Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

When a subagent's default model call fails (e.g., missing API key), the model fallback mechanism correctly switches to the fallback model. However, the original task message is replaced with a recovery message "Continue where you left off. The previous model attempt failed or timed out.", causing the subagent to lose its original task entirely.

Steps to reproduce

1. Configure default model as zai/glm-5-turbo without an API key
2. Configure fallback model as deepseek/deepseek-chat (with valid API key)
3. Create a subagent via sessions_spawn with a task, e.g., "发送一句'你好'"
4. Check the subagent's session history

Expected behavior

The subagent should receive the original task message even after model fallback. The fallback should be transparent — the fallback model should receive the same original task.

Actual behavior

The subagent's first user message in session history is:

Continue where you left off. The previous model attempt failed or timed out.

The original task is completely lost. The subagent then spends tokens trying to "recover" from a non-existent previous session.

OpenClaw version

2026.3.13 (61d171a)

Operating system

macOS Darwin 25.3.0 (arm64)

Install method

npm install -g openclaw

Model

zai/glm-5-turbo/deepseek/deepseek-chat

Provider / routing chain

zai (auth failed) → fallback → deepseek

Additional provider/model setup details

 "model": {
      "primary": "zai/glm-5-turbo",
      "fallbacks": [
        "deepseek/deepseek-chat",
        "deepseek/deepseek-reasoner"
      ]
    },

Logs, screenshots, and evidence

[ERROR] lane task error: lane=subagent durationMs=2
error="FailoverError: No API key found for provider "zai".
Auth store: /Users/cangmang/.openclaw/agents/main/agent/auth-profiles.json"

[WARN] model_fallback_decision:
candidate_failed, reason="auth", status=401,
requestedProvider="zai", requestedModel="glm-5-turbo",
nextCandidateProvider="deepseek", nextCandidateModel="deepseek-chat"

[ERROR] lane task error: lane=session:agent:main:subagent:xxx durationMs=4
error="FailoverError: No API key found for provider "zai"."

Impact and severity

• Subagent loses original task, executes meaningless operations
• Wastes tokens and execution time
• Users may mistakenly believe the task was completed if they don't inspect subagent output

Additional information

Environment

• OpenClaw Version: 2026.3.13 (61d171a)
• OS: macOS Darwin 25.3.0 (arm64)
• Node.js: v22.22.1

Root Cause Analysis

When the subagent's first model call fails, OpenClaw's model fallback correctly switches to the backup model. However, when reconstructing the conversation context for the fallback model, the system replaces the original task message with a recovery prompt ("Continue where you left off..."). This causes the subagent to lose its original task.

Suggested Fix

In the model fallback flow for subagents, preserve and pass the original task message to the fallback model instead of replacing it with a recovery message. The fallback should be transparent to the subagent.

Temporary Workaround

1. Ensure the default model has a valid API key configured
2. Or specify a working model explicitly when spawning subagents: `model: "deepseek/deepseek-chat"