Symptom

The Server.LLM node owned an interactive chat UI and expected the model to respond with structured JSON (CREATE_NODES, UPDATE_NODE, etc.) to manipulate the swarm directly — bypassing verb dispatch. This contradicted every Krill architectural rule: nodes are woken by sources, not by user taps; state changes flow through ServerNodeManager, not through model output parsed in a processor.

Root cause

The original design treated the LLM as a chat surface rather than a transform node. The processor implemented a bespoke planner schema and emitted raw actions rather than calling executeSources(). There was no prompt field on LLMMetaData, no RESET verb path, and no downstream fan-out. The node was architecturally isolated from the rest of Krill’s source-trigger chain.

Fix

• Rewrote ServerLLMProcessor as a pure verb-driven processor: onSourceTrigger(EXECUTE) runs Ollama inference, appends [user, assistant] to meta.chat, emits an LLM SSE event, then calls nodeManager.executeSources(). RESET clears the chat without fan-out. post() and process() are no-ops.
• Added LLMMetaData.prompt (krill-sdk 0.0.27) and extended LLMEventPayload with chatCleared: Boolean so the client can clear its local chat state on RESET.
• Rewrote EditLLM.kt to surface prompt, context-node picker, Test Connection button, and a read-only conversation replay; removed the interactive chat input and planner.
• Added POST /api/llm/{nodeId}/test route backed by ServerLLMProcessor.testConnection().
• Added LlmApiClient in shared for the client-side Test Connection call, registered via Koin single<LlmApiClient> in SharedModule.
• Updated KrillApp.Server.LLM.json with llmRole: "transform" and verb-driven behavior descriptions; rewrote the LLM blog post to match.

Prevention

• Any new node processor that breaks the onSourceTrigger → executeSources fan-out pattern should be caught in code review. The pattern is now documented in KrillApp.Server.LLM.json under llmBehavior and in the blog post, making the contract explicit.
• Regression tests in ServerLLMProcessorTest cover EXECUTE (chat grows, executeSources called), RESET (chat cleared, executeSources not called), and Ollama HTTP error (error state set, no fan-out), preventing a future revert to the old model.