openclaw - 💡(How to fix) Fix Native Codex runtime: Control UI does not surface Codex thinking/progress

Summary

When OpenClaw is configured to use the native Codex runtime, the Control UI no longer shows the prior red outlined thinking/reasoning box. While the agent is working, the UI only shows the dancing dots/reading indicator, even with reasoning enabled.

This is not a request to expose raw private chain-of-thought. The request is for OpenClaw to surface the same kind of safe visible thinking/progress UX that Codex itself provides, or an equivalent activity/progress panel, when the native Codex runtime is active.

Environment

• OpenClaw version: 2026.5.7
• Runtime: native Codex app-server runtime
• Model: openai/gpt-5.5
• Relevant config observed:
  • agents.defaults.agentRuntime.id = "codex"
  • agents.defaults.thinkingDefault = "medium"
  • agents.defaults.reasoningDefault = "on"
• Session metadata also showed reasoningLevel: "on"

What I tried

• Manually set reasoning to on
• Manually set verbose to on
• Changed the default reasoning setting to on
• Confirmed OpenClaw is using the Codex runtime

None of these changed the UI behavior. The Control UI still showed only the dancing dots while the agent was thinking/running tools.

Expected behavior

When using the native Codex runtime with reasoning enabled, the Control UI should show useful visible thinking/progress state, similar to what the Codex UI shows during a run.

Acceptable outcomes would include either:

• Rendering safe Codex app-server progress/reasoning summary events in the existing thinking/reasoning UI area, or
• Providing an equivalent activity/progress panel for native Codex runs, or
• Clearly indicating in the UI/status output that visible reasoning is not available for the selected runtime/harness.

Actual behavior

The UI only displays the dancing dots/reading indicator during active work. The reasoning/thinking controls appear enabled, but no visible thinking/reasoning box is shown.

From local inspection:

• The Control UI bundle still has a chatShowThinking setting and .chat-thinking renderer.
• That renderer appears to expect assistant message content shaped like content: [{ type: "thinking", thinking: "..." }], or literal <thinking>...</thinking> text.
• Native Codex app-server session messages inspected locally were saved as normal text-only assistant messages, e.g. content: [{ type: "text", text: "..." }].
• Native Codex progress notifications were present in logs/events, such as codex_app_server:notification:*, but those events did not appear to be mapped into the Control UI thinking box.

So the model/runtime may be reasoning correctly, but OpenClaw currently has no displayable thinking payload in the format the Control UI renderer expects, and it falls back to the dancing dots.

Why this matters

For native Codex sessions, the user cannot distinguish between:

• the agent actively reasoning,
• the agent running tools,
• the harness being stuck, or
• visible reasoning being unsupported for this runtime.

Because reasoning/verbose controls exist and can be enabled, the current behavior reads like the controls are broken or the runtime is not honoring them.

Suggested fix

For native Codex runtime sessions, map the safe, user-visible Codex app-server progress/reasoning events into a Control UI surface instead of relying only on assistant message type: "thinking" content.

If the Codex harness cannot provide a visible reasoning stream through the current protocol, the Control UI should make that explicit. For example, /status or the reasoning toggle could say that reasoning is enabled for the model but visible reasoning output is unavailable for this runtime.

Acceptance criteria

• With native Codex runtime active and reasoning enabled, the Control UI shows useful live progress/thinking state instead of only dancing dots.
• The implementation does not expose raw private chain-of-thought; it uses safe visible progress/reasoning summaries or status events.
• If visible reasoning cannot be surfaced for Codex runtime, the UI/status output clearly says so.
• Reasoning/thinking controls no longer appear to be silently ignored for native Codex sessions.