v1.6.0

Fixed

• The weekly limit no longer shows a reset that hasn't happened. When a limits
  reading was older than the window it described, the extension assumed the window
  had rolled over and reported it as 0% used, with a fresh countdown — visually
  identical to a live reading. Since Claude Code stopped refreshing
  ~/.claude/usage-cache.json, that stale file could be months old and still be
  presented as current. A window is now only reported empty when the reset and the
  reading are both recent; otherwise it is hidden rather than guessed at.
• Context % no longer falls as your context grows. The window size was inferred
  from the token count itself, so a conversation at 190k tokens showed 95% and
  one message later, at 205k, showed 20%. The window now comes from the model:
  1M on Sonnet 4.6 / Opus 4.6 and later, 200K on Haiku and every earlier model.
• Context % now describes your conversation, not its subagents. Subagent turns
  carry the parent session's details and are usually the most recent entry, so the
  headline jumped by tens of points every time a subagent ran and back again when it
  finished. Their usage still counts everywhere costs are reported.
• The 5-hour pace meter no longer says "on pace" when you are burning far too fast.
  Early in a window the projection is deliberately damped so a thin slice cannot look
  alarming — but the "on pace" verdict was being read off the damped number. Using 36%
  of a 5-hour window in its first 30 minutes is a pace that exhausts it in under an
  hour, and the status bar called that "on pace to use this window fully".
• Live limits can no longer show a different account's usage. After signing out and
  back in as another account, a cached access token kept being used — successfully, so
  nothing ever corrected it — and the status bar could report the previous account's
  limits for hours.
• Claude Fable 5.1 cache reads were billed at 4× the real rate ($1.00 rather than
  $0.25 per million), inherited from Claude Fable 5. Cache reads dominate token volume
  in Claude Code sessions, so this noticeably overstated Fable 5.1 usage.
• Opus fast mode was priced at half its real cost. Fast mode bills at premium rates
  and is reported per message; that signal was being discarded.
• "Explain with AI" works again on current Claude models. The token budget was small
  enough that newer models, which reason before answering, could spend all of it before
  producing any visible text — returning "The model returned no text." on a billed request.
• Warning popups no longer fire twice. A momentary dip in a reported percentage could
  be mistaken for the window resetting, re-arming every alert and immediately re-firing it.
• A window at 100% is never shown in green. Clearing the warning-threshold list
  silently disabled all colouring and every alert.
• The dashboard no longer rebuilds itself on every log write. With the dashboard open
  during an active session it recomputed everything up to several times a second, which
  could make the whole editor feel sluggish.
• "Explain with AI" can no longer get stuck on "Analyzing your usage…" after the
  dashboard is closed and reopened mid-request.
• Live limits are shared between windows more reliably. Publishing the shared reading
  could silently fail when another window was reading it at that moment — on Windows,
  often. Each window then fell back to making its own request, which is what the shared
  reading exists to avoid. A rate-limit pause is now shared too, instead of every window
  discovering it separately.
• The dashboard stays where you put it instead of jumping into the active editor group.
• A log file locked mid-write no longer skips a refresh.

Changed

• Your context % will read differently after this release. On current models
  (Sonnet 4.6 / Opus 4.6 and later) a conversation under 200k tokens now reads far
  lower, because it is measured against the full 1M window. On Haiku and pre-4.6
  models it is unchanged. The new figures are the correct ones.
• The AI Advisor now defaults to Claude Sonnet 5, which is both cheaper and stronger
  than the previous default. Set advisor.ai.model to an Opus id for deeper analysis.
• The extension now runs in Restricted Mode windows. It only ever reads usage data
  from your home directory and never executes anything from a workspace.
• An AI answer cut short by the token limit now says so, instead of appearing complete.
• If Anthropic rejects your stored API key, the extension now offers to update it rather
  than repeating the same error.

Security

• Prompt samples are limited to the folders open in your window. The opt-in
  "include prompt text" setting collected from every Claude Code project on the
  machine, so prompts from an unrelated repository could be sent under a consent dialog
  that mentioned only "your recent prompts". Samples are now matched to the open
  workspace, and none are sent when there is no folder open. The confirmation dialog
  now names the workspace it is drawing from.