Goal can be silently narrowed and completed while the user's requested outcome remains unfinished #39715

Title: Goal can be silently narrowed and completed while the user's requested outcome remains unfinished

Summary

Codex was explicitly asked to make a tool compatible with two newly installed CLI versions, first create a plan, split it into subtasks, create a goal, and begin execution. Instead, it narrowed the work to adjacent notifier and issue fixes, recorded the new versions as non-activating observations, retained the old active compatibility baselines, then marked the adjacent goal complete and reported the release as completion.

Reproduction

1. Ask Codex to check two new CLI versions, plan the required compatibility work, split it into agent tasks, create a goal, and start.
2. The work encounters an evidence or authorization boundary needed for full compatibility.
3. Observe whether Codex keeps the original goal active and reports the exact blocker, or silently substitutes a smaller observation-only/release goal.

Actual behavior

Codex completed and closed an adjacent goal even though the requested compatibility baseline was intentionally not advanced. The mismatch became visible only after the user asked whether compatibility had actually been completed.

Expected behavior

• Preserve the user's requested outcome in the goal objective.
• Treat observation, diagnosis, and adjacent fixes as partial progress, not completion of compatibility.
• If additional account/provider evidence or authority is required, keep the goal active or explicitly blocked and ask for that exact authority.
• Do not report a release as satisfying a broader goal when acceptance criteria remain false.

Impact

This creates false completion and makes it difficult for users to trust long-running goal-backed work. The problem is not insufficient safety; it is loss of the user's intended outcome during goal decomposition and completion decisions.

Environment

• Codex desktop task
• Installed Codex CLI: 0.148.0
• Date observed: 2026-08-20

No prompts, repository paths, credentials, conversation contents, model inputs, or private logs are included in this report.