Supervisor
The LangGraph supervisor: routing, history limits, and sliding-window summarisation.
AnthropicOpenAIOllamaGoogleThe supervisor is the single entry point of the Orchid LangGraph graph. Every user message flows through it first. The supervisor never calls MCP servers or reads from the vector store — its only job is intent analysis, mode selection, and routing.
What the supervisor does
On each turn the supervisor:
- Analyses the user's intent via a structured-output LLM call
- Chooses an execution mode:
parallel,sequential, orskill - Routes to one or more specialist agents (or responds directly for trivial queries)
- After all agents have replied, synthesises a final coherent response
In parallel mode the supervisor fans out to multiple agents simultaneously using LangGraph's Send() primitive. In sequential mode agents run one at a time, with each agent's output visible to the next via mcp_context.
Configuring the supervisor
All supervisor options live under the supervisor: key:
supervisor:
assistant_name: "Sports AI"
history_max_turns: 20
history_max_chars: 1000
history_summary_enabled: true
history_summary_model: "ollama/llama3.2"
history_summary_recent_turns: 10
routing_model: "ollama/llama3.2"| Field | Default | Purpose |
|---|---|---|
assistant_name | "AI assistant" | Name used in the routing system prompt |
history_max_turns | 20 | Maximum user/assistant pairs passed to the LLM |
history_max_chars | 1000 | Per-message character cap before truncation (appends …) |
history_summary_enabled | true | Activate sliding-window summarisation |
history_summary_model | supervisor model | Cheap model for compression calls |
history_summary_recent_turns | 10 | Verbatim turns kept at the tail of the window |
routing_model | supervisor model | Separate fast model for routing and sequential-advance phases |
Multi-turn context extraction
The supervisor calls OrchidAgent.extract_conversation_history() before every routing, synthesis, and sequential-advance step. This static method:
- Filters out internal supervisor routing messages (any message starting with
[Supervisor) - Strips agent-name prefixes from AI messages (e.g.
[Basketball Agent]\n…→…) - Excludes the current user query (added separately as the trigger)
- Caps the result to
history_max_turns * 2messages
Sliding-window summarisation
When history_summary_enabled: true, turns older than the history_summary_recent_turns window are compressed into a single summary paragraph via an LLM call. The most recent exchanges are kept verbatim. On LLM failure the system falls back to returning only the recent turns — no crash, no data loss.
Enable summarisation for cost-conscious LLMs
Set history_summary_enabled: true and point history_summary_model at a fast, cheap model (e.g. gemini/gemini-2.5-flash-lite or ollama/llama3.2). For a 40-turn conversation this typically reduces the token count passed to the routing and synthesis calls by 60–80 % while preserving the full conversational thread.