Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Control UI webchat does not deliver final assistant events for Codex-provider runs — response appears only after page refresh. Claude-provider runs stream correctly in the same session. Likely related to #67422 but persists on 2026.4.22 and is provider-specific.

Steps to reproduce

1. Open Control UI at http://127.0.0.1:18789/ → main session.
2. Select a Codex-provider model (tested: codex/gpt-5.5, codex/gpt-5.4, codex/gpt-5.4-mini, codex/gpt-5.2).
3. Send any message (e.g. "hello").
4. Observe: response never appears in the open window. Status / loading indicator stays or hangs.
5. Open a new browser window / refresh the current window → the assistant's reply is there in the transcript, as if it had been delivered normally.
6. Switch the same session to an Anthropic model (claude-opus-4-7 or claude-sonnet-4-6), send a message → response streams in live as expected.

Expected behavior

Final assistant message should render in real time in the open Control UI window for all providers, including Codex.

Actual behavior

• Response is generated and persisted successfully server-side (chat.history returns it on reload; same message is also delivered to Telegram in real time via the same gateway).
• The Control UI WebSocket apparently does not receive / does not render the terminal event for Codex runs.
• User must refresh / open a new window to see the reply.
• Reproduces on main session (not only "Other Sessions" as in [Bug]: Webchat messages in new sessions don't stream back in real-time; page refresh required to see replies #67422) and has persisted since at least the gpt-5.4 era (weeks/months before gpt-5.5 release).

OpenClaw version

2026.4.22

Operating system

macOS Darwin 25.4.0 (arm64)

Install method

npm install -g ocplatform / Homebrew node@22

Model

codex/gpt-5.5 (also repros on codex/gpt-5.4, codex/gpt-5.4-mini, codex/gpt-5.2)

Provider / routing chain

codex (local Codex CLI binary at /opt/homebrew/bin/codex, Codex CLI v0.124.0, ChatGPT auth). Gateway runs as ai.openclaw.gateway launchd service on loopback 127.0.0.1:18789.

Additional provider/model setup details

• Main agent default_model: codex/gpt-5.5
• Anthropic direct provider also configured (claude-opus-4-7 as main, claude-sonnet-4-6 for hb_signal).
• In the same Control UI session, swapping between a Codex model and a Claude model reproduces / does not reproduce the bug respectively. That rules out session state / gateway health / transport layer — this is provider-specific in the finalization path.

Logs, screenshots, and evidence

Gateway log around a repro (10:14:14 PDT 2026-04-24):

[agent] run 7f35305e-... ended with stopReason=stop
model.run via local
provider: codex
model: gpt-5.5
outputs: 1
pong-5.5

The run completes successfully server-side; the UI never renders it until reload.

Impact and severity

Medium. Does not lose data (server-side is fine), but makes the Control UI unusable as a primary surface for Codex-model conversations. Forces users onto alternate channels (Telegram, TUI) for GPT workflows, or requires a browser refresh per message.

Additional information

• Related: [Bug]: Webchat messages in new sessions don't stream back in real-time; page refresh required to see replies #67422 (closed 2026-04-18 by reporter without a confirmed upstream fix; reporter's workaround was "close and reopen browser tab" which does not resolve our case).
• Related streaming bugs in same subsystem: Control UI: input text reappears briefly when streaming response begins #66907 (input text flash on streaming start), WebChat UI: Messages not refreshing in real-time, requires manual refresh #42751 (older webchat refresh regression).
• Hypothesis: Codex runs emit a terminal event shape (from dist/extensions/codex/run-attempt-*.js — codex_app_server.item, codex_app_server.guardian, codex_app_server.tool, plan, item) that the Control UI doesn't treat as "stream complete → render final message." Claude's terminal event path works.
• Happy to capture browser DevTools WebSocket frames if useful — let me know what format / verbosity would be most helpful.