v0.4.2

Added

• Copying text out. ctrl+y (or /copy) puts the last reply on the
  clipboard and /copy all the whole conversation; thoth asks the terminal
  to do it with OSC 52, so it works over ssh and inside tmux and pulls in no
  clipboard library. ctrl+t hands the mouse back to the terminal so
  click-and-drag selects text the way it does in every other program, and
  ctrl+t again gives thoth the scroll wheel.
• Slash commands complete as you type them. A / at the start of the
  message opens the list, with what each command does beside it, narrowing
  with every letter; tab or enter takes the highlighted one. It is the
  same picker @path already used.
• thoth --view <name> in a debug build draws any screen of the interface
  with made-up contents and prints it as text: the start screen, a session
  mid-task, a permission prompt, the chooser, plan mode, both pickers, the
  four permission modes, the profile screen, and the whole thing at 46
  columns, a turn mid-stream, a scrolled-back transcript, hostile content,
  markdown, and a chooser with more options than rows to draw them in.
  thoth --view lists them and --view-size WxH sets the
  terminal. Screens that only appear after a permission prompt or an
  ask_user call were minutes of model time away from being looked at,
  which is where their bugs were living. cargo test every_view_draws
  renders all of them at two sizes.

Fixed

• A shell command could draw on top of the interface on Windows. Redirecting
  stdout and stderr is not enough there: a console program inherits its
  parent's console and can write to it through the console api, which is what
  Write-Progress, Invoke-WebRequest's progress bar and a coloured
  Write-Host do, and a child that changes the console mode takes the
  terminal's escape handling with it. Commands now run with no console of
  their own.
• PowerShell wrote its output in the console codepage, so anything past ascii
  came back through the pipe as mojibake: Thai, Japanese and accented text in
  a command's output was unreadable, to the user and to the model both. The
  pipe is UTF-8 now. Commands also run -NonInteractive, so one that would
  have asked a question fails instead of waiting for an answer nobody can
  give.
• Escape sequences from tool output reached the terminal. cargo test,
  git and npm emit colour the moment they think a terminal is listening
  and the shell tool captures it verbatim, so a \x1b[31m in a result
  repainted the rest of the screen and \x1b[2J wiped it. Every line thoth
  draws now has its escapes, control characters and bidi overrides taken out
  first, wherever it came from: tool output, a file, a web page, the model's
  own reply, the model name the server gave, or a paste.
• A question with more options than the terminal had rows drew as many as
  fitted and no more: the rest could be walked onto with the arrows but
  never seen, and the question itself could be pushed off the screen. The
  list scrolls with the highlight now, says how many are below it, and never
  takes the last row of the transcript.
• A tool call with a permission prompt open said "running" underneath it.
  It has not run, and may never run: it now says it is waiting for you.
• "plan ready" and its three keys were written into the transcript and drawn
  again on the state line one row below. The transcript keeps the moment,
  the state line keeps the keys.
• Tabs in tool results and error messages were drawn as a single character,
  so a stack trace lost its indentation.
• On a narrow terminal the status line dropped "scrolled up" and "selecting"
  to make room for the token counts, so the two states that explain why the
  screen has stopped behaving were exactly the ones not shown. They win over
  the numbers now.

Changed

• The input sits in a box of its own instead of under a rule. It is the one
  part of the screen written into rather than read, and the border says so;
  it also carries whether thoth is listening, in the colour, which saves
  saying it in words somewhere else.
• A tool's result hangs off the call on a corner (⎿) with the rest of it
  lined up past the corner. A result and the next line of prose were both
  just indented text before.
• A shell command is drawn once, not twice: the preview under the call
  repeated what the header line already said. It comes back the moment the
  header line is too narrow to hold the whole command, and a command that
  does not fit now wraps at its spaces instead of being cut, because a
  permission prompt is answered on what is drawn.
• The model's question is drawn above the input rather than below it: it is
  asked before anything is typed, while @path completes what is being
  typed, and the two belong on opposite sides of the box.
• Markdown markers are obeyed instead of printed: ## in front of a heading
  is dropped rather than shown next to a heading that is already bold, -
  becomes a bullet with the wrapped rest of the item lined up under it, >
  becomes a bar down the side, a ``` fence becomes a rule carrying the
  language, and --- becomes a rule.
• Each tool call is marked with a dot coloured by how it went: still going,
  done, failed. A screenful of them reads down the left edge.
• The start screen names the three things nobody discovers by typing a
  sentence at thoth: /, @ and shift+tab.
• The chooser for a question from the model is a list you move through with
  up/down and take with enter, drawn under the input with the
  highlighted row marked. 1-9 still picks one straight off.
• Its last row is not one of the model's options: it opens the input box so
  you can write the answer in your own words when none of them is what you
  wanted. The model is told plainly that you took none of its options, so it
  does what you asked rather than the offer yours sounded closest to. esc
  goes back to the list, and a message half-typed when the question arrived
  is put back afterwards.

Full changelog: v0.4.1...v0.4.2