daemon-backend-codex

# Codex Daemon Backend — Flag Discovery Entrypoint The installed CLI's own help is the authority for Codex flags; this page is only the entrypoint. Conversion rules, key safety, and persistence live in the parent [`reference/cli-backends/SKILL.md`](../../../SKILL.md). ## Discover flags from the installed CLI 1. Run, in bash: `codex --version`, `codex --help`, and `codex exec --help`. The daemon backend wraps `codex exec`, so `codex exec --help` is the relevant flag surface. These are local read-only commands; no session is started. 2. Translate what you found into `backend_options` with the parent's generic conversion rules. Nothing Codex-specific is added to that contract here. ## Example: reasoning effort via the generic `config` route Codex exposes most of its tunables as repeated `-c, --config <key=value>` overrides (see `codex exec --help` for the key=value syntax). Through `backend_options`, a list value repeats the flag once per item: ```jsonc { "backend": "codex", "tasks": [{ "task": "Implement and validate the change.", "tools": [], "backend_options": { "config": ["model_reasoning_effort=\"ultra\""] } }] } // argv: --config model_reasoning_effort="ultra" ``` The effort vocabulary belongs to the installed CLI and the selected model — LingTai does not validate, enumerate, or simulate effort levels. A value like `ultra` passes through and the CLI/model decides its semantics (or rejects it). ## Installation Use [OpenAI's current official Codex documentation](https://developers.openai.com/codex/) for installation and upgrades; this manual maintains no vendor install/update commands or pinned version. Any actual install or update still needs its owner's authorization. Verify the resulting `codex --version` and `codex exec --help` before relying on a flag. ## Configuration ### Auth & API key The external CLI uses its existing authorized account and credential context. Consult installed help and OpenAI's documentation for subscription/API-key or other provider configuration; do not infer CLI account selection from a LingTai preset or pool label. This manual grants no login/logout, account/profile switch, credential read/copy or configuration-repair authority. Never print credentials or whole auth/config files while diagnosing; escalate an unavailable approved account rather than silently switching it. ### Models Model names and reasoning levels belong to the installed CLI and selected account. Query its help/vendor documentation instead of copying a "latest" model list or vendor configuration schema here. Per-run model selection passes through `backend_options` (`"model": "<name>"` → `--model <name>`), under the parent's generic conversion rules and the task's actual model/authority contract. ## Style axis: Codex vs Claude Code Codex and Claude Code are both LingTai daemon backends; the choice between them is about *style*, not just run shape: - **Codex** — tightly-scoped diffs, deterministic refactors, mechanical validation sweeps. More conservative; the right choice when the change is well-specified and the scope is clear. - **Claude Code** — exploratory code reading, multi-file edits, skill/doc work, PR composition. See `reference/backends/claude-p/SKILL.md`. ## Capability note Codex CLI is open source and built in Rust. Beyond the headless `codex exec` surface the daemon wraps, the CLI also ships an interactive TUI (native Vim editing, plugin management, hooks) and a Chrome extension, plus a headless remote-control app-server. Those are interactive features; the daemon backend only exercises `codex exec`, whose live surface is `codex exec --help` — not a static catalog here. Official docs: https://developers.openai.com/codex/ ## Harness boundary Codex currently declares no reserved-flag list at the validation layer, so nothing is refused for this backend beyond the generic key/value safety rules. Still, do not re-set harness-owned surfaces (`--json`, sandbox/approval bypass, or `mcp_servers.daemon_common.*` config keys): breaking them silently breaks progress/result extraction and completion enforcement. ## Troubleshooting Installation failures, missing/invalid credentials, and plugin-marketplace issues are the installed CLI's own concerns — use installed help and OpenAI's official docs above; do not print or `cat` a credential file/env var into agent-visible output while diagnosing. The one failure mode that is actually LingTai's contract to diagnose: 1. **Agent appears stuck while `codex exec` runs.** You likely used synchronous CLI for work that should have been daemon-backed or supervised in the background. Inspect the child process and worktree; if needed, kill the child so the blocking call returns, then resume with the Codex daemon backend or a supervised background wrapper that records logs, timeout, cancellation path, and recovery notes.