feat(tui): collapse tool runs into Claude Code's brief line

[ericleepi314]

The transcript listed every tool call on its own row. A turn that read three
files and ran three commands cost six ⏺ rows plus their results, and the
"grouping" of same-type calls was just adjacency — nothing actually grouped.

Real Claude Code folds a run of calls into one muted summary line and puts
the full list back under ctrl+o. This ports that.

Ground truth

Claude Code 2.1.228 was installed and driven through a pty (pty.fork +
pyte) on a fixture project, and its rendering captured for each tool shape.
Everything below is from those captures.

Claude Code, default view

⏺ I'll read the three files first.

  Read 3 files, listed 1 directory, ran 1 shell command

⏺ DONE

Claude Code, ctrl+o

⏺ Read(/…/src/alpha.py)
  ⎿  Read 8 lines

⏺ Read(/…/src/beta.py)
  ⎿  Read 8 lines

clawcodex before

 ⏺ Read(src/alpha.py)
   ⎿  Read 7 lines
 ⏺ Read(src/beta.py)
   ⎿  Read 7 lines
 ⏺ Read(src/gamma.py)
   ⎿  Read 3 lines
 ⏺ Bash(ls src)
   ⎿  alpha.py
      beta.py
      gamma.py
 ⏺ Bash(echo hello)
   ⎿  hello
 ⏺ Bash(wc -l docs/numbers.txt)
   ⎿  12 docs/numbers.txt

clawcodex after (same prompt, same fixture)

   Read 3 files, listed 1 directory, ran 2 shell commands

clawcodex after, ctrl+o — every call back, blank-line separated as upstream
renders them:

 ⏺ Read(src/alpha.py)
   ⎿  Read 7 lines

 ⏺ Read(src/beta.py)
   ⎿  Read 7 lines

 ⏺ Bash(echo hello)
   ⎿  hello

While the turn runs, gerund + ellipsis, tally bold — byte-identical in
shape to upstream's Reading \e[1m1\e[22m file…:

 ⏺ Listing 1 directory, running 2 shell commands…

What the brief says

ui-tui/src/domain/toolBrief.ts owns the vocabulary. Two rules carry the look:
only the first clause is capitalized, and a live run uses the gerund with a
trailing ellipsis. Clause order and wording follow upstream: searched for N patterns, read N files, listed N directories, ran N shell commands.

Shell calls bucket on the command, not the tool — ls reads as a listing,
grep as a search, cat as a read — which is what upstream does and what
keeps a shell-driven session reading as exploration. The match is deliberately
a prefix on the bare command, so anything wrapped (sudo ls, cd x && ls, a
pipeline) falls through to "shell command"; under-claiming reads fine, a
cat > file.py heredoc counted as a read would not.

What does not fold

Only the read/search band and shell commands. That is what upstream's message
type has been saying by name all along — collapsed_read_search — and the
captures agree: a lone Bash folds to Ran 1 shell command, but a lone
WebSearch keeps its row and its result, and so does an Agent:

⏺ Web Search("Rust 1.90 release date")
  ⎿  Did 1 search in 2s

⏺ Agent(Count .py files under src)
  ⎿  Done (2 tool uses · 48.0k tokens · 11s)

⏺ Write(src/delta.py)
  ⎿  Wrote 2 lines to src/delta.py

So edits, delegations, answer-bearing tools (AskUserQuestion, advisor,
vision_analyze, ExitPlanMode) and the catch-all bucket all keep their own
block. briefRuns() preserves order, so a trail that interleaves them reads
brief / row / brief. Reading, searching, listing and shelling out are what
bury a transcript; they are what folds.

Failures also break out — a deliberate divergence. Upstream tallies a
failed call away like any other; a captured cat of a missing file renders as
a plain Read 1 file with the error nowhere on screen. clawcodex runs against
models that do not reliably narrate a failure, so the failing call keeps its
block while the successes around it still fold:

   Read 1 file

 ⏺ Bash(cat missing.txt)
   ⎿  Error: cat: missing.txt: No such file or directory
      Command failed with exit code 1

   Ran 1 shell command

Defects found in review

Three rounds of adversarial review, each of which found something real.

The brief painted bold from its first tally onward. Bold and faint share
SGR close code 22, so wrapping bold counts in a dim parent rewrites each
count's \e[22m into the parent's \e[2m — faint reopened, bold never
cleared. Every column after the first tally rendered bold: the opposite of
what the line is for. muted alone carries the quiet now. stripAnsi cannot
see this class of bug, so the test asserts on the bytes.

The height estimator disagreed with the paint, five ways. The virtualized
transcript positions rows from estimatedMsgHeight before Yoga measures
anything, so a wrong estimate shows up as scrollbar drift and blank gaps.
It counted raw trail lines where the renderer counts painted blocks; never
counted the ⏺ call row; wrapped the brief at the prose width (a trail has no
role gutter, so it is three columns wider); counted lines rather than wrapped
rows, so an Edit's Updated <full path> with N additions… — capped at twelve
lines, never at columns — was short by one or two on any normal terminal;
missed the ∴ Thinking… header row, which a one-row text floor had been
paying for by accident; wrapped the reasoning body three columns short of its
└─ rail; and missed the marginBottom on every prose row's details
wrapper, which predates this branch.

The fix that matters is the test: it renders thirteen message shapes at two
widths through both ToolTrail and MessageLine and asserts the estimate
equals the painted row count. The numbers in it are measured, not asserted
by hand — which is how the prose-row margin surfaced at all, and why
virtualHeights.test.ts's response-separator expectation moves from +3 to +5.

/details tools expanded collided in the height cache key, so pinning it
changed the paint without changing the key and served the brief's heights to
the expanded layout.

Also fixed: the ⎿ gutter is three columns wide and a formatter had quietly
collapsed the bare JSX whitespace to two — it is a string literal now, with a
test that can see the difference.

Testing

• ui-tui suite: 1819 passing, 8 failures all pre-existing on origin/main.
• 75 new tests in toolBrief.test.ts — classification, vocabulary, run
  splitting, render shape, byte-level styling, and estimate-vs-paint agreement
  across trail and prose rows at 60 and 100 columns.
• Verified end to end against a live backend (deepseek-v4-pro) through the
  same pty harness: collapsed, ctrl+o, live, and failure cases all captured
  above.
• ui-tui vitest is not wired into CI (the Node job covers ui-desktop), so
  the suite above was run locally against origin/main for the baseline.

🤖 Generated with Claude Code