MDX-Tom/gpt-instruct

A Codex jailbreak prompt and test pack for gpt. 针对 gpt 系列的 Codex 破甲提示词与测试包。

gpt‑instruct – Prompt‑and‑evaluation toolkit for OpenAI Codex

What it is – A small, open‑source utility that bundles versioned prompt files (called instructions) for the Codex code‑generation model together with a reproducible test harness. The repository maintains two parallel product lines:

Line	Status	What it contains
gpt‑5.6‑sol‑v45	Stable production release	The original prompt bytes (≈8 KB) packaged as a ZIP and a Markdown instruction file ready to be dropped into a Codex installation.
gpt‑6‑astra‑v1‑rc1	Early‑pre‑release	A snapshot of the best‑performing prompt draft from the “e1b5” epoch, intended for evaluation only.

Both lines share the same test suite, failure‑analysis pipeline, and artifact‑verification rules, but results are only compared when the model, reasoning level, and execution context match.

Core ideas

Prompt versioning – Each epoch can have up to 20 numbered versions (gpt‑6‑astra‑v1‑e<epoch>b<attempt>). The repository keeps the raw prompt bytes and a human‑readable Markdown description.
Gate‑kept release process – New prompts must pass three increasingly strict test tiers (A → B → C) before they are promoted:
- A – 3 user‑feedback cases, all turns succeed.
- B – 66 regression issues (74 turns) all succeed.
- C – 120 medium‑difficulty test cases, run only after A and B are clean.
Reproducible evaluation – Scripts in scripts/ unzip the prompt archives, run the Codex model with the specified reasoning level (medium), score the outputs, and verify them against stored manifests.
Safe deployment & rollback – codex‑instruct.py can apply a prompt version, snapshot the existing model_instructions_file entry in config.toml, and later restore it with --reset or --restore-snapshot.

Quick start (stable line)

# clone the repo
git clone https://github.com/MDX-Tom/gpt-instruct.git
cd gpt-instruct

# dry‑run – just preview what would be written
python3 codex-instruct.py --apply --version gpt-5.6-v45 --dry-run

# actually install the stable prompt set
python3 codex-instruct.py --apply --version gpt-5.6-v45

You can point the script at a custom Codex home with --codex-dir ~/.codex or install a ZIP/Markdown directly with --file ./my‑prompt.zip.

Project layout

gpt-instruct/
├─ codex-instruct.py          # CLI for applying, resetting, snapshotting prompts
├─ sync-archives.py           # Keeps source Markdown and ZIP in sync
├─ gpt-5.6-sol-v45.*          # Stable prompt (MD + ZIP)
├─ gpt-6-astra-v1-rc1.*       # Early‑pre‑release prompt
├─ historical-versions/      # Archived older releases
├─ scripts/                   # Evaluation, scoring, and verification scripts
├─ tests/                     # A/B/C test case manifests
├─ docs/                      # Architecture diagrams, comparison reports
└─ reports/                   # Local run evidence (not committed)

Who might use this?

Researchers exploring prompt engineering for code‑generation models and needing a disciplined, version‑controlled workflow.
Developers who want to replace the default Codex instruction set with a community‑vetted one while retaining the ability to roll back safely.
Safety engineers interested in the structured A→B→C gate process and the explicit separation of model‑failures from network or policy interruptions.

License & acknowledgements

Licensed under the MIT License.
Built on top of the open‑source work from yynxxxxx/Codex-5.5-codex-instruct-5.5.

All details are taken directly from the repository’s README; no additional features are inferred.