Skip to main content

​
songbird-agent

Agentic orchestration for Songbird’s copilot: single converged pipeline with complexity-based model tier routing and multi-round tool execution.

​
What it owns

  • Prompt complexity classification (ML-based with heuristic fallback)
  • MiniLM-L6-v2 embedding + logistic regression head for simple/complex routing
  • Model tier selection: Simple → Speed (fast, no thinking) / Complex → Thinker (reasoning, planning)
  • Intent classification for tool filtering (Chat, BirdEdit, Mixer, etc.)
  • DAW tool registry (maps agent actions to sync-engine commands)
  • Chat orchestrator for streaming multi-round conversations with tool execution

​
Architecture

User prompt
  → classify_complexity() → Simple | Complex
  → select model tier    → Speed (flash-lite) | Thinker (pro + thinking)
  → classify_intent()    → filter tool subset
  → streaming LLM call   → text/thinking/tool-call events
  → tool execution loop   → dispatch via sync engine, feed results back
  → repeat (max 5 rounds) or complete

​
Classification

The classifier routes user prompts to Speed (simple) or Thinker (complex):
  • ML path (ml-classifier feature): MiniLM-L6-v2 sentence embeddings → trained logistic regression head. ~15ms on CPU. Trained on 345 synthetic DAW prompts (196 simple, 149 complex) with 100% test accuracy.
  • Heuristic fallback: keyword pattern matching against the tool registry. Used when the MiniLM model isn’t available.
Weights are bundled at compile time via include_bytes! (models/classifier_weights.bin).

​
Public API

  • ChatOrchestrator — primary entry point: classify complexity/intent, streaming multi-round chat with tool execution
  • ModelTier — Speed (fast) or Thinker (reasoning) tier, selected by complexity classification
  • Complexity — Simple or Complex, from the classifier
  • AgentConfig — model assignments, planner depth, thinking settings, fallback routing
  • ChatOrchestrationResult — content, thinking, tool calls executed, model, rounds
  • RoundOutcome — Complete / NeedsToolResults / Error
  • OrchestratorEvent — streaming events (TextDelta, ThinkingDelta, ToolCallRequested, Done, Error)
  • ToolRegistry — DAW tool definitions with intent-based filtering and sync-engine command mappings
  • AgentPipeline — deprecated, retained as reference for future local-model planner integration

​
Feature flags

  • ml-classifier (default): enables MiniLM embedding-based classification. Requires songbird-inference/embeddings.

​
Where tests live

cargo test -p songbird-agent — classifier accuracy, weight parsing, plan parsing, tool registry, orchestrator, model tier routing.