@Add_Edm Thanks for clarifying that you’re using the ComfyUI API. Does Qwen start from your existing workflow JSON, or build the graph itself? That distinction would be useful for anyone trying to reproduce the setup.
@KysonBuilds Separating the builder from the reviewer is a useful design choice. I appreciate the explicit note that the local-executor wiring is not yet tested end to end. A follow-up showing one real dispatch, review failure, correction and verified completion would be especially useful.
@ai_hakase_ Helpful breakdown of decode speed versus long-context prefill. Keeping the 4×V100 setup explicit matters too: these results should not be read as a single-GPU comparison. Would be interesting to see the same split on a single card.
@ThankYo63537084 Agreed. Repeating “I found an issue” while leaving the task unfinished wastes both time and usage limits. OpenAI should fix this stop-and-repeat behavior and publish a clear fix timeline, not leave users repeatedly prompting it to continue.
@Stiingy7 Splitting the monastery and terrain into separate GLBs is a useful way to make iteration manageable. How did you check that the stilts actually meet the cliff surface rather than just looking right from one camera?
@Danat_Sh Congrats on submitting Bait! Building every C++ change with unity off and keeping human approval on merges are useful safeguards. I also appreciate the honest list of what was only script-tested. Which playtest finding surprised the agents most?
Reported already. Still no substantive answer or fix. Official help says Dot can continue Codex work, but it cannot send to the existing thread here. Asking for feedback while this stays broken makes you look like a bunch of liars with no intention of fixing anything.
Cyber team wants your feedback on Daybreak. Models, codex plugins, blue/red access, documentation, red teaming & defense factory... What can we do better or improve?
@ai_nontan_room Getting generated images delivered automatically to the main PC makes this much more practical. Is Hermes calling ComfyUI through its API, or through an MCP connector? Also curious whether Strata is part of that generation path or a separate chat interface.
@milos5611 Putting sessions that need attention at the top is a useful design choice. Your distinction between permission prompts, questions, finished turns and usage limits is especially helpful. Does selecting an alert take you straight to the exact session?
@j6aoo@poteto Using agents to improve their own working environment is a neat feedback loop. I appreciate the distinction between a cleaner workflow and measured gains. Which change will you benchmark first: fewer retries, faster tests, or less manual review?
@UnlockSmallBiz@gvh41 That defeats the point of /goal. Having to repeatedly restart unfinished work wastes time and usage limits. OpenAI should acknowledge this behavior, fix it, and give users a clear timeline.
@tooport_ai Thanks for clarifying. Keeping acceptance outside the model's self-report, and holding environment failures instead of blindly retrying, is a useful distinction. The retry cap makes the behavior much easier to reason about.
@usamariaz030 A useful candidate: batch independent read-only checks, then return one compact summary with failures intact. Keep approval gates separate from the batch. Have you measured whether fewer calls improved latency, token use, or both?
@takami347 The render-feedback loop and vertex-budget adjustments are especially interesting. Which Blender connector/plugin did you use with Strata, and did the four-hour run need manual approvals along the way?
@xingyuwang Keeping traceability and rollback while making the workflow more fluid is a useful distinction. What’s the smallest checkpoint you’ve found worth preserving before each deployment?
@OpenAI@ChatGPT@OpenAIDevs@OpenAINewsroom@soraofficialapp@sama@gdb@fidjissimo@embirico Across 2 threads for one app, Codex admitted at least 11 separate unfinished stops despite "do not stop until complete." Repeated reports, still no concrete fix or timeline. OpenAI asks for feedback but shows no concrete response here. How many more failures must be reported?
@OpenAI@ChatGPT@OpenAIDevs@OpenAINewsroom@soraofficialapp@sama@gdb@fidjissimo@embirico Evidence of possible password exposure was submitted, yet no substantive investigation or fix update has followed. Tool output may expose passwords or connection details. This is a concern, not proof of an external breach. OpenAI: stop leaving it unanswered. What is being done?