Codex CLI examples - Yelu AI API

These examples assume the yelu provider is configured as described in the Codex CLI setup guide. Run them from a Git repository with YELU_API_KEY available in the current environment.

Explore in a read-only sandbox

macOS WSL

codex --model gpt-5.4 --sandbox read-only \
  "Explain this repository's architecture, entry points, external dependencies, test coverage, and operational risks. Cite the files you inspected. Do not modify anything."

Inside the interactive UI, use /status to verify the Yelu provider and /permissions to inspect or change execution boundaries.

Implement and verify a change

macOS WSL

codex --model gpt-5.4 --sandbox workspace-write \
  "Add input validation to the user creation endpoint. Preserve the response contract, add focused tests for valid and invalid requests, run the relevant checks, and show the final diff summary."

workspace-write allows edits inside the workspace while retaining sandbox boundaries. Review the diff and test output before committing.

Review local changes

Review all staged, unstaged, and untracked work without asking Codex to modify it:

Terminal

codex review --uncommitted \
  "Prioritize correctness, authorization bugs, data loss, concurrency hazards, and missing regression tests. Include file and line references."

Review a branch against its base:

Terminal

codex review --base main

Run a non-interactive audit

Terminal

codex exec --model gpt-5.4 --sandbox read-only --ephemeral \
  "Audit authentication and authorization boundaries. Return concrete findings with severity, evidence, and remediation. Do not edit files."

--ephemeral avoids persisting the session rollout. The final response is written to standard output; progress is written to standard error.

Save a final report

codex exec --model gpt-5.4 --sandbox read-only \
  --output-last-message audit.md \
  "Summarize test gaps and propose the smallest high-value test plan."

codex exec --model gpt-5.4 --sandbox read-only `
  --output-last-message audit.md `
  "Summarize test gaps and propose the smallest high-value test plan."

Use an image as context

Attach an error screenshot, UI reference, or architecture diagram:

codex --model gpt-5.4 --image ./error.png \
  "Diagnose the error in this screenshot, find the responsible code, and propose a minimal fix."

codex --model gpt-5.4 --image .\error.png `
  "Diagnose the error in this screenshot, find the responsible code, and propose a minimal fix."

Resume the latest session

Terminal

codex resume --last

Use the session picker with codex resume when several tasks exist for the same repository.

Add repository instructions

Use /init in Codex or commit a reviewed AGENTS.md:

AGENTS.md

# Repository guidance

## Validation
- Run `npm test` for unit tests.
- Run `npm run lint` and `npm run typecheck` before completion.
- Add regression coverage for every fixed bug.

## Boundaries
- Do not read `.env`, credential files, or production exports.
- Do not change lockfiles unless dependencies intentionally change.
- Do not run deployments or irreversible database operations.
- Keep changes focused on the requested task.

Keep sandboxing and approvals enabled for normal local work. Never use --dangerously-bypass-approvals-and-sandbox unless the process already runs inside a separately secured disposable environment.

​Explore in a read-only sandbox

​Implement and verify a change

​Review local changes

​Run a non-interactive audit

​Save a final report

​Use an image as context

​Resume the latest session

​Add repository instructions

Explore in a read-only sandbox

Implement and verify a change

Review local changes

Run a non-interactive audit

Save a final report

Use an image as context

Resume the latest session

Add repository instructions