acp-forge

This Skill addresses the lack of reproducibility, delayed conflict detection, and weak skill composition capabilities in multi-agent collaboration. On platforms like Grok Build, multiple sub-agents work in parallel but cannot replay sessions, detect file conflicts in advance, or orchestrate complex skill pipelines. ACP Forge fills these critical gaps by introducing a content-addressable execution DAG, pre-merge conflict detection for working trees, and composable skill pipelines.

Usage is straightforward: simply clone the repository and run cargo build to compile, then initialize a session via the CLI, specify a base branch, and let multiple sub-agents execute in parallel within isolated git working trees. You can then load skills in TOML and SKILL.md formats using cargo run -- skill list, compose pipelines with cargo run -- skill render, and verify session integrity with cargo run -- session v. It also integrates seamlessly with Grok Build's ACP agent pipeline—just add an endpoint in the configuration file.

It is ideal for teams or individuals requiring multi-agent coordination, especially those already using Grok Build for divide-and-conquer development but seeking stronger observability and preemptive conflict handling. It also suits scenarios needing reproducible experimental workflows and templated skill reuse, such as CI/CD pipelines, automated code reviews, and multimodal data validation. For projects prioritizing auditability and team collaboration transparency, this is a powerful addition.

It is recommended to integrate it as an external ACP agent in Grok Build projects to maximize its orchestration capabilities. Note that it primarily relies on Rust 2024 edition and ACP SDK 0.14, currently only supports integration with Grok Build's ACP agent, and may require additional adaptation for other platforms. Additionally, it defaults to a maximum of 8 parallel sub-agents, aligning with Grok Build's native limit; resource consumption should be evaluated for large-scale scenarios.