Claude Code Usage

Track Claude Code and OpenAI Codex usage locally in VS Code. See today's tokens in the status bar, remaining quota, cache usage, project and session trends, and optional usage advice in one dashboard. Install from the VS Code Marketplace or Open VSX.

The extension reads local provider logs. Processed Codex tokens include input and output; uncached usage excludes cached input, and quota is the last local observation, not a live account balance. Claude costs and Codex API-equivalent prices are estimates, not bills or subscription charges. Provider accounts remain the source of truth for billing. The 简体中文说明 covers the same features in detail.

Open the dashboard by clicking either usage indicator in the VS Code status bar. See features, privacy, installation, and contributors below.

🌐 Multi-language documentation: English · Deutsch · 繁體中文 · 简体中文 · 日本語 · 한국어 · Português (Brasil) · Bahasa Indonesia

• Claude Code and Codex, side by side: provider-specific Today, Last 30 days, All time, Sessions, and Projects views; Compare keeps their accounting separate.
• Know what used tokens: input, cached input, uncached usage, output, models, effort, and month → day → hour drill-down without rereading logs on chart clicks.
• See quota without confusing it with spend: Claude's official /usage utilisation and Codex's last-observed remaining allowance; estimates are labelled.
• Keep sharing under your control: preview and export local SVG/Markdown; sharing is on by default but can be disabled, and GitHub publication is separate.
• Local first: source logs stay provider-owned; optional AI advice requires an explicit request preview and separate Send action.

Codex uses a compact Today token usage item and a separate remaining quota item: observed 36% weekly utilisation displays wk 64%. The hover card retains utilisation progress bars, reset times, and wrapped explanations. Local Codex quota is last-observed evidence, not a live account balance.

Today's cost · current-session cost · 5-hour and weekly quota utilisation.

Hover the quota indicator for a breakdown:

Real /usage data: utilisation percent, plus time left and the wall-clock reset for every window. Every weekly cap your plan meters gets its own row, per-model ones included (Anthropic supplies the name, so the row follows whichever model is capped), plus usage credits when you have them enabled.

⚙ Settings offers a compact quota-format dropdown: Built-in (default), 5-hour only, Weekly only, or Custom. Only Custom reveals the template field.

Click the status bar to open the full dashboard. Stacked token-composition chart, hourly breakdown, cache hit rate, cost composition by token type, plus per-model and per-day tables below.

Estimated breakdown of which content consumes tokens — your prompts vs. tool results (by tool) vs. assistant output / thinking. This is the lever for optimising your usage. Scoped to the last 30 days (advice.promptWindowDays).

v2.3 keeps one readable path from a local observation to its evidence, recommendation, action, feedback, and guarded result. It is off by default. Local evidence appears before any model is involved; Helpful, Not helpful, and Applied stay on this device. Once enough reliable, similar before/after tasks exist, the card reports the frozen comparison result; otherwise it says that the evidence is insufficient.

Each recommendation can be snoozed for a bounded period; it leaves the default summary and returns after expiry or when you choose to show it again.

AI personalisation is a separate choice. Aggregate-only is the default and prompt samples remain off until separately allowed. The extension prepares the complete request once and shows its exact JSON, byte count, and SHA-256. Preview sends nothing; Send this exact request is a second explicit action, using the same canonical bytes and your own configured key/endpoint. Claude Code OAuth credentials are never used as a generative backend.

A flavour of what it returns (illustrative):

Write more complete instructions

• Several prompts open with "fix the bug" but don't name the file or the symptom, so the first turn is spent searching. Lead with the file + expected vs. actual behaviour.

Cut waste where it doesn't cost clarity

• ~38% of your tokens are spent above 150k context. /clear between unrelated tasks keeps each request cheaper.

Paste a rough, half-formed request; get back one clean, paste-ready prompt (plain text, no Markdown) plus a recommended reasoning effort / thinking / model shown as chips. Three optional toggles refine it (flag vague references · condense long pastes · suggest a style direction). Experimental, off by default; only the text you paste is included — never your files or the terminal. It now uses the same full-request preview and separate explicit Send action as AI advice.

Production renderer with synthetic usage and VS Code theme variables, not a real account or an installed-VSIX capture.

Version 2.5.0 combines local achievements with Codex .jsonl.zst compatibility in sessions and archived_sessions. Zstd is lossless compression, not an LLM summary: smaller disk files preserve logged token counts. Reads stay local without changing source logs or saving decompressed copies; same-directory, same-basename .jsonl takes priority. Native Zstd (Node 22.15+ / 23.8+) adds no dependency; unsupported hosts report the limitation and retain last verified totals.

Session dates use the same uncapped admission as the canonical index; changed logs require recalibration. Verified real usage keeps its native first-recorded receipt even if calibration is cancelled. Partial saves are reported separately.

Expand all / Collapse all saves only card disclosure state in the existing local UI state; it does not change sources or awards. Ranks use localized Bronze / Silver / Gold labels without changing IDs, pins or reference pricebooks. Identical conditions and units are shown once across levels. One shared money-reference section shows each enabled provider's current book/date and the non-billing note. Different historical books retain their exact award associations there; cards keep scope and pricing coverage without repeated price/basis notes. Fully earned cards show the uncapped current statistic; stale or missing evidence is never replaced with a target value.

Expanded rank rows use per-rank theme colors for the grade name (foreground fallback in high contrast), with a bold, right-aligned threshold on the same header row. Only the next rank above the highest earned shows “In progress” when proof is ready; later ranks, fully earned cards and pending proof show no such label. Verified retained-log dates appear as bare ISO dates, with their meaning explained once in Evidence details and provenance retained in title/ARIA text. When both dates exist, only the verified date is visible; first-recorded details remain in title/ARIA text. A first-recorded date alone keeps its explicit label; unknown dates stay unknown. This presentation change leaves awards, rules, date calculations and saved records intact.

Independent root tasks counts verifiable Codex root ownership with recognized positive model usage (input or output). A root and its children count once; this is not a count of completed tasks. Existing rule IDs and thresholds are unchanged.

The navigation and page label is Share & compare. Badge choices use equal-width alignment and a separate rank line. The checkbox-icon repair uses the same local-resource source as the main webview CSP. Exported badge SVGs include the fixed public repository link, with wrapping/spacing for long labels. Preview/export adds no automatic network request or personal metadata and retains the accepted immutable SVG and explicit selection. Broad sharing-page visual unification/beautification and a durable usage archive are deferred to the next version; neither is implemented in 2.5.0.

Per-level log dates are derived in memory and mean the earliest date verifiable from retained logs, not complete account history. They do not replace saved first-record dates or unread state. Daily totals supply supported volume, output, cache, model/family, activity and weekly dates. Manual recalibration alone also reduces already-held Claude records for output-session, Fable-session and same-session Fable/other-family dates; ordinary warm refreshes do not regroup those records. Unsupported conditions retain unknown dates or their saved first record.

Recalibration is local, incremental and single-flight. It neither forces a full scan nor clears badges, and makes no OAuth, pricing, BYOK or other network request. It cannot force awards; busy work, an unaccepted current index census or failure is not success. Verified partial history retains lower bounds or unknowns; completion does not mean every award scope is fully covered. Cancellation before the historical commit adds no historical awards; after commit it does not undo saved records or change the successful result. Historical reconciliation stays quiet. See date and storage boundaries.

Collaboration/rainbow model recognition is separate from pricing, so recognized but unpriced labels stay unpriced; retained monetary records keep their published pricebook labels, and Claude-only peaks survive unrelated Codex version changes. Loading, unavailable, changed and disabled record states replace empty counts; verified Claude cumulative usage shows ≥ while enabled Codex evidence remains unverified. Next targets show the target name and rank, earned cards keep their earned rank, and SVG exports show each rank once.

Badge details link to existing current model/token statistics for the same provider, with a keyboard-safe return. Shared and historical-period badges do not substitute one provider's current totals for award evidence; navigation does not scan logs.

Removing a legacy pin keeps keyboard focus within its theme, without refreshing data. Exploration details show three saved local steps: collection, share preview and successful export. Missing step records are not inferred from old awards; advice feedback and setting changes remain separate, optional activities. The journey emblem has no embedded counts. New foreground notices reveal only the badge once for 440 ms; text stays still, with static reduced-motion and high-contrast alternatives. No sound or focus change.

Model experiences distinguish models recorded within one task from repeated multi-model tasks. Fable same-session requires output from Fable and another recognized Claude family. These names do not imply a sequence or hand-off.

Turn off Show achievement notices (showAchievementNotices) to keep badges and unread records without floating notices; re-enabling does not replay them.

Proven new Claude/Codex Token tier crossings save unread first-record days even when dashboard updates are paused. Supported rules cover model-family volume, output, cache reads, Claude cache creation and cumulative volume. Imports and explicit rebuilds stay quiet. Proven cumulative API-reference crossings now have their own live receipts; other usage scopes still need their own origin proof.

Awards use a local envelope capped at 64 KiB. Old badges keep unknown dates; new saved-action badges retain unread first-record days even if a notice is missed. Saved awards survive restarts and are shared across windows. Awards under older rules retain their history without proving newer conditions. Clearing removes the award records without changing provider source logs. Badge notices appear only in the visible foreground panel, at most once per five minutes across windows/restarts. Display confirmation reads only that batch; missed notices stay unread. Hovering or keyboard focus pauses dismissal. Missed awards appear in a folded New awards summary with at most six exact levels per group. Only Mark this group read confirms that group after a successful local save; browsing, pinning and sharing do not. Other windows' later awards remain unread. Recently earned shows at most three new series; Recent records folds up to thirty saved levels and links to their rules. My badges supports up to six manually pinned themes, with pin icons on closed cards and a searchable selection editor with explicit ordering. Manual mode shows only chosen pins; automatic mode recommends at most three. Pins and order survive restart and never select sharing. Pins hidden by a disabled provider can be removed from the folded hidden-pin list without re-enabling that provider.

Explicit sharing can include retained earned badges from a disabled provider. The gallery and pin choices follow enabled providers; retained badges do not supply current progress or authorize new awards. Dates mean first recorded, not inferred attainment; historical dates remain unknown. Browsing does not mark records read or select sharing.

Imported plugin-experience records stay quiet. A newly saved action can unlock its badge even with the dashboard closed; it never opens or refreshes a hidden panel.

Flat agent logs without a path-proven parent do not count toward session-based achievements. Their Token totals, pricing, daily aggregates and visible session rows stay unchanged; nested-agent ownership and saved awards are retained.

The 2.5.0 icon-first gallery has 50 display themes and 178 award conditions, using 52 artwork groups. Fourteen-day and first shared-day milestones live inside their growing themes, with distinct award and sharing choices. The gallery has bronze, silver and gold tiers, plus platinum/legendary sustained-use and weekly-money ranks. Fifteen sustained Claude/activity series add Platinum and Legendary ranks without changing their first three targets. These use existing verified counters and artwork; concentrated three-day weeks can grow through activity days, output and sessions without requiring Codex or higher spending. Consecutive-day tiers are not extended. One card shows each series and its compact next target; expand it for every level's rule and evidence. Series and unlocked levels are counted separately. Local All / Next level / Earned filters survive reload; upgrading a tier keeps its disclosure open. It records usage experience, not skill, productivity, billing or account entitlement. Each series has distinct artwork; tiers share that emblem and use level colors. Achievement numbers use integers. Rules and data caveats are available on demand. Pending or outdated evidence is labelled, not shown as zero progress, and excluded from Next level filtering; earned badges remain saved. Closed cards show current / next-level progress. Long Token ladders include intermediate stages, without extra awards, notices or sharing choices. Small / Medium / Large sizes change badges and card frames while preserving text size and click areas. Next goals highlights at most three distinct themes before the complete collection. Goals require positive, current evidence; unproven goals and unused models are not promoted. Selection stays fixed during a visit; opening a badge focuses its existing rule without reading logs or selecting it for sharing.

• Token-led progression: Claude Sonnet/Opus/Haiku/Fable and Codex Astra/Sol/Luna, output, cache, model/effort exploration, agents, active days and repeat experiences. Cumulative processed tokens share one series: 100M → 1B → 10B → 100B → 1T.
• Peaks and sustained use: a calendar-week peak of 500M / 2B / 10B tokens; the highest average across four completed calendar weeks has 500M / 2B / 5B tiers. Weeks start Monday in the configured timezone; incomplete coverage cannot prove zero use.
• Small local discoveries: deliberately changing safe settings, previewing/exporting a share card, and saving local advice feedback. Credentials, network consent and public GitHub publication are never achievement requirements.
• API-equivalent milestones: 24 conditions in six separate Claude/Codex themes cover cumulative priced subtotals, completed Monday-start weeks in the current data timezone, and one Claude session or one independently verified Codex task tree. Both weekly series have five ranks: Bronze $100, Silver $500, Gold $2,500, Platinum $10,000 and Legendary $25,000. Only the last two per provider are new; old IDs/thresholds, completed-week requirements and frozen-price provenance remain unchanged. These use fixed USD reference prices (api-reference-2026-10-{backend}-v1, dated 2026-10-09), exact known models and complete component rates, including Claude's known 5-minute/1-hour cache-write TTL. Codex counts uncached input, cached input and output once. Unknown models stay unpriced; a verified priced subset can prove a lower bound (≥), not full coverage. This is neither a bill, quality score nor income. History is recognized quietly. Only cumulative money milestones now save unread first-record days from verified new appends under the same reference book; completed-week and single-session/task awards stay quiet with unknown dates. Token and money receipts remain independent and use one existing notice batch; failed saves never show optimistic dates.
• Recalibrated experiences: model, team, effort and rainbow rules combine output with repeated use and usage scale. Duplicate aliases, input-only activity and unverified task structures cannot substitute for the joint evidence.
• Codex effort details: expand the existing effort-breadth card to see low, medium, high, xhigh, max and ultra rows with processed/output Tokens, qualifying root tasks and progress toward its existing next tier. Per-row tier criteria are processed 100M / 500M / 2B, output 100K / 500K / 2M and roots 5 / 10 / 30, jointly required. Only each root × effort population's own positive output counts; approval-reviewer usage is excluded. An informational USD subtotal and coverage use api-reference-2026-10-openai-v1 (2026-10-09) and do not qualify effort awards. Ambiguous model × effort attribution is not divided proportionally; partial money shows a lower bound, and invalid/overflow money is withheld. Empty and unpriced states are explicit. Incomplete evidence shows unknown, not six zero rows. This adds no rules, targets, navigation or preferences; Claude effort attribution is not currently available.
• On by default, easy to stop: turn off Local achievements in ⚙ Settings (enableAchievements). On upgrade, with release announcements enabled, one non-modal reminder offers View achievements and Turn achievements off. Existing opt-outs remain off; historical unlocks stay quiet. Later unlocks use a brief in-dashboard notice, without sound or focus theft. Night/24–48-hour badges stay silent.
• Preview-first sharing: explicitly select up to nine earned badges, preview, then export that exact SVG. Old responses cannot enable export for a new selection; failures offer retry guidance. A return action leads back to the gallery, and its settings shortcut focuses the opt-out without changing it. No auto-export, sign-in, profile lookup or background upload.

Evaluation reuses verified numeric aggregates; opening cards never reads logs. After complete current indexing, independently verified Codex lineages support Token/agent milestones, excluding conflicting branches and descendants.

With current local history verified and fully indexed, trusted partial dates can show ≥ for activity, model-family and weekly Token progress. Weekly money bounds use only verified selected files whose usage is fully dated; pricing coverage is independent of date coverage. Partial dates do not establish complete history or verified zero, and four-week means/retained peaks remain pending until full coverage. Cumulative/task scopes are unchanged, with no extra source reads, storage or network calls.

An older Codex index may need a bounded, resumable pass for positive-output hour proof, without resetting its primary totals. Turning achievements off stops this extra historical work, not the ordinary usage tracker. A refresh failure revokes new-award authority while retaining the last displayed statistics.

Shared accounts and multiple sign-ins in one local history can contribute to these badges; they are not proof of one person's activity. Continuous output means a positive-output record in each UTC hour, not continuous human work. Clearing all extension-derived data removes the local award ledger, never provider logs. See local storage and clearing.

The 2.5.0 sharing heatmap offers Claude / Codex / Both and Tokens / API-equivalent amount controls. Tokens are the default; amounts require an explicit choice. Amount mode uses current daily pricing estimates, separate from achievements' fixed USD references: Claude is estimated, and Codex prices only exact known models. Missing pricing shows an unavailable amount or a priced lower bound with coverage. Both keeps two monetary panels; amounts are never summed. Hiding amounts switches the metric, colors, labels, tooltips, metadata and Markdown back to Tokens and retires the previous preview. Export writes the exact accepted SVG bytes, checking the document, controls, source, pricing and display revisions before and after the save dialog. Preview/export reads only bounded materialized daily aggregates, with no new log scan or network request. Selected-range Claude coverage uses memoized in-memory model/component rates: the generic unknown-model Sonnet fallback does not count as known coverage, missing rates stay unpriced, and an explicit zero rate is valid. Stored daily/model costs are not recomputed; family/runtime prices remain estimates. Hidden/Token mode performs no rate or money lookup.

Weekly quota pace (showWeeklyQuotaPace) is off by default and shared by both provider settings. When quota tracking is already enabled, it compares the original account-wide weekly observation with uniform progress from reset minus seven days: 29% used at 25% elapsed projects 116% for the cycle. There is no forecast in the first six hours, after thirty minutes without a fresh observation, or after reset. Credential invalidation withdraws it immediately, including while dashboard auto-refresh is paused. Counter drops, changed accounts/sources and unexpected resets suspend trust until a verified ordinary seven-day rollover. Current Codex local-log evidence cannot verify an account, so its forecast remains unavailable. The dashboard and status-bar tooltip share localized copy; compact bar text and custom quota formats are unchanged. This adds no requests, timers or stored history.

Current-cycle total and unused API-equivalent values remain low-confidence local-history estimates when account attribution is unclear. Accurate per-account separation is deferred to future multi-account support; these values are not account balances. This approximation is separate from the weekly quota-pace forecast.

The overview uses bounded horizontal cards: artwork, title, rank and progress stay together, with pins inside the badge card and a thin progress bar for next targets. A few badges no longer spread across the entire desktop width; narrow views wrap without shrinking text. Gold rank labels stay legible in light themes while retaining a gold accent.

Lightweight inline SVG icons accompany the existing localized share, settings, filter, size, showcase-editor and information labels; counts appear as compact chips. Above 960 px, the badge column fits its content up to two overview-card widths plus their internal gap; the next-target column fills the remaining space, with 16 px between columns. At 960 px or below they stack directly with a 12 px gap. Auxiliary history sections follow both.

Gallery size controls precede the overview: Small / Medium / Large use 56 / 80 / 112 px collection emblems and 56 / 72 / 96 px showcase emblems. Pin errors stay near their controls, and keyboard focus returns to a stable summary when a control disappears or becomes disabled.

The 2.3 series introduced local Codex tracking, provider-aware comparisons, request-level token attribution, and shareable activity views. It later added model/pricing updates, the project activity matrix, and month → day → hour drill-down. See CHANGELOG.md for patch-by-patch details.

Search for Claude Code Usage in the Extensions view (Ctrl+Shift+X), or:

ext install GrowthJack.claude-code-usage

Same extension is published at the Open VSX Registry: GrowthJack.claude-code-usage.

Ctrl+Shift+P → Extensions: Install from VSIX... → pick the downloaded .vsix.

Most settings live in the dashboard now. Open the dashboard (run Show Usage Details, or click the ⚙ in its header) and use the ⚙ Settings tab — grouped into General, Status bar, Data & refresh, and AI advice & Optimizer. Changes apply immediately.

To keep VS Code's own Settings UI uncluttered, only three ordinary settings stay there. The BYOK advice key is entered in the dashboard's ⚙ Settings tab and stored in SecretStorage; it does not travel with Settings Sync. Open VS Code Settings (Ctrl+,) and search for Claude Code Usage:

Everything else — refresh interval, status-bar items, number/date formatting, project grouping, content analysis, and all the AI advice / Optimizer options — is in the dashboard's ⚙ Settings tab. Upgrading keeps your existing values: a one-time migration copies them out of settings.json on first launch.

For a custom quota status-bar layout, choose Custom in ⚙ Settings. The template accepts {5h.pct}, {wk.pct} (or {7d.pct}), and {model:Fable.pct}; each window also supports .label and .reset. Reset styles include :decimal, :units, :clock, and :at, for example {5h.pct} | {wk.reset:at}. Missing windows and their separators are omitted. The built-in choice leaves the existing quota options unchanged.

The status-bar cost is Σ (tokens × per-million rate) across input, output, cache-write and cache-read, summed by model.

• Per-million rates come from the bundled pricing table, which is verified against the public Anthropic pricing page and supplemented with reference rates for non-Anthropic models that may appear in proxied setups.
• Refresh Model Pricing (command + button in the dashboard) pulls live prices from LiteLLM's public dataset as runtime overrides.
• Unknown model snapshots are priced against the current tier of their detected family (Opus / Sonnet / Haiku / GPT / Gemini / DeepSeek / Kimi / GLM / Qwen) instead of falling back blindly.

Claude transcript totals are counted by response identity, not by JSONL row. One response can produce both a thinking row and a text row carrying the same messageId, requestId, and complete usage vector. The extension keeps the largest vector for that response once; Claude Code's stats-cache adds the rows. The two totals can therefore differ.

What the status bar does not know:

• Your actual Anthropic invoice (discounts, free credits, plan caps).
• Whether your proxy provider charges different rates.
• Anything not recorded in your local .jsonl log files.

The 5h / weekly quota indicator is different — it queries Claude Code's real /usage endpoint via the OAuth session and shows the actual percentage Anthropic is tracking for your account. That number is authoritative.

The complete user-facing inventory, retention rules, clearing behavior, and remote boundaries are in Local data and privacy (简体中文).

A timezone switch can reuse a complete, proven same-source checkpoint read-only; only date/hour buckets migrate, without restarting primary history.

• All Claude token / cost / session analysis runs locally by reading your ~/.claude/projects/**/*.jsonl files.
• Codex usage records are discovered only from .jsonl and .jsonl.zst regular files under sessions and archived_sessions below your Codex home. Separately, the extension streams exactly $CODEX_HOME/session_index.jsonl for the id → thread_name mapping used by truthful thread titles. Absolute paths in those titles are redacted and the titles remain memory-only. Credentials, databases, and unknown files are not read. Usage-record JSONL lines are streamed and temporarily parsed only for allowlisted metadata; prompt, response, command, and tool-argument fields are not inspected or used for analysis and are never retained. Deterministic insights make no network request.
• The Codex persistent index stores machine-salted pseudonymous keys, numeric and structural aggregates, and sanitized project, directory, agent, model, effort, role, time, and quality metadata. It never stores raw IDs, full paths or repository URLs, thread titles, or conversation bodies.
• The quota indicator calls api.anthropic.com/api/oauth/usage using Claude Code's existing OAuth token. If that token has expired, the extension sends the existing refresh token to console.anthropic.com/v1/oauth/token and writes the refreshed credentials back to the selected Claude credential file or macOS Keychain item. See Local data and privacy.
• AI advice and the Usage Optimizer are the only features that call a model — and only after you preview and explicitly send a prepared request. Advice defaults to allowlisted aggregates; prompt samples and optional user context require separate consent. The Optimizer includes only the text you paste into it (never your files or terminal). Both use the exact previewed bytes, the endpoint in advice.apiUrl, and your own advice.apiKey. Bring your own key; no key or generative OAuth credential is shipped.

• A reset absent from an official response or local structured event cannot be reconstructed; day-only evidence lowers confidence.
• One Codex home may contain several sign-ins. The current period may therefore show a low-confidence blended estimate from the latest real observation; ambiguous completed periods remain used-only rather than inventing an account split.
• API-equivalent values depend on current known API prices and visible pricing coverage. They are not bills or subscription prices.
• Source-log retention belongs to Claude Code and Codex. Uninstall may leave host-managed extension storage behind, so the explicit clear controls are the reliable deletion route.

"No Claude Code Data"

• Make sure Claude Code is installed and you have used it at least once.
• Check the dataDirectory setting; auto-detection looks at ~/.claude/projects and ~/.config/claude/projects.

One-shot Claude CLI activity is missing

• Calls made with --no-session-persistence can leave a prompt-history entry but no project transcript and no token usage fields. The extension does not invent token or cost totals from prompt history. Run future audited calls without that flag if they should appear; past unpersisted token usage cannot be reconstructed locally.

Quota row shows 5h:--% wk:--%

• Claude Code's OAuth token is missing or expired. Log in to the active Claude profile once. Credentials follow explicit dataDirectory, then the first valid CLAUDE_CONFIG_DIR, then ~/.claude; the single global macOS Keychain item is never substituted for a selected custom profile.

Get AI Usage Advice returns 404

• Match the API format, address and model. For DeepSeek's OpenAI-compatible format use https://api.deepseek.com/chat/completions; its explicit Anthropic compatibility base is https://api.deepseek.com/anthropic (official documentation). The request preview shows the resolved destination; the extension does not silently switch providers or remove a configured proxy prefix.

Send this exact request is unavailable

• Enable the default-off advice-effectiveness setting, allow aggregate data, and configure your own claudeCodeUsage.advice.apiKey. Previewing is always local; without a key the request remains unsent.

High CPU or sluggish refresh on a large history (Linux included)

• V2.2.1 removes the hidden 8-second active polling override and bounds the first-timestamp scan. Until you install it, set Live refresh delay to Off, set Refresh interval to 300–900 seconds, and optionally turn Content analysis off. Turning Dashboard auto-refresh off by itself does not stop status-bar parsing.
• If V2.2.1 still runs hot, open Show Diagnostic Logs and attach only the anonymous refresh: lines to issue #70; they contain counts and timings, not prompts, paths, session IDs, credentials, or raw log lines.

Usage history disappears or is missing older months

• Claude Code automatically deletes conversation logs older than cleanupPeriodDays (default: 30 days). Once deleted, those records cannot be recovered. To retain more history, add this to your ~/.claude/settings.json:

  { "cleanupPeriodDays": 365 }

  This only affects logs kept from now on; already-deleted logs cannot be restored. Thanks to @nickearnshaw for documenting this (PR #21).

Token counts are lower than Claude Code's stats-cache

• A single response can be written as separate thinking and text transcript rows with the same messageId, requestId, and complete usage vector. The extension counts that response identity once and keeps its largest vector; Claude Code's stats-cache sums the rows. The extension does not apply a multiplier to make these different mechanisms agree.

Token counts appear lower than the model provider's own dashboard

• If you use Claude Code with a third-party proxy that routes requests through sub-agents or background workflows (e.g. ultracode / dynamic workflows), each agent writes its own .jsonl log file inside a sub-directory. The extension reads all these files, but some proxy configurations may not write agent-level records at all. Until native workflow attribution is added in a future release, the total shown here may be lower than the provider's upstream count. Your actual spend is always on your provider's billing page.

Created by @jack21 and maintained by @Carl723000, with community PRs, issue reports, reviews, and translations. Thank you to everyone who helped shape the project. The contributor credits below are reconciled for v2.5.0; a linked PR is not a claim that every proposed line shipped. See CONTRIBUTING.md for the release-credit convention.

Development assistance: Claude Code and OpenAI Codex are credited separately from human contributors. The project is MIT-licensed.

See CHANGELOG.md and the versioned GitHub Releases for detailed change history and release-specific attribution.

Issues and pull requests are welcome on the GitHub repository.

MIT