Every worker. Every transcript. Replayable.
Worker sessions are the receipts. Each AI worker gets its own durable session — what it was told, what it did, what it produced — and you can read all of them from one hub.
Every agent gets its own tracked session, so its work doesn't vanish into chat history — you always know which worker touched which files.
Each worker gets a durable JSON session plus an append-only JSONL event log under OPENCLAW_WORKSPACE, tracking harness, tool profile, branch/worktree, transcript ref, evidence, and touched files.
dig in
The moment you run agents in parallel, you need provenance — who edited what, on which branch, which output got promoted. Skip it and parallel agents become loose chats you can't audit or replay. We store card_ref to join workers to plan cards without mutating the Compass schema — the kind of boring decision that keeps the system from breaking when it grows.
card_ref = {"plan_id": "...", "step_id": "..."}
