UsageMeterv1.3.2

• Novo

  Full compatibility pass for claude.ai's 2026 "dframe" redesign (the new app shell with the Chat ⇄ Cowork switcher) — every on-page module was re-anchored and re-verified against the new DOM

• Correção

  The usage stats panel opened far to the right of the chip instead of under the cursor — the redesigned shell creates containing blocks for position:fixed elements, so a tooltip nested inside the chip resolved its viewport coordinates against the wrong ancestor. It now portals to the page body (the same cure the /code chip already used), with viewport clamping so it never spills off the right edge, an arrow that keeps pointing at the chip, and a flip-below mode when the chip sits in the top half of the screen

• Correção

  Switching Chat ⇄ Cowork froze the page — UsageMeter's bar re-anchoring ran synchronously on every DOM mutation during the switch's render storm: each pass forced a reflow, scanned the whole document, and could move the bar mid-render, fighting React move-for-move. It now coalesces to one pass per animation frame, throttles full scans to one per 150ms, and only relocates the bar when it is genuinely misplaced

• Correção

  On the home page the "Project / Skip permissions" strip tucks under the composer with a negative margin; in Chat mode it collapses to zero height but KEEPS that margin, which pulled the UsageMeter bar up into the composer box. The bar now measures the strip's real layout contribution and pads its own margin dynamically — correct in both modes, glides along Anthropic's 280ms collapse animation, and re-settles after the transition ends

• Melhorado

  Export survives the redesign's virtualized conversation feed — only a screenful of messages is mounted in the DOM now, so exports of long chats would silently truncate to the visible window. When UsageMeter detects an incomplete window (via the feed's aria-setsize turn count) it fetches the full transcript from Claude's own conversation API instead, keeping thinking blocks and still extracting fenced code blocks as files, with the DOM scrape as fallback

• Melhorado

  More redesign adaptations: in-row chips (non-ring mode) now sit BEFORE the model selector wrapper (the new composer slides the wrapper sideways over anything placed after it while the send button is hidden); the disclaimer bar is found via the redesign's semantic role="note" hooks; the accounts button re-anchored into the new sidebar footer next to the user menu; auto-reply gained extra send-button label variants plus a synthetic-Enter fallback

• Novo

  Notification system rebuilt from scratch: ONE notification instead of eleven. UsageMeter now only notifies when your 5h session reaches 100% — Pro plans only, at most once per cycle, and clicking it jumps back to claude.ai. A single toggle in Settings → Notifications turns it off

• Melhorado

  Removed with the old notification pack: 80/90/95% threshold alerts, high-usage-rate warnings, reset reminders and "reset completed", the Monday weekly recap, x2 promo start/end alerts, streak milestones, "Claude finished replying", the SmartQueue sent notification and the whole in-page notification manager content script — one less script injected into every page, less background work, zero notification spam

• Novo

  The on-page usage panel got a full visual refresh — frosted editorial card (backdrop blur, layered shadows, 14px radius), a soft pop-in animation (honoring reduced-motion), refined progress bars, and a plan chip in the header. Free accounts get one discreet "Upgrade to UsageMeter Pro" row at the bottom of the panel; Pro accounts get their PRO chip and nothing else. Bonus fix: the flipped-arrow color used a raw HSL triple as a color, so the arrow silently lost its fill when the panel rendered below the chip

• Novo

  The Pro page was redesigned to actually sell: editorial ink hero with the live "users online" count (real heartbeat data, hidden when unknown), six concrete feature rows that finally use their accent colors, and an honest Free vs Pro comparison table where every line maps to a real gate in the code (3→10 accounts, 50-message md/txt exports → unlimited + PDF/DOCX/JSON/HTML, 24h & 7d insights, the 100% session alert, priority support). The Pro hint under the notification toggle is now a real door to that page instead of a dead-end sentence

• Correção

  The "upgrade to see the chart" button on usage detail pages did nothing — the component emitted show-upgrade but nothing listened. It is now wired to the Pro page, and the six identical bland lock overlays were replaced by one shared teaser (crown, "Deep usage insights" pitch over the blurred real chart, trial CTA) that actually sells the unlock

• Melhorado

  Every UI string now exists in all 34 languages — the redesign/Pro/notification additions were translated into the 32 remaining locales, and a full audit filled every historical gap (league tier names & descriptions, model breakdown labels, import/export strings, the Esperanto and Klingon stragglers). Missing-key count across all locales: zero

• Correção

  Strict Pro gating for the quota-full notification: content.js now mirrors the VERIFIED license verdict to storage and the background checks that mirror (or the signed-in plan) — a leftover cached license key is no longer treated as proof of Pro, so free accounts can never receive the notification. The setup wizard shows the toggle with a PRO pill but keeps it locked on free

• Novo

  Usage engine hardening (phase 0): the /usage parser now understands Claude's new limits[] schema COMPLETELY — session, weekly, and every per-model weekly limit (Opus, Sonnet, Fable) are mapped onto the fields all displays read, so the day Claude removes the legacy payload nothing goes blank. One shared implementation (core/usage-normalize.js) replaces the two divergent copies that background and content each carried

• Novo

  The per-model weekly limit (e.g. "Fable (7d)") is now a selectable bucket everywhere the 5h/Sonnet/Weekly choices exist: on-page chips per page, the toolbar badge, Settings → Usage Indicator (labelled with the live model name) and the setup wizard

• Correção

  Send-button detection no longer depends on the UI language — auto-reply, SmartQueue and scheduled sends matched the button's localized aria-label ("Envoyer"/"Send"), which silently broke in the other ~40 locales. They now resolve it structurally from the stable chat-input anchor (label match kept as a fast path, ProseMirror Enter dispatch as the last resort)

• Novo

  Conversation length in tokens, computed 100% locally — a vendored o200k tokenizer runs in the background worker (never on the page), reads the same conversation-tree endpoint the exporter already uses, walks the CURRENT branch only, skips ephemeral thinking blocks, and estimates images (capped) and PDFs from their metadata. Shown as "≈N tok" next to the composer, red past 50k, marked ≈ when tool calls make the true weight unknowable. Zero external API calls, works offline

• Novo

  "Messages remaining" the honest way: instead of the invented quota caps other trackers hardcode, UsageMeter measures the REAL % jump each of your messages causes on each limit (snapshot before send, /usage re-check after the reply), keeps a sliding window of 20 observations per limit, and extrapolates from the median of the cheapest limiting factor. Shows an em dash until it has 5 real observations — no false precision — then "· M msg" beside the token count, red under 15. Auto-calibrated to your plan, your models, your usage

• Melhorado

  The new metrics chip is mounted through an idempotent anchor system (compare-then-insert, state kept in JS, 750ms repaint tick) — it survives claude.ai's SPA re-renders and works in any UI language, and never touches the DOM when nothing changed

• Melhorado

  Conversation metrics polished and made configurable: a dedicated toggle in Settings → Features (and the setup tour) shows/hides the chip live; the chip itself got the house chip styling (hover, hairline separator, tabular numerals) with graduated amber→red states (30k/50k tokens, 30/15 messages). The usage panel now shows "≈M msg" remaining PER window (5h / weekly / per-model) once each has enough observations, plus a Conversation section with the measured token length and the ≈context headroom left before the 200k window fills

• Novo

  Opt-in "Approximate usage": token counts on your quota windows (≈used / ≈cap under each bar, ≈tokens left on the binding limit in the composer chip) derived EXCLUSIVELY from your own measured messages — Δ% per limit vs the message's token cost, cap estimated on the 15th-percentile FLOOR of the ratio cloud so parallel Claude Code / other-device consumption (which can only inflate deltas) gets rejected instead of skewing the number. No hardcoded caps, no invented "credits", no per-model weight tables — the exact percentage always stays the primary figure

• Melhorado

  The estimator plays it safe by design: a cross-tab single-flight lock discards any measurement taken while another claude.ai tab was answering (self-healing against crashed tabs), attribution windows wider than 8s are dropped, model switches between snapshots invalidate the per-model sample, each model keeps its own estimate, and nothing shows before 8 clean observations per window — the line stays dimmed until 25. Off by default; when off, zero token numbers appear anywhere

Todas as versões