Glossary¶
Donna-specific terms and their definitions.
| Term | Definition | See Also |
|---|---|---|
| Action handler | A chat engine plugin that intercepts natural-language commands (e.g., "show my tasks") and translates them into API calls. Registered via the ActionRegistry. |
Chat |
| Automation | A cron-dispatched recurring skill execution with cadence policy and pause/resume lifecycle. | Skill System |
| BudgetGuard | The pre-call gate that checks spend against daily (\(20) and monthly (\)100) thresholds before allowing an LLM invocation. | Cost, Handle Budget Breach |
| Capability | A unit of agent functionality (e.g., task_create, email_draft) with defined tool access and autonomy level. |
Agents, Skill System |
| Challenger | A second model invocation used to verify high-stakes outputs via a compare-and-decide pattern. | Model Layer |
| Circuit breaker | A resilience pattern that stops retrying a failing service after consecutive failures, entering an open state until a cooldown expires. | Resilience |
| Collection | The subsystem that captures full LLM request/response payloads to disk for forensic analysis via Claude Inspector. | Collection |
| Correction | A user-initiated change to a task field (priority, due date, etc.) that the preference system observes and learns from. | Preferences |
| CorrectionSubscriber | The event-driven listener that detects task field changes and logs them as corrections for preference learning. | Preferences |
| Dedup | Deduplication of incoming tasks via fuzzy title matching and LLM semantic comparison against existing tasks. | Task System |
| Escalation gate | The over-budget decision UI (Approve $X / Manual / Pause / Cancel) that replaces the simple pause-only behavior. | Cost |
| Evictor | The PayloadEvictor component that enforces disk budget by deleting oldest payload files when storage exceeds thresholds. |
Collection |
| GPU tracker | Monitors Ollama GPU memory usage on the RTX 3090 to inform model routing decisions. | LLM Gateway |
| Invocation log | The invocation_log SQLite table that records every LLM call with model, latency, tokens, cost, and task context. |
Observability, Model Layer |
| Memory vault | A Git-backed knowledge store that chunks, embeds, and indexes content from tasks, chat, and corrections for semantic retrieval. | Memory Vault |
| Model alias | A logical name (e.g., task_parse, chat_respond) mapped to a specific model + parameters in config/donna_models.yaml. |
Model Layer |
| ModelRouter | The central abstraction that routes all LLM calls through alias lookup, budget checking, and structured logging. | Model Layer |
needs_scheduling |
Task state for a time-bound task the scheduler could not place before its deadline — surfaced explicitly instead of stranding in backlog. | Task System |
| Orchestrator | The central dispatcher that receives user messages, classifies intent, and routes to the appropriate skill or handler. | Orchestrator |
| Payload | The full JSON capture of an LLM request (prompt, parameters) and response (output, tokens, latency) stored by the collection subsystem. | Collection |
| Priority engine | Computes dynamic task priority from urgency, importance, dependencies, and user corrections. | Scheduling |
| Routing gate | The deterministic, LLM-free gate that routes a freshly captured task to the scheduler (time-bound), automation (recurring), or backlog (undated) based on its time intent and priority. | Scheduling |
| Rule applier | The preference subsystem component that applies learned rules (from corrections) to new tasks at creation time. | Preferences |
| Shadow preference | A candidate preference rule extracted from corrections that needs validation before becoming active. | Preferences |
| Skill | A YAML-defined unit of work with steps, tool access, model routing, and validation. The atomic unit of Donna's capability system. | Skill System, Run a Skill |
| Slice | A build increment with defined acceptance criteria and spec references, used to scope implementation work. | Development: Slices |
| State machine | The configurable task lifecycle defined in config/task_states.yaml that governs valid transitions between states. |
Task System |
| Supabase sync | Async write-through replication from local SQLite to a Supabase Postgres instance for cloud access and backup. | Integrations |
| TaskEventBus | The pub/sub system that broadcasts task lifecycle events (created, updated, completed) to subscribers like CorrectionSubscriber. |
Task System, Preferences |
| Time intent | The structured representation of when a task should happen (exact / window / constrained / recurring / none), emitted by the parser and stored as time_intent_json. Drives the routing gate and derives deadline / deadline_type. |
Task System |
| Tool dispatch | The validation layer where models propose tool calls and the orchestrator validates and executes them. Models never call tools directly. | Agents, Skill System |
| WAL mode | SQLite Write-Ahead Logging mode, enabling concurrent reads during writes. All Donna SQLite databases use WAL. | Architecture: Overview |