How does tool call grouping work?

how's does the tool call grouping work

Now I have a complete picture. Here's how tool call grouping works:

##Tool Call Grouping

The system lives entirely in apps/web/app/[owner]/[repo]/[postNumber]/comment-content.tsx and operates at two levels: first grouping the raw message parts, then rendering each group.

###1. groupParts() — Collecting consecutive tool calls (line 614)

The groupParts function iterates over all AgentUIMessage parts and produces a flat array of GroupedPart items, which can be one of three types: "text", "reasoning", or "tool-group".

The grouping logic is straightforward:

• Tool parts (anything where part.type.startsWith("tool-") and has a "state" property) are accumulated into a running currentToolGroup array. Consecutive tool parts keep getting pushed into the same group.
• Text parts (non-empty) and reasoning parts (non-empty) break the current tool group. When encountered, the accumulated tool group is flushed to result as a single "tool-group" entry, and the text/reasoning is added as its own entry.
• Other part types like step-start, step-finish, etc. are silently ignored — they do not break the tool grouping. This is noted explicitly in the code comment on line 664.
• After the loop, any remaining unflushed tool group is pushed.

So if the AI makes 5 tool calls in a row, they all end up in one "tool-group". If it makes 2 calls, writes some text, then makes 3 more calls, you get two separate tool groups with a text part between them.

###2. ToolGroup component — Rendering a group (line 491)

The ToolGroup component renders a group of tool call parts. It has special handling depending on the group's state:

###3. countToolsByName() — Summary labels (line 412)

The countToolsByName helper counts tools by type name (e.g., "GREP", "READ") to generate the collapsible trigger text. It strips the "tool-" prefix, uppercases it, counts occurrences, and joins them with " - ".

###4. "Thinking" collapse — Outer-level grouping (line 690-710)

There's a second layer of collapsing in the CommentContent component. When not streaming, everything before the last text part is considered "thinking" and can be collapsed into a single collapsible section (if it contains non-tool parts like text or reasoning). This aggregates tool counts from all thinking-phase tool groups into a single summary trigger.

###Visual Summary

Message Stream:
  [Grep] [Read] [Read] [text: "Found it"] [Bash] [text: "Here's the answer"]

groupParts() produces:
  1. tool-group: [Grep, Read, Read]    →  collapsed: "1 GREP - 2 READ"
  2. text: "Found it"
  3. tool-group: [Bash]                →  single tool, no collapsible
  4. text: "Here's the answer"

CommentContent then collapses #1-#3 as "thinking" (since #4 is the last text).