Paseo-Agy-Acp v2.3.0: Building the Agent Adapter
Version 2.3.0 of paseo-agy-acp lands on npm with admission fencing, blank-turn safeguards, and SKILL.md command discovery.
Stock photo for illustration only, not from the actual event
- Version 2.3.0 of paseo-agy-acp released on npm to connect Paseo with Antigravity kernel.
- Introduces admission fencing to serialize prompt writes and prevent request collisions.
- Adds MCP endpoint rewriting and workspace skill discovery from SKILL.md metadata.
The maintainer of the open-source project paseo-agy-acp has published version 2.3.0 on npm. This community-driven adapter enables Paseo to utilize Google's official Antigravity ACP kernel as a fully functional multi-agent provider. Users can launch the open-source package directly via npx by supplying the required binary path in the environment variables.
While standard wire protocols handle basic connection initialization, session creation, and output streaming, generic protocol connections do not automatically define product behavior. The adapter acts as a lightweight NDJSON proxy that addresses specific product questions, leaving core responsibilities such as authentication, inference, and tool execution to the official kernel.
Stock photo for illustration only, not from the actual event
Regarding context management, Paseo supplies agent identity and working directory parameters to the provider processes. The adapter bridges this gap by injecting the operating context into official prompts whenever the PASEO_AGENT_ID variable is present, ensuring agents operate correctly with controller context. Furthermore, the release maps mode names appropriately, routing unsupported modes like plan back to the default execution path.
Protocol format discrepancies are also handled at the proxy level. The adapter rewrites Paseo's HTTP-based MCP endpoint declarations into the Server-Sent Events (SSE) format expected by the Antigravity kernel. Additionally, a blank-turn safeguard tracks visible activity during execution, returning an explicit JSON-RPC error if the kernel terminates without producing any visible output or tool interaction.
"This is a small transformation with a large behavioral impact: protocol success is not the same as a correctly contextualized agent."
Tiezbro
Building such adapters highlights the engineering effort required to integrate open-source controllers with proprietary software kernels. By deploying a thin proxy layer, developers can enforce necessary product logic, error handling, and state coordination without modifying or redistributing proprietary provider binaries, establishing a robust pattern for multi-agent ecosystem integration.
To handle concurrency safely across multiple connecting agents, the release implements an account-wide Admission fence queue. This mechanism serializes prompt writes to the provider, preventing neighboring processes from triggering empty turns, hangs, or internal protocol failures simultaneously.
The latest version also introduces a secondary integration layer for local slash-command hints derived from SKILL.md metadata, parsing frontmatter forms while maintaining workspace isolation. Version 2.3.0 was validated across 37 test files with 216 passing tests, and architecture checks confirmed full operational readiness.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment