Agents make claims. Kiwi Code checks them against git.
Agent narrative is untrusted input. Every run ends in a handoff — branch, SHA, PR — reconciled against the worktree, not the model’s word.
What changed, where it landed, what was explicitly not done.
HEAD, dirty state, PR status — read mechanically. Agents never grade their own homework.
Stated confidence vs. how that band actually held up. Recomputed from merge outcomes, not vibes.
Verification is never agentic — no model ever judges whether work happened. The board is derived from an event log with three sources, and only one of them is allowed to define reality.
A mechanical scan of the worktree: HEAD, dirty state, branch, PR status. Read from git, never from the model. This stream defines reality.
What agents report in structured handoffs — files touched, tests run, PR opened. Treated as untrusted input until reconciled.
A deterministic reconciler diffs claims against truth and emits MATCH or MISMATCH. No inference, no judgment — a comparison.
A lane showing MISMATCH cannot hand off. The claim stands corrected by the repo, not the other way around.
The variance ships with evidence: the claim, the git truth, and the diff between them — a decision, not a debugging session.
Agents report confidence with a track record attached — "88%, and this agent's 88% band has held 91% across 240 runs." Numbers earn trust or lose it.
Kiwi Code is onboarding enterprise teams now.
› Contact Us