Skip to Main Content
DocsReferenceGlossary

Glossary

All platform-specific terms, abbreviations, and English-Chinese mappings in one place. Use it as a reference when other docs get dense.

All proper nouns, abbreviations, and common terms on the platform in English and Chinese. Sort by first letter. If you have any doubts when reading other documents, come back and check them.

A

agent_based
A step type. An orchestrator agent reads the task description and dynamically decides which worker agent to assign the subtask to. It is suitable for scenarios where tasks have multiple branches and each branch requires LLM judgment.
Agent
An entity that encapsulates an LLM session and its available tools. Agents in Braidrun are declared by preset and overrides as needed. It is not a step itself - some type of step (single / group_chat / classifier, etc.) will refer to the Agent.
API Key
Authentication key issued by LLM provider. It is stored as Credential in the platform and is always encrypted and never appears in the log.
Artifact(Artifact)
Files generated during the step run (Markdown, Excel, images, JSON, etc.). Save it to the artifact directory of the platform and can be downloaded from the execution details page.
Audit Log / Audit Log
A structured log of every user action and system event on the platform. The platform records it continuously; the viewer is unlocked on Enterprise. The default retention is 180 days hot and 730 days archived, and the operator can configure both.

B

Breakpoint / Breakpoint
Pause locations you can set in the debugger — 9 types: before a step, after a step, on step error, a state-machine state, a group-chat round, an iteration item, an agent_based dispatch, a sub-workflow entry, and a sub-workflow exit. A breakpoint can carry a conditional expression; on hit, the execution pauses and you can inspect and edit variables.
BYOK
Bring Your Own Key。Users bring their own LLM Provider’s API Key, and the token cost is charged to the user’s own account and does not account for the platform’s quota.

C

classifier
A step type. LLM assigns a category label to the input, and the output can be passed {{digest_topic}} Used as routing conditions in subsequent steps.
code
A step type. Scripts in seven languages — Python / JavaScript / TypeScript / Bash / Ruby / Lua / CLI — run in an isolated sandbox. The first choice whenever determinism beats an LLM.
Condition
Step-level Boolean expression. When false, the step is marked SKIPPED and execution is skipped; execution details will show this status. Downstream, do not assume that skipped steps must have output.
Third-Party AuthorizationsConnection
Credentials established for a specific external service. The platform comes with the field specifications, application guidelines and verification methods of the service. When saving, the upstream confirmation account will be authenticated; the official module will automatically reference it through the contract.
Credential(Credentials)
An external secret stored encrypted with AES-256-GCM. At run time it resolves personal first, then team. It never appears in logs or in a YAML export.
cron
A type of scheduled trigger. It accepts a standard 5-field cron expression (minute hour day-of-month month day-of-week). Each schedule entry carries its own timezone setting; in a multi-instance deployment, a distributed lock ensures it fires only once per trigger point.

D

DAG
Directed Acyclic Graph。Visual modeling view of the platform. Each node is a step, and the edges represent the execution sequence/data dependency. Bidirectional synchronization with YAML byte level.
DTO
Data Transfer Object。The payload format of JSON API uses camelCase; the corresponding YAML uses snake_case. Automatic two-way conversion when the platform is running.
Dry-run
The "dry run" button in the execute dialog. It only builds a static plan — agents, steps, dependencies, condition markers and missing variables — without creating an execution or calling any LLM, tool or third-party API.

E

Execute / Execution
One concrete run of a workflow. The state machine is PENDING → RUNNING → (COMPLETED / FAILED / CANCELLED / INTERRUPTED). While a manual approval is pending the execution stays RUNNING — it is the single step that waits.
extract
A step field. It pulls values out of that step's output with a regular expression or JSON Path and writes them into the runtime variable named in `variable`, so later steps can reference them by name.

F-G

FeatureGate
A service that controls feature availability and resource quotas by subscription plan (Free/Pro/Team/Enterprise). Determine how many workflows, how many schedules you can build, and whether you can use the debugger.
group_chat
A step type. Multiple Agents speak in turns and refer to each other's context. Suitable for brainstorming and critical review.

H-I

HMAC
Webhook 触发的兼容认证方式。外部调用方用共享密钥对 raw body 算 HMAC-SHA256,以 sha256=<hex> 格式放进 X-Webhook-Signature 头。新接入建议直接用带 WEBHOOK_TRIGGER 范围的 API Key。
idempotent
step field. Promise "same input → same output, no side effects". When auto-resume is turned on, completed idempotent steps can be skipped from being redone.

M

manual_approval
A step type (also available as an enhancement to any step). The execution flow is stuck here waiting for administrator approval/denial. Configurable notification channel and approvers list.
MCP
Model Context Protocol。Standard protocols across tools. Braidrun can act as an MCP Client to consume tools from external MCP Servers, and it can also expose its own capabilities in MCP Server mode.
Module / Module
A workflow that implements WorkflowModuleContract, so other workflows can call it as a black box through a sub_workflow step. The platform ships 492 built-in modules, and more can be installed from the marketplace.
Data Migration
The data structure migration process is automatically executed when the platform is upgraded; distributed locks are used across instances to ensure that only one node executes.

P

Preset
An agent template that bundles model, default tool set, system prompt, max_iterations and strategy under one name. Nineteen ship with the platform — universal / coder / researcher / writer / data_analyst and others — while the Free plan opens only universal / lightweight / chat.
Provider
LLM service provider. OpenAI / Anthropic / DeepSeek / Kimi / OpenRouter etc. A provider can be bound to multiple API Keys (credentials).

R

RAG
Retrieval-Augmented Generation。A type of Agent toolset that performs vector retrieval on the built-in knowledge base. The platform automatically inserts the search results into the Agent's context.
Recovery Policy
Strategy options for automatic continuation: ABORT (not continuing) / RESUME_FROM_LAST_INCOMPLETE (continuing from the interrupted step) / RESUME_FROM_START (restarting from the beginning).
Round-trip(Two-way Sync)
The two-way conversion YAML → DAG → YAML: structure, fields and values survive the round trip through the workflow model. Comments, blank lines and original formatting are not part of the save contract — put anything important in a real field such as description.

S

single
A step type. A single Agent + a task description. The most lightweight LLM calling primitive.
Sandbox(Sandbox)
The execution environment of the code step. Use a one-time isolation sandbox in production, least privileges / read-only root / outbound network.
Schedule
An entry that fires a workflow on a timetable. Four kinds are supported: CRON, INTERVAL, ONE_TIME and DELAYED_COMPLETION (fires a set delay after an upstream workflow finishes). Each schedule can carry preset inputs and its own time zone.
Skill
A capability package for agents: a directory built around SKILL.md that holds instructions, scripts and reference material. There are system skills preinstalled by the platform and skills you install from the skill market or import as a ZIP. Not to be confused with a tool.
SSE
Server-Sent Events。A protocol for one-way push events from the server to the browser. Real-time updates of current execution details primarily use WebSockets, with downgrade polling if the connection fails.
Step
The smallest execution unit of workflow. Each step can only select 1 main mode (single / group_chat / agent_based / code / classifier / state_machine / sub_workflow / manual_approval), and can be superimposed with enhancements such as condition / retry / extract / aggregate / idempotent / on_success / on_failure.
state_machine
A step type. Embedded with multiple states + state transition conditions, it is suitable for multiple iterations or workflows containing small processes.
sub_workflow
A step type. Call another workflow (or built-in module) as a black box. With contract verification, loop detection, and variable isolation.

T-U

Team
A group of users sharing one credential namespace and one schedule allowance. Only the owner — the person who created it — can invite or remove members and manage the subscription; who may edit which workflow is decided by share permissions.
Tool
Functions that the Agent can call during LLM inference. Built-in + MCP extension.
TPM
Tokens Per Minute。The current limiting indicator of LLM provider. There is a self-limiting TPM inside the platform to avoid blowing up the upstream.

V-W

Variable Reference
{{...}} Syntax. Common forms include var:name, steps.step_name.output, and runtime variables written by nodes such as extract and classifier. Credential plaintext is not read via template expressions.
Webhook
An endpoint that lets an external system POST to trigger a workflow execution. API Key authentication is primary, with HMAC signatures as a compatibility option. The request body's top-level fields map to workflow variables automatically by default, or you can specify the mapping explicitly with variableMapping.
Workflow
A collection of steps organized by DAG. Can be triggered manually by /cron/webhook/API. Use YAML to persist storage.

Y

YAML
The platform's workflow persistence format. Fields go snake_case (automatically mapped with DTO's camelCase). It can be run directly in the CLI (braidrun-agent), and the same YAML can also be edited in the Web UI.

Last Updated · 2026-08-06

Was this page helpful?