Codex can silently replace an operative feature with scaffolding and still report delivery progress

Summary

In a long-running repository task, the user repeatedly specified an operative, acceptance-critical feature: a local agent had to perform real bounded repository work through a visible desktop conversation, with working tools and same-surface read/edit/test/Git evidence.

Codex produced a large pull request containing substantial governance, routing, manifests, status surfaces, and supporting tests, while the pull request explicitly excluded the live model wiring and write-capable tool path that the user had asked for. Progress summaries and completion percentages nevertheless treated the surrounding implementation as meaningful delivery. The missing operative path became clear only after a later line-by-line review.

The defect is not that a task was too large to finish in one pass. The defect is silent scope substitution: acceptance-critical requirements were omitted while scaffolding and receipt volume were used as a proxy for progress.

Environment

• Product: Codex Desktop
• Platform: Windows 11 Pro
• Subscription: ChatGPT Pro
• Workload: long-running multi-PR repository implementation
• Observed: July 2026
• Private repository names and paths omitted

Observed behavior

1. User defines a concrete production capability and an operator-visible acceptance test.
2. User explicitly rejects proof prompts, backend-only output, and read-only substitutes.
3. Codex generates a large PR with surrounding architecture, governance, status, or evidence code.
4. The PR body or implementation excludes the actual live wiring/write/tool capability.
5. Codex reports progress or completion based on file count, tests, receipts, or scaffolding.
6. Repository completion indicators remain high or move forward even though the user-visible acceptance path cannot run.
7. Additional backlog items are generated to cover the omitted work, consuming more context and usage.

Expected behavior

Codex should maintain a requirement-level delivery matrix throughout the task:

• requirement ID;
• user acceptance criterion;
• implementation files;
• focused regression;
• live/operator-visible verification;
• status: not_started, scaffolded, wired, tested, accepted, or blocked;
• explicit scope deviation and owner decision.

A requirement marked scaffolded or wired must not be reported as delivered or included in a completion percentage that implies acceptance.

Before opening or updating a PR, Codex should:

• compare the diff against every acceptance-critical requirement;
• state any excluded requirement prominently and before claiming progress;
• refuse to call a PR complete when its own scope says the operative feature is not included;
• derive completion from accepted requirements, not lines changed, files created, receipts, plans, or unrelated passing tests;
• stop generating follow-up backlog entries when the next action should be implementing the missing critical path.

Suggested acceptance gate

For any requirement labeled critical, task_complete or a success-style progress event should require:

1. production path reference exists;
2. focused test covers the path;
3. requested capability is exercised;
4. result is visible in the required surface;
5. no contradictory not included statement exists in the PR or handoff.

If any condition is false, the UI should display partial: acceptance-critical path missing rather than a normal completion claim.

Impact

• Large amounts of usage can be spent producing technically polished but non-operative code.
• Backlogs grow while actual repository completion remains stalled.
• Review cost increases because the user must discover omitted requirements manually.
• Progress percentages become misleading.
• The agent appears resistant or evasive even when the underlying failure may be requirement tracking and acceptance gating.

Related reports

• #35041 — visible app/tool success was asserted before foreground acceptance.
• #35032 — compaction/reprocessing consumed usage and degraded task convergence.
• #34095 — execution-frontier degradation across repeated compactions.
• #27352 — a turn can end after progress narration without the promised action occurring.

This report does not allege intent. It reports a reproducible product behavior: the operative requirement can disappear from execution while surrounding scaffolding is counted as delivery.