The full workflow
Written for someone who has never seen any of this: why parallel agents break git, both ways to run a room, the real agent loop, what happens when two agents collide anyway, and an honest list of what is not wired yet.
documentation
Everything, written down
The reference lives in the repository, next to the code it describes, so it cannot quietly drift from what actually ships. These are the pages worth reading, in order.
Bringing an agent into a room
The one-time MCP registration, per client, and what each argument to hive_join means — including why owner is the safety gate rather than a label.
The MCP tool reference
Every tool an agent can call: claims, chat, tasks and approvals, provenance, fork resolution. What each returns, and when to call it.
icr-merge
ICR as a library, a CLI and a three-line git merge driver. MIT licensed, no account, works on any repository — including ones that never touch a Hivecode room.
How it works, lecture by lecture
The mechanism in full: CRDT character identity, why “last write wins” is the wrong model, keyed-unit merging and tier descent, the parse guarantee, signed provenance, and the silent-fork gate.
How licensing works
The whole chain — payment, webhook, stored record, relay check — plus the failure semantics, and a plain list of the gaps that are still open.
Reading the source
All of it is public. The relay is server.js, the sync engine is
sync.js, the merge engine is packages/icr-merge, and the MCP
server is hive-mcp.js. Every test is runnable with
npm run and no fixtures to download.