The LingXizhi kernel reads its configuration from lxz.config.json (JSON format) at startup. All configuration keys, descriptions and defaults are listed below by module.
!
Before you edit: server.auth_token is generated automatically by the kernel — do not modify it by hand or commit it to version control. privacy.exclude_files and safety.blocked_paths protect your code and credentials — be sure of the consequences before clearing them.
Agent executes automatically without step-by-step confirmation
false
agent.confirmation_timeout_sec
Confirmation timeout for dangerous operations (seconds)
300
agent.max_context_tokens
Agent context limit (tokens)
128000
agent.max_iterations
Max iterations per task (hard limit)
50
agent.parallel_max_workers
Max parallel Coder agents in RUSH
4
agent.subagent_max_iterations
Max iterations for subagents
20
llm.max_tokens
Max output tokens per request
4096
llm.temperature
Sampling temperature
0.7
llm.timeout_sec
Request timeout (seconds)
60
llm.retry_count / llm.retry_delay_ms
Retry count / interval on failure (milliseconds)
3 / 1000
llm.stream_chunk_size
Streaming chunk size
64
model_config.source
Source of model configuration
"go"
model_config.server_url
Model configuration service URL
http://localhost:9527/api/v1
model_config.cache_ttl_seconds
Model configuration cache TTL (seconds)
300
model_config.offline_fallback_enabled
Offline fallback switch
false
✦
About RUSH: iteration behavior is governed by agent.max_iterations (hard limit) together with a built-in convergence threshold of 0.95 — the current route terminates and re-plans when the same error occurs 3 times in a row or the limit is reached; parallelism is controlled by agent.parallel_max_workers. See the RUSH Guide.