| name | codex-cli |
| description | Delegates large multi-step work to OpenAI Codex CLI like a subagent—include full goal, prior decisions, scope, and constraints in prompts so isolated sessions stay aligned with the main thread (context handoff per Cognition-style delegation). Use for heavy edits, exploration, and reasoning-intensive tasks. Skip for trivial one-shot tasks or when everything is already in context. |
| allowed-tools | ["Bash","Read","Write","Grep","Glob"] |
OpenAI Codex CLI (subagent/task delegation)
Use OpenAI Codex CLI to run a separate long-horizon pass over the repo: multi-step implementation, broad refactors, batch file writes, or deep exploration—similar to handing a task to a subagent. You stay orchestrator: smaller prompts, less context burn, often lower spend than doing the same work entirely in-session.
Codex CLI (powered by the latest GPT-5 architectures) excels at high-level reasoning, multi-turn autonomous problem solving, and native environment interaction.
When to use Codex CLI
- Large or multi-step work: several files, phases, or checkpoints (feature slice, migration, test suite, docs sweep).
- Heavy code generation or editing: Codex drives
--full-auto tool use while you summarize outcomes and merge.
- Reasoning-heavy tasks: Successor to the o-series; use when PhD-level logic and native "Thinking" modes are required.
- Parallel mental lane: you continue planning or reviewing while Codex runs a bounded task (optionally background—see reference.md).
- User explicitly asks for Codex or “use Codex for this.”
When not to use
- Small / single-step tasks answerable with one or two edits or a short explanation.
- Tight feedback loops where the user wants rapid back-and-forth refinement in one thread.
- Secrets or policy-sensitive flows—avoid piping credentials; redact before delegating.
- Already-loaded context where duplicating the whole plan adds no value—handle locally.
- Low ROI (Return on Investment): If the task is "needle-in-a-haystack" (requires high precision over a single line) or if the time to compose the Handoff Table exceeds the time to simply edit the file locally. Delegation should only be used when the "mental offloading" outweighs the "handoff overhead."
Delegation and context (critical)
Isolated subagent context saves tokens but splits the story: Codex does not see the main session's full thread. Poor handoffs cause misread subtasks, conflicting assumptions (stack, style, APIs), and wasted edits—see Cognition’s discussion of sharing context and implicit decisions in multi-step setups.
When composing the single Codex prompt (using exec), treat it as passing enough shared state, not just a title:
| Include | Why |
|---|
| Original goal | Same north star as the user—not only the immediate micro-task. |
| Decisions already made | Framework, patterns, naming, auth approach, “use X not Y”—anything that would otherwise be guessed wrong. |
| Scope | Paths, modules, and explicit out of scope / do-not-touch areas. |
| Constraints | Performance, a11y, compatibility, review gates, “no new deps,” etc. |
| Verification | Explicit command (e.g. npm test, lint) the subagent must run and pass before returning. |
| Expected output | e.g. “summarize then list files changed,” “report only—no edits,” or “apply edits with minimal diff.” |
After Codex returns, pull decisions and constraints back into the main thread (what it assumed, what it changed, open risks). Prefer sequential delegations with explicit carry-over over parallel runs that might diverge unless they share the same briefing.
If the delegation would need a long transcript to be safe, summarize the relevant parts into the prompt (compressed “state of the union”) rather than a one-line subtask.
Model Selection & Discovery (Mandatory)
MANDATORY: Search the web for latest OpenAI model names and pricing before selecting a model. You should also consult Artificial Analysis for the most up-to-date benchmarks, pricing, and model performance data. GPT-4o is legacy/old for Codex CLI.
- Default (Reasoning/Complex):
gpt-5.4-thinking (Successor to o3; PHd-level logic).
- Standard (Speed/Implementation):
gpt-5.4 (Balanced, incorporates GPT-5.3-Codex logic).
- Strategy: Default to
gpt-5.4 for standard tasks and gpt-5.4-thinking for complex architecture. Verify current version via search to avoid stale names.
Programmatic usage (required)
You MUST use Codex CLI programmatically via the exec command. Do NOT start interactive sessions.
| Requirement | Flag |
|---|
| Non-interactive | exec "prompt" |
| Auto-approval | --full-auto or -a full-auto |
| Model Selection | --model [model_name] |
| Ephemeral Run | --ephemeral (Prevents local session bloat) |
Command pattern
codex exec "[prompt]" --full-auto --model gpt-5.4 2>&1
After Codex returns
- Review diffs and security-sensitive areas (XSS, injection, auth)—do not merge blindly.
- Run project checks (
lint, test, typecheck) as appropriate.
- Compress results for the user: summarize results for the user instead of pasting huge logs unless asked.
- Reconcile context: note decisions, files touched, and remaining risks so the main session stays aligned.
Quick prompts
- Delegate implementation:
codex exec "GOAL: [goal] | DECISIONS: [decisions] | SCOPE: [paths] | CONSTRAINTS: [constraints] | VERIFICATION: [test_command] | OUTPUT: [format]" --full-auto
- Logic Check:
codex exec "GOAL: Analyze logic for edge cases and potential race conditions | SCOPE: [paths] | VERIFICATION: [check_command] | OUTPUT: detailed report" --model gpt-5.4-thinking
- Audit:
codex exec "GOAL: Perform a security audit of the authentication layer | SCOPE: [paths] | OUTPUT: audit report" --full-auto
More detail
