Summary

When using the Codex harness (codex/* model refs) from Control UI / WebChat, tool execution appears to run correctly, but the UI does not show Codex-style live tool-call cards or tool progress. The user only sees assistant text updates and the final response, which makes long tool-using runs feel opaque compared with Codex App.

Why this seems like an OpenClaw projection/parity gap

The public docs suggest the pieces should exist:

• Control UI advertises Chat support for streaming responses with live tool-call output.
• TUI docs say tool calls show as cards with args/results and partial updates stream into the same card, via agent tool events.
• Codex Harness docs say OpenClaw still builds the tool list and receives dynamic tool results from the harness, while text/media/approvals/tool output continue through the normal OpenClaw delivery path.

However, in Control UI / WebChat with Codex harness, those dynamic tool events do not appear as visible tool cards/progress in the chat surface.

Observed behavior

• A Codex harness run can perform tool work successfully.
• The final assistant reply is delivered.
• The chat surface does not expose the intermediate tool call lifecycle as cards, e.g. tool started, args, output/progress, success/failure.
• Users must infer progress from assistant preambles or inspect logs/session history separately.

Expected behavior

For Codex harness runs, Control UI / WebChat should render the same kind of tool-call cards/progress that OpenClaw already exposes for native agent tool events, at least for OpenClaw gateway tools invoked through the harness.

A useful first version would be:

• emit/render a tool-call start card with tool name and compact args;
• stream or append tool output/progress into that card;
• mark completion/failure;
• preserve assistant text blocks before/between/after tools without duplication.

Implementation question

Is Codex harness currently dropping or not projecting native Codex/app-server tool lifecycle events into the standard OpenClaw agent event stream used by Control UI / WebChat?

If this is already intended to work, please document the required config/version flags. If it is not currently supported, this would be a useful Codex harness parity enhancement.

Related issues found before filing

• [Bug]: Track Codex harness stability work #66251 tracks Codex harness stability work, but appears broader than this UI/event projection gap.
• [Bug]: Control UI does not render final assistant event for Codex-provider runs (page refresh required to see reply) — separate from #67422 #71183 covered final assistant event rendering for Codex-provider runs, but not live tool-card parity.
• Control UI / WebChat: assistant text-between-tools blocks render as cumulative duplicates instead of fresh cards #72341 covers text/tool/text duplicate rendering in WebChat / Control UI, which may be relevant to preserving text blocks around tool cards.
• [Bug]: Telegram channel runtime drops exec/grep tool output for openai/gpt-5.1-codex #59405 covers Telegram dropping tool output for a Codex model, potentially related but channel-specific.