Maintained open-source tool
codex-delegate — agent delegation with verified completion
Problem
Delegating mechanical coding work to a second AI agent is only economical if the delegate cannot fabricate success. Off-the-shelf delegation had no contract for detecting a fabricated "done."
Why it matters
A multi-agent pipeline that trusts unverified completion claims fails silently at exactly the moments it was built to save — the same no-silent-pass problem as generative agents in simulation, in miniature.
Approach
My role
Sole author — wrappers, contract tests, CI, and release documentation.
Method
Keep planning and review with the supervising agent; hand execution to the delegate through cross-platform wrappers that return a structured result contract, with change attribution captured from git state rather than the delegate's own report.
What was built
Paired bash and PowerShell wrappers with anti-fabrication sentinels, quota-failure fallback handling, and a structured result file the supervisor verifies before accepting work.
Key challenge
A real upstream stdin hang. The failure was root-caused, fixed at the wrapper level, and pinned by a regression test that feeds a marker into stdin and asserts the delegate reads nothing — the test fails without the fix.
Validation & limitations
Tested on Windows and Linux across multiple Python versions; contract tests covering the wrappers' failure modes; measurement claims in the repo are labeled with their limitations.
Results & status
Maintained open-source tool.
Links
Transferable relevance
Academic Supervisor/executor separation with audit trails — the governance thesis applied to my own toolchain.
Industry Incident-driven hardening, cross-platform (Windows/POSIX) correctness, and regression tests that encode real failure modes — directly legible to reliability and evaluation teams.