fix(agent): strip <think> blocks from stored assistant content

Inline reasoning tags in an assistant message's content field leak to every downstream consumer: messaging platforms (#8878, #9568), API replay of prior turns, session transcript, CLI recap, generated session titles, and context compression.  _extract_reasoning() already captures the reasoning text into msg['reasoning'] separately, so the raw tags in content are redundant.

Stripping once at the storage boundary in _build_assistant_message() cleans the content for every downstream path in one place — no per-platform or per-path stripper needed.  Measured impact on a real MiniMax M2.7-highspeed session (per @luoyejiaoe-source, #9306): 55% of assistant messages started with <think> blocks, 51/100 session titles were polluted, 16% content-size reduction.

3 new regression tests in TestBuildAssistantMessage: closed-pair strip with reasoning capture, no-think-tag passthrough, and unterminated-block strip.

Resolves #8878 and #9568.

Originally proposed as PR #9250.
This commit is contained in:
Tranquil-Flow 2026-04-18 19:18:03 -07:00 committed by Teknium
parent 9489d1577d
commit ec48ec5530
2 changed files with 49 additions and 0 deletions

View File

@ -7294,6 +7294,20 @@ class AIAgent:
if reasoning_text:
reasoning_text = _sanitize_surrogates(reasoning_text)
# Strip inline reasoning tags (<think>…</think> etc.) from the stored
# assistant content. Reasoning was already captured into
# ``reasoning_text`` above (either from structured fields or the
# inline-block fallback), so the raw tags in content are redundant.
# Leaving them in place caused reasoning to leak to messaging
# platforms (#8878, #9568), inflate context on subsequent turns
# (#9306 observed 16% content-size reduction on a real MiniMax
# session), and pollute generated session titles. One strip at the
# storage boundary cleans content for every downstream consumer:
# API replay, session transcript, gateway delivery, CLI display,
# compression, title generation.
if isinstance(_san_content, str) and _san_content:
_san_content = self._strip_think_blocks(_san_content).strip()
msg = {
"role": "assistant",
"content": _san_content,

View File

@ -1142,6 +1142,41 @@ class TestBuildAssistantMessage:
result = agent._build_assistant_message(msg, "tool_calls")
assert "extra_content" not in result["tool_calls"][0]
def test_think_blocks_stripped_from_content(self, agent):
"""Inline <think> blocks are stripped from stored content (#8878, #9568).
The reasoning is captured into ``msg['reasoning']`` via the inline
fallback in ``_extract_reasoning``; the raw tags in ``content`` are
redundant and leak to messaging platforms / pollute titles /
inflate context if left in place.
"""
msg = _mock_assistant_msg(
content="<think>internal reasoning</think>The actual answer."
)
result = agent._build_assistant_message(msg, "stop")
assert "<think>" not in result["content"]
assert "internal reasoning" not in result["content"]
assert "The actual answer." in result["content"]
# Reasoning preserved separately via inline extraction fallback
assert result["reasoning"] == "internal reasoning"
def test_think_blocks_stripped_preserves_normal_content(self, agent):
"""Content without reasoning tags passes through unchanged."""
msg = _mock_assistant_msg(content="No thinking here.")
result = agent._build_assistant_message(msg, "stop")
assert result["content"] == "No thinking here."
def test_unterminated_think_block_stripped(self, agent):
"""Unterminated <think> block (MiniMax / NIM dropped close tag) is
fully stripped from stored content."""
msg = _mock_assistant_msg(
content="<think>reasoning that never closes on this NIM endpoint"
)
result = agent._build_assistant_message(msg, "stop")
assert "<think>" not in result["content"]
assert "reasoning that never closes" not in result["content"]
assert result["content"] == ""
class TestFormatToolsForSystemMessage:
def test_no_tools_returns_empty_array(self, agent):