Codex Agents covers requests like “ask Codex”, “check with Codex”, handing Codex a build or review, or being stuck between two framings. It treats Codex as a peer with the same standing as Claude: argue decisions with it, let it build a stage, and let it review code it did not write.
The wrapper codex.py resumes the recorded session per directory and re-sends model, effort and sandbox on every turn, since a resume otherwise falls back to the config defaults.
What You Get
- codex.py, a Python wrapper that records session id, model and sandbox and resumes them
- Dispatch rules for briefs, instruction-file budgets and installed skill packs inside Codex
- Guidance on resume versus fresh sessions and one worktree per parallel build
- Warnings about sandbox limits for git and browser tests, observed and dated
- Binary lookup order: CODEX_BIN, codex on PATH, then the VS Code extension bundle
Common Use Cases
- Debating a load-bearing decision with Codex before building
- Handing Codex a build stage in its own worktree
- Running Codex as the fresh reviewer for a release
- Resuming an earlier Codex session by id
What Makes It Unique
The skill is built around what goes wrong silently: a resume that drops the model, a resume that drops the sandbox so a build lands read-only and changes nothing, an instruction file cut by Codex’s context budget, and a round that argues a verdict while writing no files. The wrapper prints the reported model and sandbox after each run, and the skill requires checking git status before believing a build round happened.
Under The Hood
The skill is instructions, references and a Python wrapper that uses the standard library. The wrapper runs the Codex command-line tool as a subprocess in the repository you name, with the brief read from a file, and stores a marker file with the active session. Depending on the sandbox you choose, Codex can write files and run commands, including tests, in that repository. The skill also tells the agent to run codex plugin commands to check or install skill packs. It does not fetch web pages itself.