviewed eight fixes in the 2026-10-02 tree against source/tests/official protocols/native data. Corrected earlier claims about no currency conversion, the 658-count composition and whole Kilo database readiness. Production database/settings read-only; write isolated copies only. No commit/push/deployment this round.
Run information
| Item | Value |
|---|---|
| Date | 2026-10-03 |
| Environment | Windows 11 x64; Node 24.21.0; Rust 1.98.1; headless Edge. Desktop dependencies restored from lockfile: Tauri API/CLI 2.12.0, Svelte 5.57.1, TypeScript 6.0.3, Vite 8.3.1 |
| Revision | Uncommitted frontend charts/settings/cost cards and core pricing/adapters/health changes |
| Design references | Interactions, repairs, pricing, data rules, adapters, architecture |
Problems and repairs
| # | Feedback | Cause | Repair |
|---|---|---|---|
| 1 | Today chart became bars | TodayHourly.renderTotal() calls series type:bar | Calls/tokens both smooth lines; agent comparison within one hour retains bars |
| 2 | Cache lifetime days shows 0 | Derived PricingSettings Default sets online_cache_ttl_days=0; 0 ?? 3 stays 0 | Manual default=3, DEFAULT_TTL_DAYS; load_settings resets zero/out-of-range to 3 |
| 3 | API reference is text only | Previous repair left it outside metric grid | Eighth bordered today/trend card; short ten-language API reference cost heading, full hover explanation. Eight/four/two/one columns, quota separate. Widths 1920/1280/1024/760/480: no overflow, equal-width columns, one-line headings |
| 4 | Are all currencies USD; is conversion needed? | models.dev and native CNY rates differ | models.dev USD/million tokens; native snapshots USD/CNY. Rates stored in hundredths of a cent/million tokens; 1M input+1M output regressions yield USD 18/CNY 120. CNY amounts/rates show approximate USD alongside, with ECB date/source; currencies/history retained |
| 5 | k3-256k has no applicable rate | Native provider=kimi-code-owent, bare model_raw=k3-256k; reference_model_key only knew prefixed kimi-code/k3-256k | reference_model_key/official_providers also map bare k3/k3-256k to kimi-k3/moonshot, using official USD reference |
| 6 | 658 unpriced/9 partial estimates | Earlier explanation mixed current reference/occurrence-time prices and treated known-zero usage as unpriced | Distinguish no tokens/missing price/partial coverage; known zero prices as zero with applicable rate. Original snapshot after repair: 622 unpriced/9 partial, detailed below |
| 7 | Source needs 38 file checks | Cumulative mismatch treated as read failure; valid null info treated as invalid fields | codex-rollout-3 replays old cursors automatically; mismatch/regression retained without independently worsening read health. New individual records and legacy required total/last fields handled separately. Invalid rows/times survive batches. Copy’s 38 files all reevaluated active |
| 8 | Kilo reads one file compatibly | 7.8.1 unregistered; highest DB version incorrectly applied to every message | Native 34-call sample registers 7.8.1 and reevaluates old version metadata/cursors. kilo-message-tokens-2 uses each message’s session.version, preserving mixed-DB compatibility across incremental reads. Empty 7.8.3 session cannot verify format; notice explains automatic checks/review |
Accurate explanation of 658 and 9
Counts refer to current API reference in the original application snapshot, without substituting occurrence-time price history.
- 564 codex-auto-review records: route name is not a known actual model; no guessed GPT-series price.
- 34 k3-256k records: missing bare profile alias, fixed; all priced with official USD reference.
- 58 Copilot round markers: no token usage, retain calls without zeros/invented amounts.
- Two GLM known-zero records: old condition misclassified zero amounts as no priceable components, fixed.
Read-only native route files contain turn_context codex-auto-review plus individual call identity/usage, without underlying model fields. Token counts/quota models cannot establish the actual model. Same snapshot after repair: 622 unpriced, 564+58. Nine partial estimates are five VS Code usage observations and four Visual Studio calls: output/some cache known, without complete uncached-input split. Whole-turn input_total cannot use uncached rates directly. Unknowns remain unknown; newly collected records can change these counts.
Official references and automatic-check limits
Sources reread on 2026-10-03:
| Reference | Applied fact |
|---|---|
| Kimi Code models, global prices | k3/k3-256k are K3 profiles; per million input USD 3/cache read 0.30/output 15/cache writes 5m 3, 1h 6, tax excluded; membership quota is not API rate |
| models.dev maintainers | cost is USD/million tokens; a provider’s CN name does not justify converting again |
| ECB daily XML | Effective 2026-10-02, EUR/USD=1.1225, EUR/CNY=7.5259; CNY→USD=1.1225/7.5259 |
| Codex 0.144 official protocol, quota-event issue | TokenCountEvent.info is Option; local 0.144.0-alpha.4 info:null/rate_limits:object valid. token_count can repeat during quota refresh, without another call |
Exchange rates are code-versioned reference snapshots, without automatic network requests. Amounts/unit prices use their integer units and same-date cross rates/rounding; original value/approximation/date/source shown. Historical values/original-currency subtotals unchanged. Historical pricing rule official-reference-3 corrects only retained, unarchived dates.
Original 38 Codex files all reevaluated active; mismatch retained. Healthy reading does not establish complete source coverage. New 0.159.0-alpha.12.1 still latest-compatible, without verification based on same-shaped fields. Five regressions cover consumed unchanged bytes, monotonic revisions/history, reads split by byte limits, null info and invalid records.
Kilo 7.8.1 native sample’s 34 calls independently match 3,189,308 tokens; automatic replay does not double count. Live DB includes empty 7.8.3 session without assistant usage. Full check of 256 message-containing sessions still has one historical cumulative mismatch, retaining diagnostics/degraded health. Whole database readiness is not established. Empty/mixed-version tests ensure highest version does not verify other sessions or replace the basis of already verified messages.
Commands and results
Repository-root cwd unless specified:
| # | Command | Exit | Result |
|---|---|---|---|
| 1 | npm –prefix desktop ci –ignore-scripts | 0 | Restore locked dependencies, no lockfile changes |
| 2 | npm run verify | 0 | Documents/assets/scripts/UI/Svelte/fmt/Clippy/Rust/frontend build all pass |
| 3 | npm run test:rust within verify | 0 | 71 test binaries; 811 pass/0 fail/3 ignored; dashboard_repair 14, Codex replay 5, Kilo 8, pricing settings 5 |
| 4 | npm run check, svelte-check | 0 | Zero errors/warnings |
| 5 | npm run build:web | 0 | Frontend builds |
| 6 | npm run test:ui | 0 | 16 pass; ten languages, exchange units/rounding, coverage reasons |
| 7 | npm run test:browser, Edge | 0 | Curves/cost cards/resolutions/TTL=3/CNY amount-rate/reasons/compatibility/existing regressions |
| 8 | npm run fmt:check / npm run clippy | 0/0 | Formatting/lint pass |
| 9 | npm run lint:md | 0 | 171 files/0 issues |
| 10 | python -X utf8 build/review-feedback/run_live_probe.py | 0 | Native read-only sources into isolated DB; 38 files reevaluated, 34 K3 calls priced, version/history diagnostics retained |
| 11 | python -X utf8 build/review-feedback/audit_summary.py | 0 | Permitted-field redaction passes; 34 native backup records and sample have equal independent five-component totals |
| 12 | git diff –check | 0 | Final tree; untracked code/test data inspected separately |
Failures and unexecuted checks
| Item | Status | Reason | Next condition |
|---|---|---|---|
| codex-auto-review pricing | Unexecuted | Route has no public rate, actual model unknown | User mapping/configuration can enable pricing; no guessing |
| Production degraded/compatible file replay | Isolated copy verified, production unchanged | Original configuration/DB read-only | First collection after update automatically replays, no clearing required |
| New native GUI acceptance | Unexecuted | New GUI not launched here | Desktop checks alongside build:desktop |
Initial assets check failed installed Tauri 2.11.5 versus lockfile. Restoring dependencies passed 67 derived assets/82-file checks, without regenerating/editing icons. Node sandbox child-process EPERM retried through authorized execution. fmt found missing –all; Clippy found new syntax incompatible with repository MSRV; both fixed. Final logs: build/review-feedback/verify-final.log and browser.log. No release bundling, default-ignored remote models.dev smoke or actual telemetry-configuration write checks.
Validation files
- Read-only consistent snapshots/scripts/isolated DBs/logs in ignored build/review-feedback/. SQLite mode=ro plus Online Backup, without immutable skipping WAL.
- New desktop/src-tauri/crates/core/tests/fixtures/kilo/session-7.8.1-k3.sanitized.json: IDs/paths/message text removed; tokens/time/finish/modelID/providerID fields preserved. session-7.8.1-k3._expectations.md contains independent calculations.