little-loops on Codex CLI¶

little-loops integrates with OpenAI Codex CLI as a first-class host. Hook intents, orchestration CLIs, and skill/command discovery all work under Codex — with a small set of limitations documented below.

New to Codex + little-loops? Start with Getting Started.

What works¶

Hook intents¶

Intent	Status
`session_start`	✓ wired (matcher: `startup` only)
`pre_compact`	✓ wired
`user_prompt_submit`	✓ wired
`post_tool_use`	✓ wired (fire-and-forget, ≤5s timeout)
`pre_tool_use`	opt-in (see Usage → Opt-in pre_tool_use)
`stop`	deferred — no current consumer
`post_compact`	deferred — no current consumer
`permission_request`	deferred — no current consumer

Hook intents reach the same host-agnostic Python dispatcher (scripts/little_loops/hooks/) as Claude Code and OpenCode. The Codex adapter (hooks/adapters/codex/) is a thin Bash shim with no logic of its own.

Orchestration CLIs¶

ll-auto, ll-parallel, ll-action, ll-loop, and ll-sprint all route through the host runner abstraction and work with Codex. Set LL_HOST_CLI=codex or let the runner auto-detect the codex binary on PATH. See Usage → Running orchestration CLIs.

Skill and command discovery¶

After running ll-adapt-skills-for-codex --apply once, all ll skills and commands appear in Codex as ~/.codex/skills/<name>/SKILL.md entries, making /ll:* slash commands discoverable from the Codex TUI.

What is deferred¶

stop / post_compact / permission_request: Codex fires these events, but little-loops has no current consumer for them. Hooks for these intents are not wired in .codex/hooks.json by default.
--agent (persona selection, ll-orchestrated): Now partially supported via prompt injection (ENH-1533). When ll's orchestration layer (ll-auto, ll-parallel, ll-loop) requests a persona via --agent <name>, CodexRunner.build_streaming reads .codex/agents/<name>.toml, extracts developer_instructions, and prepends a [Persona: <name>] block to the prompt. Generate the TOML files once with ll-adapt-agents-for-codex --apply. When the TOML is absent, CodexRunner falls back to emitting CapabilityNotSupported plus a stderr notice and proceeds with Codex's default configuration. Permanent native-flag gap: Codex has no --agent CLI flag (openai/codex#10067); the prompt-injection workaround is the only way to apply ll-defined personas. Behavioral note for ll-doctor/CI consumers: agent_select is now "partial" instead of "unsupported", so Codex hosts that previously failed ll-doctor solely on agent_select will now exit 0. See Usage → Current Limitations.
--tools (sandbox modes): Partially supported; only basic sandbox modes pass through.

Run ll-doctor (or ll-doctor --json) to see exactly which capabilities and hook intents are wired for your active host.