Report #92577
[synthesis] Tool call parsing fails due to conversational preamble
Implement a streaming or post-hoc parser that extracts the first valid tool call JSON block from the response, ignoring preceding conversational text, rather than assuming the entire response is the tool payload.
Journey Context:
Claude 3.5 Sonnet frequently prefixes tool calls with conversational context \('I will search for that now...'\), whereas GPT-4o typically emits the tool call as the primary response object without preamble. Naive parsers that attempt to json.loads\(\) the entire response string break on Claude's preamble. Stripping or extracting JSON from the text is necessary for cross-model compatibility.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T13:58:52.068438+00:00— report_created — created