Tsukaima 0.9.3

Fixed

• Clarify questions: choices are full-width rows whose text wraps, so a long choice is read in full (they were
  one-line chips cut with "…", which hid what was being picked). The question notification shows up to 300 characters
  instead of 120.
• Tool rows: tapping a row open unfolds its whole command or summary instead of keeping the one-line "…" cut, and a
  row that came without a summary shows its arguments above the output. Risk and interruption notes are no longer cut
  at two lines. Subagent rows unfold their goal and detail the same way.
• Reasoning and tool rows keep their arrival order (reasoning, tool, reasoning, tool, reply). The reasoning that led
  to a tool call was moved below the tool each time one started, so tool rows appeared to hop above it; a loaded
  history now shows those reasoning rows in their place too, instead of dropping them.

Tsukaima 0.9.2

0.9.2 — 2026-10-21 (versionCode 22)

Fixed

• Stop and queued sends: Stop no longer leaves a stale "stopping" state after a turn that had already ended (also
  after a reconnect that lost the stopped turn's end), and a Stop that races a turn's end is much less likely to mark
  a queued send "Not sent" when it actually ran (a narrow race while the server drains its queue remains). A message
  sent between turns, while an earlier queued one waits, is queued too ("Send when done" matches what happens).
• A queued send that Stop purged keeps its "Not sent" bubble (with Retry) across a reconnect or history reload instead
  of vanishing or showing twice.
• Long runs of _ in a reply no longer freeze the app while it parses; very long text with a stray _ parses in
  milliseconds instead of seconds.
• Dollar amounts and shell variables ($HOME, a$b$c, "costs $5 and $6") are no longer read as formulas.
• A backslash-escaped ] inside link text ([a\]b](url)) now makes a link; an empty $$ $$ block reads as the text it
  shows.
• Formulas: they no longer block the screen while they draw (a cached one shows in its first frame), an inline formula
  too wide for the line scales down to fit or, if still too wide, shows clipped with a "…" button that opens it in
  full, and TalkBack reads the formula in words ("x to the power 2", "a over b") instead of raw TeX, inline ones once,
  as part of their sentence.
• Connection card: the countdown rounds down ("2 min left", then seconds), the buttons lock while an answer is being
  sent so it goes out once, a target with a sign-in link and settings sends the settings first and then offers
  "Continue in browser", and a "Hermes suggests connecting: …" hint shows for directory requests.
• Task card: an older update arriving late (or a cleared list coming back from a stale reload) no longer overwrites a
  newer list; the transcript keeps scrolling to a card that changed in place; children of a folded done task no longer
  stay indented.
• Images: the image cache no longer loses pictures when many fail to load, one tap on a foreign image loads it
  everywhere it appears (reply and tool row), and long reply images are no longer shrunk to the tool-row size
  limit.
• A failed edit or regenerate no longer drops side notes and background-task results that arrived in the meantime.
• Report a problem: a report that finishes after you left the screen no longer pops the share sheet on a later visit;
  it says "Report ready — tap Report a problem again".
• Read-aloud with a server voice pauses less between sentences: the next sentence is synthesised while the current
  one plays (the first sentence, and one that is slow to synthesise, still wait).