Glossary
The vocabulary used across these docs. It is also the translation contract: the Portuguese pages use exactly the equivalents in the last column, and never alternate between two words for the same thing.
Core concepts
| Term | What it means | Portuguese |
|---|---|---|
| Agent | A class decorated with @Agent, pairing a provider, a set of tools and a markdown prompt. One step of a workflow. | Agente |
| Tool | A class decorated with @Tool. An action the model may ask for, described by a name, a description and a Zod schema. | Tool |
| Workflow | A class decorated with @Workflow. Declares the order of the steps and the shared state. | Workflow |
| Provider | What talks to the model. OllamaProvider, OpenAIProvider, or your own. | Provider |
| Run | One execution of a workflow, from app.run(...) to its result. Has its own context, budget, cancellation and recorder. | Execução |
| Step | A unit inside @Workflow({ steps }): an agent, a parallel block or a loop block. | Passo |
| Turn | One round of the agent: a model call plus the tool it may have triggered. Summarised in ctx.turn. | Turno |
Execution
| Term | What it means | Portuguese |
|---|---|---|
Context (ctx) | The object that crosses every step of a run. Carries state, output, turn, data, signal and the run controls. | Contexto |
| Prompt | Two different things, and the docs always say which. The agent's prompt is @Agent({ prompt }), a markdown file, fixed per class, sent as the system message. The run's prompt is run({ prompt }), a string, one per execution, sent as the first user message. | Prompt |
| State | The class declared in @Workflow({ state }), instantiated once per run and injected with @state(). | Estado |
| Memory | The agent's working memory: run({ state }), read by the model in every turn of this run, gone when it ends. | Memória |
| Vector memory | Long-term memory: VectorMemory, injected by @memory(), searched by similarity and outliving the run. The databases behind it are ThenaConfig.stores. | Memória vetorial |
| Data | run({ data }). The run's own data channel. Never goes to the model, never reaches the report. | Data |
| Budget | A ceiling for the whole run — time, calls, tokens, cost. | Orçamento |
| Hook | An optional method on the agent class (beforePrompt, beforeTool, afterTool, afterResponse, onError). | Hook |
| Middleware | A function wrapping every tool run (tool) or every model call (chat), registered through app.use. | Middleware |
| Plugin | An object passed to app.use. Observes (onEvent) and/or intercepts (tool, chat). | Plugin |
| Runtime | The layer that compiles a workflow and runs its steps. Reachable as WorkflowRuntime for nested runs. | Runtime |
Observability
| Term | What it means | Portuguese |
|---|---|---|
| Report | The HTML + JSON record written at the end of a run. | Report |
| Recorder | What builds the execution tree while the run happens. | Gravador |
| Execution event | An ExecutionEvent: the start or end of a step, delivered to onEvent, to log and to plugins. | Evento de execução |
| Node | An entry in the execution tree — workflow, agent, chat, tool, loop, parallel. | Nó |
| Redaction | Masking known secrets in captured content before it reaches the report, the log or a plugin. | Mascaramento |
| Rescue | Recovering a tool call the model wrote as text instead of a structured field. Shows up as toolCallSource: "rescued". | Resgate |
Vector memory
| Term | What it means | Portuguese |
|---|---|---|
| Vector store | The backing database. VectorStore is the contract; @thenajs/qdrant-client is one implementation. | Banco vetorial |
| Collection | Where documents live in the store. Holds one embedding size. | Collection |
| Dataset | A logical partition inside a collection, chosen per remember / recall. | Dataset |
| Embedding | The vector representation of a text, produced by the provider's embed(). | Embedding |
| Recall | Retrieving by similarity. | Recall |
| Remember | Writing a document to the store. | Remember |
Words we deliberately do not use
- "Chain" — the framework has workflows and steps. No chains.
- "Prompt engineering" — the agent's prompt is a markdown file you edit. We say writing a prompt.
- "Prompt" without saying which one — see the entry above. In body text, say the agent's prompt or the run's prompt.
- "LLM" in body text — we say model.
LLMappears only where it names the category, as in "a framework for LLM agents". - "Session" — a run is a run. What persists between runs is memory.
Notes for translators
Portuguese pages keep the English form for terms that are already the name of an API symbol — Tool, Workflow, Provider, Hook, Middleware, Plugin, Runtime, Report, Dataset, Embedding. Translating those would break the link between the prose and the code the reader is looking at.
Terms that name a concept rather than a symbol are translated: Agent → Agente, Run → Execução, Step → Passo, State → Estado, Context → Contexto, Budget → Orçamento.
Agente is the one exception to the rule above, and it is deliberate: the word is overwhelmingly used in Portuguese technical writing, and @Agent stays in English in the code anyway.