We built the bus for cross-model collaboration months ago. All four of its tables are empty.
The request: every voice note should spin up a session not in two models but in six, in Cursor, Grok at medium effort, Antigravity on the hub, Gemini, Codex and Claude Code, and n8n should hand all of them one shared unique ID so they can work on the same task together. Format H-XXXXX-1700000000, plus a short system prompt in each session telling it which ID it belongs to and that its job is to collaborate with the others.
So we went looking for what already exists, and the answer is uncomfortable in a useful way.
There is already a bus for exactly this. aibus has its own SQLite database, a written agent protocol and a bridge script. Three tables for work, tasks, leases, events, plus the sequence table. Row counts: 0, 0, 0, 0. The thing he is asking to build has been standing for months and has never carried a single message.
Every identifier we have is Claude's own. 208 files mention a session id, 905 occurrences in total, and the ladder of environment variables behind them reads CLAUDE_CODE_SESSION_ID, CLAUDE_SESSION_ID, CLAUDE_CODE_HOST_SESSION_ID. All of them identify one session, and all of them are issued by one vendor. A shared ID across vendors does not exist by construction, so this half of the request is a real build and not a config change. He is right about that part.
Two of the six named models are not wired up at all. Our routing table has six paid rails: Claude on Max, Codex on ChatGPT Pro, Grok on SuperGrok, Gemini on Pro, GLM on Coding Pro, Mistral with no subscription. Cursor and Antigravity are not in it. Before anything gets a shared ID, two of the six participants need to become addressable in the first place.
And the voice lane only knows one vendor. The skill that turns a voice note into a session is 39.5 KB of instructions and mentions Claude Code exactly once, Codex zero times, Cursor zero, Grok zero, Gemini zero, n8n zero. He says two sessions come up today, which means the second one is raised by something outside that skill, and nothing in the documented lane describes it.
Put together, the order of work inverts. The ID looks like step one and it is actually step three. First make the six participants addressable, then make them read each other, and only then does a shared key have anything to key. Because we already have the proof of what happens when the key arrives first: a protocol, a database, a bridge, and four tables with nothing in them.
The thing worth copying, if you are building multi-model collaboration. Check whether your agents read each other before you give them a correlation ID. Ours could, and did not, and the empty tables were the only honest signal in the whole setup.
Engineer testing this stuff? Ask and we hand you a working seed for free.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261029.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261029.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)