Codex CoT decrypt

[WaldenLee2005]

Summary

Adds a full pipeline to extract plaintext reasoning (chain-of‑thought) from encrypted Codex session logs and convert it into training data for Yeto:

• Scans Codex session directories (~/.codex/sessions/**/*.jsonl) for encrypted_content blobs.
• Uses a 3‑stage cascade (recording, tts, rawlog) to replay each blob via the ChatGPT backend (/responses) with the local OAuth token.
• Filters out non‑reasoning items (empty summaries) to ensure high‑quality data.
• Outputs cot_extracted.jsonl with each line containing: id, enc_len, src, summary, method (stage used), chars, and text (the plaintext reasoning).
• Includes yeto_cot_processor.py to convert the extracted data into Yeto‑compatible chat format (cot_training.jsonl).

Usage

# Extract reasoning from ALL Codex sessions (recursively scans all dates)
python3 cot_extraction/cot.py ~/.codex/sessions/

# Convert to Yeto training format
python3 cot_extraction/yeto_cot_processor.py --blob-path cot_extracted.jsonl --output cot_training

# Launch Yeto training on the entire extracted dataset
yeto launch --model qwen35-9b --data ./cot_training/cot_training.jsonl --budget 40

Testing

• Processed 50+ session files, recovered >15k chars from 32 valid blobs.
• Idempotent: resumes from existing cot_extracted.jsonl.