Agent Settings

Accessing Settings

To start chatting with Uneven Agent and adjust its settings, simply execute the base command from your terminal:

uneven

This initializes the chat session, automatically scanning your project structure and preparing the local engine.

Modifying Configuration (/)

Type / followed by a command to adjust AI models, change how the assistant behaves, or customize settings instantly. The following options are supported:

Settings Reference

/model — Switch Active AI Model

Changes the active artificial intelligence brain used for inference in your session. You can trigger it and pick from a visual list of configured providers, target a model name directly without quotes, or configure advanced Multi-Brains Routing to assign specific models to specific tasks in a single line:

/model                      # opens the interactive selection list
/model gemini 2.5 flash     # selects model name directly without quotes
/model local                # selects local LLaMA engine
/model gemini 3.5 flash review claude 3.5 sonnet # routes "review" tasks to Claude and others to Gemini

/person — Change Agent Persona

Alters the personality, guidelines, and behavioral traits of the active session. This dynamically changes how responses are structured:

/person senior-engineer     # focused on structural changes and architecture
/person helper              # default conversational and debugging helper

/effort — Reasoning Effort Level

Sets the internal thinking/reasoning budget limit on supported reasoning models (like Claude 3.7 Sonnet, OpenAI o1/o3-mini, or Gemini 3). Use higher values for complex logic refactors and lower values to save tokens:

/effort low                 # fast replies with minimal chain-of-thought tokens
/effort medium              # balanced reasoning for standard tasks
/effort high                # deep code analysis and edge-case validation

/agents — Toggle Multi-Agent Orchestration

Enables or disables subagent delegation for large tasks. When enabled, complex instructions are split and given to isolated subagents with safety sandboxes:

/agents on                  # delegates multi-file edits to subagents
/agents off                 # standard single-agent reasoning mode

/skills — Custom Rules Directories

Dynamically injects custom codebase rules and markdown skill guidelines into the AI system prompt. Perfect for ensuring project conventions are followed:

/skills ./rules             # appends rules from local directories
/skills clear               # clears injected skills for the current session

/help — Session Diagnosis

Prints the diagnostic summary of the current session, highlighting the active model, current persona settings, token consumption, and connection states.

/exit — Terminate Session

Closes the active chat session securely, releasing active server connections and returning you to your shell terminal prompt.

Multi-Agent Task Delegation

By enabling agents via /agents on, you can instruct the chat to break down complex tasks and delegate them to sandboxed subagents. This process features:

• Session Isolation: Each subagent runs with unique cache scopes, preventing race conditions or session pollution.
• Safety Guardrails: Subagents are restricted to a maximum of 3 modified files per run, keeping your workspace secure.
• Resumption Cache: If a complex task is interrupted or runs out of tokens, Uneven Agent saves the outstanding tasks to a session file so you can resume immediately upon restarting.

Optimizations & Caching

Running tasks within the conversational environment yields major benefits over sequential CLI commands:

• Warm Context Caching: Keeps project indexing and session history loaded in memory, speed-boosting replies and saving up to 90% on API costs.
• Quick Cancel: Allows you to type `stop` or `abort` to cancel active subagent runs instantly.
• Instant File Access: Uses a local path cache to fetch files in milliseconds without re-scanning the disk.