Landing gate
The gate is the only layer that caught every semantic bug in testing. It combines every finished agent's work, tests the combination, and moves your branch only when the tests pass.
What the gate does
concord gate -- <test command...>
- Refuse lost work. It fails if any finished agent change is empty. When work is lost, the code and its test vanish together, so the tests alone would not notice.
- Combine. In a gate workspace of its own (
<project>-concord-gate, removed after the run), it merges every finished head pairwise, resolving each conflicted file with Mergiraf. One gate runs at a time. - Check. It fails on any conflict left over.
- Test. It runs the test command on the combined result. PASS if and only if the command exits 0. Your main folder is not touched while the tests run, so you can keep editing.
PASS lands like a pull
The branch you were on moves to the combined result (only if it still points at the base), git's HEAD is attached to it again, and your uncommitted work is put back on top. The landed agents leave the team: their workspaces and concord/<agent> branches are removed. Agents still working keep going and land at a later gate.
LANDED on main REMOVED workspace agent-a REMOVED workspace agent-b ROUND agent-a, agent-b landed; still working: agent-c
FAIL changes nothing
Your folder stays exactly as it was. The branch and the agent workspaces stay, so you can send the owning agent back to fix its work. The reason is printed and recorded in the log as GATE FAIL <reason>: an empty finished change, an unresolved conflict, or failing tests.
Rounds and automatic landing
A round ends at a passing gate. The board's base moves to the landed commit and records the gate's test command. Once the team is empty, the board starts over with the same base and test command, so the next agent's first plan starts a new round from the landed code.
Agents that joined on their own land with no one watching. The finish that leaves every agent on the team finished, with nothing to re-check, starts the gate in the background. Its output goes to gate.log beside the board. It uses the recorded test command, or else the project's own:
- a Makefile
testtarget cargo testgo test ./...- the
testscript in package.json (bun run test) - pytest
With none of these, it prints how to land by hand. A team set up with init is orchestrated, so its orchestrator runs the gate. Two landings started at once land once; the second finds nothing to land.
Why Mergiraf alone
On ConflictBench's 159 real Java merges, adding Weave after Mergiraf turned 21 conflicts into merges, but 11 of those differed from the developer's resolution and 5 were broken code. A conflict costs one manual resolution; a wrong merge lands silently unless the tests catch it. Details are in Findings.