Codex: requesting-code-review assumes Task/subagent dispatch that native skill install does not provide

[Nickonomic]

Summary

In Codex, the requesting-code-review workflow still assumes a Task/subagent dispatch path that the current native-skill installation does not provide.

The result is that Codex reaches a required review step, then emits a fallback message like:

I can’t dispatch the requesting-code-review subagent here because this environment doesn’t expose the Task/subagent tool it expects.

and continues with a degraded branch-finishing flow instead of performing the intended review.

Why this looks like a Superpowers bug

The current Codex install/docs path only exposes skills/ via native skill discovery:

• .codex/INSTALL.md
• docs/README.codex.md

Those instructions symlink ~/.codex/superpowers/skills into ~/.agents/skills/superpowers, but they do not install agents/code-reviewer.md or describe any Codex-compatible substitute for it.

At the same time, skills/requesting-code-review/SKILL.md still says:

• "Dispatch superpowers:code-reviewer subagent..."
• "Use Task tool with superpowers:code-reviewer type..."

So the Codex-native installation path appears inconsistent with the skill's runtime assumptions.

Reproduction

1. Install Superpowers in Codex using the current native skill instructions from .codex/INSTALL.md / docs/README.codex.md.
2. Complete implementation work in Codex and invoke a workflow that reaches requesting-code-review before merge/finalization.
3. Observe that Codex cannot execute the requested review step because there is no Task/subagent tool available in this environment.
4. The assistant falls back to a manual merge/finish flow and skips the intended reviewer dispatch.

Expected

One of these should happen in Codex:

1. requesting-code-review has an explicit Codex fallback that runs the review inline in the current session using skills/requesting-code-review/code-reviewer.md as the checklist/template.
2. The Codex install/docs also install and wire up whatever agent/review mechanism the skill expects.
3. The skill clearly marks this path as unsupported in Codex and routes to an alternate verification/review workflow instead of failing at runtime.

Actual

The Codex-native install path makes the skill discoverable, but the skill still requires a dispatch mechanism that Codex does not provide, so the review step silently degrades into "can't do that here" behavior.

Extra context

This also seems to conflict with the Codex direction mentioned in RELEASE-NOTES.md for Codex support:

• "Tool mapping for Codex: TodoWrite→update_plan, subagents→manual fallback"
• "Simplified subagent handling (manual work instead of delegation)"

requesting-code-review does not appear to have received that Codex fallback treatment.

Related:

• Officially support Codex #193
• RFC: Prerequisites header for tool-dependent skills #589

[fingolfin]

There is nothing like "hooks" for codex. But the documentation suggest to use $skill-name for explicit skill invocation.

So perhaps instead of writing e.g. "Open new session with executing-plans" it should be "Open new session with $executing-plans" -- at least for a Codex version of this? (I don't know how Claude reacts to those dollars?

(In some of the plans I find notes like "Reference relevant skills with @ syntax" so perhaps claude uses @ instead of $ -- so perhaps one needs a modified version of these skills for codex? Possibly generated from a common "ground truth" ?

[Nickonomic]

Thanks, I tested this locally in Codex CLI 0.111.0 both with a plain requesting-code-review request and with explicit $requesting-code-review.

In both cases Codex loaded the same skill, but it still could not execute it exactly as written because the environment does not expose the Task / superpowers:code-reviewer subagent-dispatch path the skill expects. So this looks like a capability-gap issue rather than an invocation-syntax issue.