The five v2.3 images below are reproducible captures of the production dashboard renderer with synthetic fixtures and VS Code Light+/Dark+ theme variables, not personal usage or billing evidence. Native VSIX installation is verified separately.

Today and Last 30 days share the configured calendar timezone. CLI usage is counted only when normal persistent sessions leave usage-bearing local logs; calls made without session persistence cannot be reconstructed.

The corrected Codex overview uses one configured-timezone calendar model for Today, Last 30 days, months, models, effort, and all-time totals.

Observed quota windows retain reset evidence locally. A valid used fraction can produce labelled total and unused subscription-durability estimates, including for the current window. Approximate evidence remains visible with low confidence. Period details start collapsed; expand them to inspect the numeric evidence.

Compare combines provider daily activity without double-counting Codex cached input or reasoning. Its preview-first share studio offers an Academic Violet default, curated/custom colors, deterministic local SVG, and privacy-safe Markdown. It is enabled by default and can be hidden with the single sharing workspace setting. Card settings sit below the preview; intensity can use quantile, logarithmic, or linear scaling. The metric is activity volume, not productivity or billing.

Projects now adds a Token-only 30/90-day project × day heatmap and stacked daily trend for both providers. Exact tooltips, explicit coverage, bounded rows, and an Other-projects tail keep the view auditable without rereading source logs.

• Bounded pricing diagnostics — unknown-model warnings are deduplicated and capped for each Extension Host lifetime, removing the per-record warning flood reported in #122. Already-expired content-analysis files retain their empty contribution rather than disrupting live-tail refreshes.
• Current model prices — dedicated Standard/cache rates for Opus 5.5, Sonnet 5.5, GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna. Unknown Codex IDs remain unpriced; family fallback estimates are not promoted to exact-price coverage.
• Defensive refreshes — malformed model metadata no longer aborts a whole index. Unsafe object keys cannot mutate shared prototypes; aggregate buckets are copied once per update, worker results have a bounded apply window, and manual price refreshes have byte, catalog-size, concurrency and time limits.
• Less disruptive dashboard updates — pausing auto-refresh freezes both provider pages, not background indexing or the status bar. Manual refresh remains available; failures retain verified data and show a compact recovery message. Primary-log coverage, hourly backfill and retry waits are separate. Completed Codex history stays quiet during normal appends/new files; identity ambiguity remains a quality warning, not a repeating backfill attempt. Unchanged hidden panels and weekly usage aggregates reuse bounded caches.
• Predictable preview and settings — AI previews show the actual endpoint, protocol and model; incompatible settings fail without redirecting to another host. Share Card export writes the accepted SVG and requires a new preview after control changes. Ordinary defaults reset preserves the API key, and the Codex directory remains editable even when provider detection fails.
• Large-history refreshes — unchanged Content attribution reuses its data without freezing AI controls. Display settings no longer restart completed cold reads, and late index results cannot restore superseded prices. Today attribution is cached by calendar day: unchanged minute-spaced polls update countdowns without rescanning the full history within the hour. The hourly hidden-panel refresh still recomputes Content attribution once.
• Upgrade safety — unchanged polls retain accepted sharing previews; changing the Claude data directory clears the old source even while updates are paused. Existing API keys without an explicit protocol retain the prior Anthropic format. If preview reports a mismatch, choose the intended API format and URL in Settings, then preview again before sending. Upgrade/default reset never silently reroutes that key to the new DeepSeek default; new installs use the matching OpenAI-compatible format.
• Codex status and scrolling — the default compact Token item shows today's processed amount; the separate weekly indicator shows remaining capacity. Live panel updates wait briefly for scrolling to pause, with a bounded delay.
• One preview-first sharing workspace — a full-width export preview now leads the surface, with controls below it and one presentation selector for the Combined activity heatmap, Claude Share Card, and Claude token heatmap. The combined presentation appears only with real data from both providers; both legacy presentations remain Claude-only.
• Compatibility without duplicate panels — exportShareCard, exportHeatmap, and publishHeatmapToGitHub remain available and open the matching presentation instead of bypassing preview. enableShareCard remains the single visible sharing switch and still defaults to on; the retired showHeatmap value is retained only for one-release state compatibility and bounded clearing.
• Strict local-artifact boundary and provider truth — switching, previewing, and local SVG/Markdown export use materialized aggregates only and perform no network request, GitHub sign-in, or profile/avatar/name lookup. Only the separate Publish to GitHub action for the Claude heatmap can connect; it remains public-repository-only and confirms the exact target and create/ overwrite action before writing. Combined activity is not billing, productivity, capability, or cross-provider equivalence.

Claude Share Card preview · 360 px dark-theme view. Both use synthetic fixtures, not a real account.

Models, token accounting, and cost semantics

• Refined throughout the v2.3 line — GPT-6 Astra and Fable 5.1 model metadata, optional AWS Bedrock pricing, a fixed-reference display-currency selector, a Token-only 30/90-day project activity matrix, complete month/day/hour drill-downs, state-preserving refresh, accessible charts, and lower watcher/title-index overhead. Patch-level details stay in the changelog and GitHub Releases.
• Codex Beta, enabled by default — usage records are discovered only from sessions/**/*.jsonl and archived_sessions/**/*.jsonl; credential, database, and unknown files stay excluded. Separately, the extension streams exactly $CODEX_HOME/session_index.jsonl to map id to thread_name for truthful thread titles. Absolute paths are redacted and titles stay memory-only. Usage-record JSONL lines are streamed and temporarily parsed only to extract allowlisted usage and structural metadata; prompt, response, command, and tool-argument fields are not inspected or used for analysis, and are never retained or persisted. Disable Codex at any time in Settings → Providers.
• Codex-native metrics — processed = input + output; uncached usage = uncached input + output; cached input is a subset of input; reasoning is a subset of output. The overview also shows input cache hit rate as cached input / input. No Codex billing cost is shown. The first Codex summary card is a clearly labelled API-equivalent cost estimate for the selected scope; the All-time view also shows the weekly trend using the same pricing basis.
• Calendar-day Today and cost trends — Codex Today means the current day in your configured timezone and adds exact hourly API-equivalent cost beside a separate token-composition view. Daily and monthly primary charts default to API-equivalent cost while token composition stays separately visible. Only exact known-model prices contribute; unknown models remain unpriced and every row keeps pricing coverage visible. The schema-3-compatible hourly sidecar keeps sparse buckets for the rolling last 30 days, is checkpointed and resumable, evicts day 31, and serves date expansion with zero JSONL reads on click. Codex monthly charts and tables list months oldest-first.
• Request-level token attribution — valid last_token_usage components are preferred, while its total_tokens remains an active-context measurement, not request usage. A full numeric total-plus-last signature suppresses only proven replay; missing last snapshots fall back to cumulative lineage high-water. Upgrading triggers one automatic reindex, with the indexed subtotal still visible throughout the pass.

Weekly estimates and provider boundaries

• Weekly allowance-value trend — Claude and Codex All-time / Compare views calculate historical used equivalents directly from local token logs. The newest valid official reset observation anchors one sequence of unique, non-overlapping weekly periods, and each usage event belongs to exactly one period; without a usable observation, usage-only rows fall back to Monday-to-Monday UTC calendar weeks. A genuinely different quota series with an overlapping, non-aligned future reset is a conflict; same-series observations remain one series and may be shown as a low-confidence approximation. It cannot create a second current period, and period ranges are shown separately from reset times. Codex usage is stored in daily slices: when one crosses an official intraday reset, its tokens are still counted once, the affected period is labelled a boundary approximation. Codex's account-wide codex observations can decorate both historical and current buckets. If an observed reset drifts from the seven-day grid, the sample is mapped to the display period containing its observation time; file-source uncertainty, reset drift, and daily boundary crossings lower confidence and are labelled as approximate. File keys are not account identities, so eligible local files in one home are included together. A current period still uses the latest real observation for a low-confidence blended estimate when local quota series overlap. Ambiguous completed periods, or periods without a usable observation, remain usage-only rather than inventing an account split. Any coherent observed window—including the current one—can show total and unused durability estimates; attribution or boundary uncertainty lowers confidence instead of silently replacing the values with dashes. This display rule neither changes the index schema nor triggers a rebuild. Current official API rates are applied consistently across history. This is a proxy, not a bill or an official subscription price. The panel is enabled by default and can be hidden in Settings with showWeeklyEquivalentValue.

Dashboard, indexing, and privacy

• One dashboard render stack — switching to Codex keeps the established Today / Last 30 days / All time / Sessions / Projects / Content / Settings structure, relabelled where Codex semantics differ. The same render functions, HTML classes, charts, tables, spacing, and responsive rules are used for both providers. Claude and Codex time-series charts stay width-aligned while dense content scrolls within its own region; Codex recommendations use indexed 30-day structural evidence.
• Truthful names, no invented concepts — root tasks use the latest real thread title after path redaction. Child rows prefer their own real thread title; when it is missing, they use the reported nickname and display the parent/root title; if those are also missing, they receive a localized neutral fallback. Projects use the Git repository name or, outside Git, the directory basename. The Codex view does not manufacture Branches or Workflows that cannot be measured reliably.
• Claude / Codex / Compare modes — keep each provider's meaning intact. Compare shows input, output, and cache side by side; it never adds unrelated costs or quota windows together.
• Auditable time and coverage — rolling 7-day and 30-day totals use exact event-day slices in your configured timezone. During an incomplete migration or rebuild, every Codex card, table, project, session, recommendation, and status value is visibly an indexed subtotal; unverified legacy totals are excluded. Once a Codex home is detected, its provider tab appears immediately; the page shows exact indexed-file, percentage, and byte progress during the first build, while Compare waits until both providers have real data. After the first atomic checkpoint, the complete subtotal dashboard stays usable while indexing continues; progress never replaces its cards or tables. A selected Codex home is account-agnostic, so logs left by multiple sign-ins in that same home are combined. Local logs expose no reliable account identity, so limit cards remain last-observed and are never summed.
• Private, scalable local index — stores machine-salted pseudonymous keys; numeric and structural aggregates; and sanitized project, directory, agent, model, effort, role, time, and quality metadata. It never persists raw IDs, full paths or repository URLs, thread titles, or conversation bodies. A background worker scans large histories with a default 30-second watcher delay (Off / 10 / 30 / 60 / 120 / 300 seconds). A first-time index or incomplete legacy migration gets one bounded 64 GiB / 16,384-file-pass streaming ceiling; it does not reserve that amount of memory and remains cancellable and resumable. After convergence, background work returns to 128 MiB / 64 file passes and the always-visible Refresh action uses 2 GiB / 512 file passes. Unchanged warm refreshes still read zero usage-record JSONL body bytes.
• Provider-aware controls — the shared Settings tab shows only common and Codex-effective controls when Codex is selected. Codex collection and local Codex recommendations can each be disabled.
• Exact-version release notice — the upgrade message only describes the installed release. It is on by default and can be disabled in Settings.

• Usage share card (opt-in, enableShareCard) — a themed, configurable one-page SVG of your usage: pick a range × scope (overall / project / session) × which metrics to show, and a theme (Claude Classic / Cream / Aurora Dark / Auto). Preview and local export never fetch GitHub identity data. Self-contained and deterministic; no prompts, paths or ids ever leave your machine. Chinese locales use 万/亿 units.
• Read-only conversation viewer (Sessions tab, on by default) — a "view" button re-opens a past session's prompts and Markdown-rendered answers so you can jog your memory without loading it back into the model's context (unlike resume). Thinking and tool traffic sit behind toggles; opens on the last rounds.
• Token heatmap (opt-in, showHeatmap) — a GitHub-style yearly token heatmap on the All tab, plus Export / Publish to your GitHub profile as a self-contained SVG with a one-click Markdown embed.
• Experimental insights (opt-in, showInsights, Content tab) — labelled estimates from your local logs: a cache-churn bill ($ spent re-writing cache after model switches / idle gaps), cache warmth by model, big one-shot turns, your active hours, and skill ROI (output tokens returned per $).
• Top-10 costliest messages (opt-in, showCostliestMessages) — ranks single turns by cost, splitting a cache miss from a long answer, with the cache-hit rate and the time since the last turn.
• Sessions "Active" column — estimated hands-on time per session (idle gaps capped at 1.5 h), far more meaningful than the raw first-to-last span.
• Cache-hit-rate column in the All-time (monthly) and This-month (daily) tables and their drill-downs, so cache efficiency is visible per row.
• Live-refresh delay (fileWatchSeconds: Off / 1 / 2 / 5 / 10 / 20 / 30 s) replaces the on/off live-watch toggle; it only re-reads your local logs.
• Timezone-aware bucketing — Today / day / month / hour totals all key off your configured IANA zone (full UTC-offset dropdown), so they agree with each other and the Anthropic console.
• "What's new" prompt after upgrades — a single dismissible nudge the first time you run a new major.minor version, so opt-in features stay discoverable.
• Fixes — Sonnet 5 reports a 1M context window (#50); 1-hour cache writes are priced at 2× base input when the log carries the split (#62); background windows refresh on focus instead of going stale (#55); the Timezone setting is a validated dropdown so a bad value can't crash the dashboard (#51); German (de-DE) and Brazilian Portuguese (pt-BR) are selectable everywhere.

• Sessions: resume / copy / delete — each row can copy the session id, resume it (official Claude Code extension in-tab for this project, or a terminal claude --resume <id> for other projects), or delete it (to the trash, with confirm). A Current project / All filter defaults to the current project.
• Quota display options — quotaFiveHourOnly (show only the 5h window) and showResetInStatusBar (append the reset countdown, e.g. 5h:50%:2.3h | wk:30%:3.2d), both in ⚙ Settings.
• Wider dashboard — detail page widened to 1600 px, still fluid on narrow screens.
• Workflows tab — every multi-agent run in one place: dynamic-workflow runs (ultracode) and ad-hoc sub-agent batches, with per-run cost, agent count, models used, cache hit rate (the "is my provider workflow-ready" diagnostic) and a per-agent drill-down labelled by each agent's task.
• Usage tracking panel — the official /usage "what's contributing" view, but multi-provider and with five scopes (Day / Week / Month / session / project): >150k-context share, 8h+-session share, subagent-heavy share, workflow share, plus Skills / Subagents / Plugins / Models breakdowns. Compact card on the Today tab.
• Thinking share per session (Sessions column + Today card) with an /effort hint when it runs high.
• Workflow quota guard — a dismissible banner before you start a run the remaining 5-hour window can't finish (claudeCodeUsage.workflowQuotaWarnPercent).
• Settings in the dashboard — a new ⚙ Settings tab manages most options in place; VS Code's own Settings keeps the language and provider data directories. The advice API key is entered in the dashboard and stored in SecretStorage, not Settings Sync. Header buttons trimmed to ✨ AI advice and ⚙ Settings (both jump to their tab); the auto-refresh toggle moved into Settings (a manual ↻ appears when paused). If you hide the cost, quota and context items, the status bar keeps a small icon as a way back into the dashboard.
• Status-bar metric (statusBarMetric) — keep showing today's cost, or switch the first item to today's total token count (compact k/M).
• Model-scoped weekly limit (showScopedWeekly, opt-in) — adds the weekly cap actually named by Anthropic, such as fable 17%; migrated from the original model-specific contribution in PR #38 by @wheelbarrel00.
• AI advice 2.0 — bring your own key for Anthropic or an OpenAI-compatible endpoint (advice.apiFormat). The v2.3 line places local evidence and an exact full- request preview before the separate Send action. Aggregate-only is the default; prompt samples and advice.userContext require independent prompt- personalisation consent and appear verbatim in the preview. The keyless Claude Code subscription backend is not shipped and is unreachable.
• Usage Optimizer (experimental, advice.optimizer.enabled, default off) — a Content-tab card where you paste a rough request and get back one tightened prompt as plain text (paste-ready, no Markdown) plus a recommended effort / thinking / model. Three optional lenses (flag ambiguous references · condense long pastes · suggest a style direction). Only the text you paste is included, and the full provider request must be previewed and explicitly sent.
• Context-window indicator (experimental, off by default) — opt in via Settings to show the current session's context fill in the status bar. A "~" marks a guessed window; set contextWindowOverride for proxied/custom models.

• Real 5-hour and weekly quota in the status bar — reads the OAuth session from the same Claude profile as the window: explicit dataDirectory, then the first valid CLAUDE_CONFIG_DIR, then ~/.claude. The macOS Keychain is used only for the default profile. Adapted from upstream [PR #9](https://github.com/ClaudeCodeUsage/ClaudeCodeUsage/pull/9) by @Dobidop.
• Four new tabs: Sessions, Projects, Content, Branches — all sortable.
• Token-composition stacked chart with Y-axis and reference lines.
• AI advice command — now routes to the unified local-evidence surface; a missing API key leaves the request unsent instead of opening a separate demo or transport path.
• Multi-vendor pricing: Opus 4.x, Sonnet 4.x, Haiku 4.5 (verified against Anthropic's public pricing); reference rates for proxied setups (OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen) with family-aware fallback. Refresh Token Pricing pulls live LiteLLM data as runtime overrides.
• Custom timezone for date display (claudeCodeUsage.timezone).
• Light-theme tab readability fixed.
• Locale-aware numbers and dates throughout (German ., English ,).
• Real-time status bar via fs.watch (1.5 s debounce) + idle-aware refresh + non-blocking loader (yields every 25 files).

Full changelog: CHANGELOG.md. Closes upstream issues #7, #10, #11, #13.

• @Alfiefe10 — quota format; test stability.
• @akapti — AWS Bedrock regional pricing.
• @Carl723000 — Claude/Codex dashboard; release and maintenance work.
• @Dobidop — OAuth usage limits.
• @e7d — Opus 5 context window.
• @henrique-carvalho-dev — Brazilian Portuguese.
• @jack21 — original project; Codex adaptation guidance.
• @jackyyangjq — dashboard date-label performance.
• @mxzinke — model pricing; German translation.
• @nickearnshaw — localization; loading/quota fixes; test, CI, and docs foundations (PRs).
• @oxsean — session actions and quota resilience.
• @PhisicsLollo0 — monthly cost in the status bar.
• @projectronic — reset countdown; Indonesian UI and timezones (PRs).
• @ptweezy — unreadable-log resilience.
• @rsyuzyov — incremental indexing and date-formatting performance (PRs).
• @ScherbakovAl — status-bar cost toggle and context fill.
• @UfukTanriverdi8 — Sonnet 5 context window.
• @wheelbarrel00 — model-scoped weekly quota.
• @YuboZhang — model pricing; daily date labels.
• @zeyutang — timezone/cache pricing; weekly quotas and credits tooltip (PRs).

• @jordanvalnet — unknown-model warning flood and Windows OOM evidence.
• @dreamerhyde #1 · @skyprawngo #2 · @tjx666 #3
• @andrew-west-empromptu #7 · @faangbait #10 · @leisn #11 · @dmathisen #13
• @Rissew #17 · @zhaoxiao9302 #18, #105
• @danielboxer #26 · @lyrzxz #27 · @BarisKuzu #45, #87
• @ekimminau #51 · @dbookstaber #54 · @eduardogomezgvp #55 · @chuccv #70
• @nsevene #80 · @arhneu #82 · @jcastillooo #84
• @HoangJN #89 · @Bozodragon #91 · @aaroncvan #94 · @mkgaskin-ops #99, #126
• PR authors also reported issues: @nickearnshaw #25, #33 · @oxsean #40 · @projectronic #74. The maintainer's own issues are in the issue history.

Thanks also to @huanglune (#12), @brenoneill (#14), and @Ailuras (#28) for proposing changes. These PRs were closed without merge; their ideas are credited without claiming that the PR code shipped.

v1.0.8 (2025-11-28)

• Converted code comments from Traditional Chinese to English.
• Improved internationalisation standards.
• Pricing: added Opus 4.5 / Haiku 4.5 (thanks @mxzinke).
• Added German (de-DE) translation (thanks @mxzinke).

v1.0.7 (2025-11-28)

• Multilingual translation for hourly usage labels.
• Removed hardcoded Chinese text; switched to i18n.

v1.0.6 (2025-08-10)

• Added support for Claude Opus 4.1 pricing.

v1.0.5 (2025-01)

• Hourly usage statistics + visualisation.

v1.0.4 (2025-01)

• All-time data calculation; "All Time" translations.

v1.0.3 (2025-01)

• Repository URL migration + README image link fixes.

v1.0.0 (2025-01)

• Initial complete release.