entropic::StreamThinkFilter Class Reference

Streaming filter that removes <think> blocks from output. More...

#include <entropic/core/stream_think_filter.h>

Public Member Functions
	StreamThinkFilter (TokenCallback cb, void *ud)
	Construct with consumer callback.
	StreamThinkFilter (TokenCallback cb, void *ud, std::string open_marker, std::string close_marker)
	Construct with family-specific reasoning delimiters (gh#108).
void	set_raw_callback (TokenCallback cb, void *ud)
	Set optional raw callback (receives ALL tokens unfiltered).
void	on_token (const char *chunk, size_t len)
	Process a chunk of tokens.
void	flush ()
	Flush any buffered partial tag content.

Detailed Description

Streaming filter that removes <think> blocks from output.

Processes tokens character-by-character. Content inside <think>...</think> tags is suppressed from the consumer callback but always forwarded to the raw callback (if set).

Version

2.0.1

Definition at line 36 of file stream_think_filter.h.

Constructor & Destructor Documentation

StreamThinkFilter() [1/2]

entropic::StreamThinkFilter::StreamThinkFilter	(	TokenCallback	cb,
		void *	ud
	)

Construct with consumer callback.

Parameters

cb	Consumer token callback (receives filtered output).
ud	Consumer user_data. @req REQ-LOOP-008

Version

2.0.1

Parameters

cb	Consumer callback.
ud	Consumer user_data. @req REQ-LOOP-008

Version

2.0.1

Definition at line 22 of file stream_think_filter.cpp.

StreamThinkFilter() [2/2]

entropic::StreamThinkFilter::StreamThinkFilter	(	TokenCallback	cb,
		void *	ud,
		std::string	open_marker,
		std::string	close_marker
	)

Construct with family-specific reasoning delimiters (gh#108).

Thinking format is a property of the chat template, not the transport: Qwen/Nemotron use <think>…</think>, Gemma-4 QAT uses <|channel>…<channel|>. Before v2.10.3 this filter hardcoded the <think> pair, so a gemma4 stream passed raw reasoning straight to the consumer — and on the agent-loop streaming path this filter is the ONLY defense, because ResponseGenerator::generate_streaming builds content from its own token accumulator rather than the parsed result.

Markers come from ChatAdapter::thinking_markers(); they are passed as plain strings so that core does not depend on inference.

Parameters

cb	Consumer token callback (receives filtered output).
ud	Consumer user_data.
open_marker	Reasoning-block opening delimiter.
close_marker	Reasoning-block closing delimiter. @req REQ-LOOP-008

Version

2.10.3

Parameters

cb	Consumer callback.
ud	Consumer user_data.
open_marker	Opening delimiter.
close_marker	Closing delimiter. @req REQ-LOOP-008

Version

2.10.3

Definition at line 34 of file stream_think_filter.cpp.

Member Function Documentation

flush()

void entropic::StreamThinkFilter::flush

(

)

Flush any buffered partial tag content.

Flush buffered partial tag and UTF-8 content.

Call at end of generation to ensure no content is lost in the tag accumulator.

@req REQ-LOOP-008

Version

2.0.1

@req REQ-LOOP-008

Version

2.0.2

Definition at line 182 of file stream_think_filter.cpp.

on_token()

void entropic::StreamThinkFilter::on_token	(	const char *	chunk,
		size_t	len
	)

Process a chunk of tokens.

Process an incoming token through the think-tag filter.

Parameters

chunk	Token data.
len	Token byte length. @req REQ-LOOP-008

Version

2.0.1

Parameters

chunk	Token byte data.
len	Byte length of chunk. @req REQ-LOOP-008

Version

2.0.2

Definition at line 168 of file stream_think_filter.cpp.

set_raw_callback()

void entropic::StreamThinkFilter::set_raw_callback	(	TokenCallback	cb,
		void *	ud
	)

Set optional raw callback (receives ALL tokens unfiltered).

Set raw (unfiltered) callback.

Parameters

cb	Raw callback.
ud	Raw user_data. @req REQ-LOOP-008

Version

2.0.1

Definition at line 49 of file stream_think_filter.cpp.

---

The documentation for this class was generated from the following files:

• entropic/core/stream_think_filter.h
• core/stream_think_filter.cpp