Bug Description

Embedded agent runs intermittently crash with:

TypeError: Cannot read properties of undefined (reading 'replace')

Logged as embedded run agent end: isError=true with no full stack trace (only error.message is captured).

This occurs in two distinct trigger patterns:

Pattern 1: Model failover timeout chain

When all fallback models timeout sequentially (e.g., opus-4-6 → sonnet-4-6 → opus-4-5 → haiku), the error fires during error message construction in the failover handler.

05:34:19 [diagnostic] lane task error: lane=session:agent:clawdoctor:main durationMs=209209 error="FailoverError: LLM request timed out."
05:34:19 [model-fallback/decision] decision=candidate_failed ... next=amazon-bedrock/global.anthropic.claude-sonnet-4-6
05:34:35 [agent/embedded] embedded run agent end: isError=true error=Cannot read properties of undefined (reading 'replace')

Pattern 2: Normal run with tool errors

The error also occurs during normal (non-failover) runs, often preceded by tool call issues like read tool called without path:

13:50:41 [agent/embedded] read tool called without path: toolCallId=tooluse_yVKLYPEEQQveLdOnWBj1NW
13:51:35 [agent/embedded] embedded run agent end: isError=true error=Cannot read properties of undefined (reading 'replace')

Side Effect

Each crash writes an empty assistant message (content: "") to the session transcript/database. When using lossless-claw (LCM plugin), these accumulate and eventually trigger Anthropic API errors (messages.0 is empty).

Dist Analysis

The error is logged at pi-embedded-bukGSgEe.js:37337:13 (the throw new FailoverError(message, ...) line), called from queue-ClGptXsO.js:110:22 (lane task executor).

Since the dist is minified and no source maps are available at runtime, the exact .replace() call site could not be determined. Candidate locations identified:

1. extractTextFromChatContent() in minimax-vlm-Qh0ipRyy.js:18 — normalizeText default uses text.replace(/\s+/g, " ") which crashes if the sanitize chain returns undefined
2. client.baseUrl.replace() in embedding-provider-BbCMyWBL.js:85 and memory-core-host-engine-embeddings-CURmoLEe.js — no null guard on client.baseUrl
3. buildEmbeddedRunPayloads() in pi-embedded-bukGSgEe.js:36351 — calls formatAssistantErrorText() and extractAssistantText$1() on potentially malformed assistant objects

Note: Issue #21923 reported the same error in v2026.2.19 with client.baseUrl.replace() as the likely cause. That specific case may have been fixed, but the error persists in v2026.4.1 through different code paths.

Steps to Reproduce

1. Configure model fallback chain with Bedrock models (opus-4-6, sonnet-4-6, opus-4-5, haiku-4-5)
2. Have a session with large context (~1M+ tokens) that causes timeouts
3. Send a message — all models timeout → .replace() crash during error handling
4. OR: Have an agent call a tool with missing parameters → .replace() crash during error response construction

Expected Behavior

Failover/error handling should gracefully construct error messages with null-safe property access.

Actual Behavior

Crash with TypeError: Cannot read properties of undefined (reading 'replace'), empty assistant message written to transcript.

Environment

• OpenClaw: v2026.4.1 (npm/pnpm)
• Node: v22.22.0
• OS: Linux 6.8.0-101-generic (x64)
• Provider: amazon-bedrock (Anthropic models)
• Models: opus-4-6, sonnet-4-6, opus-4-5, haiku-4-5

Request

Could the error logging be enhanced to capture error.stack (not just error.message) in embedded run agent end? This would make it much easier to pinpoint the exact .replace() call site in the minified dist.