fix: preserve tool-call history across thread hydration (#568) (#670)

Prevent model re-attempts and data inconsistencies when rebuilding
  conversation context from persisted tool-call records.

  - Remove raw tool parameters from persisted tool_calls JSON to prevent
    unredacted sensitive data from being stored in the database. The LLM
    context rebuild only needs call_id + name + result.
  - Make record_tool_error/record_tool_result mutually exclusive in all
    three execution paths (dispatcher, approval, deferred). Previously
    error cases called both methods, violating the TurnToolCall invariant
    and sending contradictory outcomes to the LLM.
  - Unify call_id format to turn{N}_{i} between live sessions and
    persisted hydration to eliminate ID mismatch in the LLM context.
  - Auto-close </tool_output> XML tags after truncate_preview truncation
    to prevent malformed tool output reaching the LLM.

  [skip-regression-check]
This commit is contained in:
Reid
2026-03-09 08:36:07 -07:00
committed by GitHub
parent d73e35cfb0
commit 290d925c7f
4 changed files with 666 additions and 87 deletions
+18 -17
View File
@@ -743,23 +743,6 @@ impl Agent {
.await;
}
// Record result in thread
{
let mut sess = session.lock().await;
if let Some(thread) = sess.threads.get_mut(&thread_id)
&& let Some(turn) = thread.last_turn_mut()
{
match &tool_result {
Ok(output) => {
turn.record_tool_result(serde_json::json!(output));
}
Err(e) => {
turn.record_tool_error(e.to_string());
}
}
}
}
// Check for auth awaiting — defer the return
// until all results are recorded.
if deferred_auth.is_none()
@@ -799,6 +782,7 @@ impl Agent {
}
// Sanitize and add tool result to context
let is_tool_error = tool_result.is_err();
let result_content = match tool_result {
Ok(output) => {
let sanitized =
@@ -812,6 +796,23 @@ impl Agent {
Err(e) => format!("Tool '{}' failed: {}", tc.name, e),
};
// Record sanitized result in thread so messages()
// and persist_tool_calls() use cleaned content.
{
let mut sess = session.lock().await;
if let Some(thread) = sess.threads.get_mut(&thread_id)
&& let Some(turn) = thread.last_turn_mut()
{
if is_tool_error {
turn.record_tool_error(result_content.clone());
} else {
turn.record_tool_result(serde_json::json!(
result_content
));
}
}
}
context_messages.push(ChatMessage::tool_result(
&tc.id,
&tc.name,
+318 -11
View File
@@ -16,6 +16,7 @@ use chrono::{DateTime, Utc};
use serde::{Deserialize, Serialize};
use uuid::Uuid;
use crate::channels::web::util::truncate_preview;
use crate::llm::{ChatMessage, ToolCall};
/// A session containing one or more threads.
@@ -320,7 +321,13 @@ impl Thread {
}
}
/// Get all messages for context building.
/// Get all messages for context building, including tool call history.
///
/// Emits the full LLM-compatible message sequence per turn:
/// `user → [assistant_with_tool_calls → tool_result*] → assistant`
///
/// This ensures the LLM sees prior tool executions and won't re-attempt
/// completed actions in subsequent turns.
pub fn messages(&self) -> Vec<ChatMessage> {
let mut messages = Vec::new();
for turn in &self.turns {
@@ -332,6 +339,42 @@ impl Thread {
turn.image_content_parts.clone(),
));
}
if !turn.tool_calls.is_empty() {
// Build ToolCall objects with synthetic stable IDs
let tool_calls: Vec<ToolCall> = turn
.tool_calls
.iter()
.enumerate()
.map(|(i, tc)| ToolCall {
id: format!("turn{}_{}", turn.turn_number, i),
name: tc.name.clone(),
arguments: tc.parameters.clone(),
})
.collect();
// Assistant message declaring the tool calls (no text content)
messages.push(ChatMessage::assistant_with_tool_calls(None, tool_calls));
// Individual tool result messages, truncated to limit context size.
for (i, tc) in turn.tool_calls.iter().enumerate() {
let call_id = format!("turn{}_{}", turn.turn_number, i);
let content = if let Some(ref err) = tc.error {
// .error already contains the full error text;
// pass through without wrapping to avoid double-prefix.
truncate_preview(err, 1000)
} else if let Some(ref res) = tc.result {
let raw = match res {
serde_json::Value::String(s) => s.clone(),
other => other.to_string(),
};
truncate_preview(&raw, 1000)
} else {
"OK".to_string()
};
messages.push(ChatMessage::tool_result(call_id, &tc.name, content));
}
}
if let Some(ref response) = turn.response {
messages.push(ChatMessage::assistant(response));
}
@@ -353,13 +396,16 @@ impl Thread {
/// Restore thread state from a checkpoint's messages.
///
/// Clears existing turns and rebuilds from message pairs.
/// Messages should alternate: user, assistant, user, assistant...
/// Clears existing turns and rebuilds from the message sequence.
/// Handles the full message pattern including tool messages:
/// `user → [assistant_with_tool_calls → tool_result*] → assistant`
///
/// Also supports the legacy pattern (user/assistant pairs only) for
/// backward compatibility with old checkpoint data.
pub fn restore_from_messages(&mut self, messages: Vec<ChatMessage>) {
self.turns.clear();
self.state = ThreadState::Idle;
// Messages alternate: user, assistant, user, assistant...
let mut iter = messages.into_iter().peekable();
let mut turn_number = 0;
@@ -367,18 +413,58 @@ impl Thread {
if msg.role == crate::llm::Role::User {
let mut turn = Turn::new(turn_number, &msg.content);
// Check if next is assistant response
if let Some(next) = iter.peek()
&& next.role == crate::llm::Role::Assistant
{
// iter.next() is guaranteed Some after a successful peek()
if let Some(response) = iter.next() {
turn.complete(&response.content);
// Consume tool call sequences (assistant_with_tool_calls + tool_results).
// A single turn may contain multiple rounds of tool calls, so we
// track the cumulative base index into turn.tool_calls.
while let Some(next) = iter.peek() {
if next.role == crate::llm::Role::Assistant && next.tool_calls.is_some() {
let call_base_idx = turn.tool_calls.len();
if let Some(assistant_msg) = iter.next()
&& let Some(ref tcs) = assistant_msg.tool_calls
{
for tc in tcs {
turn.record_tool_call(&tc.name, tc.arguments.clone());
}
}
// Consume the corresponding tool_result messages,
// indexing relative to this batch's base offset.
let mut pos = 0;
while let Some(tr) = iter.peek() {
if tr.role != crate::llm::Role::Tool {
break;
}
if let Some(tool_msg) = iter.next() {
let idx = call_base_idx + pos;
if idx < turn.tool_calls.len() {
// Store as result — the error/success distinction
// is for the live turn only; restored context just
// needs the content the LLM originally saw.
turn.tool_calls[idx].result =
Some(serde_json::Value::String(tool_msg.content.clone()));
}
}
pos += 1;
}
} else {
break;
}
}
// Check if next is the final assistant response for this turn
let is_final_assistant = iter.peek().is_some_and(|n| {
n.role == crate::llm::Role::Assistant && n.tool_calls.is_none()
});
if is_final_assistant && let Some(response) = iter.next() {
turn.complete(&response.content);
}
self.turns.push(turn);
turn_number += 1;
} else {
// Skip non-user messages that aren't anchored to a turn
continue;
}
}
@@ -1035,4 +1121,225 @@ mod tests {
ThreadState::Processing
);
}
// Regression tests for #568: tool call history must survive hydration.
#[test]
fn test_messages_includes_tool_calls() {
let mut thread = Thread::new(Uuid::new_v4());
thread.start_turn("Search for X");
{
let turn = thread.turns.last_mut().unwrap();
turn.record_tool_call("memory_search", serde_json::json!({"query": "X"}));
turn.record_tool_result(serde_json::json!("Found X in doc.md"));
}
thread.complete_turn("I found X in doc.md.");
let messages = thread.messages();
// user + assistant_with_tool_calls + tool_result + assistant = 4
assert_eq!(messages.len(), 4);
assert_eq!(messages[0].role, crate::llm::Role::User);
assert_eq!(messages[0].content, "Search for X");
assert_eq!(messages[1].role, crate::llm::Role::Assistant);
assert!(messages[1].tool_calls.is_some());
let tcs = messages[1].tool_calls.as_ref().unwrap();
assert_eq!(tcs.len(), 1);
assert_eq!(tcs[0].name, "memory_search");
assert_eq!(messages[2].role, crate::llm::Role::Tool);
assert!(messages[2].content.contains("Found X"));
assert_eq!(messages[3].role, crate::llm::Role::Assistant);
assert_eq!(messages[3].content, "I found X in doc.md.");
}
#[test]
fn test_messages_multiple_tool_calls_per_turn() {
let mut thread = Thread::new(Uuid::new_v4());
thread.start_turn("Do two things");
{
let turn = thread.turns.last_mut().unwrap();
turn.record_tool_call("echo", serde_json::json!({"msg": "a"}));
turn.record_tool_result(serde_json::json!("a"));
turn.record_tool_call("time", serde_json::json!({}));
turn.record_tool_error("timeout");
}
thread.complete_turn("Done.");
let messages = thread.messages();
// user + assistant_with_calls(2) + tool_result + tool_result + assistant = 5
assert_eq!(messages.len(), 5);
let tcs = messages[1].tool_calls.as_ref().unwrap();
assert_eq!(tcs.len(), 2);
// First tool: success
assert_eq!(messages[2].content, "a");
// Second tool: error (passed through directly, no wrapping)
assert!(messages[3].content.contains("timeout"));
}
#[test]
fn test_restore_from_messages_with_tool_calls() {
let mut thread = Thread::new(Uuid::new_v4());
// Build a message sequence with tool calls
let tc = ToolCall {
id: "call_0".to_string(),
name: "search".to_string(),
arguments: serde_json::json!({"q": "test"}),
};
let messages = vec![
ChatMessage::user("Find test"),
ChatMessage::assistant_with_tool_calls(None, vec![tc]),
ChatMessage::tool_result("call_0", "search", "result: found"),
ChatMessage::assistant("Found it."),
];
thread.restore_from_messages(messages);
assert_eq!(thread.turns.len(), 1);
let turn = &thread.turns[0];
assert_eq!(turn.user_input, "Find test");
assert_eq!(turn.tool_calls.len(), 1);
assert_eq!(turn.tool_calls[0].name, "search");
assert_eq!(
turn.tool_calls[0].result,
Some(serde_json::Value::String("result: found".to_string()))
);
assert_eq!(turn.response, Some("Found it.".to_string()));
}
#[test]
fn test_restore_from_messages_with_tool_error() {
let mut thread = Thread::new(Uuid::new_v4());
let tc = ToolCall {
id: "call_0".to_string(),
name: "http".to_string(),
arguments: serde_json::json!({}),
};
let messages = vec![
ChatMessage::user("Fetch URL"),
ChatMessage::assistant_with_tool_calls(None, vec![tc]),
ChatMessage::tool_result("call_0", "http", "Error: timeout"),
ChatMessage::assistant("The request timed out."),
];
thread.restore_from_messages(messages);
// restore_from_messages stores all tool content as result (not error),
// because it can't reliably distinguish errors from results that happen
// to start with "Error: ". The content is preserved for LLM context.
let turn = &thread.turns[0];
assert_eq!(
turn.tool_calls[0].result,
Some(serde_json::Value::String("Error: timeout".to_string()))
);
}
#[test]
fn test_messages_round_trip_with_tools() {
// Build a thread with tool calls, get messages(), restore, get messages() again
// The two message sequences should be equivalent.
let mut thread = Thread::new(Uuid::new_v4());
thread.start_turn("Do search");
{
let turn = thread.turns.last_mut().unwrap();
turn.record_tool_call("search", serde_json::json!({"q": "test"}));
turn.record_tool_result(serde_json::json!("found"));
}
thread.complete_turn("Here are results.");
let messages_original = thread.messages();
// Restore into a new thread
let mut thread2 = Thread::new(Uuid::new_v4());
thread2.restore_from_messages(messages_original.clone());
let messages_restored = thread2.messages();
// Same number of messages
assert_eq!(messages_original.len(), messages_restored.len());
// Same roles
for (orig, rest) in messages_original.iter().zip(messages_restored.iter()) {
assert_eq!(orig.role, rest.role);
}
// Same final response
assert_eq!(
messages_original.last().unwrap().content,
messages_restored.last().unwrap().content
);
}
#[test]
fn test_restore_multi_stage_tool_calls() {
let mut thread = Thread::new(Uuid::new_v4());
let tc1 = ToolCall {
id: "call_a".to_string(),
name: "search".to_string(),
arguments: serde_json::json!({"q": "data"}),
};
let tc2 = ToolCall {
id: "call_b".to_string(),
name: "write".to_string(),
arguments: serde_json::json!({"path": "out.txt"}),
};
let messages = vec![
ChatMessage::user("Find and save"),
ChatMessage::assistant_with_tool_calls(None, vec![tc1]),
ChatMessage::tool_result("call_a", "search", "found data"),
ChatMessage::assistant_with_tool_calls(None, vec![tc2]),
ChatMessage::tool_result("call_b", "write", "written"),
ChatMessage::assistant("Done, saved to out.txt"),
];
thread.restore_from_messages(messages);
assert_eq!(thread.turns.len(), 1);
let turn = &thread.turns[0];
assert_eq!(turn.tool_calls.len(), 2);
assert_eq!(turn.tool_calls[0].name, "search");
assert_eq!(turn.tool_calls[1].name, "write");
assert_eq!(
turn.tool_calls[0].result,
Some(serde_json::Value::String("found data".to_string()))
);
assert_eq!(
turn.tool_calls[1].result,
Some(serde_json::Value::String("written".to_string()))
);
assert_eq!(turn.response, Some("Done, saved to out.txt".to_string()));
}
#[test]
fn test_messages_truncates_large_tool_results() {
let mut thread = Thread::new(Uuid::new_v4());
thread.start_turn("Read big file");
{
let turn = thread.turns.last_mut().unwrap();
turn.record_tool_call("read_file", serde_json::json!({"path": "big.txt"}));
let big_result = "x".repeat(2000);
turn.record_tool_result(serde_json::json!(big_result));
}
thread.complete_turn("Here's the file content.");
let messages = thread.messages();
let tool_result_content = &messages[2].content;
assert!(
tool_result_content.len() <= 1010,
"Tool result should be truncated, got {} chars",
tool_result_content.len()
);
assert!(tool_result_content.ends_with("..."));
}
}
+291 -58
View File
@@ -20,7 +20,7 @@ use crate::channels::web::util::truncate_preview;
use crate::channels::{IncomingMessage, StatusUpdate};
use crate::context::JobContext;
use crate::error::Error;
use crate::llm::ChatMessage;
use crate::llm::{ChatMessage, ToolCall};
use crate::tools::redact_params;
impl Agent {
@@ -66,16 +66,7 @@ impl Agent {
.await
.unwrap_or_default();
msg_count = db_messages.len();
chat_messages = db_messages
.iter()
.filter_map(|m| match m.role.as_str() {
"user" => Some(ChatMessage::user(&m.content)),
"assistant" => Some(ChatMessage::assistant(&m.content)),
// tool_calls rows are UI metadata (tool name + preview),
// not part of the LLM conversation context.
_ => None,
})
.collect();
chat_messages = rebuild_chat_messages_from_db(&db_messages);
} else {
msg_count = 0;
}
@@ -340,10 +331,10 @@ impl Agent {
};
thread.complete_turn(&response);
let tool_calls = thread
let (turn_number, tool_calls) = thread
.turns
.last()
.map(|t| t.tool_calls.clone())
.map(|t| (t.turn_number, t.tool_calls.clone()))
.unwrap_or_default();
let _ = self
.channels
@@ -355,7 +346,7 @@ impl Agent {
.await;
// Persist tool calls then assistant response (user message already persisted at turn start)
self.persist_tool_calls(thread_id, &message.user_id, &tool_calls)
self.persist_tool_calls(thread_id, &message.user_id, turn_number, &tool_calls)
.await;
self.persist_assistant_response(thread_id, &message.user_id, &response)
.await;
@@ -464,6 +455,7 @@ impl Agent {
&self,
thread_id: Uuid,
user_id: &str,
turn_number: usize,
tool_calls: &[crate::agent::session::TurnToolCall],
) {
if tool_calls.is_empty() {
@@ -477,14 +469,24 @@ impl Agent {
let summaries: Vec<serde_json::Value> = tool_calls
.iter()
.map(|tc| {
let mut obj = serde_json::json!({ "name": tc.name });
.enumerate()
.map(|(i, tc)| {
let mut obj = serde_json::json!({
"name": tc.name,
"call_id": format!("turn{}_{}", turn_number, i),
});
if let Some(ref result) = tc.result {
let preview = match result {
serde_json::Value::String(s) => truncate_preview(s, 500),
other => truncate_preview(&other.to_string(), 500),
};
obj["result_preview"] = serde_json::Value::String(preview);
// Store full result (truncated to ~1000 chars) for LLM context rebuild
let full_result = match result {
serde_json::Value::String(s) => truncate_preview(s, 1000),
other => truncate_preview(&other.to_string(), 1000),
};
obj["result"] = serde_json::Value::String(full_result);
}
if let Some(ref error) = tc.error {
obj["error"] = serde_json::Value::String(truncate_preview(error, 200));
@@ -807,19 +809,33 @@ impl Agent {
let mut context_messages = pending.context_messages;
let deferred_tool_calls = pending.deferred_tool_calls;
// Record result in thread
// Sanitize tool result, then record the cleaned version in the
// thread. Must happen before auth intercept check which may return early.
let is_tool_error = tool_result.is_err();
let result_content = match &tool_result {
Ok(output) => {
let sanitized = self
.safety()
.sanitize_tool_output(&pending.tool_name, output);
self.safety().wrap_for_llm(
&pending.tool_name,
&sanitized.content,
sanitized.was_modified,
)
}
Err(e) => format!("Error: {}", e),
};
// Record sanitized result in thread
{
let mut sess = session.lock().await;
if let Some(thread) = sess.threads.get_mut(&thread_id)
&& let Some(turn) = thread.last_turn_mut()
{
match &tool_result {
Ok(output) => {
turn.record_tool_result(serde_json::json!(output));
}
Err(e) => {
turn.record_tool_error(e.to_string());
}
if is_tool_error {
turn.record_tool_error(result_content.clone());
} else {
turn.record_tool_result(serde_json::json!(result_content));
}
}
}
@@ -841,21 +857,6 @@ impl Agent {
return Ok(SubmissionResult::response(instructions));
}
// Add tool result to context
let result_content = match tool_result {
Ok(output) => {
let sanitized = self
.safety()
.sanitize_tool_output(&pending.tool_name, &output);
self.safety().wrap_for_llm(
&pending.tool_name,
&sanitized.content,
sanitized.was_modified,
)
}
Err(e) => format!("Error: {}", e),
};
context_messages.push(ChatMessage::tool_result(
&pending.tool_call_id,
&pending.tool_name,
@@ -1060,15 +1061,31 @@ impl Agent {
.await;
}
// Record in thread
// Sanitize first, then record the cleaned version in thread.
// Must happen before auth detection which may set deferred_auth.
let is_deferred_error = deferred_result.is_err();
let deferred_content = match &deferred_result {
Ok(output) => {
let sanitized = self.safety().sanitize_tool_output(&tc.name, output);
self.safety().wrap_for_llm(
&tc.name,
&sanitized.content,
sanitized.was_modified,
)
}
Err(e) => format!("Error: {}", e),
};
// Record sanitized result in thread
{
let mut sess = session.lock().await;
if let Some(thread) = sess.threads.get_mut(&thread_id)
&& let Some(turn) = thread.last_turn_mut()
{
match &deferred_result {
Ok(output) => turn.record_tool_result(serde_json::json!(output)),
Err(e) => turn.record_tool_error(e.to_string()),
if is_deferred_error {
turn.record_tool_error(deferred_content.clone());
} else {
turn.record_tool_result(serde_json::json!(deferred_content));
}
}
}
@@ -1090,18 +1107,6 @@ impl Agent {
deferred_auth = Some(instructions);
}
let deferred_content = match deferred_result {
Ok(output) => {
let sanitized = self.safety().sanitize_tool_output(&tc.name, &output);
self.safety().wrap_for_llm(
&tc.name,
&sanitized.content,
sanitized.was_modified,
)
}
Err(e) => format!("Error: {}", e),
};
context_messages.push(ChatMessage::tool_result(&tc.id, &tc.name, deferred_content));
}
@@ -1169,13 +1174,13 @@ impl Agent {
match result {
Ok(AgenticLoopResult::Response(response)) => {
thread.complete_turn(&response);
let tool_calls = thread
let (turn_number, tool_calls) = thread
.turns
.last()
.map(|t| t.tool_calls.clone())
.map(|t| (t.turn_number, t.tool_calls.clone()))
.unwrap_or_default();
// User message already persisted at turn start; save tool calls then assistant response
self.persist_tool_calls(thread_id, &message.user_id, &tool_calls)
self.persist_tool_calls(thread_id, &message.user_id, turn_number, &tool_calls)
.await;
self.persist_assistant_response(thread_id, &message.user_id, &response)
.await;
@@ -1490,3 +1495,231 @@ impl Agent {
}
}
}
/// Rebuild full LLM-compatible `ChatMessage` sequence from DB messages.
///
/// Parses `role="tool_calls"` rows to reconstruct `assistant_with_tool_calls`
/// and `tool_result` messages so that the LLM sees the complete tool execution
/// history on thread hydration. Falls back gracefully for legacy rows that
/// lack the enriched fields (`call_id`, `parameters`, `result`).
fn rebuild_chat_messages_from_db(
db_messages: &[crate::history::ConversationMessage],
) -> Vec<ChatMessage> {
let mut result = Vec::new();
for msg in db_messages {
match msg.role.as_str() {
"user" => result.push(ChatMessage::user(&msg.content)),
"assistant" => result.push(ChatMessage::assistant(&msg.content)),
"tool_calls" => {
// Try to parse the enriched JSON and rebuild tool messages.
if let Ok(calls) = serde_json::from_str::<Vec<serde_json::Value>>(&msg.content) {
if calls.is_empty() {
continue;
}
// Check if this is an enriched row (has call_id) or legacy
let has_call_id = calls
.first()
.and_then(|c| c.get("call_id"))
.and_then(|v| v.as_str())
.is_some();
if has_call_id {
// Build assistant_with_tool_calls + tool_result messages
let tool_calls: Vec<ToolCall> = calls
.iter()
.map(|c| ToolCall {
id: c["call_id"].as_str().unwrap_or("call_0").to_string(),
name: c["name"].as_str().unwrap_or("unknown").to_string(),
arguments: c
.get("parameters")
.cloned()
.unwrap_or(serde_json::json!({})),
})
.collect();
// The assistant text for tool_calls is always None here;
// the final assistant response comes as a separate
// "assistant" row after this tool_calls row.
result.push(ChatMessage::assistant_with_tool_calls(None, tool_calls));
// Emit tool_result messages for each call
for c in &calls {
let call_id = c["call_id"].as_str().unwrap_or("call_0").to_string();
let name = c["name"].as_str().unwrap_or("unknown").to_string();
let content = if let Some(err) = c.get("error").and_then(|v| v.as_str())
{
format!("Error: {}", err)
} else if let Some(res) = c.get("result").and_then(|v| v.as_str()) {
res.to_string()
} else if let Some(preview) =
c.get("result_preview").and_then(|v| v.as_str())
{
preview.to_string()
} else {
"OK".to_string()
};
result.push(ChatMessage::tool_result(call_id, name, content));
}
}
// Legacy rows without call_id: skip (will appear as
// simple user/assistant pairs, same as before this fix).
}
}
_ => {} // Skip unknown roles
}
}
result
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_rebuild_chat_messages_user_assistant_only() {
let messages = vec![
make_db_msg("user", "Hello"),
make_db_msg("assistant", "Hi there!"),
];
let result = rebuild_chat_messages_from_db(&messages);
assert_eq!(result.len(), 2);
assert_eq!(result[0].role, crate::llm::Role::User);
assert_eq!(result[1].role, crate::llm::Role::Assistant);
}
#[test]
fn test_rebuild_chat_messages_with_enriched_tool_calls() {
let tool_json = serde_json::json!([
{
"name": "memory_search",
"call_id": "call_0",
"parameters": {"query": "test"},
"result": "Found 3 results",
"result_preview": "Found 3 re..."
},
{
"name": "echo",
"call_id": "call_1",
"parameters": {"message": "hi"},
"error": "timeout"
}
]);
let messages = vec![
make_db_msg("user", "Search for test"),
make_db_msg("tool_calls", &tool_json.to_string()),
make_db_msg("assistant", "I found some results."),
];
let result = rebuild_chat_messages_from_db(&messages);
// user + assistant_with_tool_calls + tool_result*2 + assistant
assert_eq!(result.len(), 5);
// user
assert_eq!(result[0].role, crate::llm::Role::User);
// assistant with tool_calls
assert_eq!(result[1].role, crate::llm::Role::Assistant);
assert!(result[1].tool_calls.is_some());
let tcs = result[1].tool_calls.as_ref().unwrap();
assert_eq!(tcs.len(), 2);
assert_eq!(tcs[0].name, "memory_search");
assert_eq!(tcs[0].id, "call_0");
assert_eq!(tcs[1].name, "echo");
// tool results
assert_eq!(result[2].role, crate::llm::Role::Tool);
assert_eq!(result[2].tool_call_id, Some("call_0".to_string()));
assert!(result[2].content.contains("Found 3 results"));
assert_eq!(result[3].role, crate::llm::Role::Tool);
assert_eq!(result[3].tool_call_id, Some("call_1".to_string()));
assert!(result[3].content.contains("Error: timeout"));
// final assistant
assert_eq!(result[4].role, crate::llm::Role::Assistant);
assert_eq!(result[4].content, "I found some results.");
}
#[test]
fn test_rebuild_chat_messages_legacy_tool_calls_skipped() {
// Legacy format: no call_id field
let tool_json = serde_json::json!([
{"name": "echo", "result_preview": "hello"}
]);
let messages = vec![
make_db_msg("user", "Hi"),
make_db_msg("tool_calls", &tool_json.to_string()),
make_db_msg("assistant", "Done"),
];
let result = rebuild_chat_messages_from_db(&messages);
// Legacy rows are skipped, only user + assistant
assert_eq!(result.len(), 2);
assert_eq!(result[0].role, crate::llm::Role::User);
assert_eq!(result[1].role, crate::llm::Role::Assistant);
}
#[test]
fn test_rebuild_chat_messages_empty() {
let result = rebuild_chat_messages_from_db(&[]);
assert!(result.is_empty());
}
#[test]
fn test_rebuild_chat_messages_malformed_tool_calls_json() {
let messages = vec![
make_db_msg("user", "Hi"),
make_db_msg("tool_calls", "not valid json"),
make_db_msg("assistant", "Done"),
];
let result = rebuild_chat_messages_from_db(&messages);
// Malformed JSON is silently skipped
assert_eq!(result.len(), 2);
}
#[test]
fn test_rebuild_chat_messages_multi_turn_with_tools() {
let tool_json_1 = serde_json::json!([
{"name": "search", "call_id": "call_0", "parameters": {}, "result": "found it"}
]);
let tool_json_2 = serde_json::json!([
{"name": "write", "call_id": "call_0", "parameters": {"path": "a.txt"}, "result": "ok"}
]);
let messages = vec![
make_db_msg("user", "Find X"),
make_db_msg("tool_calls", &tool_json_1.to_string()),
make_db_msg("assistant", "Found X"),
make_db_msg("user", "Write it"),
make_db_msg("tool_calls", &tool_json_2.to_string()),
make_db_msg("assistant", "Written"),
];
let result = rebuild_chat_messages_from_db(&messages);
// Turn 1: user + assistant_with_calls + tool_result + assistant = 4
// Turn 2: user + assistant_with_calls + tool_result + assistant = 4
assert_eq!(result.len(), 8);
// Verify turn boundaries
assert_eq!(result[0].content, "Find X");
assert!(result[1].tool_calls.is_some());
assert_eq!(result[2].role, crate::llm::Role::Tool);
assert_eq!(result[3].content, "Found X");
assert_eq!(result[4].content, "Write it");
assert!(result[5].tool_calls.is_some());
assert_eq!(result[6].role, crate::llm::Role::Tool);
assert_eq!(result[7].content, "Written");
}
fn make_db_msg(role: &str, content: &str) -> crate::history::ConversationMessage {
crate::history::ConversationMessage {
id: uuid::Uuid::new_v4(),
role: role.to_string(),
content: content.to_string(),
created_at: chrono::Utc::now(),
}
}
}
+39 -1
View File
@@ -3,6 +3,10 @@
use crate::channels::web::types::{ToolCallInfo, TurnInfo};
/// Truncate a string to at most `max_bytes` bytes at a char boundary, appending "...".
///
/// If the input is wrapped in `<tool_output …>…</tool_output>` and truncation
/// removes the closing tag, the tag is re-appended so downstream XML parsers
/// never see an unclosed element.
pub fn truncate_preview(s: &str, max_bytes: usize) -> String {
if s.len() <= max_bytes {
return s.to_string();
@@ -12,7 +16,14 @@ pub fn truncate_preview(s: &str, max_bytes: usize) -> String {
while end > 0 && !s.is_char_boundary(end) {
end -= 1;
}
format!("{}...", &s[..end])
let mut result = format!("{}...", &s[..end]);
// Re-close <tool_output> if truncation cut through the closing tag.
if s.starts_with("<tool_output") && !result.ends_with("</tool_output>") {
result.push_str("\n</tool_output>");
}
result
}
/// Build TurnInfo pairs from flat DB messages (user/tool_calls/assistant triples).
@@ -162,6 +173,33 @@ mod tests {
assert_eq!(truncate_preview("hello", 0), "...");
}
#[test]
fn test_truncate_preview_closes_tool_output_tag() {
let s = "<tool_output name=\"search\" sanitized=\"true\">\nSome very long content here\n</tool_output>";
// Truncate so it cuts before the closing tag
let result = truncate_preview(s, 60);
assert!(result.ends_with("</tool_output>"));
assert!(result.contains("..."));
}
#[test]
fn test_truncate_preview_no_extra_close_when_intact() {
let s = "<tool_output name=\"echo\" sanitized=\"false\">\nshort\n</tool_output>";
// The string is short enough not to be truncated
let result = truncate_preview(s, 500);
assert_eq!(result, s);
// Should not have a duplicate closing tag
assert_eq!(result.matches("</tool_output>").count(), 1);
}
#[test]
fn test_truncate_preview_non_xml_unaffected() {
let s = "Just a plain long string that gets truncated";
let result = truncate_preview(s, 10);
assert_eq!(result, "Just a pla...");
assert!(!result.contains("</tool_output>"));
}
// ---- build_turns_from_db_messages tests ----
fn make_msg(role: &str, content: &str, offset_ms: i64) -> crate::history::ConversationMessage {