MineRatings.ts: ARG_MAX risk when passing large prompts as CLI args to Inference.ts

[virtualian]

Summary

MineRatings.ts shells out to bun Inference.ts passing both the system prompt and user prompt as CLI positional arguments. With 750+ rating entries, the JSON summary in the user prompt can approach OS ARG_MAX limits (256KB per arg on macOS), causing silent failures — the process exits with zero output and no error.

Reproduction

bun ~/.claude/skills/Utilities/PAIUpgrade/Tools/MineRatings.ts --all

With a sufficiently large ratings.jsonl (750+ entries), the spawned bun Inference.ts <system_prompt> <user_prompt> call can silently fail due to argument length.

Root Cause

MineRatings.ts runInference() passes prompts as argv:

const proc = spawn('bun', [
  INFERENCE_PATH,
  '--level', 'standard',
  '--timeout', '300000',
  systemPrompt,  // ~350 words
  userPrompt,    // JSON summary of all entries — grows with dataset
], { ... });

Meanwhile, Inference.ts itself already handles ARG_MAX correctly internally — it pipes the user prompt via stdin to claude --print (line 96-98). The problem is in MineRatings' invocation of Inference.ts, not in Inference.ts itself.

Fix

Replace the subprocess spawn with a direct import of inference():

import { inference } from '../../../../PAI/Tools/Inference.ts';

async function runInference(systemPrompt, userPrompt) {
  const result = await inference({
    systemPrompt,
    userPrompt,
    level: 'standard',
    timeout: 300000,
  });
  return { success: result.success, output: result.success ? result.output : (result.error || 'Inference failed') };
}

This eliminates the subprocess entirely and reuses the stdin-piping that inference() already does.

Priority

Low — only manifests with large datasets (750+ entries) and is already fixed in virtualian/pai#49.