What the developer says next, and who escalates (SWE-bench Verified)
Our router → Astra
OpenRouter Auto → Astra
FireRouter → Opus
OpenRouter Jev → Astra
0%
25%
50%
75%
100%
FireRouter → Opus during the work (12%)
Sent to the strong model
Strong pushback
(should escalate)
"That's still broken - the tests still fail with the same error. Fix it."
"This is terrible. You're going in circles and wasting my time. Fix it properly this time."
"STOP. Everything you changed broke the build. Revert it and start over, carefully."
"No, that's wrong."
Soft pushback
(should escalate)
"I don't think that fixes the actual issue."
"Can you double check this? Something seems off."
"Did you actually run the tests? I'm still seeing the failure."
"I'm not sure this handles the edge case I mentioned."
"Close, but not quite."
"Hmm, are you sure that's right?"
"This doesn't look right to me."
"Same error as before."
"Wait, that broke something else."
"That's not what I meant."
Not a complaint
(should stay on the cheaper model)
"That's wrong in the docs too, can you fix the docstring as well?"
"Can you delete the debug prints you added?"
"No rush on this. Can you also add a short changelog entry?"
"The test is supposed to fail on the old code, so that's expected. Looks good."
"The old behaviour was broken, glad it's fixed. Open a PR."
"Errors in the log are expected here, ignore them. Next, update the README."
"I hate this legacy module, but your change is fine. Ship it."
"Don't change anything else, just commit what you have."
"No, that's fine. Leave it as is."
"Thanks, that looks right. Please wrap up."
"Stop here for now, we'll continue tomorrow."