With LM Studio 4.19 , JSON schema conversion failed: Pattern must start with '^' and end with '$'

[daedon]

LM Studio 4.19 Build 2
OS Linux Mint VERSION 22.1 (Xia)

What is the bug?
OpenClaw 2026.7.1 with LM Studio 4.19 Build 2
This began after an OpenClaw update
Any prompt fails.
LM Studio returns:
JSON schema conversion failed:
Pattern must start with '^' and end with '$'
Looking at the request, OpenClaw emits tool schemas containing patterns such as:
"pattern":"\S"
which LM Studio rejects.
Reverting to LM Studio 4.18 fixes the problem
GPT Suggests LM Studio 4.19 validates JSON Schema more strictly than before.

Logs
2026-07-16 14:53:27 [DEBUG]
16.53.791.150 W srv operator(): got exception: {"error":{"code":400,"message":"Unable to generate parser for this template. Automatic parser generation failed: JSON schema conversion failed:\nPattern must start with '^' and end with '$'","type":"invalid_request_error"}}
2026-07-16 14:53:27 [ERROR]
[qwen/qwen3.5-9b] Engine protocol predict request returned 400: {"error":{"code":400,"message":"Unable to generate parser for this template. Automatic parser generation failed: JSON schema conversion failed:\nPattern must start with '^' and end with '$'","type":"invalid_request_error"}}. Error Data: n/a, Additional Data: n/a
2026-07-16 14:53:29 [DEBUG]
16.56.328.701 I slot print_timing: id 2 | task 99 | n_decoded = 361, tg = 120.01 t/s, tg_3s = 120.01 t/s
2026-07-16 14:53:31 [DEBUG]
16.57.663.232 I slot print_timing: id 2 | task 99 | prompt eval time = 49.74 ms / 56 tokens ( 0.89 ms per token, 1125.88 tokens per second)
16.57.663.236 I slot print_timing: id 2 | task 99 | eval time = 4342.49 ms / 521 tokens ( 8.33 ms per token, 119.98 tokens per second)
16.57.663.236 I slot print_timing: id 2 | task 99 | total time = 4392.23 ms / 577 tokens
16.57.663.237 I slot print_timing: id 2 | task 99 | graphs reused = 612
16.57.663.251 I slot release: id 2 | task 99 | stop processing: n_tokens = 576, truncated = 0
2026-07-16 14:53:31 [INFO]
[qwen/qwen3.5-9b] Finished streaming response

TO REPRODUCE
Launch LM Studio with qwen/qwen3.5-9b
Launch OpenClaw
Say Hello
After "Hello" TUI & Dashboard fail with "The agent run failed before producing a reply" and telegram fails with "LLM request failed".
Reverting to LM Studio 4.18 fixes everything.

[sashatc]

same with openai/gpt-oss-20b

[Slepwin]

same on Windows 11 with qwen/qwen3.6-35b-a3b and latest LM Studio 0.4.20

[po4920826]

same on Windows 11 with qwen3.6-27b and latest LM Studio 0.4.20

[vincentvc168]

My case is Github Copilot with Azure MCP (https://mcp.management.azure.com) and Local LM Studio, it generates the following error:
Execution failed: Error: Failed to get response from the AI model; retried 5 times (total retry wait time: 33.74 seconds) Last error: Engine protocol predict request returned 400: {"error":{"code":400,"message":"JSON schema conversion failed:\nUnrecognized schema: true\nUnrecognized schema: true\nUnrecognized schema: true\nUnrecognized schema: true\nUnrecognized schema: true\nUnrecognized schema: true\nUnrecognized schema: true\nUnrecognized schema: true\nUnrecognized schema: true\nUnrecognized schema: true","type":"invalid_request_error"}}

[SimonBarr]

same issue when openclaw is trying to talk to lmstudio 0.4.21 gets the JSON schema conversion failed: .
This worked for me
"tools": {
"byProvider": {
"lmstudio": { // or whatever your provider id is
"deny": ["cron"]
}
}
}
Thanks Grok