待翻譯：LetItLoop: Resume crashed AI agent loops with 0% token waste (<1ms resume)

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Notifications You must be signed in to change notification settings Fork 3 Star 4 BranchesTags Open more actions menu Latest commit History 205 Commits 205 Commits Folders and files NameName Last commit message Last commit date .github .github assets assets conformance conformance deploy deploy docs docs examples examples letitloop-action letitloop-action letitloop letitloop orchestrator orchestrator recipes recipes schemas schemas scripts scripts skill skill templates templates tests tests .editorconfig .editorconfig .env.example .env.example .gitattributes .gitattributes .gitignore .gitignore .pre-commit-config.yaml .pre-commit-config.yaml AGENTS.md AGENTS.md CHANGELOG.md CHANGELOG.md CLAUDE.md CLAUDE.md CODE_OF_CONDUCT.md CODE_OF_CONDUCT.md CONTRIBUTING.md CONTRIBUTING.md Dockerfile Dockerfile LICENSE LICENSE MANIFEST.in MANIFEST.in OPENCODE.md OPENCODE.md README.md README.md SECURITY.md SECURITY.md docker-compose.yml docker-compose.yml fast_test_runner.py fast_test_runner.py pyproject.toml pyproject.toml pytest.ini pytest.ini requirements-ci.txt requirements-ci.txt uv.lock uv.lock Repository files navigation Autonomous macro-task orchestration & verification control loop for AI coding agents: 2ms WAL crash durability, source-span AST self-evolution, and deterministic proof gates. Official Website • DCP-2.0 Benchmark • GitHub Action • Quickstart • Cookbook The LetItLoop Tripartite Ecosystem LetItLoop solves the central failure mode of autonomous AI coding agents: the lack of deterministic verification, uncatchable mid-task SIGKILL crashes, and destructive whole-file rewrites. graph TD subgraph "The Tripartite Ecosystem" LL["letitloop (Core Engine) Deterministic WAL plumbing, AST node splicer and FastSandbox"] LLA["letitloop-action (Marketplace) Drop-in CI gate signing proof bundles on Pull Requests"] ADB["agent-durability-bench (DCP-1.0) Open benchmark measuring agent recovery under SIGKILL faults"] end LL -.->|"bridges to"| ADB LL -.->|"scaffolds"| LLA Loading letitloop: The core engine providing sub-2ms Write-Ahead Logging (WAL) state journals, source-span AST node splicing (0% comment loss), in-memory Zero-Copy fast sandboxing, and deterministic verification gates. letitloop-action: Standalone GitHub Action for CI that validates AI pull requests, enforces strict AST signatures, and posts machine-verifiable proof bundles directly to PR comments. agent-durability-bench: An open benchmark suite implementing Durability Challenge Protocol 1.0 (DCP-1.0) with zero-API synthetic simulation to measure how well agents recover from uncatchable SIGKILL crashes. Key Capabilities Source-Span AST Node Splicer: Replaces targeted functions and class methods with surgical precision. 0% Comment Loss: Guarantees module docstrings, file comments, licensing headers, and class indentation are never stripped or altered. In-Memory Fast Sandbox: Zero-Copy sys.modules evaluation and Windows Job Object containment that verifies code hypotheses in-memory before writing anything to disk. Fault-Tolerant WAL Supervisor Loop: State journal with WAL (Write-Ahead Logging), crash recovery, atomic Win32/POSIX file locking, and bounded 3-strike retries with strategy mutation. Cognitive Feasibility Gate & Multi-Source Research: Deliberates whether a refactor is safe to perform autonomously or requires background research across arXiv, GitHub, and DuckDuckGo. Human-in-the-Loop Proposal Ledger: Automatically stages deferred, high-risk architectural proposals as structured markdown artifacts (PROP-*.md) for human review rather than executing unverified mutations. Zero-Trust Verification Engine: Deterministic acceptance check kinds (AST syntax parsers, command exit-code assertions, regex matchers, file validators, size bounds, and undeclared output detectors). 12 Pluggable Worker Adapters: Native interfaces for Claude Code, OpenAI Codex, Google Antigravity (agy), OpenCode, Hermes Agent, Cline, Aider, Docker Sandboxes, Local LLMs (Ollama/vLLM), Omniroute gateways, local scripts, and direct LLMs. Native Model Context Protocol (MCP) Server & Client: 8 stdio JSON-RPC tools connecting directly with Claude Code, OpenAI Codex, Cursor, Google Antigravity, and Hermes Agent. Cross-Platform Process Orphan Guard: Windows Job Objects (win32job) and POSIX session process-group containment ensuring complete cleanup of child/grandchild processes. Prometheus Observability & Signed Webhooks: Native Prometheus metrics exporter, lifecycle event bus, SSE streaming, and HMAC-SHA256 signed webhook dispatcher. Quickstart 1. The @durable Python Decorator (Drop-in Durability) Embed LetItLoop's crash-resilient WAL kernel directly into your Python functions, LangGraph nodes, or CrewAI agents: from letitloop import durable, step, atomic_marker @durable(goal_id="customer_sync") def main(): # If your script crashes or gets SIGKILLed midway, # completed steps are skipped on resume. Zero duplicate tokens wasted. user = step("fetch_user", fetch_crm_record, user_id=123) summary = step("llm_summarize", call_claude, user) # Guard external mutations against duplicate execution with atomic_marker("slack_notification") as should_execute: if should_execute: step("post_slack", notify_team, summary) if name == "main": main() ⚠️ Resume Semantics Notice: LetItLoop provides zero-server, single-file WAL durability. Completed steps are never re-executed (0% duplicate token waste on finished work). In-flight steps re-execute at-least-once. Design steps to be idempotent or protect external API mutations using LetItLoop's atomic_marker primitive. 2. Installation # Install letitloop core engine pip install letitloop # Or install with development and conformance tooling pip install "letitloop[dev]" 3. Basic CLI Usage # Execute DCP-2.0 durability conformance matrix lil bench --matrix # Fast pre-push code repair and AST check lil heal --target orchestrator/state.py # Inspect supervisor status, WAL journal, and active checkpoints lil status # Scaffold a production GitHub Action PR verification workflow lil action --init Supported Worker Adapters & Gateways Worker Adapter Identifier Description Tier Google Antigravity CLI antigravity-cli Invokes the official agy agent runner safely Tier-1 (Core) Claude Code CLI claude-code Autonomous task execution via Claude Code CLI Tier-1 (Core) OpenAI Codex CLI codex Autonomous task execution via OpenAI Codex CLI Tier-1 (Core) Mock Worker mock Deterministic simulation worker for CI and offline tests Tier-1 (Core) OpenCode CLI opencode Autonomous execution via OpenCode agent CLI Tier-2 (Contrib) Hermes Agent CLI hermes Autonomous execution via Nous Research Hermes agent CLI Tier-2 (Contrib) Cline CLI cline Headless execution via Cline autonomous coding runner Tier-2 (Contrib) Aider Pair Programmer aider Pair programming execution via Aider CLI Tier-2 (Contrib) Docker Sandbox Worker docker Isolated execution inside container runtime with workspace scoping Tier-2 (Contrib) Local LLM Tool Caller local-tool Local tool-calling model adapter for offline Ollama/vLLM loops Tier-2 (Contrib) Omniroute Gateway omniroute Multi-model fallback routing through local/remote gateways Tier-2 (Contrib) Script Worker script Executes local shell/Python automation scripts with env isolation Tier-2 (Contrib) Direct LLM APIs direct In-process calls to Gemini, OpenAI, Anthropic, DeepSeek, or Ollama Tier-2 (Contrib) Model Context Protocol (MCP) Integration LetItLoop runs natively as an MCP server providing 8 JSON-RPC tools to AI agent runners: { "mcpServers": { "letitloop": { "command": "letitloop-mcp", "env": { "WORKER_MODEL": "gemini:gemini-3.7-flash", "QC_MODEL": "gemini:gemini-3.1-pro" } } } } Add to Claude Code claude mcp add letitloop -- python -m orchestrator.mcp_server Recipes & Cookbooks The recipes/ cookbook provides end-to-end, schema-validated walkthroughs: Recipe Focus 01 - Legacy Codebase Refactor Refactor under pytest + ruff acceptance gates with scope fencing and bounded retries 02 - FastAPI CRUD Generator Feature decomposition into a 4-contract DAG chained with depends_on 03 - Offline Local LLM Loop Zero-cloud-key runs via Ollama (local-tool) and the docker sandbox adapter 04 - Multi-Agent QC Audit Multi-lens quality plane: panels, arbitration, budgets, and quality_spec Validate all embedded example contracts anytime with pytest tests/test_recipes.py -q. Living Architecture Decision Records (ADRs) Following the Michael Nygard ADR convention, all core design invariants and architectural decisions are codified: ADR Focus Status ADR-0001 Write-Ahead Logging (WAL) & Zero-State Recovery accepted ADR-0002 Deterministic AST, Regex & Exit-Code Verification Gates accepted ADR-0003 Zero-API-Key Headless Agent CLI Wrapper Failovers accepted ADR-0004 Format-Aware Acceptance Check & Markdown Injection accepted Security & Sandboxing Architecture letitloop operates under a zero-trust execution model: Redaction Firewall: Automatic masking of PATs, OAuth keys, AWS credentials, GCP tokens, and PEM private keys. Environment Scrubbing: Sensitive parent environment variables are stripped prior to worker execution. Scope Checking: Userland filesystem snapshot diffing (scope.py) enforcing directory bounds and declared output paths. Process Isolation: Process tree containment with Windows Job Objects (JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE) and POSIX session leadership. License Distributed under the MIT License. See LICENSE for more details. Topics Resources Readme MIT license Code of conduct Code of conduct Contributing Contributing Security policy Security policy Activity Stars 4 stars Watchers 1 watching Forks 3 forks Report repository