@open-vera/core

Stateless agent runtime foundation for the Vera framework.

Handles a single LLM call end-to-end: intent classification → model routing → tool execution → context management. No orchestration logic — pure execution primitive.

Install

npm install @open-vera/core

What's Inside

| Module | Capability | |---|---| | adapters/ | Unified LLMAdapter interface — Anthropic, OpenAI, Gemini, DeepSeek, Groq, Azure | | agent/ | streamAgent / runAgent — multi-turn loop, tool dispatch, retry, context compression | | agent/subagent.ts | agent tool — orchestrator/worker delegation, isolation modes, background jobs | | context/ | Token estimation, window trimming, progressive / micro / reactive compression, recall | | intent/ | classifyIntent / routeTarget — L0–L3 classification, domain detection | | tools/ | 7 built-in tools: read_file write_file edit_file list_dir glob grep bash | | tools/registry.ts | ToolRegistry — register, execute, lifecycle hooks | | tools/security.ts | Path boundary enforcement, tool whitelist, injection defense, read-only mode | | tools/permission-rules.ts | Persistent allow/deny rules, bash risk confirmation | | session/ | JSONL session store, cost tracking, AI-generated titles | | repl/ | React + Ink terminal UI — ConversationPanel, SessionPicker, DiffView, theme system | | project-context/ | .vera/rules.md / CLAUDE.md loading, path-scoped rule activation | | worktree/ | Git worktree creation and management |

Intent Routing

User input → [classify: ~100ms] → routing decision → [target model]

| Level | Description | Default Model | |---|---|---| | L0 | Casual chat, simple Q&A | claude-haiku / gpt-4o-mini | | L1 | Single-step tasks | claude-haiku / gpt-4o-mini | | L2 | Multi-step tasks | claude-sonnet / gpt-4o | | L3 | Complex planning, deep reasoning | claude-opus / o3 |

L3 tasks automatically activate Plan Mode.

Infinite Context

Four-layer system so long tasks never lose their goal:

| Layer | Mechanism | Trigger | |---|---|---| | Sliding window trim | Drop earliest turns, preserve task anchor | 80% token threshold | | Progressive compression | LLM summarization injected as system context | Token threshold exceeded | | Micro-compaction | Heuristic cleanup of stale tool results — no LLM call | Time-gap based | | Reactive compaction | Aggressive compression + retry on prompt-too-long | API error |

The original task definition is always preserved.

Subagent Isolation Modes

| Mode | Mechanism | Use Case | |---|---|---| | none | Shared context (default) | Standard delegation | | try | Isolated git worktree, reviewable via /merge | Experimental code changes | | remote | Pluggable external executor backend | Distributed or sandboxed execution |

Quick Start

import { streamAgent } from "@open-vera/core/agent";
import { AnthropicAdapter } from "@open-vera/core/adapters";

const adapter = new AnthropicAdapter({ apiKey: process.env.ANTHROPIC_API_KEY });

for await (const event of streamAgent({
  adapter,
  messages: [{ role: "user", content: "List files in the current directory" }],
  tools: [],
})) {
  if (event.type === "text_delta") process.stdout.write(event.delta);
}

Architecture

@open-vera/core is a strict dependency of @open-vera/openvera. Core never imports from Harness — the stateless loop is usable standalone.

@open-vera/openvera → @open-vera/core

License

MIT