Version 2.1.165

Previously the summarizer that compresses long conversations always ran with thinkingConfig: { type: "disabled" }, regardless of what model was active. In this version a new capability check is evaluated at summarization time: if the current main-loop model is listed in the server-side cedar_lagoon capability map, the summarizer now inherits the same thinking configuration used by the main conversation loop. For models that are not in the list, the old behavior (thinking disabled) is unchanged.

In practice this means users running a model that supports extended thinking may notice more coherent or accurate conversation compaction when context windows fill up, since the summarizer can now reason before producing the summary.

New capability gate function (searches clientDataCache?.cedar_lagoon); summarization call site changed from hardcoded { type: "disabled" } to hM8(K.options.mainLoopModel) ? K.options.thinkingConfig : { type: "disabled" } — search for "cedar_lagoon" to locate the check.

The x-anthropic-billing-header sent with every API request now includes cc_is_subagent=true; when the request originates from a subagent session rather than the top-level interactive session. The main CLI session is tagged isMainSession: true internally; any agent that lacks that flag gets the subagent marker appended.

This is an internal attribution change with no user-visible behavior. It helps Anthropic distinguish API usage generated directly by the user from usage generated by orchestrated subagents, which improves billing analytics and usage reporting.

Attribution header builder now appends cc_is_subagent=true; for non-main-session callers (search for "cc_is_subagent=true;" in the source); main session agent config now sets isMainSession: !0.