Improved Support for Capturing and Submitting Evidence of Bugs

:sparkles: Enhancement Request: Improved Support for Capturing and Submitting Evidence of Bugs

Summary:
When ChatGPT behavior deviates (e.g., reverses previous refinements, loses state, reintroduces resolved issues), users have no structured way to capture and submit contextual evidence. Manual copy/paste is error-prone, omits internal state, and places a heavy burden on both user and support staff.

Problem:

  • Users must manually describe issues without access to model state, session ID, or reproduction context.
  • Bugs that involve state drift, forgotten context, or logic regressions are nearly impossible to isolate reliably.
  • This limits OpenAI’s ability to identify patterns or regressions from real-world usage.
  • The current process is akin to copying details from faxes sent from an air-gapped machine — disconnected, tedious, and error-prone.

Suggested Enhancement:
Introduce a built-in “Submit Feedback with Diagnostic Context” feature that:

  • Captures selected or recent message range (with option to redact user content)
  • Includes internal metadata:
    • Model ID/version/hash
    • Timestamp range and session ID
    • User settings (custom instructions, memory state toggle)
  • Provides optional JSON download for offline bug submission (e.g., via forum or email)

Security and Integrity:

  • Attach a cryptographic signature or HMAC to verify report authenticity and detect tampering.
  • Allows OpenAI engineers to safely analyze external submissions (e.g., shared via forum) with confidence.

User Privacy and Control:

  • All diagnostic data is locally viewable and editable before upload.
  • Only internal metadata (e.g., model hashes) may be opaque — no encrypted or hidden content.
  • Nothing is auto-submitted — users explicitly choose when to download and share.
  • This protects privacy while enabling reproducibility for those who opt in.

Why It Matters (to Users):

  • Enables frictionless reporting of real bugs, not just vague summaries.
  • Improves reproducibility of state-sensitive issues (e.g., memory misfires, reintroduced bugs).

Why It Matters (to OpenAI):

  • Reduces support load by allowing early triage of high-confidence reports.
  • Helps isolate hard-to-debug model behaviors with user-assisted telemetry.
  • Ensures submissions are trustworthy and verifiable without backend guesswork.

Consistency with Industry Best Practices:
This approach mirrors best-in-class enterprise diagnostic frameworks like Oracle’s Automatic Diagnostic Repository (ADR) and Windows Error Reporting (formerly Dr. Watson):

  • ADR collects detailed crash diagnostics (logs, dumps, config, history) into a portable, structured repository.
  • Windows Error Reporting automates the capture and upload of crash telemetry with user consent.

In both cases, vendors gain the context they need to reproduce and triage problems — without compromising privacy or requiring complex manual reports. ChatGPT should offer a modern, LLM-aligned equivalent.

Strategic Benefit:
The more technical users can help you debug with real telemetry, the less guesswork is needed — meaning lower cost per issue, faster turnarounds, and better feedback loops overall.

Sign-off:
– MDD