I kept hitting the same thing: a coding agent finishes, says it's done, and the tests were either never run or still red. The summary was confident either way. So the actual work wasn't writing code it was checking up on it.
foolscap is what came out of that. You give it a task, GPT-6 Astra plans it and dispatches Claude Code, Codex or OpenCode on your machine, then a deterministic evidence engine not a model reads what the test command actually printed and decides whether the run passed. Two rules that do most of the work:
repairs are capped (two for implementation, one after review), so it can't grind forever
the review is dispatched to a different agent family, read-only, because an agent shouldn't sign off on its own work
A run from yesterday: implement failed 1 test one repair 1 passed independent review on Codex complete. Six turns, eight cents, every step on the record.
Tell foolscap what needs doing. GPT-6 Astra plans the work and routes it to Claude Code, Codex or OpenCode running on your machine. Then foolscap reads the test output and the diffs itself, so a result is evidence, not the agent's word for it. Red tests trigger a bounded repair; passing work gets reviewed by a different agent. Local-first, bring your own key, MIT.
