andre/crewAI
1
0
Fork 0
mirror of https://github.com/crewAIInc/crewAI.git synced 2026-07-01 05:08:12 +00:00
Code Issues Actions 6 Packages Projects Releases Wiki Activity

fix: summary tag handling, TUI autocomplete focus, tool output flooding

Browse Source 
1. Summary tags: Reverse the logic — for models like gpt-4.1 that wrap
   their actual response in <summary> tags (with thinking/CoT before it),
   extract the inner content instead of stripping it. Streaming uses a
   preflight buffer that waits for <summary>; if none appears, flushes
   everything normally.

2. TUI autocomplete: Change @mention accept key from Tab to right-arrow
   so autocomplete doesn't steal focus from the input widget. Only
   triggers when there's an active mention context with matches.

3. Tool output: Truncate tool results >4000 chars in LLM message history
   to prevent the model from echoing full file contents. Add soul-layer
   instruction telling the agent to summarize tool results rather than
   repeating them verbatim.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Joao Moura
2026-05-14 16:17:44 -04:00
parent 3eb1da2d9c
commit 126d0010ba
2 changed files with 97 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
lib/cli/src/crewai_cli/agent_tui.py
View File

@@ -71,7 +71,7 @@ except ImportError:
class ChatTextArea(TextArea):
"""Multiline chat input: Enter submits, Shift+Enter inserts newline, Tab completes @mentions."""
"""Multiline chat input: Enter submits, Shift+Enter inserts newline, Right-arrow completes @mentions."""
BINDINGS = [
Binding("enter", "submit", "Send", show=False),
@@ -161,10 +161,14 @@ class ChatTextArea(TextArea):
event.prevent_default()
self.action_submit()
return
if event.key == "tab":
event.prevent_default()
self.action_complete()
return
if event.key == "right":
ctx = self._get_mention_context()
if ctx is not None:
_, _, prefix = ctx
if self._get_matches(prefix):
event.prevent_default()
self.action_complete()
return
if event.key == "escape":
self._last_mention_prefix = None
self.post_message(self.MentionChanged("", []))
@@ -834,7 +838,7 @@ class AgentTUI(App[None]):
hint.display = False
return
names = " ".join(f"@{n}" for n in event.matches[:6])
hint.update(f"Tab to complete: {names}")
hint.update(f"→ to complete: {names}")
hint.display = True
# ── Message routing ──

115
lib/crewai/src/crewai/new_agent/executor.py
View File

@@ -127,7 +127,10 @@ class ConversationalAgentExecutor(BaseModel):
soul = (
f"You are {agent.role}.\n"
f"Your goal: {agent.goal}\n"
f"Background: {agent.backstory}"
f"Background: {agent.backstory}\n\n"
"When you use tools, act on the results directly. "
"Never repeat raw tool output (file contents, command output, etc.) "
"in your response — summarize findings or state what you did instead."
)
stack.add("soul", soul, source="agent.role/goal/backstory")
@@ -629,14 +632,22 @@ class ConversationalAgentExecutor(BaseModel):
pass
return ""
_INTERNAL_TAG_RE = re.compile(
r"<summary>.*?</summary>", re.DOTALL
_SUMMARY_EXTRACT_RE = re.compile(
r"<summary>\s*(.*?)\s*</summary>", re.DOTALL
)
def _strip_internal_tags(self, text: str) -> str:
"""Strip <summary> blocks that leak from the summarization prompt."""
cleaned = self._INTERNAL_TAG_RE.sub("", text).strip()
return cleaned if cleaned else text
"""Handle <summary> tags leaked from the summarization prompt.
Some models (e.g. gpt-4.1) wrap their actual response in <summary>
tags, treating the preceding text as chain-of-thought. When summary
tags are present we extract the inner content as the real answer.
"""
match = self._SUMMARY_EXTRACT_RE.search(text)
if match:
inner = match.group(1).strip()
return inner if inner else text
return text
def _detect_artifacts(self, tool_name: str, result_str: str) -> list[Artifact]:
"""GAP-67: Detect artifacts from tool results.
@@ -2064,11 +2075,19 @@ class ConversationalAgentExecutor(BaseModel):
],
}
)
_MAX_TOOL_RESULT_CHARS = 4000
display_result = result_str
if len(result_str) > _MAX_TOOL_RESULT_CHARS:
display_result = (
result_str[:_MAX_TOOL_RESULT_CHARS]
+ f"\n\n[Truncated — {len(result_str)} chars total. "
"Use the result directly; do not repeat it in your response.]"
)
llm_messages.append(
{
"role": "tool",
"tool_call_id": call_id or func_name,
"content": result_str,
"content": display_result,
}
)
@@ -2400,39 +2419,77 @@ class ConversationalAgentExecutor(BaseModel):
invoke_task = asyncio.create_task(self.ainvoke(user_message))
_streamed_chars = 0
_last_status_time = time.monotonic()
# States: "preflight" (buffering, no <summary> seen yet),
# "inside" (<summary> found, yielding inner content),
# "done" (</summary> seen, suppress the rest),
# "passthrough" (no summary tags — yield everything)
_state = "preflight"
_preflight_buf: list[str] = []
_tag_buf = ""
_suppressing = False
def _filter_chunk(raw: str) -> str:
"""Filter <summary>...</summary> blocks from streamed chunks."""
nonlocal _tag_buf, _suppressing
out = []
nonlocal _state, _tag_buf
out: list[str] = []
for ch in raw:
if _suppressing:
_tag_buf += ch
if _tag_buf.endswith("</summary>"):
_suppressing = False
_tag_buf = ""
if _state == "done":
break
if _state == "passthrough":
out.append(ch)
continue
# Accumulate into tag buffer when we might be mid-tag
if _tag_buf:
_tag_buf += ch
if len(_tag_buf) <= len("<summary>"):
if "<summary>"[: len(_tag_buf)] == _tag_buf:
if _tag_buf == "<summary>":
_suppressing = True
if _state == "preflight":
target = "<summary>"
if target[: len(_tag_buf)] == _tag_buf:
if _tag_buf == target:
_state = "inside"
_preflight_buf.clear()
_tag_buf = ""
continue
else:
out.append(_tag_buf)
_tag_buf = ""
else:
# Not a match — dump tag_buf to preflight buffer
_preflight_buf.append(_tag_buf)
_tag_buf = ""
elif _state == "inside":
target = "</summary>"
if target[: len(_tag_buf)] == _tag_buf:
if _tag_buf == target:
_state = "done"
_tag_buf = ""
continue
# Not a closing tag — flush buffered chars
out.append(_tag_buf)
_tag_buf = ""
elif ch == "<":
continue
if ch == "<":
_tag_buf = ch
else:
continue
if _state == "preflight":
_preflight_buf.append(ch)
elif _state == "inside":
out.append(ch)
return "".join(out)
def _flush_preflight() -> str:
"""If we never saw <summary>, flush buffered text."""
nonlocal _state
if _state == "preflight":
_state = "passthrough"
result = "".join(_preflight_buf)
_preflight_buf.clear()
if _tag_buf:
result += _tag_buf
return result
return ""
try:
while not invoke_task.done():
try:
@@ -2456,8 +2513,10 @@ class ConversationalAgentExecutor(BaseModel):
_streamed_chars += len(filtered)
yield filtered
if _tag_buf and not _suppressing:
yield _tag_buf
leftover = _flush_preflight()
if leftover:
_streamed_chars += len(leftover)
yield leftover
result = invoke_task.result()
self._last_stream_result = result