註冊並分享邀請連結,可獲得影片播放與邀請獎勵。

cv usk
@cv_usk
AI / Software Research Notes AI Agent, LLMOps, MLOps, Software Architecture 投稿は個人の意見です。
加入 May 2026
280 正在關注    413 粉絲
# Codex Features and Practical Usage 🤖 Stop babysitting the terminal — non-interactive mode lets you treat Codex as just another stage in your pipe, ready to drop into CI and scripts. 🏷️ Title: `codex exec` 🔗 URL: 📘 Overview `codex exec` runs Codex once, without launching the interactive UI. You pass the prompt as a single argument, the agent does the work, and only the final message is written to stdout. It is built for CI pipelines, pre-commit hooks, and ordinary shell pipelines. ⚙️ How It Works Progress logs stream to stderr while only the final agent answer goes to stdout, so it composes cleanly with pipes and redirects. Key flags: ・`--sandbox`: `read-only` (default) / `workspace-write` (allow edits) / `danger-full-access` (full access). ・`--ask-for-approval never`: suppresses approval prompts for fully unattended runs. ・`--json`: streams every event as JSON Lines (`thread.started`, `turn.started`, `item.completed`, `turn.completed`, etc.). ・`-o/--output-last-message `: writes the final message to a file. ・`--output-schema `: enforces structured output matching a JSON Schema. ・`-C/--cd `: changes the working directory before running. ・`--skip-git-repo-check`: bypasses the Git-repo requirement (normally required to prevent destructive changes). ・`--ephemeral`: keeps session files off disk. 🛠️ Practical Usage It pairs well with stdin. For example, pipe `npm test 2>&1` into `codex exec "summarize failing tests and propose minimal fix"` and tee the result to a summary file. Structured output gives you stable, machine-readable fields: pass `--output-schema ./schema.json` with `-o ./project-metadata.json`. You can chain sessions into multi-stage pipelines: run an initial `codex exec`, then continue with `codex exec resume --last` to fix what it found. 💡 Use Cases A common split is to trigger on CI failure, generate a patch read-only, then apply it and open a PR in a separate job that holds write permissions. It also works well for log triage — pipe the tail of a log and save root-cause analysis to `analysis.md`. ⚠️ Caveats Never expose API keys as job-level env vars in workflows that check out untrusted code; on GitHub, prefer the official Codex GitHub Action. The `--full-auto` flag is deprecated — use `--sandbox workspace-write` instead. If an MCP server marked `required = true` fails to start, `codex exec` exits with an error. Always pick the least-privilege sandbox for automation. #Codex# #CI#
顯示更多