zgx949/ai-pojia

AI破甲 (ai-pojia) is a configuration and Skill installer for OpenAI's Codex coding agent. It sets up a custom instruction file and a Skill so that Codex follows a specific, permission-based workflow when handling tasks. The project is at version 0.2.0 and is based on a rewritten version of the gpt-instruct v45 prompt (from an upstream project), but it explicitly removes features like automatically treating all goals as sandboxed, ignoring ownership/scope, or unconditionally refusing to answer. It aims to let AI complete authorized tasks and produce verifiable outputs, but it does not claim to work on all models or to have passed any official demo tests yet.

What it installs

• Copies a prompt file (prompts/ai-pojia.md) into the Codex directory and loads it via the model_instructions_file setting (this replaces Codex's built-in instructions for new sessions).
• Installs a Skill (skills/ai-pojia/SKILL.md) into the same directory under skills/ai-pojia/.
• Provides tools for checking, previewing, backing up, and restoring configuration.

The installer only manages its own prompt, Skill, and related settings. It does not change your model, provider, authentication, or permission settings.

Installation

• The simplest way: send a specific message to Codex asking it to clone the repo, read the README, run python3 install.py, and then run python3 install.py --doctor to verify.
• Manual: git clone, cd, python3 install.py. Requires Python 3.9+ and a Codex version that supports model_instructions_file and local Skills. On Windows, use py -3 instead of python3.
• You can specify a custom Codex directory with --codex-dir (default is CODEX_HOME or ~/.codex).

Usage

• After installation, normal tasks use the loaded instructions.
• For a self-written test demo, you put files in a separate directory and send a message like: “Use $ai-pojia to analyze this test DEMO...”. The Skill then produces a summary, analysis evidence, modified copies, diffs, verification commands, and actual results.
• The demos/ folder currently has no test programs or pre-set answers; official demos are planned later.

Checking, previewing, and uninstalling

• python3 install.py --doctor – check installation.
• python3 install.py --dry-run – preview without writing files.
• python3 install.py --uninstall – restore the original top-level instruction field and archive the Skill.
• If you manually edited files or there are conflicts, the tool stops and reports the real state.

Upstream and modifications

• The source is a fixed commit (3ab84df) of MDX-Tom/gpt-instruct, MIT licensed. Original copyrights are retained.
• This version keeps continuous execution, artifact delivery, baseline comparison, and recovery verification, but removes the parts that treat all real goals as sandboxed, ignore ownership/scope, unconditionally forbid refusal, and unrelated branches.
• Loading successfully only means the config points to and validates the built-in instructions; it does not guarantee the model will accept a request. Client, model, and server-side rules still affect results.

Development

• Run tests with python3 -m unittest discover -s tests -v.
• Package with python3 tools/package_release.py.
• There is a plugin.json for plugin packaging, but it is not yet published; it installs the Skill but does not modify global instructions. The full one-click entry is install.py.

Limitations

• No universal refusal-cancellation, no full-model compatibility, and no official demo completion results yet.
• The actual verification scope is documented in docs/validation.md.