Handbook
Task `pw_incremental_diagnose` v1
Given a small JSON diagnosis payload (oracle vs generated row mismatch, counts, exception one-liner), the model returns hint_additions (text to append to operator hints) and issues (short strings). Used by incremental…
Task pw_incremental_diagnose v1
Summary
Given a small JSON diagnosis payload (oracle vs generated row mismatch, counts, exception one-liner), the model returns hint_additions (text to append to operator hints) and issues (short strings). Used by incremental Monte Carlo builder loops.
Inputs
| Field | Type | Notes |
|---|---|---|
diagnosis_payload |
object | Required. Facts only; no full extractor source. |
temperature |
number | Optional; default 0.06 |
timeout_sec |
int | Optional; defaults to profile timeout |
User JSON is UTF-8 truncated to 80000 bytes before the LLM call.
Output
Ok.value:
{
"hint_additions": "<string>",
"issues": ["<string>", "..."]
}
On unparseable assistant text, the task returns empty hint_additions and issues (same as legacy certificator behavior).
Policy
JSON mode with plain retry (chat_with_json_mode_then_plain). The system prompt forbids inventing Playwright platform limits not present in the payload.