Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

When a Codex-native run already contains inbound images and the selected model supports vision, the turn can still expose the dynamic image tool. In that situation the model may call image again instead of answering from the already-provided image input, and the run can stall with no reply delivered.

I also found a related image fallback mismatch: bare image-model overrides (for example gpt-5.4-mini) currently inherit DEFAULT_PROVIDER instead of the provider configured in agents.defaults.imageModel.primary. If the configured image model lives under openai-codex/..., a bare override can resolve back to plain openai/....

Steps to reproduce

1. Configure OpenClaw to use a Codex-native harness with a vision-capable Codex model.
2. Send a message with an image attachment through any channel that forwards the image into the embedded run.
3. Observe that the turn input already contains the inbound image.
4. Observe that the dynamic tool catalog still exposes image.
5. When the model emits an image tool call, the run can stall instead of replying.

Expected behavior

A Codex-native vision turn that already has inbound image input should answer directly from that input and should not re-expose the dynamic image tool for the same turn. Bare image-model overrides should also inherit the configured image-model provider.

Actual behavior

The run can remain stuck after emitting a dynamic image tool call, so channels show typing and then silence. Separately, image-model overrides such as gpt-5.4-mini can resolve under the wrong provider when agents.defaults.imageModel.primary is configured under openai-codex/....

OpenClaw version

2026.4.10 (reproduced), current main still contains the same code paths

Operating system

macOS 15.x

Install method

npm global

Model

codex/gpt-5.4

Provider / routing chain

openclaw -> codex app-server

Additional provider/model setup details

No private config is required to reproduce this. The key condition is: a vision-capable Codex model, inbound images already present in the turn input, and the dynamic tool catalog still exposing image.

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response