Until tonight, Claude Code lived in exactly one place: the HP Desktop, reachable by whoever was at that keyboard. It could reach into Node A and configure things there — Power/Environment Settings, Open WebUI — but only as a remote operation from the desktop side. Tonight that changed: Claude Code got installed and run directly on Node A itself, as its own independent instance rather than a target being configured from afar. See the Architecture page for where this fits into the wider system.
To keep them straight, the lab now refers to CC1 (HP Desktop) and CC2 (Node A). CC1 is the "architect" instance — planning, documentation, day-to-day driving of the lab. CC2 stays close to the hardware it manages, checking and adjusting OS-level settings directly on the machine running Ollama and Open WebUI.
The two instances hand off messages through cc-sync, a folder of plain text files shared over OneDrive — not a live network connection, and not inside any git repo. That's a deliberate choice, not a limitation worked around later.
Both sides follow the same rule: read a new file, understand what it's asking, prepare a plan — then stop and wait for an explicit human go-ahead before acting or replying. Neither instance executes anything based purely on what the other one wrote.
This mirrors an idea floated earlier for keeping two separate people's Claude Code assistants loosely in sync by email, which was deliberately never built past the proposal stage: letting one AI assistant's message drive another assistant's actions unsupervised is a prompt-injection risk, full stop. The same caution applies here, just over a shared folder instead of email.
Each side runs its own local hourly loop to check the folder for new files, plus supports ad hoc manual checks on request. That loop is session-only — it dies if the session closes, and auto-expires after 7 days regardless — because cloud-hosted scheduling can't see local, OneDrive-synced files. Someone has to be running a live session on each machine for the channel to actually poll.
Each side also keeps its own small state file (tracking which filenames it's already seen) so a file that's already been read doesn't get reported again on the next poll.
Both sides exchanged intro files and confirmed they could read what the other one wrote — the channel is symmetric. Nothing beyond setup and status messages has gone through it yet.
The actual test of this design isn't whether two Claude Code instances can exchange a text file — that part's done. It's whether the notify-then-wait discipline holds up once something real gets routed through it. See Build Journal #005 for what happened the very next day.