@yuaone/core

Agent runtime for YUAN coding agent — skills, strategies, multi-provider LLM, tool-use loop.

npm install @yuaone/core

Usage

import { AgentLoop, createAgentLoop } from "@yuaone/core";

const loop = createAgentLoop({
  provider: "anthropic",
  apiKey: process.env.ANTHROPIC_API_KEY,
  model: "claude-sonnet-4-6",
  workDir: process.cwd(),
});

for await (const event of loop.run("refactor auth.ts to use async/await")) {
  if (event.kind === "agent:token") process.stdout.write(event.content);
  if (event.kind === "agent:completed") break;
}

Architecture

| Module | Description | |--------|-------------| | AgentLoop | Core tool-use loop — LLM ↔ tools ↔ approval | | HierarchicalPlanner | Task decomposition with dependency ordering | | ExecutionPolicyEngine | Token/cost budget per iteration | | ContinuationEngine | Checkpoint every 3 iters, recovery on error | | ReflexionEngine | Per-iteration self-reflection & insight injection | | WorldStateCollector | Tracks file changes across iterations | | QAPipeline | Quick + thorough quality checks post-execution | | Governor | Safety/security scanning before tool execution | | AutoFix | Automatic lint/type error repair | | ContextManager | Token budget management & compaction | | TaskClassifier | Classifies task type for targeted system prompt injection | | StrategySelector | Selects up to 3 execution strategies per task | | SkillLoader | Loads & scores built-in and user skills | | DAGOrchestrator | Parallel subagent execution with dependency graph |

Skill System

YUAN bundles 35 skills out of the box — 6 core skills and 29 language/domain skills.

Core Skills

| Skill | Trigger | Description | |-------|---------|-------------| | debug | auto | Reproduce → Trace → Fix → Verify | | test-driven | auto | Red → Green → Refactor | | code-review | auto | CRITICAL / HIGH / MEDIUM / LOW severity | | security-scan | auto | OWASP Top 10 pattern detection | | refactor | auto | Impact radius + minimal change | | plan | auto | Decompose → Sequence → Risk |

Language Skills (29 languages)

| Group | Languages | |-------|-----------| | Web Frontend | TypeScript, JavaScript, React, Vue, Svelte | | Web Backend | Python, Ruby, PHP, Java, Kotlin, Go, Elixir | | Systems | Rust, C, C++, Haskell | | Mobile | Swift, Dart (Flutter), Kotlin | | Scripting | Bash, Lua, R | | Data Science | Python, R, SQL | | Database | SQL | | DevOps | Docker, Terraform, Bash | | Game Dev | GDScript (Godot), C++ | | Blockchain | Solidity | | Embedded/GPU | CUDA, Verilog | | Functional | Haskell, Elixir |

Skills are stored in dist/skills/ and loaded automatically. User skills: ~/.yuan/skills/. Plugin skills: node_modules/@yuaone/plugin-*.

import { SkillLoader } from "@yuaone/core";

const loader = new SkillLoader();
const skills = await loader.loadAll();
const matched = loader.score(skills, { filePaths: ["src/auth.ts"] });

Execution Strategies

The StrategySelector picks up to 3 strategies per task and injects them into the system prompt. Strategies propagate to parallel subagents automatically.

| Strategy | When applied | |----------|-------------| | Read Before Write | Always | | Test-Driven | feature, debug, refactor, test | | Trace First | debug, security, performance | | Impact Radius | refactor, migration (DEEP/SUPERPOWER modes) | | Pattern Match First | feature, test, documentation | | Minimal Change | debug, config, infra (FAST/NORMAL modes) | | Verify Before Done | feature, refactor, migration, test, deploy |

Language Registry

All 67 supported languages are defined in a single SSOT (language-registry.ts). Language detection, system prompt verification, and skill file linking all derive from this registry.

import { LANGUAGE_REGISTRY, getLanguageByExtension } from "@yuaone/core";

const lang = getLanguageByExtension(".ts"); // { id: "typescript", name: "TypeScript", ... }
console.log(LANGUAGE_REGISTRY.length); // 67

Events

type AgentEvent =
  | { kind: "agent:thinking"; content: string }
  | { kind: "agent:token"; content: string }
  | { kind: "agent:tool_call"; toolName: string; arguments: unknown }
  | { kind: "agent:tool_result"; toolName: string; result: string; durationMs: number }
  | { kind: "agent:approval_needed"; action: { tool: string; input: unknown } }
  | { kind: "agent:completed"; message: string }
  | { kind: "agent:error"; message: string; retryable: boolean }
  | { kind: "agent:token_usage"; input: number; output: number }
  | { kind: "agent:qa_result"; stage: "quick" | "thorough"; passed: boolean; issues: string[] }

License

AGPL-3.0 — part of YUAN.