Assistants Playground (GPT-4.1) bug: structured prompt leaves thread in endless “Streaming…”—main pane blank, sidebar preview shows hidden output, logs empty

[andrewlongron]

Title: Assistant Playground returns blank output when structured system instructions are used (reproducible)

---

Steps to Reproduce

1. In the Assistants Playground, create a new assistant (model: GPT-4.1, no tools enabled).

2. In System Prompt, paste structured, multi-bullet instructions with conditional logic for generating a YouTube script.

3. Provide four plain-English inputs, e.g.

   Agency Name: X
   Topic: Condo Insurance
   Audience: Galveston Homeowners
   Tone: Educational
   

4. Click Run.

5. Refresh the thread and re-run; then replace the prompt with a one-sentence version and run again.

Expected Behavior
The assistant should parse the structured instructions every time and return the full Q&A-formatted YouTube script, regardless of formatting depth.

Actual Behavior

• With structured instructions the assistant returns blank (no content) and the message state stalls at thread.message.in_progress indefinitely.
• After a page refresh the next run may succeed once, then reverts to blank output on subsequent runs.
• Replacing the entire prompt with a simple, single sentence makes the assistant respond reliably, indicating a silent failure while parsing richer prompts.

Environment

• Windows 11
• Chrome 115
• Assistants Playground (web)
• Model: GPT-4.1
• Latest API version

Additional Observations / Logs

• Playground Logs panel shows the assistant apparently produced JSON with content, yet the main thread panel is empty.

• Issue is reproducible with any structured prompt; it disappears only when the system prompt is reduced to a single, unstructured sentence.

• No tools, no external calls, and valid JSON in the request body.

• Timestamp examples:

  • Aug 2 2025 10:18 PM CT — blank response, message stuck in_progress.
  • Aug 3 2025 1:55 PM CT — logs show output, UI shows nothing.

Attachments
(1) Screenshot of stalled thread with in_progress status.
(2) Screenshot of Logs panel showing non-empty assistant response.

---

Impact: This prevents reliable use of Assistants for any multi-step or conditional workflows. A workaround is to strip all structure from the system prompt, but that defeats the purpose.

Any insight or workaround beyond “use a one-liner” would be appreciated.

[andrewlongron]