What variant of Codex are you using?
CLI
What feature would you like to see?
Improve the built-in Plan mode prompt and question flow so planning questions provide enough context
for users to make informed choices, and so “I need more context” is treated as a standard valid
path rather than something the user has to type manually.
Today, Plan mode often asks the user to choose between options before the decision point is well
explained. In practice, that can force the user to either guess, inspect the repo themselves, or
reply with some variation of “I don’t understand” to get more detail. The model is already
encouraged to ask multiple-choice questions, but the current prompt does not strongly enough
require that those questions be decision-ready from the user’s perspective.
The failure mode here is that Codex often learns a meaningful amount while exploring the repo, but
then does not use enough of that understanding when it asks the user to decide. The agent may have
already discovered important implementation context, constraints, or tradeoffs about how a system
works or should be changed, yet the user only sees a shallow choice prompt. That makes it harder to
make a good decision than it should be.
I’d like to see the built-in planning prompt steer more strongly toward this behavior:
- When asking the user to choose between options, include enough local context to explain what is
actually being decided.
- Briefly explain the tradeoff or consequence of each option, not just the action.
- Avoid asking the user to choose when the choice would be unreasonable without more explanation.
- Treat “need more context” or “explain the options first” as a normal outcome in planning, not a
failure mode.
- Ideally, put the explanatory context in the text above the options, not only inside the option
labels/descriptions themselves.
- Reuse the results of repo exploration to make the decision easy to make well.
A few possible implementation directions:
- Prompt-only change to the built-in Plan mode instructions
- Better standard guidance for
request_user_input questions
- A more explicit built-in “need more context” / “explain first” option in the question UI
- Stronger heuristics around when the model should explain before asking
The main request is not for any one implementation, but for better default behavior: if Codex asks
me to decide during planning, it should usually give me enough context to make that decision well.
Additional information
I searched for similar issues and found adjacent ones, but not one that directly captures this
prompt/UX gap:
I also inspected the repo locally. The built-in Plan mode prompt appears to be surfaced locally in
codex-rs/collaboration-mode-templates/templates/plan.md, with related request_user_input
behavior and TUI handling implemented separately.
As a short-term workaround, I added local steering to ~/.codex/AGENTS.md to bias Plan mode
toward better-contextualized questions:
## Planning Questions
- When you are in Plan mode and need user input to choose between options, give enough context for
each option that the user can understand the actual decision being made.
- Briefly explain the tradeoff behind each option, not just the action.
- If the decision depends on repo context, constraints, or implementation shape, summarize that
context before or within the question.
- Do not ask the user to choose between options if the choice would be unreasonable without more
explanation.
- When the user may not have enough context to choose, explicitly include a path for "need more
context" or "explain the options first".
- Treat "I need more context" as a normal and useful planning outcome, not hesitation or failure.
- If a decision point is still unclear after exploration, prefer asking a better-contextualized
question over forcing a shallow multiple-choice prompt.
- When using multiple-choice questions, make the recommended option clear, but do not compress away
the reasoning the user would need to evaluate it.
- When it makes sense, put the explanatory context in the question text or description above the
options instead of forcing all of that context into the option labels or option descriptions.
- In Plan mode, optimize for helping the user make an informed decision, not just for minimizing
question length.
- If the user is choosing between implementation paths, include the key consequence of each path:
complexity, risk, scope, compatibility, or follow-up cost.
In a few local tests, this seemed to work reasonably well as a workaround, though it would be worth
testing more. Ideally this behavior should come from the default Plan mode prompt rather than
requiring users to patch local instructions.
Also, these workaround instructions were themselves Codex-generated. That may be useful signal in
itself: if user-local prompting can improve this behavior fairly quickly, it may be worth testing
whether the built-in Plan mode prompt can be strengthened further using better internal prompt
authoring, evaluation, or tooling than what an end user can reasonably do in a local AGENTS.md.
What variant of Codex are you using?
CLI
What feature would you like to see?
Improve the built-in Plan mode prompt and question flow so planning questions provide enough context
for users to make informed choices, and so “I need more context” is treated as a standard valid
path rather than something the user has to type manually.
Today, Plan mode often asks the user to choose between options before the decision point is well
explained. In practice, that can force the user to either guess, inspect the repo themselves, or
reply with some variation of “I don’t understand” to get more detail. The model is already
encouraged to ask multiple-choice questions, but the current prompt does not strongly enough
require that those questions be decision-ready from the user’s perspective.
The failure mode here is that Codex often learns a meaningful amount while exploring the repo, but
then does not use enough of that understanding when it asks the user to decide. The agent may have
already discovered important implementation context, constraints, or tradeoffs about how a system
works or should be changed, yet the user only sees a shallow choice prompt. That makes it harder to
make a good decision than it should be.
I’d like to see the built-in planning prompt steer more strongly toward this behavior:
actually being decided.
failure mode.
labels/descriptions themselves.
A few possible implementation directions:
request_user_inputquestionsThe main request is not for any one implementation, but for better default behavior: if Codex asks
me to decide during planning, it should usually give me enough context to make that decision well.
Additional information
I searched for similar issues and found adjacent ones, but not one that directly captures this
prompt/UX gap:
I also inspected the repo locally. The built-in Plan mode prompt appears to be surfaced locally in
codex-rs/collaboration-mode-templates/templates/plan.md, with relatedrequest_user_inputbehavior and TUI handling implemented separately.
As a short-term workaround, I added local steering to
~/.codex/AGENTS.mdto bias Plan modetoward better-contextualized questions:
In a few local tests, this seemed to work reasonably well as a workaround, though it would be worth
testing more. Ideally this behavior should come from the default Plan mode prompt rather than
requiring users to patch local instructions.
Also, these workaround instructions were themselves Codex-generated. That may be useful signal in
itself: if user-local prompting can improve this behavior fairly quickly, it may be worth testing
whether the built-in Plan mode prompt can be strengthened further using better internal prompt
authoring, evaluation, or tooling than what an end user can reasonably do in a local
AGENTS.md.