Enterprise-grade AI Agent System - Constrained by SDD, Powered by MCP for Controlled Perception and Execution
中文文档 | Quick Start | Architecture | Design Doc
FrontAgent is an AI Agent system designed specifically for frontend engineering, addressing core challenges faced when deploying agents in real-world engineering scenarios:
Distilled Planner Model: FrontAgent's Planner stage has been distilled into a standalone small model frontagent-planner-7B-lora. Load the LoRA adapter on top of Qwen2.5-Coder-7B to generate frontend execution plans directly, without calling large LLM APIs.
- ✅ Two-Stage Architecture - Separate planning and execution to avoid JSON parsing errors and enable dynamic code generation
- ✅ Phase-Based Execution - Steps grouped by phases with error recovery within each phase
- ✅ Self-Healing - Tool Error Feedback Loop automatically analyzes errors and generates fix steps
- ✅ Facts Memory - Structured facts-based context system for precise project state tracking
- ✅ Module Dependency Tracking - Automatic import/export parsing to detect path hallucinations
- ✅ Hallucination Prevention - Multi-layer hallucination detection and interception
- ✅ SDD Constraints - Specification Driven Development as hard constraints for agent behavior
- ✅ MCP Protocol - Controlled tool invocation via Model Context Protocol
- ✅ Minimal Changes - Patch-based code modifications with rollback support
- ✅ Web Awareness - Understand page structure through browser MCP
- ✅ Shell Integration - Terminal command execution (requires user approval)
- ✅ Pre-Planning Scan - Scan project structure before planning to generate accurate file paths
- ✅ Auto Port Detection - Automatically detect dev server ports from config files
- ✅ Remote Hybrid RAG - Full-repository indexing with submodule exclusion, combining BM25 keyword search and embedding-based semantic search
- ✅ LangGraph Engine (Optional) - Switchable graph-based execution engine with optional checkpoints
- ✅ Planner Skills Layer - Reusable planning skills for task decomposition and phase injection
- ✅ Skill Lab - Benchmark, improve, and promote content skills with local eval suites
- ✅ Repository Management Phase - Auto git/gh workflow after acceptance (commit, push, PR)
- ✅ Cross-Session Memory - Four-phase memory system (preload, runtime recall, post-task persistence, structured storage) that persists project facts, error resolutions, and dependency state across runs
FrontAgent now supports both terminal-first and VS Code desktop workflows:
- CLI: use
fa init,fa run, RAG commands, Skill Lab, and automation-friendly workflows directly from your terminal. - VS Code Extension: use the FrontAgent sidebar task console to run tasks, attach the current file or selection, provide a browser URL, review phase/step progress, approve sensitive actions, initialize/validate SDD, and open run logs from inside VS Code.
Install the VS Code extension from the Marketplace by searching for FrontAgent or the extension id ceilf6.frontagent.
# 1. Install globally via npm
npm install -g frontagent
# or using pnpm
pnpm add -g frontagent
# or using yarn
yarn global add frontagent
# 2. Configure LLM (supports OpenAI and Anthropic)
# OpenAI config
export PROVIDER="openai"
export BASE_URL="https://api.openai.com/v1"
export MODEL="gpt-4"
export API_KEY="sk-..."
# Or Anthropic config
export PROVIDER="anthropic"
export BASE_URL="https://api.anthropic.com"
export MODEL="claude-sonnet-4-20250514"
export API_KEY="sk-ant-..."
# 3. Navigate to your project directory and initialize SDD
cd your-project
fa init
# 4. Let AI help you complete tasks
fa run "Create a user login page"
fa run "Optimize homepage loading performance"
fa run "Add dark mode support"
# Use LangGraph engine + checkpoint (optional)
fa run "Add route guards and open a PR" --engine langgraph --langgraph-checkpointFrontAgent can run as a local stdio MCP Server for MCP hosts such as Claude Desktop, Cursor, Codex, and other clients that can launch a command-based MCP server.
MCP mode exposes FrontAgent's upper-level agent capabilities only. It does not expose raw internal tools such as read_file, apply_patch, run_command, browser tools, or rag_query directly to the external host.
# Use the installed CLI
fa mcp serve
# Or run from a source checkout after pnpm build
node /absolute/path/to/FrontAgent-app/apps/cli/dist/index.js \
mcp serveBy default, FrontAgent resolves the project root from the MCP host's workspace roots when the host exposes exactly one file root. If host roots are unavailable, it falls back to the MCP server process current working directory.
Use --project-root only when you want to pin the server to a specific project, or when the host exposes multiple workspace roots and FrontAgent cannot choose safely:
fa mcp serve --project-root /absolute/path/to/your-projectOne MCP server process is bound to one resolved project root.
Useful server options:
fa mcp serve \
--security-mode balanced \
--rag-repo https://github.com/ceilf6/Lab.git \
--rag-branch mainMost MCP hosts use the same mcpServers shape. Start with the simple config:
{
"mcpServers": {
"frontagent": {
"command": "fa",
"args": [
"mcp",
"serve"
]
}
}
}If your host UI has separate fields, use:
- Command:
fa - Args:
mcp,serve
Do not put fa mcp serve into the command field as one string.
If the host reports command "fa" not found or env: node: No such file or directory, the GUI host probably does not inherit your terminal shell PATH. Then use absolute paths:
which node
which fa{
"mcpServers": {
"frontagent": {
"command": "/opt/homebrew/bin/node",
"args": [
"/opt/homebrew/bin/fa",
"mcp",
"serve"
]
}
}
}If you are using a source checkout instead of a globally linked package, point node at the built CLI file after pnpm build:
{
"mcpServers": {
"frontagent": {
"command": "/opt/homebrew/bin/node",
"args": [
"/absolute/path/to/FrontAgent-app/apps/cli/dist/index.js",
"mcp",
"serve"
]
}
}
}For direct LLM fallback, pass environment variables through the host config:
{
"mcpServers": {
"frontagent": {
"command": "fa",
"args": [
"mcp",
"serve"
],
"env": {
"PROVIDER": "openai",
"BASE_URL": "https://api.openai.com/v1",
"MODEL": "gpt-4",
"API_KEY": "sk-..."
}
}
}
}Examples of where to put the config:
- Claude Desktop: add the server under
mcpServersinclaude_desktop_config.json. - Cursor: add the server under
mcpServersin your Cursor MCP config, for example.cursor/mcp.json. - Codex or other MCP hosts: use the same command, args, and env values in the host's MCP server configuration surface.
FrontAgent exposes six MCP tools:
frontagent_status: returns project root, SDD status, visible skills, LLM backend status, RAG status, and run-log directory.frontagent_run_task: runs a full FrontAgent task. Inputs includetask,type,files,url,sddPath, andsecurityMode.frontagent_plan_task: generates a FrontAgent execution plan without executing tools or writing files.frontagent_validate_sdd: validates the project SDD file.frontagent_list_skills: lists visible content skills.frontagent_init_sdd: creates an SDD template. Existing files are not overwritten unlessforce=true.
frontagent_run_task returns structured JSON text with:
{
"success": true,
"taskId": "task_...",
"output": "...",
"error": null,
"duration": 1234,
"runLogPath": "/absolute/path/.frontagent/runs/...",
"executedStepsSummary": [],
"securityDecisions": []
}MCP mode uses auto LLM backend selection:
- If the host supports MCP Sampling, FrontAgent asks the host model through
sampling/createMessage. - If Sampling is unsupported or unavailable, FrontAgent falls back to direct LLM configuration.
Direct fallback uses the same environment variables and flags as fa run:
export PROVIDER="openai"
export BASE_URL="https://api.openai.com/v1"
export MODEL="gpt-4"
export API_KEY="sk-..."Read-only tools such as frontagent_status, frontagent_list_skills, frontagent_validate_sdd, and frontagent_init_sdd do not require LLM configuration. frontagent_run_task and frontagent_plan_task require either host Sampling support or a valid direct LLM fallback.
MCP mode keeps FrontAgent's internal safety boundary:
- External MCP hosts cannot directly call internal file, shell, browser, or RAG tools.
- Internal file writes, shell commands, browser actions, and other side effects still go through
SecurityManager. - The default security mode is
balanced. - Because stdio MCP does not provide FrontAgent's interactive approval UI, any action that requires an
askdecision fails closed. frontagent_init_sddonly writes SDD files inside the configured project root.
FrontAgent now supports a full remote repository knowledge base flow for planning and code generation:
- It syncs the remote repository into
.frontagent/rag-cache/repo - It indexes the full repository by chunk, and automatically excludes Git submodule paths
- It runs BM25 keyword retrieval and embedding-based semantic retrieval in parallel
- It applies metadata filters to each candidate list, then fuses the ranked results
- Built indexes and embedding vectors are cached under
.frontagent/rag-cache
Default knowledge source:
- Repository:
https://github.com/ceilf6/Lab.git
CLI options:
fa run "Explain React setState behavior" \
--provider openai \
--base-url https://yunwu.ai/v1 \
--api-key YOUR_TOKEN \
--rag-repo https://github.com/ceilf6/Lab.git \
--rag-branch main \
--rag-keyword-candidates 40 \
--rag-semantic-candidates 40 \
--rag-keyword-weight 0.45 \
--rag-semantic-weight 0.55
# When provider=openai, RAG embeddings inherit the same base-url/api-key by default.
# Override them only if your embedding endpoint is different.
fa run "Explain React setState behavior" \
--provider openai \
--base-url https://yunwu.ai/v1 \
--api-key YOUR_TOKEN \
--rag-embedding-model text-embedding-3-small
# Use Weaviate as the semantic vector store (BM25 stays local)
fa run "Explain React setState behavior" \
--provider openai \
--base-url https://yunwu.ai/v1 \
--api-key YOUR_TOKEN \
--rag-embedding-model text-embedding-3-small \
--rag-vector-store-provider weaviate \
--rag-weaviate-url http://127.0.0.1:8080 \
--rag-weaviate-collection-prefix FrontAgentRagChunk
# Disable LLM query rewrite before retrieval
fa run "How to build a custom selector" \
--disable-rag-query-rewrite
# Cross-encoder reranking is enabled by default after BM25 + embedding candidate retrieval
fa run "Explain React setState behavior" \
--provider openai \
--base-url https://yunwu.ai/v1 \
--api-key YOUR_TOKEN \
--rag-embedding-model text-embedding-3-small \
--rag-reranker-model jina-reranker-v2-base-multilingual \
--rag-reranker-base-url https://your-reranker-endpoint/v1
# Disable reranking for a run
fa run "Explain React setState behavior" \
--disable-rag-reranker
# Disable semantic retrieval and use BM25 only
fa run "Explain React setState behavior" \
--disable-rag-semantic
# Disable remote RAG for a run
fa run "Create a page" --disable-rag
# Force a remote git sync before this query; by default FrontAgent reuses the local cache
fa run "Explain React setState behavior" --rag-sync-on-queryFrontAgent now includes a local Skill Lab workflow for iterating on content skills under skills/.
# List visible content skills
fa skill list
# Scaffold a new content skill
fa skill scaffold pricing-audit
# Generate starter trigger evals for a skill
fa skill init-evals frontend-design
# Generate starter behavior evals (binary checks for output quality)
fa skill init-behavior-evals frontend-design
# Benchmark current trigger behavior
fa skill benchmark frontend-design
# Benchmark trigger + behavior together
fa skill benchmark frontend-design --behavior
# Generate a candidate revision and compare it against baseline
fa skill improve frontend-design
# Improve with both trigger and behavior eval suites
fa skill improve frontend-design --behavior
# Promote a candidate after review
fa skill promote frontend-design 20260331T120000ZThe current Skill Lab flow supports two eval tracks for content skills:
- Trigger evals: whether the skill activates correctly.
- Behavior evals: whether the final output quality passes binary checks.
You can run trigger-only (default) or trigger + behavior (--behavior) in benchmark/improve.
Environment variables:
export FRONTAGENT_RAG_REPO="https://github.com/ceilf6/Lab.git"
export FRONTAGENT_RAG_BRANCH="main"
export FRONTAGENT_RAG_SYNC_ON_QUERY="false"
export FRONTAGENT_RAG_MAX_RESULTS="5"
export FRONTAGENT_RAG_KEYWORD_CANDIDATES="40"
export FRONTAGENT_RAG_SEMANTIC_CANDIDATES="40"
export FRONTAGENT_RAG_KEYWORD_WEIGHT="0.45"
export FRONTAGENT_RAG_SEMANTIC_WEIGHT="0.55"
export FRONTAGENT_RAG_QUERY_REWRITE_MAX_TOKENS="160"
export FRONTAGENT_RAG_QUERY_REWRITE_TEMPERATURE="0.1"
export FRONTAGENT_RAG_RERANKER_MODEL="jina-reranker-v2-base-multilingual"
export FRONTAGENT_RAG_RERANKER_BASE_URL="https://your-reranker-endpoint/v1"
export FRONTAGENT_RAG_RERANKER_API_KEY="sk-..."
export FRONTAGENT_RAG_RERANKER_CANDIDATE_COUNT="20"
export FRONTAGENT_RAG_RERANKER_MAX_DOCUMENT_CHARS="1800"
export FRONTAGENT_RAG_EMBEDDING_MODEL="text-embedding-3-small"
export FRONTAGENT_RAG_EMBEDDING_BASE_URL="https://api.openai.com/v1"
export FRONTAGENT_RAG_EMBEDDING_API_KEY="sk-..."
export FRONTAGENT_RAG_VECTOR_STORE_PROVIDER="weaviate"
export FRONTAGENT_RAG_WEAVIATE_URL="http://127.0.0.1:8080"
export FRONTAGENT_RAG_WEAVIATE_API_KEY=""
export FRONTAGENT_RAG_WEAVIATE_COLLECTION_PREFIX="FrontAgentRagChunk"If provider=openai, and FRONTAGENT_RAG_EMBEDDING_BASE_URL / FRONTAGENT_RAG_EMBEDDING_API_KEY are not set, FrontAgent will reuse the LLM base-url and api-key automatically.
Main LLM sampling controls:
fa run "Explain React createElement" \
--temperature 0.2 \
--top-p 0.9--temperatureis supported.--top-pis supported through the AI SDK call settings.--top-kis exposed, but only some providers/models support it. For example, Anthropic models can use it, while OpenAI-compatible chat models may ignore it as unsupported.repetition_penaltyis not exposed yet in FrontAgent because the current AI SDK/provider stack does not provide a stable cross-provider path for it.
Before retrieval, FrontAgent now sends the user's original request through a separate LLM rewrite step to generate a more retrieval-friendly frontend search query. This rewrite uses the same provider/base-url/model/api-key as the main agent, but the rewritten query is only used for RAG and does not replace the user's original task.
After BM25 + embedding recall, FrontAgent will by default send the top candidate chunks to a reranker endpoint (/rerank, Jina/Cohere-compatible) for cross-encoder-style final ordering when reranker model/base-url/api-key are available. Use --disable-rag-reranker to turn it off for a run.
When FRONTAGENT_RAG_VECTOR_STORE_PROVIDER=weaviate, FrontAgent keeps BM25 in the local index.json, but semantic vectors are written to and queried from Weaviate instead of embeddings.json.
Prebuilt cache bundle workflow:
- Do not commit
.frontagent/rag-cacheinto Git - Export a prebuilt cache bundle and upload it to GitHub Releases or object storage
- Other users can import the bundle locally before their first query
# Export the current cache directory as a distributable tar.gz bundle
fa rag export
# Export to a custom path
fa rag export --output ./artifacts/frontagent-rag-cache.tar.gz
# Import from a local file
fa rag import ./artifacts/frontagent-rag-cache.tar.gz --force
# Import from a remote URL
fa rag import https://example.com/frontagent-rag-cache.tar.gz --force┌─────────────────────────────────────────────────────────────────────┐
│ FrontAgent System │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ User Input │────▶│ Agent Core │────▶│ Output │ │
│ └─────────────┘ └──────┬──────┘ └─────────────┘ │
│ │ │
│ ┌─────────────────┼─────────────────┐ │
│ ▼ ▼ ▼ │
│ ┌────────────────┐ ┌────────────┐ ┌────────────────┐ │
│ │ SDD Layer │ │ Planner │ │ Executor │ │
│ │ (Constraints) │ │ (Stage 1) │ │ (Stage 2) │ │
│ └───────┬────────┘ └─────┬──────┘ └───────┬────────┘ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ MCP Layer (Trusted Interface) │ │
│ ├──────────────┬───────────────┬──────────────────────┤ │
│ │ MCP File │ MCP Web │ MCP Shell │ │
│ └──────┬───────┴───────┬───────┴──────────┬───────────┘ │
└─────────┼───────────────┼──────────────────┼────────────────────────┘
▼ ▼ ▼
┌──────────────┐ ┌──────────┐ ┌──────────┐
│ File System │ │ Browser │ │ Shell │
│ (Project) │ │(Playwright)│ │Commands │
└──────────────┘ └──────────┘ └──────────┘
User Task
│
▼
┌──────────────────┐
│ Pre-Planning │ ← Scan project structure (NEW!)
│ File Scan │ Detect dev server port (NEW!)
└────────┬─────────┘
│
▼
┌──────────────────┐
│ Planner (Stage1)│ ← SDD Constraints
└────────┬─────────┘ Project file list
│ Dev server port
│ Generates execution plan (with phase field)
▼
┌──────────────────────────────────────────┐
│ Executor (Stage 2) │
│ ┌────────────────────────────────┐ │
│ │ Phase 1: Analysis │ │
│ │ ├─ Step 1 ✓ │ │
│ │ ├─ Step 2 ✗ (error) │ │
│ │ └─ Error Recovery │ │
│ │ ├─ Analyze error │ │
│ │ ├─ Generate fix steps │ │
│ │ └─ Execute fix ✓ │ │
│ └────────────────────────────────┘ │
│ ┌────────────────────────────────┐ │
│ │ Phase 2: Creation │ │
│ │ ├─ Step 3 ✓ │ │
│ │ └─ Step 4 ✓ │ │
│ └────────────────────────────────┘ │
│ │
│ After each step: │
│ └─ Update Facts │
│ ├─ File system state │
│ ├─ Dependency state │
│ ├─ Module dependency graph │
│ └─ Project state │
│ │
│ Phase completion validation: │
│ └─ Check missing module references │
│ └─ Auto-generate fix steps ✓ │
└───────────────────────────────────────────┘
│
▼
Task Complete ✓
Default phase taxonomy:
Phase 1 Analysis -> Phase 2 Creation -> Phase 3 Installation -> Phase 4 Validation/Acceptance -> Phase 5 Startup -> Phase 6 Browser Validation -> Phase 7 Repository Management (git/gh)
Before generating the execution plan, FrontAgent now automatically scans the project structure to provide accurate file context to the LLM:
// Automatically executed before planning
const projectStructure = await scanProjectFiles();
// Returns: "Project files (245 files): src/App.tsx, src/components/Button.tsx, ..."
// LLM receives this context and generates more accurate file pathsBenefits:
- ✅ Accurate File Paths - LLM knows existing files and generates correct relative paths
- ✅ Reduced Hallucination - Fewer "file not found" errors
- ✅ Better Context - Planner understands the project structure before planning
Implementation: packages/core/src/agent.ts:217-255
FrontAgent now automatically detects the development server port from your project configuration:
// Detection sources (in order):
// 1. vite.config.ts/js: server.port field
// 2. package.json scripts: --port or -p flags
// 3. Framework defaults: Vite (5173), Next.js (3000), CRA (3000), Angular (4200)
// 4. Fallback: 5173
const devServerPort = await detectDevServerPort();
// Used in browser navigation tasksBenefits:
- ✅ Automatic Port Discovery - No manual port configuration needed
- ✅ Framework Awareness - Recognizes different framework defaults
- ✅ Browser Testing - Correct port used for browser validation tasks
Implementation: packages/core/src/agent.ts:732-793
FrontAgent uses an innovative two-stage architecture that completely solves JSON parsing errors when generating large amounts of code:
- Input: User task + SDD constraints + project context + project file list (NEW!)
- Output: Structured execution plan (descriptions only, no code)
- Tech: Uses
generateObjectto produce Zod Schema-compliant JSON - Key: No code in JSON, avoiding escape and parsing issues
{
"summary": "Create login page",
"steps": [
{
"description": "Create Login.tsx component file",
"action": "create_file",
"params": {
"path": "src/pages/Login.tsx",
"codeDescription": "Create a React component with username, password inputs and login button"
},
"needsCodeGeneration": true
}
]
}- Input: Structured execution plan
- Process: Execute each step in the plan sequentially
- Code Generation: When encountering
needsCodeGeneration: true, dynamically generate code usinggenerateText - Tech: Use MCP tools for file operations, command execution, etc.
Advantages:
- ✅ Completely avoid JSON parsing errors (code not in JSON)
- ✅ Better controllability (each step validated individually)
- ✅ Support large projects (no JSON size limit)
- ✅ More precise code generation (based on real-time context)
FrontAgent implements advanced phase-based execution and automatic error recovery:
The execution plan is automatically divided into multiple phases, each focused on a specific goal:
{
"steps": [
{
"stepId": "step-1",
"phase": "Analysis Phase",
"description": "Read existing files, analyze project structure",
"action": "read_file"
},
{
"stepId": "step-2",
"phase": "Creation Phase",
"description": "Create new component files",
"action": "create_file"
},
{
"stepId": "step-3",
"phase": "Installation Phase",
"description": "Install necessary dependencies",
"action": "run_command"
},
{
"stepId": "step-4",
"phase": "Validation Phase",
"description": "Run tests to verify functionality",
"action": "run_command"
},
{
"stepId": "step-5",
"phase": "Repository Management Phase",
"description": "Commit changes, push branch, and create/update PR with gh",
"action": "run_command"
}
]
}Advantages:
- 🎯 Clear Execution Flow - Each phase has a clear objective
- 🔄 Intra-Phase Error Recovery - Errors automatically fixed within phases
- 📊 Better Progress Tracking - Users see which phase is currently executing
- 🔀 Dependency-Aware Phase Ordering - Phase DAG scheduling reduces out-of-order skips
- 🚀 Post-Acceptance Automation - Optional repository management phase can handle git/gh flow
When tool execution fails, the system automatically analyzes errors and generates fix steps:
// 1. Error detected
Error: Cannot apply patch: file not found in context: src/App.tsx
// 2. LLM analyzes error
{
"canRecover": true,
"analysis": "File src/App.tsx not read into context, need to read it first",
"recoverySteps": [
{
"description": "Read src/App.tsx into context",
"action": "read_file",
"tool": "filesystem",
"params": { "path": "src/App.tsx" }
},
{
"description": "Reapply patch to src/App.tsx",
"action": "apply_patch",
"tool": "filesystem",
"params": { /* original params */ }
}
]
}
// 3. Auto-execute fix steps
// 4. Continue original flowFeatures:
- 🔍 Smart Error Analysis - LLM understands error causes and finds root issues
- 🛠️ Auto-Generate Fixes - No manual intervention needed
- 📝 Common Error Patterns - Built-in handling for common errors
- ♻️ Phase-Level Recovery - Errors fixed within phases without blocking overall flow
FrontAgent now supports a switchable execution engine:
native(default): Existing executor flow with phase DAG schedulinglanggraph: Runs phase flow throughStateGraphwith optionalMemorySavercheckpoint
CLI options:
# Use native engine (default)
fa run "Add login page" --engine native
# Use LangGraph engine
fa run "Add login page" --engine langgraph
# LangGraph + checkpoint + custom recovery attempts
fa run "Add login page" --engine langgraph --langgraph-checkpoint --max-recovery-attempts 5FrontAgent adds a dedicated skills layer in Planner to encapsulate reusable planning logic.
- Built-in task skills:
task.create,task.modify,task.query,task.debug,task.refactor,task.test - Built-in phase skill:
phase.repository-management(injects git/gh workflow after acceptance) - Custom task skills with the same match condition override built-ins (latest registered wins)
- Executor also supports action-level skills (for params/codegen/error policy)
- Supports runtime extension via API:
import { createAgent, type TaskPlanningSkill } from '@frontagent/core';
const agent = createAgent(config);
const customSkill: TaskPlanningSkill = {
name: 'task.security-audit',
supports: (task) => task.type === 'debug' && task.description.includes('security'),
plan: ({ stepFactory }) => [
stepFactory.createStep({
description: 'Scan for security-sensitive patterns',
action: 'search_code',
tool: 'search_code',
params: {
pattern: 'eval|innerHTML|dangerouslySetInnerHTML',
filePattern: 'src/**/*.{ts,tsx,js,jsx}',
},
}),
],
};
agent.registerTaskSkill(customSkill);
console.log(agent.getPlannerSkillSnapshot());
agent.registerExecutorActionSkill({
name: 'action.run-command.noncritical-policy',
action: 'run_command',
shouldSkipToolError: ({ errorMsg, params }) => {
if (typeof params.command === 'string' && params.command.includes('echo')) {
return true;
}
return errorMsg.includes('already exists');
},
});
console.log(agent.getExecutorSkillSnapshot());Traditional agents use logs as context, leading to information redundancy and inaccuracy. FrontAgent uses a structured "facts" system:
Traditional approach (log-based):
Executed operation log:
1. Attempted to read src/App.tsx - failed
2. Attempted to create src/components/Button.tsx - success
3. Attempted to read src/App.tsx - success
4. Installed react-router-dom - success
...(lots of redundant info)
FrontAgent approach (facts-based):
## File System State
### Confirmed Existing Files:
- src/App.tsx
- src/components/Button.tsx
- package.json
### Confirmed Non-Existent Paths:
- src/pages/Login.tsx
## Dependency State
### Installed Packages:
react-router-dom, axios
### Missing Packages:
@types/node
## Created Modules
### component (3 modules):
- src/components/ui/Button.tsx (default export: Button)
- src/components/ui/Card.tsx (default export: Card)
- src/components/layout/Header.tsx (exports: Header, Navigation)
### page (2 modules):
- src/pages/HomePage.tsx (default export: HomePage)
- src/pages/LoginPage.tsx (default export: LoginPage)
### ⚠️ Missing Module References:
- src/pages/HomePage.tsx references non-existent module: ../components/ui/Spinner
## Project State
- Dev server: Running (port: 5173) ← Auto-detected!
- Build status: Success
## Recent Errors
- [apply_patch] Cannot apply patch: file not found in contextAdvantages:
- 📊 Structured Information - Clear state categories (filesystem, dependencies, module graph, project state)
- 🎯 Deduplication - Automatic deduplication using Set/Map
- 💡 Context Awareness - LLM knows which files exist/don't exist
- 🔄 Real-time Updates - Auto-update facts after each tool execution
- 📉 Reduced Token Usage - Concise information reduces LLM input length
- 🔗 Module Tracking - Auto-parse import/export relationships for each created file
FrontAgent now implements a four-phase memory architecture that persists knowledge across task runs, so the agent no longer starts from scratch every time.
The memory system treats context as a time-phased pipeline:
- Phase 1 -- Startup Preload: Load durable memories and seed
ProjectFactsfrom the last snapshot before planning begins - Phase 2 -- Runtime Recall: Dynamically recall relevant memories during code generation based on file path, tags, and keywords
- Phase 3 -- Post-Task Persistence: Extract and persist durable learnings (created files, error resolutions, dependency changes) after each task completes
- Phase 4 -- Compaction: Deferred until multi-turn interactive mode is added
<projectRoot>/.frontagent/memory/
MEMORY.md # Index entrypoint (concise topic list)
topics/
project-structure.md # Filesystem layout, key modules
dependencies.md # Packages, versions, known issues
errors.md # Past error resolutions
snapshots/
facts-latest.json # Last ProjectFacts snapshot
All memory files are human-readable Markdown (inspectable, editable). Facts snapshots use JSON for efficiency.
The system prompt is now structured into three independent zones, each with its own token budget:
- Rules zone -- SDD constraints, behavioral instructions (immutable per task)
- Memory zone -- Durable project knowledge loaded from
.frontagent/memory/ - Context zone -- Dynamic per-task data (files, RAG results, skills, facts)
- Single-writer pattern: All writes go through
MemoryStoreto prevent conflicts - Dedup tracking: Per-session
injectedKeysset prevents re-injecting the same memory - Budget enforcement: Configurable character limits for preload (default 8000) and per-step recall (default 2000)
- Non-blocking persistence: Memory writes run off the critical path with full error swallowing
- Backward compatible: Fresh projects with no prior memory run identically to previous behavior
const agent = createAgent({
// ...other config
memory: {
enabled: true, // default: true
preloadBudgetChars: 8000, // max chars injected at startup
recallBudgetChars: 2000, // max chars per code-gen recall
maxTopicFiles: 10, // max topic files loaded at startup
},
});Implementation: packages/core/src/memory/
Specification Driven Development (SDD) as hard constraints for agent behavior:
# sdd.yaml
version: "1.0"
project:
name: "my-project"
type: "react-spa"
tech_stack:
framework: "react"
version: "^18.0.0"
language: "typescript"
forbidden_packages:
- "jquery"
- "lodash"
code_quality:
max_function_lines: 50
max_file_lines: 300
forbidden_patterns:
- "any"
- "// @ts-ignore"
modification_rules:
protected_files:
- "package.json"
require_approval:
- pattern: "src/api/*"
reason: "API layer changes require approval"Provides file operation MCP tools:
read_file- Read file contentlist_directory- List directory content (supports recursion)create_file- Create new file (two-stage: generate code from description)apply_patch- Apply code patches (two-stage: generate changes from description)search_code- Search codeget_ast- Get AST analysisrollback- Rollback changes
Provides terminal command execution (requires user approval):
run_command- Execute shell commands- Custom working directory support
- Timeout settings
- User approval required before execution
- Auto-distinguish warnings from errors
- Use cases:
npm install,git init,pnpm build, etc.
Provides browser interaction MCP tools:
browser_navigate- Navigate to URLget_page_structure- Get page DOM structureget_accessibility_tree- Get accessibility treeget_interactive_elements- Get interactive elementsbrowser_click/browser_type/browser_scroll- Page interactionsbrowser_screenshot- Page screenshotbrowser_wait_for_selector- Wait for element availability
Multi-layer hallucination detection:
- File Existence Check - Verify referenced files exist
- Import Validity Check - Verify imports are resolvable
- Syntax Validity Check - Verify code syntax is correct
- SDD Compliance Check - Verify compliance with SDD constraints
- Language: TypeScript
- Runtime: Node.js 20+
- Package Manager: pnpm
- MCP SDK: @modelcontextprotocol/sdk
- Browser Automation: Playwright
- AST Analysis: ts-morph
- LLM Integration: Vercel AI SDK
frontagent/
├── packages/
│ ├── shared/ # Shared types and utilities
│ ├── sdd/ # SDD control layer
│ ├── mcp-file/ # File operations MCP client
│ ├── mcp-web/ # Web awareness MCP client
│ ├── mcp-shell/ # Shell commands MCP client
│ ├── hallucination-guard/ # Hallucination prevention
│ └── core/ # Agent core (two-stage architecture)
│ └── memory/ # Cross-session memory system
├── apps/
│ └── cli/ # CLI tool
├── examples/
│ ├── sdd-example.yaml # SDD config example
│ └── e-commerce-frontend/ # E-commerce frontend example
└── docs/
├── architecture.md # Architecture design
└── design.md # Original requirements
cd examples
fa run "Create an e-commerce frontend project using React + TypeScript + Vite + Tailwind CSS"Agent will automatically:
- Analyze project requirements
- Generate execution plan
- Create package.json and config files
- Request to execute
npm install(requires user approval) - Generate page components and style files
fa run "Modify vite.config.ts to add path alias configuration"Agent will:
- Read existing vite.config.ts
- Understand current configuration
- Generate new config code
- Apply minimal patches
fa run "Add user authentication feature, including login, registration, and token management"Agent will:
- Analyze existing project structure
- Plan files to create
- Generate auth-related components
- Create API integration code
- Update route configuration
fa run "Analyze and optimize homepage loading performance"Agent will:
- Read relevant component code
- Analyze performance issues
- Propose optimization solutions
- Implement code-level optimizations (lazy loading, code splitting, etc.)
fa run "Add route configuration in App.tsx"Execution process shows self-healing:
Phase 1: Analysis Phase
✓ Step 1: Read package.json
Phase 2: Creation Phase
✗ Step 2: Modify App.tsx
Error: Cannot apply patch: file not found in context
🔄 Error recovery in progress...
Analysis: App.tsx not read into context
✓ Recovery Step 1: Read src/App.tsx into context
✓ Recovery Step 2: Reapply patch to App.tsx
Phase 3: Validation Phase
✓ Step 3: Run type check
✅ Task complete! Auto-fixed 1 error
Key Features:
- 🎯 Phase-Based Execution - Clear execution phases (Analysis, Creation, Validation)
- 🔄 Auto-Fix - Detected file not read, auto-insert read step
- 📊 Facts Tracking - System knows which files are read/unread
- ⚡ No Retry Needed - One-shot completion, no manual re-runs needed
fa run "Implement user profile page and open PR" \
--type create \
--engine langgraph \
--langgraph-checkpoint \
--max-recovery-attempts 5Notes:
--engine langgraphenables graph-based phase orchestration--langgraph-checkpointenables in-memory checkpointing for the run- After acceptance passes, repository management phase can run
git/ghactions
| Variable | Description | Example Value |
|---|---|---|
PROVIDER |
LLM provider | openai or anthropic |
API_KEY |
API key | sk-... |
MODEL |
Model name | gpt-4 or claude-sonnet-4-20250514 |
BASE_URL |
API endpoint | https://api.openai.com/v1 |
EXECUTION_ENGINE |
Execution engine | native or langgraph |
LANGGRAPH_CHECKPOINT |
Enable LangGraph checkpoint | true / false |
MAX_RECOVERY_ATTEMPTS |
Max recovery attempts per phase | 3 |
export PROVIDER="openai"
export BASE_URL="https://api.openai.com/v1"
export MODEL="gpt-4"
export API_KEY="sk-..."export PROVIDER="anthropic"
export BASE_URL="https://api.anthropic.com"
export MODEL="claude-sonnet-4-20250514"
export API_KEY="sk-ant-..."# Development mode
pnpm dev
# Type check
pnpm typecheck
# Build
pnpm build
# Clean
pnpm clean- Two-stage agent architecture (Planner + Executor)
- Phase-based execution
- Tool Error Feedback Loop (self-healing)
- Facts-based context system
- Module dependency graph
- Post-generation validation
- Path hallucination detection
- Multi-LLM provider support (OpenAI, Anthropic)
- Shell command execution (with user approval)
- Dynamic code generation (avoid JSON parsing errors)
- MCP tool integration (File, Web, Shell)
- Type auto-normalization (handle LLM output uncertainty)
- Unlimited steps (support complex tasks with many steps)
- LLM schema constraint optimization (multi-strategy auto-fix, smart retry)
- Pre-planning file scan (NEW!)
- Auto dev server port detection (NEW!)
- Dependency-aware phase DAG scheduling (NEW!)
- LangGraph execution engine (optional) (NEW!)
- Repository management phase (git/gh automation) (NEW!)
- Cross-session memory system (NEW!) -- Four-phase durable memory with structured Markdown storage, runtime recall, and prompt zone separation
- Distilled Planner Model -- SFT fine-tuned from FrontAgent Planner prompts, published as frontagent-planner-7B-lora (Qwen2.5-Coder-7B + LoRA, 100% JSON validity, 100% complete plan rate)
- Enhanced SDD constraints (finer-grained rule control)
- Memory-driven pattern learning (auto-extract coding conventions from past tasks)
- GUI agent auto-testing (Playwright-based)
- VS Code plugin (use directly in IDE)
- Multi-agent collaboration (decompose large tasks)
- Custom MCP server support (user-defined tools)
- Code review mode (auto-check code quality)
- Incremental update mode (only modify necessary parts)
Welcome to contribute! Submit issues, bugs, or suggestions:
- Fork the repository
- Create a new branch
git checkout -b feature/amazing-feature - Commit your changes
git commit -m 'Add amazing feature' - Push to the branch
git push origin feature/amazing-feature - Open a Pull Request
MIT
Contributions are welcome! Please feel free to submit issues, bug reports, or suggestions.
- Fork this repository
- Create a feature branch (
git checkout -b feature/amazing-feature) - Commit your changes (
git commit -m 'Add amazing feature') - Push to the branch (
git push origin feature/amazing-feature) - Open a Pull Request
MIT