Record the durable session before spawning, then run a uniform `silverwood spawn from-id <sid> --workstream <ws>` for every tab — silverwood now decides the kind and first-run vs resume. Forward CLAUDE_CONFIG_DIR in the spawn seed so it can locate Claude's transcripts. Retire the automatic orphan-detection that record-first makes obsolete: the buildNode transient-tab projection, the NO_CONVERSATION_MARKER scraping and on-exit doctoring, the POST …/doctor endpoint and its silverwood.ts wrapper, and the client's orphan-delete UI (disconnectReason/doctorKind). A never-prompted claude session now simply starts fresh on reconnect; explicit per-tab delete stays. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DY9N9STx6HMGtSEJauX1KG