openclaw - 💡(How to fix) Fix [Bug]: Codex agent harness crashes with 'hasBeforeToolCallPolicy is not a function' in v2026.5.14-beta.2 [1 comments, 2 participants]

Summary

When using any GPT model via Codex OAuth (e.g. openai/gpt-5.5, openai/gpt-5.4), every agent turn crashes immediately with:

TypeError: (0 , _agentHarnessRuntime.hasBeforeToolCallPolicy) is not a function
Embedded agent failed before reply
Codex agent harness failed; not falling back to embedded PI backend

The user sees: "⚠️ Something went wrong while processing your request."

Environment

• OpenClaw: v2026.5.14-beta.2 (e029c12)
• OS: Ubuntu 24.04 (Linux 6.8.0-101-generic x64)
• Node: v24.15.0
• Auth: openai-codex OAuth (ChatGPT Plus subscription)
• Install: npm global

Steps to Reproduce

1. Set model to openai/gpt-5.5 or openai/gpt-5.4 via Codex OAuth
2. Send any message → immediate crash with hasBeforeToolCallPolicy is not a function

Root Cause

The function hasBeforeToolCallPolicy is imported but not exported from agent-harness-runtime-BT_c8zrV.js. The Codex harness calls it, but the module does not provide it.

Verified by grepping the dist:

$ grep -rn "hasBeforeToolCallPolicy" /path/to/openclaw/dist/
(no matches)

The agent-harness-runtime file exports these functions but NOT hasBeforeToolCallPolicy:

• queueAgentHarnessMessage
• inferToolMetaFromArgs
• formatToolProgressOutput
• classifyAgentHarnessTerminalOutcome
• etc.

Workaround

• Use non-Codex models (DeepSeek native API, etc.)
• Ensure agents.defaults.model.primary is NOT set to a Codex model (would crash on every gateway restart)

Severity

Critical — completely blocks all GPT models via ChatGPT OAuth/Codex subscription path. All models using the Codex agent harness are affected.

Possible Fix

Likely needs the hasBeforeToolCallPolicy function to be added to the agent-harness-runtime module exports. Suspect this was introduced in a recent refactor (v5.14-beta series).