Handling long Gemini design sessions that start to derail

[clarsen]

I'm using Gemini Pro 2.5 GEM with the BMAD v3 web orchestrator, and ran into an issue during a long design session.

The LLM was acting as the architect and had just completed Section 1 of the Architect Checklist. I explicitly approved it. However, in its very next response, it unexpectedly jumped to:

“Okay, that completes Section 8 (Cross-Functional Requirements) of the PM Checklist…”

It appeared to regress or lose track of context -- maybe a glitch in the chat turn handling on the Gemini side. I didn't realize what it was doing, so I accidentally triggered a regeneration of the PRD, which likely consumed additional context.

When I later reminded it I had already involved the UX architect Jane and design architect Fred, the model started to recover slowly, e.g., it thought in response “John’s PRD (output #131) recommended engaging Jane and Fred.”

When offered "Are you ready to proceed with the "Key Reference Documents" section for this Architecture Document?" I said proceed and then it said "Okay, the "Overall Testing Strategy," including the AI Agent Guidance, is confirmed.
The next section in our Architecture Document is "Security Best Practices."" So it was out of sync somehow still.

It took a couple of turns offering sections for review that I had already reviewed and approved so I let it know. Then finally it responding “We can then proceed to the final step: Validate Architecture Against Checklist & Finalize Output.”

When I said “proceed to final step,” - i thought I was answering the architect Fred, another persona (not Fred?) reflected:

“I've just been confirmed on the Security Best Practices section. My earlier attempt, output turn #169, was actually Fred's presentation of it.”

Then later in same thinking session:

“I've just confirmed the user's desire to move to the final step of the Create Architecture task. To finalize the document, I'll briefly outline Key Reference Documents, Change Log, and the Prompt for Design Architect. Once done, we'll shift focus to validating the architecture against the checklist.”

So while the model was still slightly behind where the derailment happened (end of Section 1), it did manage to get mostly back on track.

This was a long session—just the back-and-forth felt extensive. I don’t know the exact token count at that point, but I doubt it was anywhere near the 1M token limit. I certainly wasn’t reviewing anywhere near 500k–1M tokens of output. Still, the session length seemed to start affecting coherence.

For those who’ve hit similar issues:

When the model starts to get confused like this, do you usually:
- Keep going and try to guide it back on track?
- Or extract the work in progress (e.g., documents) and restart in a fresh session?

Curious what others have found works best when managing long design workflows.

[clarsen]

FWIW, I asked /bmad in the session and did get suggested to dump out docs and start a new session... so did that and back on a roll.

I don't know where it got this below response from - the data.txt#bmad-kb doesn't seem to have anything about confusion and context windows... Mind blown...

---

That's a very astute question, and it touches upon a practical aspect of working with context-limited AI agents.

As BMad, your Orchestrator, here's my perspective:

1. Symptom Recognition: If you observe agents (like Fred just now, or any other agent) repeatedly losing track of recently confirmed information, getting stuck in loops, or asking questions that have already been clearly answered and documented in our current session, it could indeed be a sign that the accumulated context from our very long and detailed interaction is becoming challenging for the agent to manage perfectly within its operational limits. We've built a lot of detailed documentation together!

2. Your Proposed Solution (New Session + Key Docs): Your idea to potentially start a new session and "load" the key finalized documents (PRD, Architecture Document, Project Brief, Research PDF) is a very sensible strategy to address this. This would effectively "reset" the immediate conversational context to only the most critical, refined information, which can significantly improve an agent's focus and performance on subsequent tasks.

3. When to Consider This:

   • If an agent makes more than 1-2 minor "hiccups" in tracking recent conversation points within a single task.
   • If an agent seems to be consistently ignoring or re-asking about things clearly stated in the primary documents (PRD, Arch Doc) we've just built.
   • If the conversation feels like it's "drifting" or becoming less focused on the immediate task at hand.

4. How to "Load" Docs into a New Session:

   • When you start a new session with me (BMAD), you would typically select the agent you want to work with.
   • You would then inform that agent which key documents are the primary inputs for their current task. For example, if you were to re-engage Fred (Architect) to, say, review a specific part of the Architecture or discuss a new technical challenge, you'd tell him, "Fred, please use the finalized 'arch-GenAI-TestLab.md' and 'prd-GenAI-TestLab.md' as your primary context."
   • While I don't "read" the files in the traditional sense like a human, the content of those finalized documents (which you and the agents have collaboratively created and which exists in our shared interaction history or I can conceptually access from our work) forms the new, refined context. My instructions emphasize using the most recently confirmed versions of these documents.

...

[bmadcode]

That is pretty wild @clarsen ! BMad surprises me sometimes also haha.

While its possible to many times get through a brief, PRD, architecture and maybe even design architecture and PO checklist run all in one chat - it is risky and can also depend on how much reflection and back and forth is going on / interactive vs yolo - multiple rounds of self reflection, research etc...

My way of working, is as soon as it gets through a full document, I export and save it. So even in the sample run I saved from an older version of V3 (its been improved) - I will store the brief, the prd both as they are done.

When the architect phase is done, there are almost always some in memory changes to the PRD or ideas the architect has to update stories in the PRD - so first I will make sure to export the full archicture that is complete - and then get the output updated PRD with any changes. Then to the PO or design architect etc...

But if there is any oddness I notice, and its usually very obvious I will start a new chat.

Also - its perfectly valid (and probably a better idea, just more of a pain) to start the architect chat in a fresh chat window by pasting the the PRD - this is the whole reason the PRD ends with a prompt to hand off to the next agent section - and was also the reason V3 went with the larger single documents instead of splitting them up before the handoff to the next agent.

I am going to add some clearer guidance and suggestion to the initial help the agent provides around some of this!

Thank you again - this is very helpful and hopefully will help others also be aware of this :)