Requests and compaction gain advisorStripPlan, reminderFold and declaredToolNames

What

Before each request to the model, Claude Code rebuilds the conversation into the format the API expects. That step gains three new options:

• advisorStripPlan: removing advisor tool blocks (advisor_tool_result and advisor server_tool_use) can now be limited to a chosen set of assistant messages. Before, they were removed from every assistant message. A message left empty still gets the text "[Advisor response]". This works like the existing thinkingStripPlan.
• reminderFold: a setting for how reminder messages are folded into turns, now passed into this rebuilding step.
• declaredToolNames: added next to keptToolNames to the set of tool names whose tool_reference blocks are kept.

Compaction, which summarises a long conversation to free up room, now uses the same options. Compact and reactive-compact runs pass advisorStripPlan next to thinkingStripPlan. The compact summariser also computes a reminder fold and passes advisorStripPlan and reminderFold into the request. Its fallback, refusal and credit handling is unchanged. Reuse of the existing prompt during compaction is still controlled by a separate server flag.

Why

Advisor output can now be kept for some turns and dropped for others, instead of all or nothing. Compaction requests are now shaped the same way as normal requests, so the conversation being summarised is prepared by the same rules.