Defensive parsing of model output

The model is a probabilistic component; its output will sometimes be malformed. Guard every place you consume it:

• Tool arguments: parse defensively; if the JSON is malformed, default to empty arguments rather than throwing, and let the tool report a useful error.
• Unknown tool names: if the model calls a tool you don't have, return a "that tool isn't available" result so the model can recover, rather than erroring out.
• Emitted protocols: if the citation/structured block is missing or malformed, proceed without it; don't make a missing trailer fatal.
• Missing tool results: guarantee that every tool call the model made gets some result back (a real one or a synthesized error). Providers require this pairing; a missing result breaks the next call (Chapter 2).

The principle: the model's mistakes should become recoverable signals, not crashes. Often the best recovery is to feed the error back to the model as a tool result; it will frequently adjust and try again.