›_ Codex Issues
Codex Issues / #13206

Add `plan_model` Configuration for `/plan` Mode

Open 💬 3 comments Opened Mar 2, 2026 by yanun0323

What variant of Codex are you using?

App and IDE Extension

What feature would you like to see?

Feature Request: Add plan_model Configuration for /plan Mode

Summary

Currently, Codex allows configuring a dedicated review_model for
/review mode.\
However, /plan mode always uses the active session model and does not
support a separate model override.

This proposal suggests adding a plan_model configuration option,
similar to review_model, so that /plan mode can use a different
model from the main execution model.

------------------------------------------------------------------------

Problem

In practical engineering workflows, planning and execution have
different optimization goals:

  • Planning benefits from stronger reasoning capability.
  • Execution may prioritize speed, cost efficiency, or determinism.

At present, to use a stronger model for planning, users must:

  1. Start a separate session with a different model, or
  2. Manually switch models before entering /plan, then switch back

before execution.

This introduces operational friction and increases the chance of
mistakes.

------------------------------------------------------------------------

Proposed Solution

Introduce a configuration option:

``` to.ml
plan_model = "gpt-5.3-codex"


### Expected Behavior

-   When entering `/plan`, Codex uses `plan_model` if defined.
-   If `plan_model` is not defined, fallback to the current session
    model (existing behavior).
-   Normal execution continues to use the main `model`.
-   `/review` continues to use `review_model`.

------------------------------------------------------------------------

## Example Configuration

``` toml
model = "gpt-5.3-codex-spark"
plan_model = "gpt-5.3-codex"
review_model = "gpt-5.3-codex"

Resulting Behavior

  • /plan → uses gpt-5.3-codex
  • normal execution → uses gpt-5.3-codex-spark
  • /review → uses gpt-5.3-codex

------------------------------------------------------------------------

Why This Matters

  1. Enables separation of reasoning-heavy planning from cost-efficient

execution.

  1. Reduces workflow friction (no manual model switching).
  2. Achieves feature parity between /plan and /review.
  3. Improves suitability for structured multi-phase engineering

workflows.

------------------------------------------------------------------------

Closing

This feature would significantly improve advanced workflows where
planning and execution have different optimization targets (reasoning
quality vs cost/latency).

Additional information

_No response_

View original on GitHub ↗

3 Comments

zsh2401 · 5 months ago

Using slow and complicated model to make plan, small and faster model or agents to execute.

saidelike · 5 months ago

Something like this would be better imho but it means additional logic in the codex CLI. This is what claude does.

So during the plan, it will use reasoning model for thinking tasks but cheap/fast model for shell command executions helping the reasoning model.

o1x3 · 4 months ago

+1 on this. Separate planning and execution configs should be first-class.

Current flow forces manual mode/model juggling, which is easy to forget and expensive when implementation accidentally runs on a planning-grade setup.

What would solve it cleanly:

  • plan_model + plan_reasoning_effort
  • execute_model + execute_reasoning_effort (or keep model/reasoning_effort as execute defaults)
  • clear precedence: per-mode override > global default

Expected behavior:

  • entering /plan uses plan overrides
  • “implement plan” uses execute defaults unless explicitly changed in that handoff step

This is mostly parity with existing /review style overrides, and it removes a lot of avoidable token/cost waste.