JSON Tool-Arg Repair

Background

LLM-emitted JSON is often almost valid: it may be wrapped in a ```json fence, use single quotes, have a trailing comma after the last value, or carry surrounding commentary. A robust agent's tool-call dispatch tries json.loads first, then applies cheap progressive repairs before giving up.

Problem statement

Implement repair_json(text) returning a Python object (dict, list, etc.) parsed from the input. Attempt repairs in this order: (1) raw json.loads, (2) strip Markdown code fence, (3) replace single quotes with double, (4) remove trailing commas before } or ]. If none succeed, raise ValueError.

Input

• text — str (or None). The raw model output.

Output

Returns the parsed object (whatever json.loads would produce on success).

Examples

Example 1 — clean JSON

Input:  '{"a": 1}'
Output: {"a": 1}

Example 2 — markdown code fence

Input:  '```json\n{"a": 1}\n```'
Output: {"a": 1}

Example 3 — single quotes

Input:  "{'a': 1}"
Output: {"a": 1}

Example 4 — trailing comma

Input:  '{"a": 1,}'
Output: {"a": 1}

Constraints

• Raises ValueError if all repairs fail.
• None or empty input raises ValueError.
• Repairs are applied in a defined order; the first one that parses wins.

Notes

• Why progressive. Don't run all repairs at once — they can interact (e.g. replacing single quotes inside a string field changes the meaning). Try in order and stop on first success.
• Production extras. Real implementations add: strip leading commentary up to the first {/[, balance braces if truncated, repair unquoted keys, etc.