interlens.reasoning_visibility¶

Bases: str, Enum

Controls whether a participant's prior <think> reasoning is re-injected into views on later turns.

Reasoning is always parsed out of Message.content into metadata['parsed_think'] (so it is available for interpretability regardless of this setting). This enum only governs history re-injection:

• STRIP (default): prior reasoning is dropped from all views — matches R1/Qwen3 chat templates, keeps reasoning a genuinely private scratchpad the other participant never sees, and is always template-safe.
• SELF_RETAIN: a participant sees its own prior reasoning re-injected into its view (visible-scratchpad self-continuation experiments).
• SHARED: all participants see all prior reasoning (shared-CoT collaboration/debate experiments).

Non-STRIP modes re-inject reasoning as tagged text at render time so templates that reject a native prior <think> turn don't error.