test: cross-workspace contract tests + root orchestrator (Phase 3)

[JeremyDev87]

Parent

E2E test infrastructure initiative.

Problem

The three workspaces (MCP server, Plugin, Rules) have inter-dependencies that are only validated by convention:

• Plugin → MCP: Plugin calls `activate`, `search_rules`, `parse_mode` — but no test verifies the MCP server actually returns what the plugin expects.
• Rules → MCP: Agent JSON files in `packages/rules/.ai-rules/agents/` must match the schema the MCP server uses to parse them — currently only ajv validates the schema, not the runtime behavior.
• Root orchestration: There is no single command to run all E2E tests across workspaces.

A breaking change in the MCP server's tool response format would silently break the plugin with no test catching it.

Scope

Contract tests (~5-8 tests)

MCP ↔ Plugin contract

• `activate` tool returns `{mode, rules, primaryAgent, specialists}` structure
• `search_rules` returns `{results: [{name, content, relevance}]}` structure
• `get_agent_details` returns agent with required fields (name, description, expertise)
• Tool names match what plugin code references (no typos in tool name strings)

Rules ↔ MCP contract

• Every agent JSON in `packages/rules/.ai-rules/agents/*.json` loads successfully via MCP server's RulesService
• Rule markdown files referenced by `activate` actually exist on disk
• Agent count matches between filesystem and MCP `resources/list`

Root orchestrator

• Add `test:e2e` to root `package.json` that runs:
  1. `yarn workspace codingbuddy test:e2e` (MCP E2E — Phase 1)
  2. `python3 -m pytest tests/e2e/plugin-hooks/ -v` (Plugin E2E — existing)
  3. `yarn workspace landing-page test:e2e` (Landing E2E — Phase 2)
  4. Contract tests (this phase)

Technical approach

• Location: `tests/e2e/contracts/` at repo root (cross-workspace)
• Framework: Vitest (can import from both MCP server and rules package)
• MCP interaction: Spawn MCP server in stdio mode, send JSON-RPC calls
• CI: Add `e2e-contracts` job to `dev.yml`, depends on `build`

Test commands

# Individual
yarn test:e2e:contracts     # Contract tests only

# Root orchestrator
yarn test:e2e               # ALL e2e across workspaces

Acceptance criteria

• `tests/e2e/contracts/` with ≥5 contract tests
• MCP ↔ Plugin response shape validated
• Rules ↔ MCP agent count parity validated
• Root `test:e2e` script orchestrates all workspace E2E
• CI job for contract tests
• Contract tests fail fast on response shape mismatch

Dependencies

• Phase 1 (MCP E2E) — reuses the stdio spawn infrastructure
• Phase 2 (Landing E2E) — needed for root orchestrator completeness

Out of scope

• Plugin Python ↔ MCP server live integration (requires running both)
• CDC (Consumer-Driven Contracts) tooling like Pact — overkill for current scale