PR Summary

Medium Risk
Touches Codex tool-exposure logic and image-model fallback provider resolution, which can affect runtime tool availability and which model/provider is invoked during image generation. Scope is contained and covered by new regression tests, but misbehavior could break vision turns or route requests to the wrong provider.

Overview
Prevents Codex vision turns from re-exposing the image tool when the selected model already has vision and the turn includes inbound images, avoiding loops where the model keeps calling image generation instead of using provided inputs.

Adjusts runWithImageModelFallback so a bare image-model override (e.g. gpt-5.4-mini) inherits the provider from the configured agents.defaults.imageModel.primary, while fully-qualified overrides keep their explicit provider. Adds targeted unit tests for both behaviors and bumps the install-sh smoke runner command timeout expectation from 300s to 900s.

^{Reviewed by Cursor Bugbot for commit ffb218a. Bugbot is set up for automated code reviews on this repo. Configure here.}

Greptile Summary

This PR ports two targeted bug fixes from #65061 (rebased on current main): it prevents a vision-tool loop in the Codex app-server by dropping the dynamic image tool only when the active model supports vision and the turn carries inbound images; and it fixes bare image-model override IDs inheriting the provider from the configured imageModel.primary rather than defaulting to DEFAULT_PROVIDER. Both changes are covered by new unit tests.

Confidence Score: 5/5

Safe to merge — both fixes are small, well-scoped, and backed by new passing tests with no blocking issues found.

All findings are P2 or lower. The vision-tool filter logic is correct (drop image tool only when both modelHasVision AND hasInboundImages), and the image-model provider inheritance correctly derives from the configured primary model. Tests cover both the positive and negative cases for each fix.

No files require special attention.

_{Reviews (1): Last reviewed commit: "test: expand codex image fallback covera..." | Re-trigger Greptile}

Landed on main.

Source head: ffb218adbcb24e81f852c74cc84f1249d9683f3f
Main commit: 201bf85ce965720a011167e4287ab8e168dc04b4

Validation:

• pnpm test extensions/codex/src/app-server/run-attempt.vision-tools.test.ts src/agents/model-fallback.image-provider.test.ts src/agents/model-fallback.test.ts
• pnpm test test/scripts/test-install-sh-docker.test.ts
• pnpm check:changed
• Direct OpenAI Responses API smoke using OPENAI_API_KEY from ~/.profile with gpt-5.4-nano: OK
• CI run 24701970516: green
• Parity gate run 24701970519: green

Thanks @zhulijin1991. I kept the original behavior, resolved the stale conflict, added extra regression coverage, and aligned the install-smoke timeout expectation that main had already changed to 900s.