2.5.0 — 成就系统与压缩日志兼容

中文速览

• 成就系统：支持可核验达成日期、手动校准、一键展开／收起、固定徽章及更紧凑清晰的详情。— @Carl723000 (#131)
• 恢复压缩日志统计：兼容 Codex .jsonl.zst 会话及归档，恢复因压缩而漏掉的历史用量并避免重复计数；不修改原日志、不保存解压副本。读取失败时保留已验证结果并明确提示。— @Carl723000 (#131)
• 分享与对比：增加供应商／统计项热力图，优化徽章选择、修复图标，并在成就分享卡中加入公开仓库链接；另有默认关闭的周额度进度功能。— @Carl723000 (#131)
• 稳定性：修复大量 Claude 历史记录导致的参数栈溢出，改善成就证据恢复；栈溢出复现由 @mkgaskin-ops 在 #126 提供，不代表所有 Windows CPU／内存问题均已解决。— @Carl723000 (#131)，报告者 @mkgaskin-ops (#126)

API 等价不是账单或官方余额；成就日期以保留证据中最早可核验日期为准。压缩日志需要原生 Node Zstd 支持。

What's new

• Local achievements: verifiable unlock dates, manual recalibration, pinned badges, expand/collapse controls and clearer, compact details. — @Carl723000 (#131)
• Compressed Codex history: recover usage omitted when session or archive logs were compressed to .jsonl.zst, without double-counting. Original logs remain untouched and no decompressed copy is saved. Failed reads preserve verified results with an explicit status. — @Carl723000 (#131)
• Share & compare: provider/metric heatmaps, clearer badge selection, repaired badge thumbnails and a public repository link in exported achievement cards. — @Carl723000 (#131)
• Quota pace: an optional, default-off weekly indicator based on existing local quota observations. — @Carl723000 (#131)
• Reliability: prevent argument-stack overflow on large Claude histories and improve achievement evidence recovery. The overflow reproduction was reported by @mkgaskin-ops in #126; this does not claim to resolve every Windows CPU or memory issue. — @Carl723000 (#131), reported by @mkgaskin-ops (#126)

Notes

• API-equivalent figures are estimates, not bills, subscription charges or official balances. Achievement dates show the earliest date verifiable from retained evidence; missing history cannot establish an exact original unlock date.
• Compressed logs require native Node.js Zstd support (Node 22.15+ or 23.8+). Older extension hosts retain verified results and report the compatibility limitation.

Contributors: @Carl723000; issue report by @mkgaskin-ops.

Full changelog: v2.4.1...v2.5.0

🤖 Generated with OpenAI Codex

v2.4.1 — Performance and stability patch

中文速览

• 控制日志与内存压力：未知模型诊断去重并设置全生命周期上限，避免每条计价记录重复输出警告；补充 Opus 5.5、Sonnet 5.5、GPT-6.1 Sol、GPT-6 Sol 和 GPT-6 Luna 的内置定价。未知 Codex 模型仍保持未定价，不猜测价格。修补来自 #120、#124，感谢 @rsyuzyov @Carl723000；感谢 @jordanvalnet 在 #122 提供崩溃证据。
• 减少 Claude 增量计算：保留已过期窗口的空贡献；保持内容分析的精确滚动截止时间；新增会话时保留旧文件的相对发现排序，避免把之后所有文件的统计反复删去／重加。(#119、#121、#123) @rsyuzyov
• 降低刷新打扰：完整 Codex 历史在常规追加日志时不再反复显示“100% 补齐历史”；失败保留已验证数据，区分实际迁移、重试与暂停。滚动、展开状态及预览不因无变化轮询反复重置。(#124、#125) @Carl723000
• 保护增量索引：不同日志来源、本地盐和时区使用独立缓存；只读沿用可证明归属的旧检查点。新时区可保留同源已验证主历史，只迁移日期／小时；重建与清除使用精确目标，不修改或删除源日志。(#125) @Carl723000
• 减少操作摩擦：统一自动刷新暂停、AI 请求目标与协议校验、分享预览／导出一致性及目录切换恢复；保留既有 BYOK 密钥与协议兼容行为。(#124) @Carl723000

本版专注性能和稳定性，成就系统不进入本版。Token、缓存、API 等效成本及最后观测额度的含义保持不变；API 等效成本不是账单。修补降低已定位风险，不代表所有环境中的高 CPU 或崩溃都已被复现或排除。感谢 @mkgaskin-ops、@rsyuzyov、@jackyyangjq 与 @jordanvalnet 在 #99、#122 提供报告与性能证据。

Fixes — contribution details

• Deduplicate unknown-model diagnostics instead of warning for every priced record. (#120) @rsyuzyov
• Preserve empty contributions for files already outside the content-analysis window. (#119) @rsyuzyov
• Share one exact content-analysis cutoff across parsing and aggregation; preserve the exact rolling window rather than switch to hourly batches. (#121) @rsyuzyov
• Keep stable discovery ranks when a transcript appears, avoiding re-tagging every later file while preserving full-loader tie-breaks; fall back safely on reorder or precision exhaustion. (#123) @rsyuzyov
• Bound diagnostics, avoid repeated large-history work, cover newer known models, and fix refresh/recovery, AI-destination and share-export friction. (#124) @Carl723000
• Keep completed history quiet; isolate persisted Codex indexes by source/salt/timezone; reuse only proven read-only checkpoints, resume regressed coverage, and preserve primary history during targeted timezone migration. (#125) @Carl723000

These address confirmed paths, not all CPU or renderer OOM scenarios. Native Windows Section-handle/commit-memory validation remains user-side; deletion-triggered analysis rebuilding and some derived-row work remain optimization opportunities. Unknown-model fallback is not promoted to exact Codex pricing coverage.

Documentation

• Credit PR/issue contributors, make release history collapsible, and improve README discoverability. (#117) @Carl723000
• Keep contribution summaries visible and lead previews with usage views rather than sharing. (#118) @Carl723000
• Synchronize all nine README editions and bilingual architecture/local-data contracts with cache isolation. (#125) @Carl723000

Validation / 验证说明

• Final local regression: 1,272 Node tests passed, zero failed; one exact-mtime identity fixture skipped because of local filesystem precision. 195 UI tests passed; verified 2.4.1 VSIX packaging passed.
• #123 integration additionally verified batched head/middle/tail insertions, 70 consecutive same-gap insertions with an actual precision-exhaustion fallback, append/reorder/move/delete, and previous-snapshot immutability. Every checked step matched full-loader records and content analysis.
• 维护者确认 R3 真实本地安装通过；R4 的缺失目标检查点时区迁移／读取边界及 R5 的 #123 整合通过自动回归，本轮未另做原生 R4／R5／F5 验收。Earlier Claude review covers the progress-display portion; the maintainer explicitly waived extra cache/timezone review and this turn's #123 review. No unperformed review is claimed.
• Original PR CI and final merged-main CI passed Node/UI/Release Package gates. The draft targets verified main commit 81e204f956d58fe9a6c1a238204f71bf1b811db8, whose source tree matches the tested R5 integration. Publishing the reviewed draft triggers automatic GitHub VSIX, Marketplace and Open VSX delivery; no manual Marketplace upload is required.

Code contributors: @rsyuzyov and @Carl723000

Full changelog: v2.4.0…v2.4.1

🤖 Generated with OpenAI Codex

What's new in v2.4.0

• Codex is now a first-class provider. Claude and Codex share navigation, accessible chart controls, and supported month → day → hour drill-downs. Codex retains its own token and cache definitions: API-equivalent prices are estimates, not bills, and its weekly status-bar percentage shows remaining allowance while today's token item shows processed usage.
• One sharing workspace. Preview and export the combined activity heatmap, Claude Share Card, or Claude token heatmap in one place. It is on by default and can be turned off; existing commands still open their matching preview. Local SVG/Markdown exports require no GitHub access. Publishing a Claude heatmap to GitHub remains a separate, explicitly confirmed action.
• Smoother large-history use. Incremental Claude indexing now avoids whole-corpus rereads for ordinary multi-file appends, new transcripts, and safe rolling-window changes. Date-formatting caches and bounded retained text reduce avoidable CPU and memory work. Live dashboard updates coalesce during scrolling; opening a chart drill-down does not reread source logs.
• A more complete, easier-to-use interface. The status-bar quota format has simple presets with an optional custom template. All eight UI languages now have README editions; German and Brazilian Portuguese are new concise editions. The PR first-pass bot no longer posts a boilerplate reply when its analysis request fails.

Other reliability work

• Marketplace delivery recovery uses the verified release asset, so a transient registry failure can be retried without rebuilding a different package (#107).
• The latest incremental-index and date-label fixes were contributed through #109, #110, #111, #112, #113, #114, and #115. The provider and sharing updates are in #116.

Important: Codex allowance observations come from local logs and may be stale or incomplete, especially on shared accounts. These improvements do not establish that the reported Windows high-CPU case #99 is fully resolved.

Contributors: @Alfiefe10, @Carl723000, @rsyuzyov, and @jackieyangjq

Full changelog: v2.3.3...v2.4.0

🤖 Generated with OpenAI Codex

What's changed

• Usage stays available with a legacy workspace advice key. A key that cannot safely move into global SecretStorage no longer prevents the status bar, commands, or dashboard from starting. The old key remains untouched; AI advice stays unconfigured until you remove the legacy workspace setting and enter the key in Dashboard Settings → Advice. Fixes #105 via #106.
• Smoother, more reliable refreshes. Live dashboard updates preserve local scroll positions, focus, selection, and drill-down state. Claude's Today and rolling 30-day views advance at the configured-timezone day boundary without rereading unchanged logs. Codex refreshes coalesce overlapping triggers and avoid restarting an index during clear or rebuild. Claude content analysis keeps full-scan attribution semantics while bounding unchanged and ordinary append refreshes. #106.
• More resilient release delivery. The publish workflow attaches the verified VSIX to the GitHub Release before attempting VS Code Marketplace and Open VSX independently. It supports bounded, duplicate-safe retries for either registry. #104.

Contributors: @Carl723000
Full changelog: v2.3.2...v2.3.3

🤖 Generated with OpenAI Codex

What's changed

✨ Features

• feat: bring project activity matrix into v2.3.2 (#102) @Carl723000

🐛 Fixes

• fix: stabilize v2.3.2 dashboard interactions (#100) @Carl723000

📝 Docs / CI / chore

• docs: keep README history at minor level (#101) @Carl723000

Contributors: @Carl723000
Full changelog: v2.3.1...v2.3.2

What's changed

v2.3.1 is the release-quality pass for the new Codex Beta and Compare surfaces. It restores Claude's recent usage views, aligns shared interactions without pretending the providers expose identical metrics, and tightens the upgrade and privacy boundaries found during independent review.

Usage and interaction

• Claude Today and Last 30 days again use real configured-timezone aggregates; Today, Yesterday, rolling ranges, charts, and tables reconcile.
• Codex's token status item shows today's used tokens, while its weekly quota item shows remaining allowance. The hover card reuses Claude's progress bars, reset times, and line-broken detail layout.
• Codex day bars expand to materialized hourly detail with no JSONL reads on click, and expanded state survives reload.
• Weekly API-equivalent allowance charts can estimate total and unused durability from real quota observations. Period tables start collapsed; approximate or cross-account evidence stays visibly low confidence.
• The Compare share workspace is on by default but can be disabled. Its controls sit below the preview and support quantile, logarithmic, and linear intensity scales plus curated/custom colors.
• Codex token composition names uncached usage explicitly without double-counting cached input or reasoning.
• Continuous Codex scrolling no longer serializes state every frame, and live indexing progress patches text without rebuilding the page.
• Codex per-model green summaries now mean exact-model API-equivalent value, matching Claude's monetary hierarchy; unknown models remain unpriced and effort stays token-native.

Models and pricing

• Added exact gpt-6-astra support with a 1.05M context window and official Standard short-context pricing.
• Added exact claude-fable-5-1 and claude-mythos-5-1 support with 1M context and the new Fable 5.1 cache-read rate, without rewriting historical Fable 5 usage.
• Added the opt-in AWS Bedrock in-region Claude pricing backend from #95. Switching backend reprices unchanged local logs immediately. Sonnet 5 uses the standard rate after its launch promotion ended on 31 August 2026.

Correctness and safety

• Existing v2.3.0 Codex indexes perform one bounded, resumable token-semantics rebuild after upgrade.
• Account/profile changes can no longer lose the account-ambiguous confidence marker at a reset boundary.
• OpenAI-compatible Optimizer requests send reasoning_effort without the unsupported top-level thinking field, resolving #94.
• AI advice and Optimizer remain default-off, BYOK-only, exact-preview-first, and separately sent. Revoking consent invalidates previews and cancels matching in-flight requests immediately.
• Detailed local-data and destructive-clear documentation remains in the repository; the plugin Settings UI stays concise.
• Dollar figures and weekly allowance values are API-equivalent estimates, not bills, cash balances, or claims of provider equivalence.

Verification

• Required PR checks passed: Node Tests, UI Test, and Release Package.
• Local full gate: 940 Node tests passed, 0 failed, 1 filesystem-capability skip; Playwright 129/129 passed.
• A separately stamped v2.3.1 VSIX passed repository policy and ZIP integrity checks, then installed in a fresh isolated VS Code profile as growthjack.claude-code-usage@2.3.1.

Contributors: @Carl723000 for the v2.3.1 implementation, integration, and release preparation; @akapti for AWS Bedrock pricing (#95); and @jack21 for the Codex tool-adaptation documentation (#96).

Full changelog: v2.3.0...v2.3.1

🤖 Generated with OpenAI Codex

What's changed

v2.3.0 adds Codex Beta to the shared usage dashboard and substantially improves Claude refresh performance and profile correctness.

Codex Beta

• Adds Codex usage views for Today, the last 30 days, all time, sessions, projects, and local optimization guidance.
• Builds a privacy-safe local index from Codex session logs without reading credentials or storing raw prompts, responses, paths, or account identifiers.
• Uses exact request token snapshots when available, with lineage-aware reconciliation for copied rollouts, resets, and duplicate evidence.
• Shows API-equivalent cost estimates, cache hit rate, hourly/daily/monthly cost trends, token composition, and conservative weekly allowance-value estimates.
• Combines usage found across multiple sign-ins in one Codex home while keeping quota observations separate and last-observed rather than summing them.
• Accelerates large first-time backfills with bounded local parallelism, resumable checkpoints, and verified partial totals while indexing is still in progress.

Claude improvements

• Fixes the full-corpus refresh problem reported in #87. Claude now keeps an exact in-memory per-file index, reads only changed tails during normal appends, and updates only affected contributions for truncate, replace, move, and delete events.
• Fixes #89 by resolving quota credentials from the active Claude profile. Explicit custom profiles no longer silently fall back to the default macOS Keychain identity.
• Locks transcript response-identity deduplication with regression coverage for thinking/text pairs, monotonic snapshots, cross-file clones, distinct request IDs, and mixed request-ID presence (#92).
• Includes the official quota thresholds and improved usage-credit tooltip from #90, plus weekly and model-specific quota views from #86.

Dashboard and reliability

• Aligns Claude and Codex within one responsive dashboard while preserving provider-specific metric semantics.
• Preserves expanded rows, sorting, chart selections, destinations, and scroll position across webview reloads.
• Improves active-tab readability, chart colors, responsive tables, and horizontal access to dense 30-day views.
• Keeps first-pass issue automation responsive (#88).

Important notes

• Codex API-equivalent costs and allowance values are estimates derived from locally recorded tokens and current public API prices. They are not bills, subscription charges, or official quota values.
• Unknown Codex models remain unpriced, and the dashboard reports pricing coverage instead of manufacturing a zero-cost value.
• Reasoning tokens are already included in output tokens and are never charged twice.
• Codex remains a beta provider. Branches, Workflows, Cache Warmth, and content attribution remain Claude-only for now.

Validation

• Node release suite: 614 tests — 613 passed, 0 failed, 1 capability-gated skip.
• UI suite: 63/63 passed.
• Release-package policy, ZIP integrity, and isolated VS Code installation checks passed.

Contributors: @Carl723000 and @zeyutang

Full changelog: v2.2.2...v2.3.0

What's changed

🐛 Fixes

• fix: reduce background refresh energy use for V2.2.2 (#83) @Carl723000
• fix: one unprobeable .jsonl no longer blanks the whole usage dashboard (#79) @ptweezy
• fix: track Claude Opus 5 as a 1M-context model (#81) @e7d

📝 Docs / CI / chore

• docs: finalize v2.2.2 release notes (#85) @Carl723000

Contributors: @Carl723000, @e7d and @ptweezy
Full changelog: v2.2.1...v2.2.2

v2.2.1 — refresh stability and repository hardening

Added

• Adds an all-sessions view with persisted time-range, project, and model filters (#73).
• Adds decimal, whole-unit, and local clock/date formats for quota reset countdowns (#75, closes #74).
• Adds Bahasa Indonesia as the eighth complete UI locale, including a dedicated Indonesian README (#76).
• Adds WIB, WITA, and WIT presets to the timezone picker (#77).

Fixed

• Mitigates #70 by removing the hidden 8-second active polling override. Polling now always honors the configured 30–3600 second interval, while file watching offers quiet delays up to 300 seconds.
• Reuses one JSONL manifest per refresh, bounds first-timestamp probes to eight concurrent readers, and coalesces watcher bursts into one strongest follow-up.
• Preserves the last successful snapshot when a scan or read is incomplete, and clears the initial loading state with a localized retry/diagnostic message on cold-start failure.
• Corrects first-of-month daily labels and negative-UTC monthly labels (#54 / #71; thanks @YuboZhang).
• Defaults automatic first-pass replies to English, using English-first bilingual output only for Chinese authors (#72).

Diagnostics and maintenance

• Adds anonymous refresh: diagnostics with file/change/reuse counts, bytes, lines, watcher events, and stage timings. They never contain prompts, paths, session IDs, credentials, or raw logs.
• Establishes AGENTS.md as the canonical Codex repository policy, explicitly preserves contributor attribution, and credits Claude Code and OpenAI Codex as development tools.
• Hardens the comment-only GitHub first-pass automation with truthful provider attribution and bounded base-repository reads.
• Makes manual publish retries tag-safe and independently selectable per registry.

Merged pull requests

• #77 by @projectronic — Indonesian timezone presets
• #76 by @projectronic — Bahasa Indonesia UI locale
• #75 by @projectronic — reset countdown format options
• #73 by @Carl723000 — all-sessions view and filters
• #78 by @Carl723000 — v2.2.1 refresh stability and repository hardening
• #72 by @Carl723000 — automatic first-pass replies default to English
• #71 by @YuboZhang — daily and monthly date-label fixes

Contributors: @Carl723000, @YuboZhang, and @projectronic

Issue #70 remains open until the Linux reporter confirms the result. Persistent per-file indexing and the Codex usage-provider adapter are intentionally deferred.

Full changelog: v2.2.0...v2.2.1

🤖 Generated with OpenAI Codex

Claude Code Usage 2.2

A feature release — new ways to see, share, and understand your Claude Code usage, all computed locally.

✨ New

• Usage share card — a themed, configurable one-page SVG of your usage (pick a range × scope × which metrics to show). Chinese locales use 万 / 亿 units.
• Read-only conversation viewer (Sessions tab, on by default) — re-read a past session's prompts and answers to jog your memory without loading it back into model context (unlike resume). Markdown-rendered answers; thinking and tool traffic tucked behind toggles; opens on the last rounds.
• Token heatmap (opt-in) — a GitHub-style yearly token heatmap on the All tab, plus an Export / Publish to your GitHub profile SVG with a one-click Markdown embed.
• Experimental insights (opt-in, Content tab) — estimates from your local logs: a cache-churn bill, cache warmth by model, big one-shot turns, your active hours, and skill ROI.

📊 Sessions & data

• "Active" column — estimated hands-on time (idle-capped), distinct from the raw first→last span.
• Live-refresh delay — a delay control (Off / 1 / 2 / 5 / 10 / 20 / 30s) replaces the on/off toggle. This only re-reads your local log files; the /usage quota fetch is throttled separately.
• Timezone-aware bucketing — Today / day / month / hour totals all key off your configured zone (full UTC-offset dropdown), so they agree with each other and the Anthropic console.

🔒 Privacy & defaults

• Read-only features default on; anything that acts on your Claude Code (resume / delete, now one setting) is opt-in, default off.
• No prompt text ever leaves your machine.

🐛 Fixes & housekeeping

• Multi-window status bar — with several windows open, background windows now
  refresh on focus instead of going stale. (#55)
• Sonnet 5 now correctly reports a 1M context window. (#50, @UfukTanriverdi8)
• Cache-write cost by TTL — a 1-hour cache write is priced at 2× base input (vs the 5-minute 1.25×) when the log carries the split. (#62, @zeyutang)
• Dropped the unused @types/glob devDependency, clearing an advisory. (#63, @zeyutang)
• New @ccu-bot first-pass CI automation (inert unless enabled).
• 7-language i18n for the new surfaces; 114 unit tests.

Contributors: @Carl723000, @zeyutang, @UfukTanriverdi8
Full changelog: v2.1.1...v2.2.0