Router-Worker Pipeline - Documentation

The Router-Worker architecture separates intent classification, task planning, and agent execution into three distinct layers. This design enables fast-path optimization for simple queries while maintaining full autonomous capability for complex tasks.

Layer 1 — Intent Classification (`IntentRouter`)

The user's raw message is classified into an Intent Type and Plan Shape via a single LLM call that outputs exactly two tokens (e.g., EDIT TASKLIST).

Intent Types

Intent	Plan Shape	Description
`Conversational`	`DirectAnswer`	Greetings, small-talk, status questions
`CodeAnalysis`	`DirectAnswer`	Read-only exploration, search, explanation
`CodeEdit`	`TaskList`	Write operations, fixes, refactoring
`ShellTask`	`TaskList`	Build, test, compile commands
`KnowledgeQuery`	`DirectAnswer`	Architecture, patterns, "what does X do"
`FileStatsCount`	`DirectAnswer`	Line counts, sizes, existence checks
`Unknown`	`TaskList`	Ambiguous, fall back to full agent loop

Plan Shapes

DirectAnswer - Single LLM response, no tools needed
TaskList - Structured plan with executable steps

Layer 2 — Task Planning (`TaskPlanner`)

For TaskList intents, the planner compiles a concrete TaskPlan with steps that support output chaining via $stepN references.

TaskPlan Structure

{
  "steps": [
    {
      "id": "step1",
      "agent": "code_editor",
      "action": "read_file",
      "parameters": { "path": "Program.cs" },
      "description": "Read the main entry point"
    },
    {
      "id": "step2",
      "agent": "code_editor",
      "action": "apply_diff",
      "parameters": { "path": "Program.cs", "diff": "Listing: C:\Users\x11\Desktop\newsite\GravityLandingPage\Views\Docs (top-level only)
Found 12 items:
- [FILE] Views\Docs\Architecture.cshtml
- [FILE] Views\Docs\Article.cshtml
- [FILE] Views\Docs\Article.cshtml.bak.639273610948482039
- [FILE] Views\Docs\Article.cshtml.bak.639273615833682148
- [FILE] Views\Docs\Index.cshtml
- [FILE] Views\Docs\Index.cshtml.bak.639273610427101065
- [FILE] Views\Docs\Index.cshtml.bak.639273615329883474
- [FILE] Views\Docs\Index.cshtml.bak.639273616218142076
- [FILE] Views\Docs\Installation.cshtml
- [FILE] Views\Docs\Licensing.cshtml
- [FILE] Views\Docs\Licensing.cshtml.bak.639273620177728074
- [FILE] Views\Docs\Licensing.cshtml.bak.639273620356689749
.output" },
      "description": "Apply the fix"
    }
  ]
}

Key features:

Output chaining - Steps can reference $stepN.output from previous steps
Agent assignment - Each step targets a specific agent
Parameter validation - JSON schema validation before execution
Parallel execution hints - Independent steps can run concurrently

Layer 3 — Agent Execution (`Orchestrator`)

The Orchestrator manages an agent pool (max 5 concurrent AppEngine instances). Each engine runs a reasoning loop up to MaxSteps iterations.

Execution Loop

Observe - Receive task step or user message
Reason - Call LLM with system prompt, history, and available tools
Act - LLM outputs tool call → ReasoningRouter → target IAgent
Observe Result - Agent returns AgentResult (success/output/error)
Loop - Feed observation back to LLM until task complete or max steps

AppEngine

Each AppEngine instance maintains its own conversation history, token budget, and compression state. The Orchestrator assigns steps to available engines.

Fast-Path Bypass

Simple read/list commands are detected by regex and executed directly without LLM involvement:

"show me Program.cs" → code_editor.read_file
"list files in src" → code_editor.list_directory
"search for TODO" → code_editor.search_in_files

This reduces latency from ~2-5s to ~50-200ms for common exploration tasks.

Layer 1 — Intent Classification (IntentRouter)