Tool calls

When Claude reads a file, edits one, runs a command or searches, that is a tool call. Calls that come one after another share a single Tool calls card, with a tile for each call.

The card and its tiles

The card’s header counts its calls: 1 call, 4 calls · 1 running, 4 calls · 1 failed. Anything between two calls ends the card and the next call starts another: a reply, thinking, a hook run, a prompt, a subagent, the start of a task. Tiles go three across, then two, then one as the card narrows, and the tiles in a row are all as tall as the tallest.

A tile shows:

a status dot (running, done or failed), the tool’s icon, its name and the time of the call;
what it worked on: a file by the end of its path (…\Player\StarfallCharacter.cpp, with the whole path as its tooltip), a command, or a search pattern;
how it went: Running…, 120 lines · 0.3s, Failed · 2.1s, with an Edit’s or Write’s +added −removed beside it;
the first line of its output, and any Claude Doc link, git chips and images (Images).

A call under ten seconds shows its duration with a decimal, such as 0.4s. The result line says what the tool found: the lines a Read read (40 lines of 900), what a Grep found, a Glob’s files, a fetch’s status and size (200 OK · 14 KB), whether a Write created a file or updated it.

A question or a plan gives its tile up to its own card (Questions, Approving a plan). A subagent is a card of its own (Subagents).

Folding the calls

A long run of calls takes a lot of room. Click anywhere on a Tool calls card’s header line, which lights up under the pointer as a thinking row does, and the card folds to that line, such as Tool calls · 4 calls · 1 failed ›; a second click brings the tiles back. A call that is running or has failed still shows in the header’s words.

It works as thinking does. Settings → Conversation → Show tool calls collapsed starts new cards folded. Right-click a Tool calls card for Collapse all tool calls and Expand all tool calls: they fold or open every card in the tab, and start the ones still to come the same way, without changing Settings. The tab remembers which you chose. Opening a call from elsewhere, as find does, opens its card.

Folding a chain of steps

Claude’s thinking and tool calls between two messages can still take a lot of room, even with each Tool calls card folded. Settings → Conversation → Fold each chain of steps (off by default) folds the whole run of steps, the thinking, Tool calls cards and subagents between two messages, into one: how many steps and how long Claude worked, such as 7 steps · 18m 43s.

On a task’s row. Steps right after the row where one of Claude’s tasks started are that task’s work, so the row carries them: 1 step · 2s after the task’s name. Click it and the steps open under the row; click the rest of the row for the Tasks page, as ever. While Claude is still on the task, the step it’s on stays under the row.
In Claude’s reply. When Claude’s reply comes right after the steps, and they aren’t a task’s, 7 steps · 18m 43s sits on the reply’s header line, after Claude; point at it to see what the steps were (9 tool calls · 3 thinking). Click it and the steps open inside the reply’s card, under its text. The reply stays where it was on screen while you read, rather than the conversation jumping to its end.
A line of its own otherwise, such as while Claude is still working, with what the steps were beside it. Click anywhere on it, or its chevron.
The list is the steps, a line each: Thought for 7s, Ran check for UE4SS.pdb, Created analyze_dump.py, Edited disasm.py, with how long a call took at the right of its line. A long list scrolls, with a scroll bar of its own.
Open a step with its chevron. A thinking line opens to what Claude thought, and a call to its detail, as a tile of a Tool calls card opens: the input and the output, and Open for the file it worked on.
Fold them again with another click, and the list is gone.
While Claude is working, the line says Working · 6 earlier steps, and the last step stays under it: the one Claude is on. Once Claude has moved on, the whole run is in the line.
A single step, such as a lone Thinking row before a reply, folds the same way once Claude has moved on from it. If a step’s time isn’t known, or the run took under a second, the line counts the steps instead of saying how long they took.

Find and Export conversation… still reach a folded step, and going to one opens the list at its line.

One tab its own way

The setting is what every tab does, and a tab can fold its steps or show them all for itself:

Steps in the tab’s right-click menu: Default, which says what Settings does, such as Default (folded), Folded or All. Steps in its Tab settings… is the same choice.
Steps: folded in the message box’s bar, while the setting is on: click it for Steps: all in this tab, and again to fold them.
Show every step in a folded line’s right-click menu shows every step in its tab.

A tab with its own choice keeps it when the setting changes, and Settings says 1 tab has its own by Fold each chain of steps.

Opening a call

Click a tile to open its detail under the tiles. One tile is open at a time, so opening another closes it, and the × at the detail’s right closes it. The detail shows the call in full:

Edit and Write. An inline diff of the change. Open diff opens the file in the diff view: from before Claude’s first change to it in this session to the file as it is now, the way Changed files shows it.
Bash. The command, and its output in a monospace font. What git did shows as chips under the tile: Committed 1a2b3c4 on main, Pushed main, Rebased onto origin/main, Opened PR #42 (a link to the pull request). A command that reached its time limit says so, and that it carries on in the background; one in the background says Running in the background or Moved to the background.
Read, Grep, Glob and others. What the tool was given and what it returned.
Open. For a call that worked on a file, Open opens the file in Claudette’s editor or the OS’s app (Opening files from the conversation).

An output over 20,000 characters shows its start, with Show all (n KB) to show the rest. A call with nothing more to say reads Nothing more to show for this call.

Find opens the tile a match is in.

A call in a window

Open in a window, the button in a call’s detail, opens the call in a window of its own, which follows the call: its title and state, and its running time, ticking each second. It has two tabs, Call and Details.

Call depends on the tool:

Bash. The command and its output. Claude Code sends a command’s output only when it ends, so a running command shows its time, not its output.
Read. The file as Claude read it, numbered from where it started, in the diff view’s colors.
Edit and Write. This call’s own change, with Open diff for the whole session’s.
Grep and Glob. What they found, file by file, with a Grep’s matching lines.
Anything else. Its arguments and its result.

Details has the arguments and the result as the tool got and gave them. There is one window per call: opening it again brings it to the front. Esc closes it.

Settings

Show tool calls collapsed (Settings → Conversation, off by default) starts new Tool calls cards folded to their header line. Fold each chain of steps (off by default) folds a run of steps between two messages into one, as Folding a chain of steps says.

Tool calls are one of the in-between steps that Quieter in-between steps (Settings → Conversation, on by default) draws without a tint and outline until you point at them. See Reading the conversation.