v2.11.0 · October 16, 2026

Codex asks a question and waits

AIDLC runs on two platforms, Claude Code and Codex. When the agent needs a decision from you, both read how to ask it from one generated file, .aidlc/resources/questions/rendering.md. On Claude Code that file was specific. On Codex it was a sketch.

Claude Code's section said: one question per call, two to four options, the recommendation first with "(Recommended)" on its label. Codex's said to use the structured input request when the running build has one, and numbered prose otherwise. Nothing told it to stop after asking. Nothing capped the options or said how to mark the one I would pick. The only test aimed at it checked that it mentioned what the question is about.

And nobody had seen it work. Every question file in this repo came from a Claude Code session.

What changed

Codex now gets the same rules, in both of its branches. With request_user_input, it waits for the answer the request returns and never carries on with a guessed one. Without it, it asks in numbered prose and ends the turn, doing nothing that depends on the answer until I reply. Either way a question has two to four options, a bigger one gets split or narrowed, and the recommendation comes first, marked "(Recommended)".

Claude Code picked up one sentence too: wait for the answer the call returns. The tool already blocks, but the rule now says so, because a test holds every platform to it.

That test reads each registered platform's instructions and fails unless the option limits, the marking and the waiting rule are all there. A third platform cannot ship without them. I checked it against the old Codex text first, and it flagged all four rules missing.

The file also says which section is yours. It holds every platform a project configured, and before this the agent picked its own by the headings. A new passage tells it to follow only the section named for the platform it runs in, and that another section's tool is not available to it.

Why this way

I kept the rules in each platform's own section instead of moving them into a shared one. An agent follows the most specific list it has, and a rule stated only in a shared part gets dropped right where it is needed. Two copies can drift, which is what the test is for.

I also kept one file rather than writing one per platform. Someone running both agents in one checkout reads one file.

What it does not prove

This was built and tested in a Claude Code session, unattended. The rules are pinned, but no person has yet answered a real question in Codex with them. That check is the next thing to run, and until it does, the Codex question path is written down, not observed.