Codex Desktop starter cards submit only one-word prompts without their displayed intent
What version of the Codex App are you using (From “About Codex” dialog)?
26.820.60940 (build 7119)
What subscription do you have?
ChatGPT paid subscription; exact tier was not displayed in the reproduction.
What platform is your computer?
Darwin 25.6.0 arm64 arm
What issue are you seeing?
The Codex Desktop new-task screen displays four descriptive starter cards:
- Explore and understand code
- Build a new feature, app, or tool
- Review code and suggest changes
- Fix issues and failures
However, selecting a card and submitting the prefilled composer sends only one word:
ExploreBuildReviewFix
The created task receives no card identity, displayed description, target, or other semantic context. For example, a task created from Review code and suggest changes receives only Review, so the agent cannot know what should be reviewed. The same behavior was reproduced with Fix issues and failures, which sends only Fix.
This creates a context-handoff and discoverability problem: the UI presents the cards as complete task intentions, but the resulting task receives only a contextless verb.
Feedback was submitted through the app from both reproduced tasks, with one Feedback ID per task:
- Review task:
01a03b2e-c86e-71c3-bca2-700ec462593e - Fix task:
01a03b2e-4473-7322-b876-0e4f6209ae9c
What steps can reproduce the bug?
- Open Codex Desktop.
- Start from the new-task screen.
- Select Review code and suggest changes.
- Submit the prefilled composer without adding text.
- Inspect the new task's first user message.
- Observe that it contains only
Review. - Repeat with Fix issues and failures and observe that the task receives only
Fix.
What is the expected behavior?
The resulting task should receive the selected card's displayed intent or equivalent semantic metadata. Alternatively, the UI should clearly indicate that the card inserts only an editable one-word starter and that the user must add a target before submitting.
Additional information
The official Learn overview displays the four cards, but the current documentation does not explain their exact click behavior or disclose that they submit only one-word prompts:
Related issue #22000 concerns starter prompts being hidden by responsive layout. It does not cover this semantic context-handoff behavior.
No screenshots are attached because the reproduction screenshots contain private Project and task names; the relevant behavior is fully described in text.