| name | codex-cli |
| description | When handing work to the Codex CLI earns its cost, and how to size the run: second-model review, bounded implementation hand-offs, sandbox permissions, model and reasoning effort. Use when the user asks for Codex or `codex exec`, when a change is complex or high-stakes enough that an independent reviewer would change the outcome, or when a delegated run needs its model, effort, or permissions chosen. For agents other than Codex itself. |
| license | Apache-2.0 |
| metadata | {"author":"scarletkc","source":"https://github.com/scarletkc/agents","summary":"Reach for the Codex CLI when a task is hard enough to earn it: second-model review, bounded hand-offs, sandbox permissions, and a model and effort matched to the difficulty."} |
Codex CLI
Codex is a second agent on the same machine, with its own model behind it.
That is the entire reason to reach for it: a model that did not write the code
has no memory of intending it to work. Everything below follows from that one
asymmetry — who wrote it, who reads it, and how much thinking each step is
worth paying for.
This skill is for the supervising agent, not for Codex. If you are Codex,
this does not apply — calling yourself buys nothing but a second opinion from
the same mind. Claude Code is the intended caller.
It also assumes the machine is already set up: codex on PATH, the user
logged in, and whatever MCP servers and tools they want available configured
in ~/.codex/config.toml. If the binary is missing, auth has expired, or a
run dies on permissions, report that plainly and stop — quietly falling back
to doing it yourself hides the fact that the review the user asked for never
happened.
Concrete flags belong to codex --help, which is authoritative and moves
faster than this file. What follows is the judgment.
When to reach for it
Every invocation is a second model spending the user's money and your
wall-clock time. It earns that when the problem is hard enough that another
model changes the outcome — not as a reflex after every edit. Doing the work
yourself and checking it with the project's own tests remains the normal path.
- When the user asks for it. They have already made the call; don't
re-litigate it. Match the model and effort to the task and go.
- After writing something complex or expensive to get wrong. Your own
review of your own diff is the weakest review available, because you are
checking the code against the intent you already have in your head rather
than against what it says. That weakness only matters when the defect would
be costly — concurrency, migrations, security-adjacent paths, platform
assumptions, anything on a compatibility surface. A routine edit that the
suite already covers is not worth a review pass. Counter-example: an agent
changed one side of a path comparison to a normalized form and left the
other side platform-native; every test it wrote passed, because it wrote
them against the same wrong mental model.
- When a demanding change is bounded well enough to describe in a prompt.
A hand-off is worth it when you can state the goal, the files, and the
acceptance check in a paragraph. That paragraph is also the honest test of
whether you understand the change — if you cannot write it, delegating it
just moves the confusion downstream. What "bounded" means is
scoped-change,
and it binds Codex exactly as it binds you: pass the boundary along in the
prompt, because Codex cannot infer where the user drew it.
- When the work is long, mechanical, and verifiable. Wide renames,
repetitive migrations, and mass edits with a green suite proving them buy
throughput rather than insight, and they are cheap to check.
- When you are stuck. After two failed attempts on the same defect, a
third attempt from the same context tends to repeat the second. A fresh
agent with the symptom and the reproduction, and none of your accumulated
theory, is a better use of the next few minutes.
When to keep it
- Ordinary work you can verify yourself. Most changes are this. Writing
the prompt, waiting for the run, and reading the diff costs more than the
edit, and a delegated pass over a small change mostly returns items you
already knew. Absent a reason above, just do it.
- Judgment about words. User-facing copy, documentation, naming, and
release notes need the taste and the context of the session that has been
talking to the user, and they survive delegation badly. See
ux-writing.
- Anything you cannot check afterwards. Delegating work you have no way
to verify converts an unknown into a confident-sounding report, which is
worse than the unknown. Establish the check first.
- Decisions the user reserved. Choosing the approach, committing,
pushing, opening or merging a PR — those stay where the user put them. A
permissive sandbox makes it possible for a delegated run to do all of
them, which is a reason to scope the prompt tightly, not a licence to let
it decide.
Choosing the model
Pick from the task's difficulty, not from habit. Reaching for the strongest
model every time wastes the user's money on renames; reaching for the cheapest
on a subtle bug wastes the user's afternoon.
gpt-5.6-sol — the frontier model. Worth it for reviews that must not
miss anything, root-cause hunts, concurrency and lock-ordering questions,
cross-platform semantics, and any change whose failure mode is silent.
gpt-5.6-terra — balanced, and the sane default for ordinary feature
work inside a boundary you have already defined.
gpt-5.6-luna — fast and cheap with a lower ceiling. Right when a test
suite or a compiler, not the model, is what actually decides whether the
result is correct.
Choosing the effort
Reasoning effort buys deliberation, not knowledge, and it multiplies both
latency and cost. Scale it with how subtle the failure would be:
medium — mechanical work with an immediate, objective check.
high — real implementation work and routine reviews. The usual pick.
xhigh — subtle bugs, unfamiliar subsystems, anything one attempt has
already failed at.
max — the hardest problems, when a wrong answer costs far more than
the extra minutes. Deliberate, not habitual.
The pairing that matters most: review the code at least as high as you wrote
it. A cheap review of an expensive change finds the typos and misses the
reason you delegated it.
Running it
Drive the non-interactive surface — codex exec for work, codex exec review
for review. Model and effort are per invocation: -m <model> and
-c model_reasoning_effort=<level>, both overriding the user's config.toml
defaults for that run only. Prefer codex exec review over the top-level
codex review, which is equally non-interactive but takes the model through
-c model="..." rather than -m. Review scope is --base <branch> for
a branch, --uncommitted for the working tree, --commit <sha> for one
commit. Everything else — --json, output files, resuming a session — is in
codex --help.
Give it the permissions the task needs
This is the step that most often turns a delegated run into a wasted one.
codex exec is sandboxed, and its own default — before the user's config is
applied — is read-only: the model reads the repository, plans the change,
and every write is refused. Failures inside the sandbox are handed back to the
model rather than raised to you, so what returns is a fluent description of a
change that never reached disk.
- Let the user's configuration apply, and reach for
-s mainly to
narrow. Their config.toml already encodes the permission level they are
willing to run at, and on many machines it is the setting that actually
works. -s read-only is a sound narrowing for a review or an
investigation, since nothing should be written anyway.
--ignore-user-config discards their settings wholesale and is rarely what
you want.
- Raising the mode is a request, not a guarantee.
-s workspace-write
asks for a writable workspace; whether it is granted depends on the
platform's sandbox backend and on any .rules policy in effect, and on a
host without a working backend the writes are refused anyway. So confirm
with git status on the target tree rather than with the run's summary.
When the answer is "nothing changed", the fix lives in the user's
configuration or their host setup — say so, rather than rerunning the same
command or escalating the flag yourself.
- Widen the reach deliberately, not by default.
--add-dir makes another
directory writable, which matters when the work spans a worktree and its
main checkout; -C sets the working root; --skip-git-repo-check allows
running outside a repository. workspace-write does not imply network
access — that is a separate setting
(sandbox_workspace_write.network_access), so dependency installs inside
it fail until it is enabled.
- Treat full access as the user's call.
-s danger-full-access and
--dangerously-bypass-approvals-and-sandbox remove the boundary that keeps
a delegated agent inside the task. Some users configure exactly that
globally and are happy with it; inheriting their setting is different from
escalating to it yourself on a task they scoped narrowly.
Write the prompt like a brief
- Put the acceptance check in it. State the goal, the files in scope, the
test or command that proves it, and the boundary it must not cross. Codex
cannot see your conversation with the user, so anything the user said that
constrains the change has to be restated.
- Ask review prompts for specifics. "Review this" returns prose. Naming
what you are unsure of — the migration path, the error handling, the
platform assumption — returns findings you can act on.
Reading the result back
A Codex run is a proposal, not a merge. You asked for a second model precisely
because a single model's confidence is not evidence, and that cuts both ways.
- Read the diff, not the summary. The report describes what Codex meant
to do. Only the diff says what it did, and the gap between the two is where
the surprises live — the unrelated file it touched, the test it relaxed to
make something pass, the fallback it added to keep an error from surfacing.
- Run the suite yourself. "Tests pass" from the agent that changed the
tests is a claim about the same run that produced them.
- Findings are input, not a verdict. A review from a strong model still
produces items that are wrong about this codebase or out of scope for this
change. Judge each one, fix what is real, and say plainly which ones you
dismissed and why — an unexplained dismissal reads as an oversight later.
- Report the division of labour. When the user reads the result, they
should know which parts another agent wrote and what you verified. That is
what makes the supervision worth anything.