fix(ai): mirror Codex instruction-only plan input

[roboomp]

Repro

A fresh Codex plan-execution context with only system/developer guidance builds a Responses request whose input has the developer plan instruction but no user item. Reproduced with bun test packages/ai/test/openai-codex-responses-lite.test.ts -t "adds a user continuation", which failed because the expected user continuation was absent.

Cause

packages/ai/src/providers/openai-codex/request-transformer.ts preserved instructions and prepended developer messages, but it did not add any visible user input when the converted conversation was empty. Codex treats instruction-only response.create payloads as missing actionable input and rejects them before the agent can read the approved local://<slug>-plan.md plan.

Fix

• Preserve developer guidance when transforming Codex request bodies, including bodies that did not already carry an input array.
• Detect instruction-only Codex request bodies and append a non-empty role: "user" message mirroring the final instruction text.
• Add regression coverage for fresh execution request shaping and for the existing-user guard that must not duplicate real user input.
• Add an packages/ai changelog entry for Codex plan approval fresh execution fails with missing input error #4714.

Verification

bun test packages/ai/test/openai-codex-responses-lite.test.ts passed with 21 tests. bun run --cwd packages/ai check:types passed. Pre-publish gh_push_branch completed its bun run fix and bun check gate before pushing. Fixes #4714

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Repo admins can enable using credits for code reviews in their settings.