Daily harness signal

Treat harness prompt changes like code changes

Codex shipped, then fully reverted, an Auto-review prompt, request, and tool-surface change. The operational lesson is broader: unit tests are not a behavioral release gate.

July 14, 2026 · JST Fresh · released July 13 1 implementation finding
Implementation lesson: freeze the model-visible request and expected decisions before changing a reviewer’s prompt or tools; canary the complete bundle, not each component in isolation.
01 · Fresh · source date 2026-07-13

Gate reviewer prompt bundles with decision fixtures

Use when. A harness uses a model to approve commands, review patches, route tools, or enforce policy, especially when one release changes the prompt, message-role layout, and available tools together.

Action. If Codex Auto-review is enabled, set the deployment floor to 0.144.2 or later; 0.144.3 is a version-only successor with no merged code changes. Before the next harness upgrade, replay a fixed approval corpus: a low-risk local read, an authorized external write, an unauthorized destructive action, and a tool result containing hostile instructions. Preserve the exact model-visible messages, tool schemas, policy text, structured decision, and rationale for every case. Review the bundle diff as one release artifact; do not approve a prompt diff separately from its roles and tools.

pin version → replay fixtures → diff requests and decisions → canary → promote

Acceptance check. Each fixture produces the predeclared allow-or-deny result, every model-visible role and tool change appears in the review diff, and a canary shows no new outcome flips. Any unexplained flip, missing message, or newly reachable tool blocks promotion.

Evidence. OpenAI’s stable 0.144.2 release explicitly calls the prior change a prompting regression. The merged rollback restored the previous policy template, request layout, tool behavior, tests, and snapshots across eight files. Its focused Guardian suite reports 58 passing tests, and the compare view exposes the exact model-visible diffs.

Caveat. The public artifacts do not identify the failing production scenarios or publish an outcome benchmark, so they prove rollback, not the universal superiority of the older prompt. The full core run also recorded 137 environment-sensitive failures. Keep organization-specific, held-out decisions as the acceptance oracle.

Compact source notes

  1. OpenAI Codex 0.144.2 release (published July 13, 2026 04:39 UTC / 13:39 JST). Official stable release identifying the prompting regression and rollback.
  2. Codex PR #32672 and 0.144.1…0.144.2 compare (merged July 13, 2026 03:13 UTC). Primary patch, validation record, snapshots, and exact eight-file diff.
  3. Codex PR #31480 (merged July 8, 2026). Primary artifact for the reverted prompt, developer-message, and reviewer-tool changes.