dyoshikawa/agent-loop-flow

최근 기록된 소스 활동: 2026년 3월 20일 00:55

SkillsMP 카탈로그 업데이트: 2026년 11월 5일 19:57

수집된 skills: 1

GitHub 스타: 1
GitHub 포크: 0

AI 어시스턴트로 설치

이 프롬프트를 복사해 사용 중인 AI 어시스턴트에 입력하세요.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/dyoshikawa/agent-loop-flow.

저장소 README

README 원본 파일: README.md

README 원본 보기

README 전체 펼치기

# agent-loop-flow [![npm version](https://img.shields.io/npm/v/agent-loop-flow.svg)](https://www.npmjs.com/package/agent-loop-flow) [![license](https://img.shields.io/npm/l/agent-loop-flow.svg)](https://github.com/dyoshikawa-claw/agent-loop-flow/blob/main/LICENSE) AI coding agent utility CLI that orchestrates prompt flows with transitions and loops, defined in JSONC files. ## Overview Agent-loop-flow lets you compose agent workflows by chaining prompts: ``` prompt A -> prompt B -> prompt C ... ``` Define flows in JSONC files with support for: - **Sequential execution** -- run prompts in order - **Flat transition rules** -- route execution via `next` rules on steps (no nesting for if/else) - **Loops** -- while-loops and for-each iteration - **JSON Schema validation** -- validated flow definitions with IDE autocompletion ## Quick Start ### Install ```bash npm install -g agent-loop-flow ``` Or use with npx: ```bash npx agent-loop-flow run my-flow.jsonc ``` For use as a library: ```bash npm install agent-loop-flow ``` ### Define a Flow Create a `.jsonc` file (e.g., `my-flow.jsonc`): ```jsonc { "$schema": "./flow-schema.json", "name": "my-flow", "description": "Analyze and fix code", "defaultTool": "opencode", "defaultModel": "github-copilot/claude-opus-4.6", "variables": { "targetFile": "src/index.ts", "hasIssues": true, }, "steps": [ { "type": "skill", "name": "analyze", "skill": "code-analysis", "prompt": "Analyze {{targetFile}} for issues", "next": [{ "condition": "hasIssues", "step": "fix" }, { "step": "done" }], }, { "type": "skill", "name": "fix", "skill": "code-fix", "prompt": "Fix issues found in {{targetFile}}:\n{{previousResult.output}}", }, { "type": "skill", "name": "done", "skill": "reporter", "prompt": "Report that {{targetFile}} is clean", }, ], } ``` ### Run a Flow ```bash # Run a flow agent-loop-flow run my-flow.jsonc # Run with variables agent-loop-flow run my-flow.jsonc --var targetFile=src/app.ts # Validate without executing agent-loop-flow validate my-flow.jsonc ``` ## Flow Definition ### Step Types #### Skill Step Executes a single skill with a prompt. Supports `{{variable}}` interpolation and optional `next` transition rules. ```jsonc { "type": "skill", "name": "analyze-code", "skill": "code-analysis", "prompt": "Analyze {{targetFile}}", "next": "verify-step", // unconditional jump } ``` #### Transition Rules (`next`) The `next` field controls which step runs after the current one. It can be: - A **string** for an unconditional jump: `"next": "step-name"` - An **array of rules** for conditional branching: ```jsonc { "type": "skill", "name": "check", "skill": "checker", "prompt": "Check code", "next": [ { "condition": "hasIssues", "step": "fix-issues" }, { "step": "report-clean" }, // default (else) branch ], } ``` Supported conditions: - Variable truthiness: `"variableName"` - Equality: `"variable == \"value\""` - Inequality: `"variable != \"value\""` - Last result: `"lastResult.success"` - Output contains: `"lastResult.output contains \"error\""` If no `next` is specified (or no rule matches and there is no default), the engine falls through to the next step in array order. #### While Loop Repeats steps while a condition is true. ```jsonc { "type": "while-loop", "name": "retry-loop", "condition": "shouldRetry", "maxIterations": 5, "exitCondition": "lastResult.success", // break when tests pass "steps": [ // steps to repeat ], } ``` The `condition` is checked before each iteration (loop stops when false). The optional `exitCondition` is checked after each iteration body completes -- when it evaluates to true the loop breaks immediately. `maxIterations` defaults to 100 if omitted. #### For-Each Loop Iterates over items in a variable. ```jsonc { "type": "for-each", "name": "process-files", "items": "fileList", "as": "currentFile", "maxIterations": 10, "exitCondition": "lastResult.success", // stop on first success "steps": [ { "type": "skill", "name": "process", "skill": "processor", "prompt": "Process {{currentFile}}", }, ], } ``` The optional `maxIterations` caps the number of items processed. The optional `exitCondition` is checked after each iteration body completes -- when it evaluates to true the loop breaks immediately. ### Variables Variables can be defined at the flow level and overridden via CLI: ```jsonc { "variables": { "targetFile": "src/index.ts", "mode": "strict", }, } ``` ```bash agent-loop-flow run flow.jsonc --var targetFile=src/app.ts --var mode=lenient ``` #### Previous Result Variables Each skill step can reference the previous step's output using `{{previousResult.output}}` or the flat camelCase form `{{previousResultOutput}}`. All fields from the previous result are available: ```jsonc { "steps": [ { "type": "skill", "name": "analyze", "skill": "code-analysis", "prompt": "Analyze {{targetFile}}", }, { "type": "skill", "name": "fix", "skill": "code-fix", "prompt": "Fix the issues:\n{{previousResult.output}}", }, ], } ``` Available variables: `previousResult.output`, `previousResult.success`, `previousResult.stepName`, `previousResult.skill`, `previousResult.error` (and their flat camelCase equivalents like `previousResultOutput`). ## Programmatic API ```typescript import { parseFlowFile, createFlowEngine } from "agent-loop-flow"; import type { SkillExecutor } from "agent-loop-flow"; // Define how skills are executed const skillExecutor: SkillExecutor = async ({ skill, prompt, variables }) => { // Integrate with OpenCode SDK, Claude Agent SDK, etc. return { output: "result", success: true }; }; // Parse and run a flow const flow = await parseFlowFile({ filePath: "my-flow.jsonc" }); const engine = createFlowEngine({ skillExecutor }); const result = await engine.executeFlow({ flow }); console.log(result.success); console.log(result.results); ``` ## Examples See the `examples/` directory for sample flow definitions: - `simple-sequential.jsonc` -- Basic prompt chain - `conditional-fix.jsonc` -- Conditional branching via next rules - `loop-processing.jsonc` -- For-each and while loops ## Development ```bash pnpm install pnpm check # fmt + lint + typecheck pnpm test # run tests pnpm build # build for distribution ``` ## License MIT

저장소 README를 불러오는 중…

이 저장소의 skills

소프트웨어 개발자

직업 카테고리 1개 · 100% 분류됨

skill

직업 분류

설명

업데이트

rulesync

소프트웨어 개발자

Rulesync CLI tool documentation - unified AI rule file management for various AI coding tools

원문 언어: 영어

2026-03-20

수집된 skill 1개 중 1개를 표시합니다.