Weave: A Graph Workflow Task Tracking System for AI Programming Agents

Weave is a graph-based workflow management system designed for AI coding agents. The original author/maintainer is AGM1968, source platform is GitHub, original link: https://github.com/AGM1968/weave, release date: 2026-06-04. Its core features include: replacing traditional vector retrieval with explicit graph traversal, using SQLite and tmpfs for efficient token-optimized task tracking, supporting multi-agent integration and semantic task relationship management. It aims to solve problems such as traditional task tools lacking complex relationship expression, and RAG systems struggling to capture explicit dependencies.

With the popularity of AI programming assistants like GitHub Copilot and Claude Code, task flow and context management have become new challenges. Traditional tools treat tasks as isolated entries, lacking the ability to express complex relationships; RAG systems rely on implicit vector similarity, making it difficult to capture explicit dependencies and logical connections between tasks. Weave's design philosophy: use graph structures to model tasks and workflows, express task relationships through 8 semantic edge types; focus on token efficiency, reduce token consumption in LLM interactions through the context package mechanism.

Graph Data Model

Uses SQLite + tmpfs for high-performance in-memory access. Each code repository has an independent graph database. Nodes represent entities like tasks/decisions, edges encode 8 semantic relationships (e.g., implements, blocks, depends_on, etc.).

Token-Optimized Retrieval

Obtains relevant context through explicit graph traversal, combines BM25 + cosine similarity with RRF fusion for local hybrid search, avoiding irrelevant content from occupying the context window.

Code Search Capability

Built-in code search function uses RRF fusion to index source code locally, supports function definition and call point queries, complementing graph node search.

• Repository Initialization: wv init-repo generates configuration, creates the .weave directory, and sets up Git hooks to associate task submissions.
• Task Lifecycle: wv add (create), wv work (claim), wv done (complete) manage tasks, including unique Weave ID and learning point records.
• Task Hierarchy: wv link creates semantic edges, wv tree visualizes the hierarchy, supporting epic task decomposition.
• Context Recovery: wv bootstrap --json provides session snapshots, replacing git status + wv status to get work status.

• Claude Code: Integrate via hooks and skills files; pre-commit hooks ensure code submissions are associated with tasks.
• VS Code Copilot: Provides an MCP server, injects tool definitions via .vscode/mcp.json configuration, allowing Copilot to directly query and operate tasks.
• Unified Workflow: Different agents share task status, learning points, and context; seamless tool switching without losing information.

• Learning Point Capture: When tasks are completed, use --learning to record key insights, forming an organizational knowledge base to avoid repeated problems.
• Impact Analysis: wv impact analyzes the scope of changes, calculates the explosion radius of affected nodes, risk scores, etc., helping with refactoring risk assessment.
• Semantic Query: wv query supports queries based on attributes (status, edge type, etc.), such as finding blocked tasks, core tasks, etc.

Suitable for teams and individuals with complex development workflows:

• Long-term Project Maintenance: Graph structure preserves task history associations, helping new members understand system evolution.
• Exploratory Development: Learning point capture saves experimental results.
• Team Collaboration: Unified task view reduces communication friction; wv ready views executable tasks, wv context gets complete context. For AI agents: Solves context window management problems, improves decision quality and consistency.

Weave is an evolution direction of AI-assisted programming toolchains. As a semantic layer connecting humans and AI agents, it explicitly encodes workflow relationships through graph structures, enabling agents to understand task context and dependencies. Its token-optimized design reflects a pragmatic consideration of LLM interaction costs. In the future, such workflow infrastructure will become a key component of AI-native development toolchains, managing the boundaries and protocols of human-AI collaboration.