@open-vera/openvera

Stateful orchestration kernel for the Vera framework.

Drives multi-step autonomous tasks through a principled Plan → Act → Critique → Replan loop. Every flow transition is validated; the runtime cannot drift into an inconsistent state.

Install

npm install @open-vera/openvera

Requires @open-vera/core as a peer dependency.

What's Inside

| Module | Capability | |---|---| | runtime/flow-state.ts | Flow state machine — 11 states, legal transition enforcement, illegal jumps throw | | runtime/runtime.ts | HarnessRuntime — drives Plan → Act → Critique → Replan loop | | runtime/planner.ts | planFromPrompt — LLM → ExecutionPlan, with retry and JSON repair | | runtime/critique.ts | critiquePlan / critiqueStep — confidence < 0.7 triggers automatic replan | | runtime/approval.ts | High-risk operation gates — pause flow, await human confirmation | | skill/ | Skill loading from Markdown, SkillResolver — intent-driven activation | | evaluator.ts | exact / contains / tool_match / llm_judge evaluation, concurrent execution |

Flow State Machine

intaking
  → planning
    → dispatching
      → executing
        → waiting_tool
        → waiting_approval   ← human-in-the-loop gate
        → critiquing
          → replanning → dispatching  (loop)
          → completed
          → failed
      → paused

Every transition is validated. An invalid state jump throws immediately — the runtime cannot silently drift.

Cognitive Role Separation

Planner      Reads context → generates ExecutionPlan
Role Agent   Executes steps to defined exit criteria → produces deliverables
Challenger   Independently scores every output → accumulates lessons
Orchestrator Schedules agents → manages context resets → enforces approval gates

A Role Agent never decides whether its own work is complete — that right belongs exclusively to the Challenger. This prevents the most common agent failure mode: acting as implementer, evaluator, and judge simultaneously.

Challenger Learns Over Time

After every run the Challenger appends discovered failure patterns to .flow/challenger/lessons/{step}.md. On the next run it reads those lessons and applies them as attack angles — the system becomes harder to fool as it accumulates institutional knowledge about where this specific workflow tends to fail.

Quick Start

import { HarnessRuntime } from "@open-vera/openvera";
import { AnthropicAdapter } from "@open-vera/core/adapters";

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

await runtime.run("Refactor the authentication module and add unit tests");

CLI

# Run a flow
openvera flow run

# Launch REPL with Harness
openvera repl --dir .

Architecture

@open-vera/openvera → @open-vera/core

Harness owns flow orchestration. Core handles single LLM call execution. Dependency is strictly one-way — Core is usable standalone.

License

MIT