admin/openfang
1
0
Fork 0
mirror of https://github.com/RightNow-AI/openfang.git synced 2026-07-30 15:01:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: capture text from intermediate tool_use iterations

Browse Source 
When an LLM produces text alongside tool_use blocks (e.g., a chat
message followed by memory_store calls), the text was lost if the
final EndTurn iteration returned empty text. The empty-response guard
would activate and return "[Task completed — the agent executed tools
but did not produce a text summary.]" even though the agent DID
produce text in an earlier iteration.

This is a common pattern when agents are instructed to respond to
users AND persist state via memory_store in the same turn.

Fix: accumulate text content from all ToolUse iterations. When the
final EndTurn has empty text, use accumulated text as fallback before
triggering the empty-response guard.

Applied to both sync and streaming agent loop paths.
This commit is contained in:
default
2026-04-05 11:17:10 +00:00
parent 07963779be
commit 0408d65d8f
1 changed files with 69 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
crates/openfang-runtime/src/agent_loop.rs
+69 -24
View File
@@ -332,6 +332,10 @@ pub async fn run_agent_loop(
let mut total_usage = TokenUsage::default();
let final_response;
// Accumulate text from intermediate iterations (tool_use turns may include text
// alongside tool calls — this text would otherwise be lost when the final
// EndTurn iteration has empty text).
let mut accumulated_text = String::new();
// Safety valve: trim excessively long message histories to prevent context overflow.
// The full compaction system handles sophisticated summarization, but this prevents
@@ -524,20 +528,30 @@ pub async fn run_agent_loop(
}
}
// Guard against empty response — covers both iteration 0 and post-tool cycles
// Guard against empty response — covers both iteration 0 and post-tool cycles.
// Use accumulated_text from intermediate tool_use iterations as fallback.
let text = if text.trim().is_empty() {
warn!(
agent = %manifest.name,
iteration,
input_tokens = total_usage.input_tokens,
output_tokens = total_usage.output_tokens,
messages_count = messages.len(),
"Empty response from LLM — guard activated"
);
if any_tools_executed {
"[Task completed — the agent executed tools but did not produce a text summary.]".to_string()
if !accumulated_text.is_empty() {
debug!(
agent = %manifest.name,
accumulated_len = accumulated_text.len(),
"Using accumulated text from intermediate tool_use iterations"
);
accumulated_text.clone()
} else {
"[The model returned an empty response. This usually means the model is overloaded, the context is too large, or the API key lacks credits. Try again or check /status.]".to_string()
warn!(
agent = %manifest.name,
iteration,
input_tokens = total_usage.input_tokens,
output_tokens = total_usage.output_tokens,
messages_count = messages.len(),
"Empty response from LLM — guard activated"
);
if any_tools_executed {
"[Task completed — the agent executed tools but did not produce a text summary.]".to_string()
} else {
"[The model returned an empty response. This usually means the model is overloaded, the context is too large, or the API key lacks credits. Try again or check /status.]".to_string()
}
}
} else {
text
@@ -658,6 +672,18 @@ pub async fn run_agent_loop(
consecutive_max_tokens = 0;
any_tools_executed = true;
// Capture any text content from this tool_use turn — the LLM may
// produce text alongside tool calls (e.g., a message to the user
// before calling memory_store). Without this, the text is lost if
// the next iteration returns EndTurn with empty text.
let intermediate_text = response.text();
if !intermediate_text.trim().is_empty() {
if !accumulated_text.is_empty() {
accumulated_text.push_str("\n\n");
}
accumulated_text.push_str(intermediate_text.trim());
}
// Execute tool calls
let assistant_blocks = response.content.clone();
@@ -1497,6 +1523,7 @@ pub async fn run_agent_loop_streaming(
let mut total_usage = TokenUsage::default();
let final_response;
let mut accumulated_text = String::new();
// Safety valve: trim excessively long message histories to prevent context overflow.
if messages.len() > MAX_HISTORY_MESSAGES {
@@ -1703,20 +1730,29 @@ pub async fn run_agent_loop_streaming(
}
}
// Guard against empty response — covers both iteration 0 and post-tool cycles
// Guard against empty response — use accumulated text as fallback (streaming).
let text = if text.trim().is_empty() {
warn!(
agent = %manifest.name,
iteration,
input_tokens = total_usage.input_tokens,
output_tokens = total_usage.output_tokens,
messages_count = messages.len(),
"Empty response from LLM (streaming) — guard activated"
);
if any_tools_executed {
"[Task completed — the agent executed tools but did not produce a text summary.]".to_string()
if !accumulated_text.is_empty() {
debug!(
agent = %manifest.name,
accumulated_len = accumulated_text.len(),
"Using accumulated text from intermediate tool_use iterations (streaming)"
);
accumulated_text.clone()
} else {
"[The model returned an empty response. This usually means the model is overloaded, the context is too large, or the API key lacks credits. Try again or check /status.]".to_string()
warn!(
agent = %manifest.name,
iteration,
input_tokens = total_usage.input_tokens,
output_tokens = total_usage.output_tokens,
messages_count = messages.len(),
"Empty response from LLM (streaming) — guard activated"
);
if any_tools_executed {
"[Task completed — the agent executed tools but did not produce a text summary.]".to_string()
} else {
"[The model returned an empty response. This usually means the model is overloaded, the context is too large, or the API key lacks credits. Try again or check /status.]".to_string()
}
}
} else {
text
@@ -1816,6 +1852,15 @@ pub async fn run_agent_loop_streaming(
consecutive_max_tokens = 0;
any_tools_executed = true;
// Capture text from intermediate tool_use turns (streaming path).
let intermediate_text = response.text();
if !intermediate_text.trim().is_empty() {
if !accumulated_text.is_empty() {
accumulated_text.push_str("\n\n");
}
accumulated_text.push_str(intermediate_text.trim());
}
let assistant_blocks = response.content.clone();
session.messages.push(Message {