Services - storage, server, model runtime

flow-storage

Flow stores data in SQLite at ~/.flow-studio/db/flow.sqlite. The Store type wraps that database.

pub struct Store { /* open(path), shared by all hosts */ }

// Runs and steps
pub fn list_runs(&self, limit: usize) -> Result<Vec<RunSummary>, StoreError>;
pub fn get_execution(&self, id: &str)
    -> Result<Option<(ExecutionRecord, Vec<ExecutionStepRecord>, Vec<InterceptionRecord>)>, StoreError>;
pub fn upsert_step(&self, rec: &ExecutionStepRecord) -> Result<(), StoreError>;

// AI decision audit trail
pub struct AiAuditRecord { pub execution_id: String, pub node_id: String,
    pub at: DateTime<Utc>, pub kind: String, pub payload: serde_json::Value }
pub fn record_ai_audit(&self, rec: &AiAuditRecord) -> Result<(), StoreError>;
pub fn list_ai_audit(&self, execution_id: &str) -> Result<Vec<AiAuditRecord>, StoreError>;

Audit kinds include ai_invocation, ai_routing_decision, ai_review_required / ai_review_resolved, ai_tool_call (coding-agent turns), and agent_edit_applied / agent_edit_rejected. Writes happen as part of the execution cycle via StorageSink, not after it.

Schedules (flow_schedules), template membership, and interception records live in the same database.

flow-server

This is the host for the browser edition. It is an axum HTTP server that exposes FlowApp using the same command names the desktop uses. As a result, the React app runs unchanged against either transport.

• POST /api/invoke - the command bridge ({ cmd, args }), kept in lockstep with the Tauri command set.
• POST /api/run - execute a graph with the event stream over SSE.
• GET /health - liveness.

To add a host-visible command, you implement it in both the Tauri commands and flow-server/src/invoke.rs. Cross-host parity is a release gate.

flow-models-server

This is the lifecycle manager for the local inference engine. It is re-exported as flow_application::llm_server.

pub struct LlmServerHandle;
pub struct LlmServerStatus { pub running: bool, pub endpoint: Option<String>,
    pub model_path: Option<String>, pub pid: Option<u32> }
pub struct LlamaParams { /* context, gpu layers, threads, flash-attn,
    kv-cache type, enable_thinking, ... */ }

It resolves the engine binary from the managed ~/.flow-studio/engines/ dir, a saved setting, or $PATH. On first use, it fetches the binary for the current OS and arch. It then starts llama-server with a chosen .gguf, discovers the OpenAI-compatible endpoint, and reports status. FlowApp::start_local_llm and stop_local_llm wrap it and keep local_ai_base_url in sync, so the local AI provider and the DSL generator always point at the managed server.