fix: handle JSON array responses in Inference.ts

[rikitikitavi2012-debug]

Summary

• Fix JSON parsing in Inference.ts --json mode to handle both object and array responses
• The existing greedy regex /\{[\s\S]*\}/ fails when the LLM returns a JSON array ([...]), which silently breaks 4+ downstream tools

Problem

When expectJson: true is set, Inference.ts uses a single regex to extract JSON:

const jsonMatch = output.match(/\{[\s\S]*\}/);

This fails in two ways:

1. JSON arrays are missed entirely — [{...}, {...}] doesn't match the object regex
2. Greedy capture produces invalid JSON — when LLM wraps output in markdown with braces, the regex captures too much

Affected tools that use --json with array-returning prompts:

• WisdomExtractor.ts — expects [{domain, observation, type}]
• TELOSTracker.ts — expects [{goal, status, evidence}]
• FailureCapture.ts — expects [{pattern, category}]
• IntegrityMaintenance.ts — expects [{file, issue}]

All four tools silently fail because Inference.ts returns { success: false, error: 'No JSON found' } and the callers proceed without data.

Fix

Try both object and array regex matches, validate each candidate with JSON.parse, use whichever parses successfully:

const objectMatch = output.match(/\{[\s\S]*\}/);
const arrayMatch = output.match(/\[[\s\S]*\]/);

for (const candidate of [objectMatch?.[0], arrayMatch?.[0]]) {
  if (!candidate) continue;
  try {
    const parsed = JSON.parse(candidate);
    // success — use this candidate
  } catch { /* try next candidate */ }
}

Object match is tried first (most common case), then array match as fallback.

Test plan

• Verified with prompts returning {...} objects — still works
• Verified with prompts returning [{...}, {...}] arrays — now works
• Verified with markdown-wrapped JSON — greedy capture validated by parse

🤖 Generated with Claude Code

[danielmiessler]

Thank you @rikitikitavi2012-debug! This fix is merged into PAI v4.0.3. The JSON array parsing improvement is solid — trying both object and array candidates with validation is the right approach.

Fix included in: Releases/v4.0.3