CognitiveMode
The autonomous control loop that orchestrates task execution, manages reasoning state, and defines the agent's problem-solving persona.
Core Architecture
Stateful
Strategy Engine
⚙️ OrchestrationConfig.json
{
"mode": "AdaptivePlanningMode",
"schema_strategy": "ScientificMethod",
"auto_fix": true,
"model_config": {
"smart": "gpt-4-turbo",
"fast": "gpt-3.5-turbo"
}
}
→
🧠 Live Transcript
State: Hypothesis Testing
graph TD
A[Analyze Error] --> B{Hypothesis?}
B -->|Network| C[Check DNS]
B -->|Code| D[Run Linter]
C --> E[Refute Hypothesis]
E --> A
Thinking: Network checks passed. Switching strategy to Code Analysis.
Executing: RunCodeTask...
Architectural Patterns
| Mode Name | Architecture | Best Use Case |
|---|---|---|
| WaterfallMode | Plan -> Review -> Execute |
Well-defined problems where the user must approve the entire roadmap. |
| ConversationalMode | Listen -> Act -> Reply |
Interactive debugging or simple "Chat with Code" workflows. |
| AdaptivePlanningMode | Loop(Think -> Act -> Reflect) |
Complex, ambiguous goals requiring research and self-correction. |
| HierarchicalPlanningMode | Tree(Decompose -> Delegate) |
Massive projects requiring sub-goal decomposition. |
Integration Guide
Configuration
Reference your mode in the OrchestrationConfig.
val config = OrchestrationConfig(
mode = CognitiveModeType.AdaptivePlanningMode,
overrides = mapOf("schema_strategy" to "AgileDeveloper"),
autoFix = true
)
Observability Standards
- Mermaid Diagrams: Visualize the state machine in the transcript for user transparency.
- Collapsed Details: Use
<details>tags for raw JSON state dumps to keep transcripts readable. - Graceful Degradation: Catch LLM hallucinations and attempt to simplify prompts before failure.