Yehowshua/ZIm
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

agent: Handle tool use without text (#28030)

Browse source 
### Context 

The Anthropic API fails if a request message contains a tool use and no
`Text` segments or it only contains empty `Text` segments. These are
cases that the model itself produces, but the API doesn't support
sending them back.

#27917 fixed this by appending "Using tool..." in the thread's message,
but this causes the actual conversation to include it, so it would
appear in the UI (we would actually display a gap because we never
rendered its markdown, but "Using tool..." would show up when the thread
was restored).

### Solution

We'll now only append this placeholder when we build the request, so the
API still sees it, but the UI/Thread doesn't.

Another issue we found is that the model starts mimicking these
placeholders in later tool uses which is undesirable. So unfortunately,
we had to add logic to filter them out.

Release Notes:

- agent: Improved rendering of tool uses without text

---------

Co-authored-by: Bennet <bennet@zed.dev>
This commit is contained in:
Agus Zubiaga 2025-04-03 14:22:59 -03:00 • committed by GitHub
parent ece4a1cd7c
commit ed3722023e
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 270 additions and 237 deletions
Download patch file Download diff file Expand all files Collapse all files

47
crates/agent/src/thread.rs
View file

@ -35,7 +35,7 @@ use crate::thread_store::{
SerializedMessage, SerializedMessageSegment, SerializedThread, SerializedToolResult,
SerializedToolUse,
};
use crate::tool_use::{PendingToolUse, ToolUse, ToolUseState};
use crate::tool_use::{PendingToolUse, ToolUse, ToolUseState, USING_TOOL_MARKER};
#[derive(Debug, Clone, Copy)]
pub enum RequestKind {
@ -85,6 +85,12 @@ pub struct Message {
}
impl Message {
/// Returns whether the message contains any meaningful text that should be displayed
/// The model sometimes runs tool without producing any text or just a marker ([`USING_TOOL_MARKER`])
pub fn should_display_content(&self) -> bool {
self.segments.iter().all(|segment| segment.should_display())
}
pub fn push_thinking(&mut self, text: &str) {
if let Some(MessageSegment::Thinking(segment)) = self.segments.last_mut() {
segment.push_str(text);
@ -131,6 +137,16 @@ impl MessageSegment {
Self::Thinking(text) => text,
}
}
pub fn should_display(&self) -> bool {
// We add USING_TOOL_MARKER when making a request that includes tool uses
// without non-whitespace text around them, and this can cause the model
// to mimic the pattern, so we consider those segments not displayable.
match self {
Self::Text(text) => text.is_empty() || text.trim() == USING_TOOL_MARKER,
Self::Thinking(text) => text.is_empty() || text.trim() == USING_TOOL_MARKER,
}
}
}
#[derive(Debug, Clone, Serialize, Deserialize)]
@ -1083,32 +1099,15 @@ impl Thread {
}
}
LanguageModelCompletionEvent::ToolUse(tool_use) => {
let last_assistant_message = thread
let last_assistant_message_id = thread
.messages
.iter_mut()
.rfind(|message| message.role == Role::Assistant);
.rfind(|message| message.role == Role::Assistant)
.map(|message| message.id)
.unwrap_or_else(|| {
thread.insert_message(Role::Assistant, vec![], cx)
});
let last_assistant_message_id =
if let Some(message) = last_assistant_message {
if let Some(segment) = message.segments.first_mut() {
let text = segment.text_mut();
if text.is_empty() {
text.push_str("Using tool...");
}
} else {
message.segments.push(MessageSegment::Text(
"Using tool...".to_string(),
));
}
message.id
} else {
thread.insert_message(
Role::Assistant,
vec![MessageSegment::Text("Using tool...".to_string())],
cx,
)
};
thread.tool_use.request_tool_use(
last_assistant_message_id,
tool_use,