Allow native review targets to accept supplemental custom instructions
What variant of Codex are you using?
CLI
What feature would you like to see?
I would like codex review / codex exec review to support combining native review targets with supplemental custom review instructions.
Today these modes are mutually exclusive:
codex exec review --base main "Use this plan as context..."
codex exec review --uncommitted "Focus on the acceptance criteria below..."
codex exec review --commit <sha> "Check whether this implements plan-N.md..."
They fail with errors like:
error: the argument '--base <BRANCH>' cannot be used with '[PROMPT]'
The desired behavior is not to replace the built-in reviewer prompt. The custom text should be additional review context/instructions while preserving the native Codex review behavior and diff targeting.
Possible CLI shapes:
codex exec review --base main --instructions "Review against this implementation plan: ..."
codex exec review --base main --instructions-file docs/superpowers/plans/foo-plan-1.md
codex exec review --uncommitted --instructions-file /tmp/review-context.md
or, if compatible with the existing UX:
codex exec review --base main "Supplemental review instructions..."
The important part is that the native target remains native (baseBranch, commit, uncommittedChanges) and the extra text is treated as supplemental review context.
Use case:
I am building automation around Codex review from another coding agent. The workflow is:
- A feature is planned in a design/plan document.
- Code is implemented on a feature branch.
- An orchestrator runs Codex as an independent reviewer.
- The orchestrator reads Codex's findings, applies or rebuts them, and loops until approval or a round cap.
For this to work well, I want to use Codex's native reviewer rather than reimplement code review with a generic codex exec prompt. The native reviewer appears to have better review-specific behavior and target handling. However, the review also needs per-invocation context, such as:
- the implementation plan the branch is supposed to satisfy
- acceptance criteria
- non-goals and intentional tradeoffs
- a requested output contract for automation, e.g. a final
VERDICT: APPROVED/VERDICT: REVISEline - round-specific context in a review/fix/review loop
Current workarounds are weaker:
codex exec review --base mainpreserves native review targeting, but cannot receive plan/context.codex exec review "custom instructions"preserves review mode, but loses native--base/--commit/--uncommittedtargeting.- Plain
codex execwith an inlinegit diffcan be fully customized, but it no longer uses the native review target path and effectively reimplements review prompting. - Writing temporary instructions into
AGENTS.mdis too coarse-grained for per-review context and mutates repo/global state for what should be an ephemeral invocation. - Running native review first and then
resumewith context is possible, but it requires two turns and the first pass can be context-free.
The ideal behavior is a single native review invocation with both:
- deterministic native diff target selection, and
- supplemental caller-provided review context.
Additional information
I found #7825, but it is closed as not planned. I am opening this more specific request because the automation use case needs native target selection plus supplemental per-invocation context, not a full system prompt override.
Related issues:
- #7825 requested prompt support with
--base/--commit; it is closed as not planned, but the CLI still exposes[PROMPT]in help while rejecting it with scoped review flags. - #5598 asks for
/reviewto use an issue summary / discussion context. - #10874 asks for custom review instructions in the Codex App.
- #13163 asks for scriptable review -> fix -> review loops using the same built-in reviewer.
- #6023 asks for more control over system prompts / reviewer behavior in exec workflows.
This request is narrower than system prompt customization. I am not asking to override the reviewer system prompt. I am asking for a native review target plus supplemental user instructions/context in the same invocation.
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗