Recent backend and connector changes are unacceptable

To the OpenAI Support / Product Team,

I’m writing to provide direct feedback as a long-term Plus user and power user who builds structured systems inside ChatGPT (notably, project-based frameworks using JSON file logic and behavioral rule sets).

Over the past several days, multiple silent backend and connector changes have effectively broken working systems that had been stable for weeks.
This email summarizes the key issues and the larger concern they raise about how these changes are being rolled out.

1. Project Files Now Load as Summarized Snapshots Instead of Verbatim JSON
   Until recently, JSON files in ChatGPT Projects were loaded verbatim, exactly as stored.
   As of this week, files intermittently load as summarized or condensed representations rather than their literal contents.
   This causes functional corruption in projects that rely on deterministic configuration and logic (for example, frameworks that load behavioral rules or structured data).

Observed symptoms:
– Large JSON files (e.g., rules__behavior.json) were reduced to short semantic digests upon load.
– Reopening the same project sometimes reuses the cached digest instead of the real file.
– Adding a "force_disk_read": true directive restores correct behavior, proving it’s a new caching behavior and not user error.

This completely undermines the reliability of structured project architectures.
If a project system is going to be viable, file reads must always return the literal bytes of the file, not a model summary.

1. Dropbox Connector Change (Loss of Direct File Access)
   The Dropbox connector was silently changed to use an “indexed search” model rather than direct file reads.
   Previously, ChatGPT could open specific files or folders on command (ChatGPTInbox/GlobalInbox.index.json.txt).
   Now, even when the connector is authenticated, file access fails unless a broad search is performed.
   This broke existing integrations that depended on deterministic, structured file paths.

There was no advance warning or changelog for this change, and it rendered previously functional workflows inoperable overnight.

1. Web Connector Change (Loss of Direct URL Access)
   ChatGPT can no longer directly open and process provided web URLs (open_url() calls now fail).
   Instead, every web link is routed through the web.search() tool, which performs a search rather than fetching the provided URL.
   This is highly unreliable for research, technical referencing, and universe-building workflows because search results don’t always return the exact page.

Like the Dropbox regression, this change occurred without warning or documentation.

1. Lack of Public Beta Testing or Staged Rollouts
   All of the above issues share one root cause: there is no public beta or staging environment for power users or business users to test changes before they go live.
   OpenAI’s backend behavior changes are being silently deployed to production, immediately affecting paying users who depend on stability for ongoing projects.

Had there been a beta track or opt-in pre-release testing channel, these issues would have been immediately caught by advanced users before release.

Requested Fixes / Improvements
1. Revert or add a toggle for literal (verbatim) file reads inside projects — summaries should never replace actual file data.
2. Provide a documented option for direct Dropbox file access (path-based, not search-based).
3. Restore direct web URL loading alongside web.search instead of replacing it entirely.
4. Create a public beta or developer test environment for upcoming backend and connector changes.
– This can be limited to Plus/Pro/Business users who opt in.
– It would allow OpenAI to identify regressions before release.
5. Publish a clear, detailed changelog whenever connector or tool behavior changes.

Summary
The core problem is not simply that features changed — it’s that they changed silently, without user visibility or a safe testing environment.
OpenAI has built an incredible platform for structured and intelligent systems, but these backend shifts are eroding trust among the advanced users and small businesses who rely on it.

Please forward this feedback directly to the Product and Engineering teams.
If needed, I can provide exact reproduction steps, affected project files, and before/after behavior logs to demonstrate the issue clearly.

Sincerely,
Dave [Last Name]
OpenAI Plus Subscriber
(Project-based ChatGPT user and early system architect)