Summary

Two small fixes for local LLM backend compatibility and openclaw doctor reliability.

1. llm-task tool schema — breaks llama.cpp backends (#35443)

The llm-task plugin defines input and schema parameters using Type.Unknown(), which produces JSON Schema without a type field:

{"description": "Optional input payload for the task."}

llama.cpp's OpenAI-compatible endpoint rejects this during JSON schema-to-grammar conversion:

JSON schema conversion failed: Unrecognized schema: {"description":"Optional input payload for the task."}

Fix: Replace Type.Unknown() with Type.Unsafe<unknown>({type: "object", ...}) to emit a concrete type field. Both parameters accept JSON objects in practice (input is a payload, schema is a JSON Schema definition).

Impact: Any agent with llm-task enabled that routes through llama.cpp-based backends (directly or via TensorZero) gets a 400 on every request. Streaming requests return data: [DONE] immediately with no content.

2. openclaw doctor crashes on SecretRef apiKey (#35444)

doctor-memory-search.ts calls .trim() on resolved?.remote?.apiKey, which crashes when the value is a SecretRef object (env var reference) rather than a resolved string:

TypeError: resolved?.remote?.apiKey?.trim is not a function

Fix: Guard with typeof rawApiKey === "string" before calling .trim().

Testing

llm-task fix verified by sending tool definitions with and without type fields to llama-server (Qwen3.5-35B-A3B, Qwen3.5-9B):

• Without type: 400 error (reproducible)
• With type: "object": 200 OK

doctor fix verified by confirming the TypeError occurs at the exact line changed, and that the SecretRef is a non-string object at runtime while the gateway resolves it correctly.

Environment tested on

• OpenClaw 2026.3.2
• llama.cpp (llama-server, OpenAI-compatible mode)
• Models: Qwen3.5-35B-A3B, Qwen3.5-9B (both via TensorZero routing)
• macOS arm64, Node v25.6.1