| name | codex |
| description | Use the local Codex CLI as an independent second agent. Two branches โ (1) proactively run `codex review` for a second opinion after completing a substantive change, before presenting it as done or committing; (2) delegate a well-defined implementation task via `codex exec`, ONLY when the user explicitly asks for Codex to do it. Also covers how to prompt Codex. |
Codex
Codex is an independent agent on PATH (codex โ invoke as command codex if
a shell alias shadows it), sharing this working tree and already authenticated. It is a second opinion, not ground truth: verify what it
reports, own what it changes. It reads the same skills your repo carries.
If codex is not installed
When codex is missing from PATH, offer to install it โ ask the user for
approval first, never install on your own initiative. On yes, follow the
current instructions at https://developers.openai.com/codex/cli. First-run
authentication is interactive โ hand that step to the user. Verify with
codex --version before proceeding.
Prompting Codex
Prompt Codex like an operator, not a collaborator: compact, block-structured
with XML tags. State the task, what "done" looks like, and the few constraints
that matter. A tighter prompt beats a bigger run โ improve the contract before
raising --effort.
- One task per run. Split unrelated asks (review, then fix, then docs) into
separate runs; a mixed prompt gets a mixed result.
- Name skills instead of restating them. Codex reads the same skills your
repo carries โ say "follow write-docs for the doc", "obey refactor-clean: no
compatibility wrappers." Don't re-explain what a skill already carries.
- Blocks, added only where the task needs them:
<task> โ the concrete job, the repo/failure context, the expected end
state. Nearly always present.
<output_contract> โ exact shape, highest-value first, compact.
<default_follow_through> โ take the low-risk interpretation and keep
going; stop only when a missing detail changes correctness, safety, or an
irreversible action.
<verification_loop> โ before finalizing, check the result against the
requirements and the changed files; revise rather than ship the first
draft. Any risky fix.
<grounding> โ ground every claim in code or tool output; label inferences
as inferences. Review and research.
<action_safety> โ keep the diff tightly scoped; no drive-by refactors.
Write tasks.
- Anti-patterns: vague framing ("take a look"), no output contract ("report
back"), "think harder" in place of a contract, mixing jobs in one run, and
demanding certainty the evidence can't support.
Models (GPT-5.6 era โ verified 2026-07-10)
The default model lives in ~/.codex/config.toml (currently gpt-5.6-sol,
model_reasoning_effort = "medium"). Codex is the only local route to GPT-5.6 โ
PAL MCP still tops out at gpt-5.5-pro.
| Model | -m value | Use for |
|---|
| GPT-5.6 Sol | gpt-5.6-sol | Flagship. Ambiguous, difficult, or high-value work: complex code changes, deep review, research. Default when unsure. |
| GPT-5.6 Terra | gpt-5.6-terra | Everyday workhorse โ GPT-5.5-level quality, cheaper/faster. Natural replacement for tasks previously sent to gpt-5.5. |
| GPT-5.6 Luna | gpt-5.6-luna | Fast, well-defined, repeatable work: extraction, classification, transformation, mechanical edits. |
| GPT-5.5 | gpt-5.5 | Prior-generation frontier; still available as a fallback. |
| GPT-5.3 Codex Spark | gpt-5.3-codex-spark | Near-instant iteration preview (ChatGPT Pro only). |
- Always pass explicit tier IDs. The bare
gpt-5.6 alias hangs codex exec
indefinitely on this setup (process alive at 0% CPU, no output), and the
interactive /model picker doesn't list the 5.6 models even though -m
works (known upstream issue). gpt-5.6-sol|terra|luna all work.
- Reasoning effort:
-c model_reasoning_effort=<value>. The API enum is
none|minimal|low|medium|high|xhigh; on top of that the CLI accepts max
(deepest single-task reasoning) and ultra (Sol Ultra โ fans out subagents
for work that splits into parallel chunks). All of these are accepted by
codex exec too, not just interactive sessions (verified 2026-07-10).
There is no exact effort mapping from 5.5 to 5.6 โ start lower than you
think and raise only if the result is shallow. Most tasks don't need
max/ultra; reserve ultra for genuinely decomposable work and expect
it to burn plan quota fast.
- Deprecated:
gpt-5.2 and gpt-5.3-codex are dead under ChatGPT sign-in โ
fix any script still passing them.
Review โ proactive
Run a Codex review whenever you have a substantive diff you'd want a second set
of eyes on โ a refactor, a tricky algorithm, renderer work, a security-sensitive
change โ before declaring it done or committing. Skip it for trivial edits
(typos, comments, doc-only).
- Pick the diff scope:
codex review --uncommitted for working-tree changes,
--base <branch> for a branch diff, --commit <sha> for a landed commit.
Scope flags and custom instructions are mutually exclusive (despite what
--help implies): codex review "<instructions>" reviews the default
scope with your framing, a scope flag takes no prompt. When you do write
instructions, scope the risk area โ never state the answer you expect
(unprimed, same discipline as
screenshot-critique).
- Triage every finding: confirm it against the code before acting. Preserve
Codex's evidence boundaries โ an inference it labelled is not a fact.
Overlap with your own doubts is high-priority evidence; a finding you
dismiss needs a stated reason, not silence.
- Report the outcome to the user โ what Codex flagged, what you fixed, what
you dismissed and why. Done when every finding is either fixed or
explicitly dismissed.
Implementation โ explicit ask only
Delegate implementation to Codex only when the user names Codex for the task.
Never hand it work on your own initiative, and never re-delegate follow-up
work without a fresh ask.
- Slice the task sharp before delegating โ goal, constraints, and how to
verify โ using the prompt discipline above. An underspecified task stays
with you until a fresh agent couldn't misread it.
- Start from a clean tree (or record the baseline commit) so Codex's diff is
separable from yours.
- Pick the sandbox by what the task must RUN:
- Pure code + typecheck/unit:
codex exec --sandbox workspace-write "<task>".
Network is off; add -c sandbox_workspace_write.network_access=true only
when the task must fetch (e.g. new deps).
- Browser verification, dev servers, or full test runs: use
codex exec --dangerously-bypass-approvals-and-sandbox "<task>". The
sandbox blocks localhost binds (listen EPERM on vite/playwright), so a
sandboxed codex ships code it never saw run. Bypass trades that blindness
for zero OS control: only in a dedicated git worktree, only with a prompt
you authored end-to-end (never relaying third-party text), and the diff
review you owe afterwards is the control.
Use -o <file> to capture the final message and background long calls;
suppress codex's stderr thinking-noise with 2>/dev/null so it doesn't
bloat your context (drop it only to debug a failing run), and add
--skip-git-repo-check to run outside a git repo. Non-interactive runs
never ask for approval either way.
- Follow up with
codex exec resume <session-id> "<follow-up>", taking the
id from the run header. resume --last means the most recent session
globally โ a review or any other codex run in between will hijack it. If
two resume rounds don't converge, stop delegating and finish it yourself โ
iterating a confused agent costs more than taking over.
- You own the result: read the full diff, run the tests, and only then report
it. "Codex says it's done" is not done.
Exec liveness โ a hang looks like work
A backgrounded codex exec can wedge at startup: process alive at ~0% CPU, but
no session file under ~/.codex/sessions/<Y/M/D>/, no network socket, no tree
changes. "Process running" is NOT "working."
- Stdin is always a file or
/dev/null, never an inherited terminal. Left
on an interactive/piped stdin with no prompt source, exec prints Reading additional input from stdin... and blocks forever โ the most common hang.
Either pass the prompt as a positional arg with < /dev/null, or feed a
prompt file with codex exec [flags] - < prompt.txt (the - makes codex
read the task from stdin, and a missing file fails the redirect loudly โ
unlike "$(cat prompt.txt)", which silently sends the fallback string as
the task). Add nohup/& as needed.
- Watchdog: put a unique marker in the prompt, then kill the exec if
grep -rl "<marker>" ~/.codex/sessions/<Y/M/D>/ finds no session within
~3 minutes. Relaunching after a kill reliably works.
- Trust the worktree. Headless exec in a directory Codex doesn't trust can
block forever on an invisible prompt. Git worktrees are separate paths from
the trusted repo root โ add
[projects."<worktree-path>"]\ntrust_level = "trusted" to
~/.codex/config.toml before exec'ing in one. This is the safe fix; the
bypass flags stay forbidden here.
- Don't launch two execs in the same instant, and kill stale hung execs
before starting a new one.
- Launch from the repo/worktree ROOT. The writable sandbox root is the
CWD at launch: exec'd from a subdirectory (e.g.
web/), every edit outside
it is rejected as "writing outside of the project" and a never approval
policy can't recover โ the run burns with zero files changed. Pass
-C <root> to set the working root explicitly instead of cd-ing into it.
Rules
- Don't touch the working tree while a Codex exec is running on it.
--sandbox read-only (the default) for consultation and questions;
workspace-write only for delegated implementation.
--dangerously-bypass-approvals-and-sandbox is reserved for tasks that must
run browsers/servers/full suites (above) โ dedicated worktree, self-authored
prompt, mandatory diff review after. --full-auto has been removed (it
aliased workspace-write) โ don't reach for it.
- Reasoning effort and model are config/flag overrides โ
-c model_reasoning_effort=high, -m <model> (the old --effort flag is gone
in current Codex). Leave both at their defaults unless the user asks or the
task clearly fits a cheaper tier (see Models above) โ tighten the prompt
first. When overriding, use explicit tier IDs, never the bare gpt-5.6
alias.