feat(terminal): an input line on phones, instead of typing into the grid - #76

The row takes a line above the message box on every pane; it now waits for
the ⚡ button, and the choice is remembered per device as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Transcripts already record what each request filled:
- Codex: token_count events with last_token_usage and model_context_window
- Claude Code: assistant usage, input + cache creation + cache read,
  subagent (sidechain) entries skipped
- omp/omo/gjc: message usage, input + cacheRead + cacheWrite

ConversationMetadata gains an optional `context { used, window }`. The
window is only what the transcript leaves no doubt about: Codex states it,
and a Claude request past 200k can only have run in the 1M window (kept
through a compaction). Otherwise the line shows the tokens used, never a
guessed percentage.

The composer shows `74% left` or `68k used`, red at 20% or less, with the
counts on hover. PaneTerminal's metadata dedupe compared only model and
effort, so it now compares the context too, or the number would never move.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…, a loading state

From comparing the chat with chatmux's:

- A failed tool call looked like a success inside "Worked for…". Tool parts
  gain `error`: Claude's tool_result is_error and omp's toolResult isError,
  and for Codex, which records no flag, the output's own verdict
  (exit code, "Script failed", "apply_patch verification failed"; a
  completed script is judged as a whole). The row reads "failed" in red,
  its output is headed Error, and the block header counts failures.
- A file path in a tool row opens the FileViewer, like paths in prose.
- Codex's apply_patch, bare or as the string an exec script passes to
  tools.apply_patch, renders as a per-file diff and its row is summed up
  by the files it touches (shared/patch.ts, used by server and client).
- The jump-to-latest button showed only for new messages; it now shows
  whenever the reader is scrolled up (a round arrow), and stays the "New
  messages" pill when there are some.
- The chat said "No conversation yet" until the first answer came back;
  it now says it is loading.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

More from comparing with chatmux's chat:

- Images the user sent: Claude transcripts keep pasted images as base64
  in the user entry. The parser now emits an image part naming it
  (`<entry uuid>:<block index>`, png/jpeg/gif/webp only) instead of the
  data, and GET /api/pane/conversation/image decodes it on request
  (immutable cache), so a 2 s poll never carries megabytes. Images this
  app attaches appear as `@…png` mentions and show through /api/fs/file.
  Thumbnails sit above the user bubble; a tap opens them.
- Edit/MultiEdit render as one line diff (src/lib/diff.ts, an LCS that
  falls back to removed-then-added past 250k line pairs).
- A compaction's summary, which was dropped, is a divider that opens to
  it.
- Markdown headings #### to ###### render as h4-h6.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… menu

- Tool outputs are cut at 4000 characters in the page. A cut one now
  keeps its call id and full length (output_ref, output_size; the cut
  moved into server/tool-output.ts, shared by the Claude, omp and Codex
  parsers). GET /api/pane/conversation/tool-output finds the call's result
  in the transcript and returns it as text (up to 2 MB, immutable cache);
  the row offers "Show the whole output (48k characters)" and shows it
  in a box that scrolls on its own.
- The `/` menu for Claude adds user and project skills (SKILL.md) and the
  skills and commands of plugins enabled in ~/.claude/settings.json, as
  /<plugin>:<name>. For Codex it adds saved prompts (/prompts:<name>) and
  its skills, which Codex takes as `$name`: a `$` where a word starts opens
  them, only in a Codex pane.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

On a phone, text typed in the terminal lens came out repeated ("널에
추가로 폰에서는채팅을 모바일 가터미널에 미널에…"): the keyboard rewrites
what it typed (dictation revising a phrase, an IME composing a Korean
syllable, autocorrect), and xterm sends each revision as new keys a
terminal cannot take back. Termius documents the same problem; Moshi and
Termius both answer it with a panel where the text is finished first and
sent whole.

- On a coarse pointer, the terminal lens gets TerminalInput above the key
  bar: a line that grows to four, Enter or the button sends it, an empty
  line's button presses Enter alone. It is written with the keyboard's own
  editing and sent whole.
- It goes as a submit with `typed`: the server types it like the keyboard
  (send_text, then Enter after the usual gap) instead of agent.prompt,
  which refuses while an agent waits, since in the terminal an answer to
  an open menu is what the user means. Several lines go as one paste.
- The grid's hidden textarea gets inputmode=none, so a tap no longer raises
  the keyboard, and key-bar taps leave focus in the line. A ⌨ key on the
  key bar switches to typing straight into the grid, remembered per device.
- Desktops (fine pointer) are unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>