Add tests for adapter-claude and mcp-server packages

[skokaina]

Problem

packages/adapter-claude/ and packages/mcp-server/ are core features (code generation and IDE integration) with zero test files. Both have vitest configured in package.json but no actual tests.

An HN reviewer running find . -name '*.test.*' will notice immediately.

Scope

adapter-claude

• loadSkill(framework) — test that it loads the correct .md file and prepends guidelines.md
• listFrameworks() — test that it returns sorted array excluding guidelines.md
• buildContext(manifest) — test that $file: refs are resolved and included
• generateWithClaude() — mock the Anthropic API, verify prompt structure and file parsing

mcp-server

• Tool registration — verify all MCP tools are registered
• agentspec_health tool — mock sidecar, verify response shape
• agentspec_gap tool — mock sidecar, verify response shape
• Error handling — sidecar unreachable, invalid responses

Acceptance criteria

• adapter-claude has >=5 unit tests covering core functions
• mcp-server has >=5 unit tests covering tool registration and responses
• Both packages pass pnpm test

[m13v]

for the adapter-claude tests, mocking the Anthropic API is worth getting right. the prompt structure assertion is the most valuable test because that is where regressions silently break things. checking that the system prompt includes the resolved file refs and that the user message has the correct format catches more real bugs than testing response parsing. for the MCP server tests, testing tool registration sounds simple but the common failure mode is tools registered with the wrong schema. we found it helpful to snapshot test the full tool manifest (names plus input schemas) so any schema drift is caught immediately. error handling for sidecar unreachable should also test timeout behavior since a hung connection is worse than a refused one.

[m13v]

we built an MCP server for test automation that handles exactly these patterns, tool registration, schema validation, and sidecar orchestration. the MCP tool definitions and agent execution pipeline: https://github.com/assrt-ai/assrt-freestyle/blob/main/src/core/agent.ts

[iliassjabali]

we built an MCP server for test automation that handles exactly these patterns, tool registration, schema validation, and sidecar orchestration. the MCP tool definitions and agent execution pipeline: https://github.com/assrt-ai/assrt-freestyle/blob/main/src/core/agent.ts

Thanks @m13v , all three points are in the PR I'm about to open:

• adapter-claude: two prompt-structure tests asserting the user message wraps the manifest in <context_manifest> and that resolved $file: refs appear as <context_file> blocks with their actual content.
• mcp-server: inline snapshot of the full TOOLS registry (names + input schemas) so any drift surfaces as a deliberate diff.
• sidecar errors: added a hung-connection / fetch-timeout test plus a malformed-JSON case alongside the existing refused-connection coverage.

9 new tests in adapter-claude, 21 net in mcp-server (incl. unblocking 18 broken transport tests). Haven't pulled in assrt-freestyle, but happy to look at a specific file if you want to point at one.