Read the trace, not the chat — find the broken turn before reading the user's complaint — step 6 of 9

The agent returned malformed JSON. The downstream parser is catching the error and returning "unknown". This hides the symptom but doesn't fix the bug — the next user gets the same wrong answer, packaged as "unknown".

The trace shows the bug is upstream: the prompt didn't tell the model to return JSON. The fix is to return a dict describing the upstream class so the caller can route to the right fix.

Fix the parser to RETURN {"class": "prompt", "fix_layer": "the prompt", "raw": <the raw model output>} instead of swallowing the error with "unknown".

Expected output:

{'class': 'prompt', 'fix_layer': 'the prompt', 'raw': 'I think the answer is 42'}