adapter-msaf

Microsoft Agent Framework adapter for Weave Protocol

The @weave_protocol/adapter-msaf package enforces WARD.md policies inside Microsoft Agent Framework (MSAF) via its middleware pipeline. Unlike the Claude Code and Antigravity adapters which use config-file hooks, this adapter is a library that the user registers programmatically in their agent setup.

When to use

• User is building an MSAF agent and wants policy enforcement at the tool-call layer
• User wants the same WARD.md they use elsewhere (Claude Code, Antigravity, MCP) to apply to their MSAF agent
• User wants typed exceptions (WardDeniedError) to catch and handle WARD denials in their code
• User wants per-agent enforcement (not global, since one process may run multiple agents with different policies)

Architecture difference from claudecode / antigravity adapters

MSAF middleware is registered programmatically, not via a config file. So:

• No init writes to disk; instead it prints the integration snippet
• Middleware is per-agent, not global
• Runs in-process (no subprocess invocation overhead)

The user must paste a small code snippet into their MSAF agent setup. The CLI helps with init (snippet generator), status (verify policy resolution), and test (dry-run a call).

Primary API

import { WardMiddleware } from '@weave_protocol/adapter-msaf';

const ward = new WardMiddleware({
  // wardPath: './WARD.md'  // or auto-resolve from cwd
  // failMode: 'closed'      // or 'open' (default)
  // toolMappings: { ... }   // custom tool name → capability
  // onAllow: (call, result) => { ... }
  // onDeny:  (call, result) => { ... }
});

agent.useFunctionMiddleware(ward.functionMiddleware());

WARD resolution order

1. wardPath option (explicit)
2. wardSource option (raw markdown)
3. $WEAVE_WARD_PATH
4. <cwd>/WARD.md
5. <cwd>/.weave/WARD.md
6. <cwd>/.msaf/WARD.md

CLI

weave-msaf init [--language=ts|csharp|python]    # print integration snippet
weave-msaf status                                 # show active WARD + mappings
weave-msaf test <tool> [--input=JSON]             # dry-run a tool call