Turnscope 0.1.2

• The panel's headline names an account by its title, never its email,
  and the line under it names the account too: "At this pace, no Claude Max
  limit runs out." A notification adds the email only when another account
  in use has the same title.
• Reading a session with turnscope sessions read or the
  read_session tool quotes what each entry says, > a line, and says
  to treat it as data, not instructions. Text from a web page or a file in
  a session can no longer pass for one of its entries, such as a request
  from you.
• Long messages read back from a session are cut at 40,000 characters,
  saying how many were left out, so one long paste stays within what an
  agent can take from a tool.
• Run from its disk image, Turnscope no longer offers to connect an
  agent: an agent connected there would lose Turnscope once the image is
  ejected. Settings → Agents turns off its switches and says to move
  Turnscope to Applications.
• Updates are checked against their signature before the disk image is
  opened, from the update after this one.
• Claude Code usage through a gateway or another backend that leaves
  out the cache counts or their split is counted. Such a response was
  skipped whole.
• A Claude Code fork read the moment it starts, while its description
  is still being written, no longer shows its parent's history as its own
  and keeps its task title. A fork of a session shows the session it was
  forked from.
• Branches: a Claude Code session that moves outside a repository keeps
  the branch it was on.
• Search no longer counts a session's messages twice once Codex
  archives it, or once a history file is replaced. A moved file is read on
  from where it was.
• doctor lists history files that can't be read and why, and counts
  the lines it couldn't read once, not again each time a file is read from
  the start.
• Handoff: the session's title, its tasks and what each tool call was
  for are in double quotes, as the rest of what the session said is
  quoted, with a quote inside escaped as \". A task a model wrote can no
  longer read as Turnscope's own words. A session of one entry says
  "1 entry", not "1 entries".
• Session titles are in quotes wherever a line shows one: turnscope sessions show and explain_usage by session quote them now, as the
  list of sessions does. A session with no title shows as (untitled),
  without quotes.
• What used a limit on one model, such as Claude's Opus weekly, counts
  only that model's responses. Other models' usage was given shares of it.
  Its column reads "% of Weekly (Opus)".
• Hours that fit (status --hours, check_limits) leave out a window
  that rolls, as OpenCode Go's 5 hours, and say so, instead of saying too
  little work was measured. One hour reads "1 more hour of work fits", not
  "fit".
• An API key held by the same agent in two folders names that agent
  once.
• --since no longer takes a negative count, which meant a time ahead,
  and usage says when --since isn't before --until.
• Limits are read with a shorter wait. A request to a provider gives up
  after 15 seconds instead of 60, and any request gives up after 5 if it
  can't connect. On a network that drops what's sent, guard in an agent's
  hook could hold the agent for a minute or more.
• "Sign in again" is no longer said when a web page in front of a
  provider, such as a bot check, refuses the request. The account reads as
  one that couldn't be read, and it's tried again in 5 minutes.
• OpenCode Go: a limit whose reset time can't be read fails the read,
  instead of showing as a limit that never resets.
• A key held by two agents, such as one OpenRouter key in OpenCode and
  Pi, is sent once each read instead of twice.
• Not up to date: when the engine sends a status or alert the app can't
  read, or doesn't answer within 30 seconds, the panel says "Not up to date ·
  trying again", the menu bar item dims, and the app connects again. Before
  the first status, the panel says "Turnscope's engine didn't answer". It
  used to keep showing the last status as current, or "Reading your
  agents' history…" with no end.
• Settings → General: the notification switches are grayed out until
  the first status arrives. Before, clicking them did nothing.
• Reading a long history, as on a first run, no longer makes other
  commands fail with "database is locked". config set, accounts hide,
  reading limits and raising alerts used to wait up to 10 s for it and
  then fail. Now they wait a second or two at most.
• Costs: when models.dev first listed a model while another process
  was reading history, the responses that process wrote could stay
  without a cost for good. They're now priced.
• Codex forks that open with a copy of the parent thread's history no
  longer take that copy as their own. A fork's title is its own first
  request, search finds the parent's words only in the parent,
  sessions read and handoffs show only the fork's conversation, and the
  parent's usage isn't counted a second time.
• A Codex thread rewound to an earlier turn no longer shows the turns
  it went back past in sessions read and handoffs, and search no longer
  finds the thread by their words. A fork that points at its parent's
  history instead of copying it is titled by its own first request, not
  "(untitled)".
• OpenCode: a reply another OpenCode process saved late is no longer
  missed. A session OpenCode left with a placeholder title, such as
  "New session - 2026-07-30T18:45:03.662Z", is titled by its first request.
• Handoff: a failed command run again under a reworded description no
  longer shows under "What failed". A grep -i that found nothing is a
  read, not a failure; sed -i.bak, find -delete and git branch -D are
  work; and a command after cd v2; counts as its own. A request is left
  out as sent again only if the agent hadn't answered it.
• Handoff: a file a Codex script patched counts as changed when the
  tests the script ran next failed. It was left out, so a session whose
  edits all came in such scripts had no list of files at all.
• statusline and guard no longer read every agent's MCP settings.
  With a 5.8 MB ~/.claude.json, each run takes a quarter of the work and
  8 MB of memory instead of 38 MB.
• The app's engine works out the status again only after reading
  something new, in a new minute, or after another Turnscope process wrote.
  While an agent writes files other than its history, it uses about a
  twentieth of the CPU it did.
• Connecting or disconnecting an agent no longer hangs when your
  shell's startup files leave a program running. It waits no longer than
  its time limit.
• The app's engine answers a line that isn't UTF-8 with a "not JSON"
  error instead of closing the connection, and keeps running when it runs
  out of file descriptors for a new connection.
• Colors: amber and red are darker in light mode, and red is lighter in
  dark mode, so text in them reads at 4.5:1 or more against the panel.
  Amber text in light mode was 2.4:1 on white. With Increase Contrast they
  reach 7:1.
• Reduce Motion: with it on, pages fade instead of sliding, a limit's
  bar doesn't grow into place, and the footer's spinner holds still.
• VoiceOver reads the menu bar item as it is now. A change only it
  would read, such as an API key's cost becoming partly unpriced, used to
  wait until the figures shown changed too.
• Dates are on the Gregorian calendar on a Mac set to another one, such
  as the Hebrew or Islamic calendar: "Oct 18", not "Heshvan 7".
• With no forecast yet, the line under the panel's headline says only
  what's left, such as "Claude Max has 99% left of 5 hours". The row below
  already says when it resets.
• VoiceOver and notifications spell out the day and month: "Back
  Saturday 9 AM", "October 18". The panel and the menu bar keep them short.
• Private keys are redacted in a JSON string, such as a service
  account's file, in a diff that removes one, and cut short with no END
  line, as head prints one, even encrypted or with spaces after its
  lines. Search and reading a session showed them.
• Keys in a tool call's input, such as a token at the start of a line
  in a file an agent wrote, or after a tab, are redacted. The input is kept
  as JSON, where the \n or \t before the token hid it, so reading a
  session and handoffs showed it.
• Your agents' history is read again once after updating, so these fixes
  apply to what was read before.