@mitii/v8

Host-neutral Mitii agent runtime: intake → understanding → decision policy → repository context → prompt construction → model/tool loop → verification.

The model is never the permission authority. Decision Policy owns routes, grants, approval mode, and verification requirements. Hosts inject ports (filesystem, SQLite, LLM, search, skills, memory).

Install

npm install @mitii/v8

Requires Node.js 20+. License: AGPL-3.0-or-later.

Most apps should use @mitii/sdk (createMitiiClient) instead of wiring pipelines by hand. Use this package when you need the V8 modules/ports directly.

Quick start

import {
  AgentEnginePipeline,
  EchoLlmPort,
  RequestIntakePipeline,
  RequestUnderstandingPipeline,
  DecisionPolicyPipeline,
  PromptConstructionPipeline,
  ToolRuntimePipeline,
  VerificationPipeline,
  composeReadOnlyAgentEngine,
} from '@mitii/v8';

// Prefer @mitii/sdk for a full client. Direct composition example:
const understandingLlm = new EchoLlmPort();
const runLlm = new EchoLlmPort();
// …inject host ports, then compose AgentEnginePipeline / composeReadOnlyAgentEngine

Module layout

src/
├── engine/
│   ├── agent-engine/        AgentEnginePipeline (run orchestration)
│   └── tool-runtime/        ToolRuntimePipeline (granted tool execution)
└── modules/
    ├── request-intake/
    ├── request-understanding/
    ├── repository-state/
    ├── repository-context/
    ├── decision-policy/
    ├── planning/
    ├── prompt-construction/
    ├── model-gateway/         Echo + OpenAI-compatible LlmPort
    ├── skills/
    ├── memory/
    └── verification/

Public pipelines

| Module | Pipeline | Input → Output | |--------|----------|----------------| | request-intake | RequestIntakePipeline | CreateUserRequestInput → UserRequestEnvelope | | request-understanding | RequestUnderstandingPipeline | envelope → understanding result | | repository-state | RepositoryStatePipeline | candidate → published RepositoryStateReference | | repository-context | RepositoryContextPipeline | state ref + query → context | | decision-policy | DecisionPolicyPipeline | envelope + understanding → ExecutionDecision | | planning | PlanningPipeline | evidence + depth → PlanArtifact | | prompt-construction | PromptConstructionPipeline | decision + context → budgeted ModelRequest | | tool-runtime | ToolRuntimePipeline | authorized call → ToolResult | | verification | VerificationPipeline | change + state + policy → verification result | | agent-engine | AgentEnginePipeline | start request → AgentRunHandle |

Import policy

Import only from the package root (@mitii/v8). Never import another module's internal/ or actions/ paths.

import {
  RequestIntakePipeline,
  DecisionPolicyPipeline,
  AgentEnginePipeline,
  EchoLlmPort,
  OpenAiCompatibleLlmPort,
  MODEL_PROVIDER_SUPPORT,
} from '@mitii/v8';

Architecture

See ARCHITECTURE.md for component boundaries, request flows, storage, security, and an end-to-end example.

Development (monorepo)

pnpm --filter @mitii/v8 typecheck
pnpm --filter @mitii/v8 test
pnpm --filter @mitii/v8 build

pnpm --filter @mitii/v8 test auto-heals better-sqlite3 ABI mismatches (Electron vs system Node). Prefer pnpm run rebuild:native for F5 — it stages the Electron binding and restores Node ABI afterward.

Links

• Repo: Mitii-dev/Mitii
• Docs: docs.mitii.dev
• Related: @mitii/sdk, @mitii/host, @mitii/cli