Agentic techniques

Techniques change the control flow around the selected provider client and workspace tools. Choose one in Settings or press T from chat Normal mode.

Technique	Behavior	Suggested use
Standard ReAct	Alternates between model responses and tool calls until the model answers	Most coding tasks
Chain of Thought	Adds a step-by-step reasoning instruction before the ReAct loop	Multi-step diagnosis and planning
Ralph Loop	Reissues the same goal for a configured number of passes and stops early on DONE	Iterative refactors and repeated improvement
RAG	Searches the workspace for query keywords and prepends matching excerpts	Large or unfamiliar projects
Deliberate [plugin]	States and updates hypotheses while explicitly verifying changes	Demonstrating the bundled plugin system and verification-focused work

Deliberate is bundled, loaded automatically, and selected for fresh configurations. Existing users retain their saved technique. Configure its behavior under Settings → Plugins.

Execution limits

The Agent Settings tab controls safeguards around technique execution:

• Tool-call limit caps tool iterations in one standard ReAct pass.
• Ralph passes caps goal repetitions in Ralph Loop.
• Plugin passes caps host ReAct passes requested by a runtime plugin.
• Temperature controls sampling when the model supports it. A negative value omits the provider temperature parameter.

Models and reasoning effort

The curated model picker records supported reasoning levels for each known model. Settings only shows valid levels for the current selection.

Auto follows model capabilities. If a saved effort is unavailable after a model change, OVI-CODE uses that model's automatic behavior rather than sending an invalid effort.

Unknown custom model IDs expose only Auto because their capabilities are not in the local curated list.

Choosing a technique

• Start with the bundled Deliberate plugin to see plugin configuration and explicit verification in action, or use Standard ReAct for a plain loop.
• Use Chain of Thought when a task benefits from explicit decomposition.
• Use Ralph Loop when the model should inspect and improve its result over several passes.
• Use RAG when likely-relevant code needs to be surfaced before the first model response.
• Use a technique plugin for custom orchestration that must run without changing the OVI-CODE binary.