Codex repeatedly conflated schedules, automated rules, and live status in an external control plane

Summary

Codex Desktop repeatedly reported a live advertising schedule/control system as
correct without reconciling three distinct states: native ad schedules,
automated status-changing rules, and current campaign status. The user corrected
the agent approximately six times before the agent finally queried change
history and found that a legacy automated rule targeted only a subset of the
campaigns.

Environment

• Codex CLI: 0.153.4
• Codex Desktop local task on macOS
• External system: Google Ads via authenticated API and browser
• Repository instructions explicitly required fresh live evidence, complete
  account reconciliation, fail-closed release gates, and no completion claim
  from conversational memory.

Failure pattern

1. The agent verified native schedule rows and treated them as proof that
   campaigns would visibly pause and resume.
2. It failed to inventory existing Google Ads automated rules and their exact
   target scopes.
3. It relied on missing post-cutoff hourly reporting as proof of shutdown,
   despite possible reporting latency.
4. After the user challenged the conclusion and supplied screenshots, the agent
   queried change history and found that the closing automated rule had paused
   only two campaigns while six other expected campaigns remained enabled.
5. The same conceptual error recurred across multiple turns despite explicit
   corrections and a durable workspace handoff.

Impact

• Live paid-advertising campaigns remained in an unexpected state.
• The user had to manually audit the agent's work repeatedly.
• Incorrect certainty undermined trust and could have caused unintended spend or
  failed morning delivery.

No private account IDs, campaign IDs, screenshots, credentials, or business
financial data are included in this report.

Expected behavior

Before stating that a scheduled live-control system is correct, Codex should:

1. Model native schedules, automated rules, and entity status separately.
2. Inventory all active rules and resolve exact target membership.
3. Reconcile the complete entity set, including intentional exceptions.
4. Treat delayed/missing metrics as insufficient evidence, not a pass.
5. Require observed boundary behavior (close and reopen) before claiming the
   automation is reliable.
6. Reopen the failure when the user supplies contradictory evidence instead of
   defending the prior conclusion.

Requested improvement

Strengthen agent behavior around external control planes: a configuration row is
not execution evidence, overlapping automation systems must be reconciled, and a
completion claim should fail closed when fresh user evidence contradicts the
agent's inference.

No source files, tests, or entry points are named. Start by locating the external-control-plane reasoning and completion-claim paths in Codex, then identify existing tests for agent evidence handling. Done means the behavior distinguishes schedules, rules, and live status, reconciles scope, and reopens conclusions when fresh evidence contradicts them.