Skip to main content

Zenii Process Flows

Table of Contents​

Chat Request Flow​

Startup Sequence​

Default Paths by OS​

Resolved via directories::ProjectDirs::from("com", "sprklai", "zenii"):

OSConfig PathData Dir / DB Path
Linux~/.config/zenii/config.toml~/.local/share/zenii/zenii.db
macOS~/Library/Application Support/com.sprklai.zenii/config.toml~/Library/Application Support/com.sprklai.zenii/zenii.db
Windows%APPDATA%\sprklai\zenii\config\config.toml%APPDATA%\sprklai\zenii\data\zenii.db

Override via config.toml:

data_dir = "/custom/data/path"        # overrides default data directory
db_path = "/custom/path/zenii.db"  # overrides database file directly

Error Handling Flow​

Database Operation Flow (async-safe)​

WebSocket Message Flow​

Identity Loading Flow​

Skill Loading Flow​

User Learning Flow​

Channel Message Flow​

Channel Registration Flow​

Desktop Boot Flow​

The desktop app uses a hybrid gateway model. By default it starts an embedded gateway; if ZENII_GATEWAY_URL is set, it connects to an external daemon instead.

Credential Flow​

Provider Management Flow​

Context Injection Flow​

Skill Proposal Flow​

Scheduler Notification Flow​

The scheduler tick loop executes payloads via PayloadExecutor (scheduler/payload_executor.rs) and delivers notifications through multiple channels. The TokioScheduler ↔ AppState circular dependency is resolved via OnceCell post-construction wiring.

Embedding Flow​

Reasoning Continuation Flow​

Channel Router Message Pipeline​

The ChannelRouter orchestrates the full message processing flow from inbound channel message to outbound response. It runs as a background task spawned during init_services(), consuming messages from an mpsc channel and using a watch signal for graceful shutdown. Lifecycle hooks (Stage 8.8) are best-effort — failures are logged but do not block the pipeline.

Plugin Lifecycle Flow​

Plugins are managed through all three client interfaces, each communicating with the gateway over HTTP:

  • CLI: zenii plugin <cmd> -- direct HTTP calls to gateway plugin endpoints
  • Web/Desktop: PluginsSettings.svelte component -- pluginsStore fetches/mutates via HTTP
  • TUI: PluginList mode -- ZeniiClient HTTP calls (keybindings: p open, j/k nav, e toggle, d remove, i install, r refresh, Esc back)

Onboarding / First-Run Setup Flow​

On first launch, all interfaces check GET /setup/status to determine if onboarding is needed. The SetupStatus response includes needs_setup, missing fields, detected_timezone, and has_usable_model. If setup is needed, a multi-step wizard collects AI provider configuration (provider, API key, model), optional channel credentials (Telegram, Slack, Discord), and user profile (name, location, timezone).

Interface Variants​

  • Desktop: 3-step OnboardingWizard component (provider setup via embedded ProvidersSettings, optional channels via ChannelsSettings, then profile fields). Next button is in the card header for visibility on long pages.
  • CLI: zenii setup command -- interactive flow using dialoguer (Select, Confirm, Password, Input prompts). Channels step uses Confirm prompt (default: skip).
  • TUI: 5-step overlay modal (ProviderSelect, ApiKey, ModelSelect, Channels, Profile) with j/k navigation. Channels step has Tab to switch between Telegram/Slack/Discord and s to skip.

Config fields: user_name: Option<String>, user_timezone: Option<String> (IANA format), user_location: Option<String> (human-readable)

Key files: onboarding.rs, gateway/handlers/config.rs (setup_status), web/src/lib/components/OnboardingWizard.svelte, crates/zenii-cli/src/commands/onboard.rs, crates/zenii-tui/src/ui/onboard.rs

Auto Fact Extraction Flow​

After each chat response, ContextBuilder::extract_facts() optionally calls an LLM to extract structured facts about the user and stores them via UserLearner::observe(). Fire-and-forget -- errors are logged, not propagated.

Output format: category|key|value per line. Categories: preference, knowledge, context, workflow.

Key files: ai/context.rs (ContextBuilder::extract_facts), user/learner.rs (UserLearner::observe)

Auto-Discovery Flow​

The context engine uses keyword matching to detect which feature domains are relevant to the user's message, then loads only pertinent agent rules and expanded context sections.

Domain-to-category mapping:

  • Channels → "channel" rules
  • Scheduler → "scheduling" rules
  • Skills/Tools → "tool_usage" rules
  • Always included → "general" rules

Key file: ai/context.rs (ContextDomain, detect_relevant_domains(), domains_to_rule_categories())

Agent Self-Learning Flow​

The agent can record behavioral rules during conversations via the agent_notes tool. These rules persist in the database and are automatically injected into future conversations based on domain relevance.

Categories: general, channel, scheduling, user_preference, tool_usage

Control: Gated by self_evolution_enabled config flag (runtime toggle via Arc<AtomicBool>)

Key files: tools/agent_self_tool.rs, ai/context.rs (load_agent_rules())