Two days into an investigation session, a large refactor is merged that moves and rewrites most of the files the agent had read. After resuming, the agent repeatedly cites functions that no longer exist. What is the most reliable way forward?

Answer: A — Start a new session seeded with a structured summary of durable findings, and let it read the current code

When most prior tool results are stale, starting a new session with an injected structured summary is more reliable than resuming. Correcting the agent repeatedly (b) leaves the stale history in place, telling it to ignore its tool results (c) still keeps them in context, and forking (d) copies the same stale analysis into the new branch.