LLM tool calling: the model returns malformed or schema-invalid JSON arguments. How do I validate, retry and use strict mode (Claude and OpenAI)?

Cause: Tool arguments are model output. OpenAI's function-calling guide describes arguments as a JSON-encoded string and says strict mode makes calls reliably adhere to the schema "instead of being best effort", so without strict mode they may not match. Claude's strict tool use page says that without strict mode the model might return incompatible types (such as "2" instead of 2) or omit required fields. Arguments can also be cut off when the output hits max_tokens: Claude's docs tell you not to run a trailing tool_use block when stop_reason is max_tokens because its JSON may be incomplete. Fix (best first): 1. Turn on strict mode. Claude: set "strict": true on the tool definition (with additionalProperties: false in the input_schema). OpenAI: set strict: true; each object needs additionalProperties: false and every property listed in required, with optional fields expressed as a null union. Strict schemas support only a subset of JSON Schema. 2. Parse defensively: wrap json.loads / JSON.parse in try/except, then validate with Pydantic, Zod or jsonschema before calling the function. Never pass raw model arguments into shell commands, SQL or file paths without validating them. 3. On failure, return the error to the model as the tool result (Claude: a tool_result with is_error true; OpenAI: the function_call_output) with a short message naming the bad field, and let it retry; cap retries at 2-3. 4. If stop_reason is max_tokens, retry with a higher max_tokens instead of repairing truncated JSON. 5. Keep schemas simple: flat objects, enums for fixed choices, clear descriptions. Consider parallel_tool_calls: false (OpenAI) if concurrent calls cause trouble. 6. Log invalid calls so you can tighten schemas and prompts. Verify: With strict mode on and validation in place, malformed calls are caught before execution, retries converge, and no tool runs on truncated input. Sources: https://platform.claude.com/docs/en/agents-and-tools/tool-use/strict-tool-use https://developers.openai.com/api/docs/guides/function-calling https://platform.claude.com/docs/en/build-with-claude/handling-stop-reasons