[Bug] Advisor call re-injects full executor transcript, doubling context window usage

Bug: Advisor call doubles context window usage, instantly hitting 1M limit

Running Opus 4.6 [1m] with Opus 4.7 as advisor via /advisor. When context reaches ~50% (~485k tokens), a single advisor call causes an instant jump to 97% (971.6k/1m). Post-bug /context breakdown:

Messages: 942.5k tokens (94.3%)
System prompt: 6.7k, System tools: 8k, MCP tools: 5.1k
Memory files: 7.5k, Skills: 1.7k, Custom agents: 56
Free space: 25.4k (2.5%)

The Messages category roughly doubled from ~470k to 942.5k in a single advisor call. This is consistent with the full executor transcript (~470k of messages) being re-injected into the conversation as part of the advisor round-trip, rather than only the small advisor_tool_result block being appended. Per the API docs, the advisor runs a separate sub-inference on the full transcript server-side — those tokens should not inflate the executor's message history. Autocompact has no opportunity to intervene since the jump is atomic. Session is effectively killed. Reproducible every time context crosses ~50% with advisor enabled. Related: #37926 (same instant-jump pattern with background agents).

Environment Info

• Platform: linux
• Terminal: konsole
• Version: 2.1.126
• Feedback ID: 26a2ffb7-0c41-4379-9717-5bddc6cec8cf