I’ve noticed that while the agent is working, it will often say some stuff I definitely want to read. Breakthroughs, reasonings, responses, etc., that will only be available during the current Working… task, then it will disappear and not be in any of the expandable history once it’s done working.
Any idea how to see the full output anywhere? The export chat .md is also missing those messages.
In simple terms, this is an anomaly inherent to Antigravity—or, if you prefer, a peculiar ‘feature.’ The client intercepts the thought: React tag designated within the agent’s system prompt, subsequently truncating and obscuring it. Should you wish to reveal this hidden content, you can adapt any React template from Hugging Face and inject it as a workflow prompt. By specifying a custom React tag, the client’s interception service will falter, thereby rendering the raw thoughts directly within the conversation. The caveat, however, is that disabling this client-side listener may inadvertently disrupt the structured output of the agent’s command invocations, ultimately causing console command executions to fail. Consequently, a more prudent approach is to refrain from viewing the agent’s raw responses. Instead, employ global or workspace rules to mandate that the agent document its session logs in a designated location, such as a work_log.md file. Bear in mind, however, that this alternative will incur a significantly higher cost in output token consumption.
Thanks for that insight, I had similar thoughts about making the agent to output its thoughts elsewhere. Will definitely play around with both options, was just wondering if there was something I was missing in the settings somewhere thank you.
+1 to this! I’ve used the reasoning of models on many occasions, and frequently have added to the rules based on what I see. And sometimes I’m just curious how it came to a particular decision, sometimes I even learn from it, and sometimes I create a new rule. 
Would love to have the ability to see the reasoning back!
They’ve known about this for weeks and haven’t done anything about it, because your hidden response used up tokens, and if you ask for a repeat, it’ll use up more tokens, and the AI will cheekily reply that something must have glitched and that’s why you didn’t see the response—because it’s already been trained for this very situation
Eh we’re a month out from their major update separating the IDE from 2.0, I wish it could be better, but I expect some issues.
I’m leaning towards this being a bug rather than a trained feature lol. Haven’t even tested whether this works better in the non-IDE version either.
edit: I have been trying out the non-IDE one for about an hour and the thoughts do stay readable in that agent view. Suppose I’ll do that for the longer tasks I give it so I can see those.