Surfaced: 2026-06-28. Cody (cloud-codex) created PR #503 end-to-end fine, but when asked to land a small follow-up (Theo's review note: extract a shared helper), two codex-exec turns ran tests in a scratch dir but did not commit/push to the working clone — the branch stayed at the original commit. I applied the one-liner directly to unblock the merge.
Pattern: codex-exec is reliable for a fresh from-scratch task but flaky on iterative follow-ups (checkout existing branch → small edit → push). The work happens in a sandbox/scratch dir and doesn't persist to the intended repo path.
Direction: investigate the cloud-codex run wrapper / codex exec workdir handling for follow-up tasks — ensure the edit lands in the checked-out branch and that git push is the terminal step. Possibly pin the workdir explicitly and verify-push in the run loop.
Refs: docs/agents/AGENT_CODING_CAPABILITY.md (footguns — "tell Cody the repo path"), PR #503.
Surfaced: 2026-06-28. Cody (cloud-codex) created PR #503 end-to-end fine, but when asked to land a small follow-up (Theo's review note: extract a shared helper), two codex-exec turns ran tests in a scratch dir but did not commit/push to the working clone — the branch stayed at the original commit. I applied the one-liner directly to unblock the merge.
Pattern: codex-exec is reliable for a fresh from-scratch task but flaky on iterative follow-ups (checkout existing branch → small edit → push). The work happens in a sandbox/scratch dir and doesn't persist to the intended repo path.
Direction: investigate the cloud-codex run wrapper / codex
execworkdir handling for follow-up tasks — ensure the edit lands in the checked-out branch and thatgit pushis the terminal step. Possibly pin the workdir explicitly and verify-push in the run loop.Refs:
docs/agents/AGENT_CODING_CAPABILITY.md(footguns — "tell Cody the repo path"), PR #503.