Codex Plan Mode: from a good idea to an implementation-ready plan

Planning earns its cost when the work crosses boundaries or the requested outcome is not yet implementation-ready.

OpenAI recommends Plan Mode for complex, ambiguous, or difficult-to-describe work. In the current Codex CLI, /plan switches the active chat into Plan Mode and can carry an inline request. It is temporarily unavailable while Codex is already working. See the current command reference.

Search results and team conversations often collapse these concepts into one. Keeping them separate makes the implementation handoff much easier to reason about.

The archived OpenAI ExecPlan guide remains useful for the principle of a self-contained, living plan. Its older model recommendations should not be treated as current Codex guidance.

A useful starting request gives the investigation a destination without pretending you already know the implementation.

/plan

Goal:
Relevant context:
Constraints and non-goals:
Done when:

Inspect the repository before asking questions. Resolve discoverable facts yourself. Ask me only
about product intent or implementation trade-offs that materially change the plan. Do not finalize
until the plan is decision-complete.

This is a useful default derived from OpenAI’s current Codex best-practices guide, not a required schema. Its job is to make the destination testable.

The current open-source Plan Mode contract directs Codex to investigate with non-mutating tools, resolve discoverable facts itself, and ask the user about intent or trade-offs that materially change the plan. That creates a useful ownership split.

“Which option do you want?” is weak when the options arrive without evidence. A useful decision prompt explains what Codex found, why the choice remains open, and what each option changes.

A long plan can still be speculative. A short plan can be complete. Approve the proposal only when these six gates have defensible answers.

Audit this proposed plan before I approve it.

Find unsupported assumptions, unresolved decisions, omitted interfaces or data flows, missing
compatibility, migration, rollback, failure, or observability concerns, weak tests or acceptance
criteria, and steps that are too broad, duplicated, or code-heavy.

Inspect the repository to resolve discoverable gaps. Ask me only for decisions I own. Then return a
revised decision-complete plan.

This six-part gate is Supercode’s compact editorial synthesis of the current Plan Mode contract and durable execution-plan guidance. It is not an official Codex checklist and cannot guarantee a correct implementation.

The plan is approved. Leave Plan Mode and implement it in the stated order.
Track execution progress separately. Run the listed verification after each meaningful stage, and
stop if repository evidence invalidates an accepted assumption.

During execution, a progress checklist can track pending, active, and completed steps. It does not replace the accepted plan. If repository evidence invalidates a decision, update the plan deliberately instead of allowing silent drift.

Public Codex issue reports show that shallow questions, plan bloat, transition friction, and an unexpected edit have all been experienced by individual users. They motivate bounded guardrails; they do not establish prevalence or reproduce every current version.

Native Plan Mode is enough for a bounded task that moves directly into implementation. When the plan must become repository knowledge, Supercode Specs & Docs provides a guided “Create implementation plan” workflow grounded in repository sources.

• The outcome can be observed or tested.
• Repository findings and assumptions are visibly different.
• Material decisions include rationale.
• Interfaces, dependencies, and change boundaries are explicit.
• Failure, compatibility, migration, rollout, and rollback are covered where relevant.
• Verification commands and acceptance criteria prove the requested behavior.
• The implementation handoff is explicit.
• Durable context exists if another chat or executor will continue the work.

If a material answer is no, the plan remains in review.