@arivie/agent

@arivie/agent builds the Mastra Agent with analytics and workspace tools attached directly. Single-agent architecture — no supervisor, no sub-agents.

`makeAgent(opts)`

import { makeAgent } from "@arivie/agent";

const agent = makeAgent({
  ownerId: "acme",
  model: google("gemini-2.5-flash"),
  semantic,           // SemanticLayer
  contextMode: "preload",  // "preload" | "indexed"
  sources,            // resolved source adapters
  sourceMetadata,     // descriptions for system prompt
  workspace,          // Mastra Workspace
  skillsProcessor,    // skill loader
  skillsMode: "auto", // "eager" | "on-demand" | "auto"
  compileMetric: true,
  limits: { rowsPerQuery: 5000 },
  hooks: { onToolCall: async (e) => audit(e) },
});

You typically don’t call makeAgent directly — defineArivie calls it for you. Use this reference to understand the parameters.

Context modes

Mode	Behavior
`preload`	Full semantic layer in system prompt. Best for fewer than ~30 entities.
`indexed`	Embedding-based retrieval. Requires vector store + embeddings config.

autoDetectMode(semantic) picks the mode based on token count.

Tools registered

Tool	Condition	Purpose
`execute_<source>`	Always (one per source)	Read-only SQL execution
`compile_metric`	`compileMetric: true`	Canonical metric definitions
`mastra_workspace_write_file`	Workspace configured	Write files
`mastra_workspace_read_file`	Workspace configured	Read files
`mastra_workspace_grep`	Workspace configured	Search file contents
`mastra_workspace_list_files`	Workspace configured	List directory
`mastra_workspace_edit_file`	Workspace configured	Find/replace in file
`mastra_workspace_mkdir`	Workspace configured	Create directory
`mastra_workspace_delete`	Workspace configured	Delete file/directory
`mastra_workspace_file_stat`	Workspace configured	File metadata
`workspace_bash`	`bash: true`	Shell access in sandbox
`finalize_report`	Sandbox FS	Report completion signal
`search_skills` / `load_skill` / `skill_read`	`skillsMode: "on-demand"`	Lazy skill loading

Prompt constants

The agent’s system prompt is built from composable constants exported by @arivie/agent:

Export	Purpose
`buildSystemPrompt(opts)`	Assembles the full system prompt
`buildSystemPromptIndexed(opts)`	Indexed-mode variant
`WORKSPACE_NAVIGATION_RULE`	Instructs agent on workspace usage
`ASSUMPTION_STATING_RULE`	”State assumptions explicitly” rule
`SELF_CORRECTION_RULES`	Zero-row / all-null / implausible checks
`SKILL_DISCIPLINE_EAGER`	Skill-following rules (eager mode)
`SKILL_DISCIPLINE_ONDEMAND`	Skill-loading rules (on-demand mode)

SQL guard

validateExecuteSql(sql) from @arivie/db-postgres uses a state-machine token scanner to enforce:

Must start with SELECT or WITH
No multi-statement queries
No CTE-DML hybrids
No DML/DDL operations

Types

ContextMode = "preload" | "indexed"
SkillsMode = "eager" | "on-demand" | "none"
MakeAgentOptions — full options interface