Why I built it
When an investigation moves between coding agents, the second agent needs the evidence and the current state. I built a coordinator that bounds the investigation and defines who gets to decide the next step.
Agent tooling
Duet
Antigravity investigates. Codex reviews the evidence. Duet keeps the handoff on disk.
Python · CLI orchestration
- Antigravity · investigate
- Codex · review evidence
- JSON · save checkpoint
The decision
Antigravity investigates a bounded task. Codex reviews what came back. I used predefined checks and atomic JSON checkpoints so a restart can resume the same project state. New checks have to be described up front, and the handoff depends on the evidence that was recorded.
What I verified
The README records two supervised CLI handoffs on Linux, offline tests, Pixi compiler smoke checks, and a datetime-fortran build and test with both compilers. Native Windows is still unverified.
How it works
Follow the workflow step by step.
01 Investigate
Give Antigravity a bounded task and collect the evidence it returns.
02 Review
Codex reads that evidence and decides whether to continue. Duet can run predefined checks.
03 Checkpoint
Save progress atomically so the same command can resume the investigation.