SuperAPI AI Directory

codex-harness

General & other

View repository

Updated 10-16 07:14

About this skill

Codex harness (OpenAI Codex CLI)

Prefer codex exec from a host agent (no TUI). Use the TUI only with a real PTY.

Verified against codex-cli 0.147.0 via the official Windows installer. Re-check codex exec --help if flags look stale.

Do not invent terminal() / process() APIs. From Grok or PowerShell, call codex. Shared job shape: harness-offload. Not the ChatGPT desktop app (codex app). Do not install winget OpenAI.Codex or jcv8000.Codex.

Install (Windows)

$env:CODEX_NON_INTERACTIVE = '1'   # skip "Start Codex now?"
powershell -ExecutionPolicy Bypass -c "irm https://chatgpt.com/codex/install.ps1 | iex"

That writes codex.exe and codex-code-mode-host.exe to %LOCALAPPDATA%\Programs\OpenAI\Codex\bin (junction into ~/.codex/packages/standalone/current/bin) and prepends that dir to User PATH. codex update works on this layout.

Refresh PATH in this process, then:

$env:Path = [System.Environment]::GetEnvironmentVariable('Path','Machine') + ';' +
            [System.Environment]::GetEnvironmentVariable('Path','User')
where.exe codex
codex --version

There is no codex-cli command. Use codex.

Auth

ChatGPT device login (desktop / headless):

codex login --device-auth
# open https://auth.openai.com/codex/device and enter the one-time code
codex login status
codex doctor --summary

API key (automation; do not persist in skills). CODEX_API_KEY is honored on codex exec. ChatGPT login may still require a paid plan that includes Codex. --oss is local (lmstudio / ollama), not ChatGPT.

Home: ~/.codex/ (CODEX_HOME). Do not commit auth.json.

Drive from this host (one-shot)

This machine is host-native. ~/.codex/config.toml:

sandbox_mode = "danger-full-access"
approval_policy = "never"

Do not set [windows] sandbox = "unelevated" or "elevated" — those pick a Windows restricted-token sandbox, not “run as me.” Stock exec with no user config is read-only; this host is not stock. Invoke-HarnessOffload -AllowTools passes --sandbox danger-full-access.

codex exec prints the final message on stdout, progress on stderr. --json is JSONL. Non-git cwd needs --skip-git-repo-check or exec refuses.

codex exec --json --skip-git-repo-check -C <dir> "Your question or task"
Flag Meaning
exec / e Non-interactive
--json JSONL events on stdout
-o / --output-last-message Also write the final text
-C / --cd Workspace root (not valid on exec resume)
-m / --model Model id
-s / --sandbox read-only | workspace-write | danger-full-access
--approve-for-me Auto-reviewed approvals (0.147+)
--dangerously-bypass-approvals-and-sandbox Per-run unsandbox (this host already is)
--skip-git-repo-check Allow a non-git dir
--ephemeral Do not persist session files
--output-schema Constrain the final message
--search Enable web search

-a / --ask-for-approval is TUI only. Passing it to exec errors. --full-auto is gone (0.147).

codex debug models (this account): gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4-mini, codex-auto-review.

Resume:

codex exec resume --json --skip-git-repo-check <session-id> "Follow-up"
codex exec resume --last --json --skip-git-repo-check "Follow-up"

--json includes thread.started.thread_id — that is the resume id.

{"type":"thread.started","thread_id":"<uuid>"}
{"type":"item.completed","item":{"type":"agent_message","text":"…"}}
{"type":"turn.completed","usage":{…}}

Smoke:

codex exec --json --skip-git-repo-check -C <dir> "Respond with exactly: CODEX_SMOKE_OK"

Success = an agent_message whose text contains CODEX_SMOKE_OK.

TUI (human or PTY)

codex
codex resume --last
codex resume <session-id>

Bare codex opens the TUI and hangs a non-PTY shell. Use codex --help and codex exec.

Structured output / review / MCP

--output-schema <schema.json> + -o out.json constrains the final message (verified: {"token":"SCHEMA_OK","ok":true}).

codex review --uncommitted          # git repo; do not also pass a prompt arg
codex review --base main
codex review --commit <sha>

--uncommitted cannot take a prompt argument (including - / stdin). Verified: flagged a wrong add() as P1.

codex mcp-server is stdio MCP (protocolVersion 2024-11-05). Tools:

Tool Role
codex New session. Required prompt. Optional cwd, model, sandbox, approval-policy. Returns threadId + content.
codex-reply Continue. prompt + threadId (legacy conversationId).

That is the peer offload path (same job as exec / exec resume).

Bundled skills live under ~/.codex/skills/.system/ (openai-docs, review-agent, skill-creator, …). codex debug prompt-input shows them injected as <skills_instructions>.

Other commands

codex login status
codex logout
codex mcp list                 # none configured on this host
codex plugin list              # openai-curated marketplace; install only if asked
codex features list
codex debug models             # JSON catalog
codex doctor --summary
codex update

Pitfalls

  • Official installer only — not winget OpenAI.Codex / jcv8000.Codex.
  • Do not --disable code_mode_host — file tools fail closed.
  • Non-git cwd needs --skip-git-repo-check.
  • --json makes stdout JSONL; parse it or use -o.
  • exec resume does not take -C; cd first (the dispatcher already does).
  • Host is PowerShell: do not put Linux 2>/dev/null in double-quoted strings (see wsl-containers).
  • Offload: harness-offload. MCP peer: codex mcp-server.