Summary

Phase B of the AI-agent integration: gives agents machine-readable access to becwright, on top of the Claude Code plugin (#17). Two additions, no impact on the minimal-deps core.

becwright check --json

A JSON summary instead of colored text — consumable without parsing. Exit code unchanged (1 if a blocking rule failed). Shares logic with a new report.py module (also used by the MCP server). No extra dependency; works from the standalone npm binary too.

{ "rule_count": 3, "checked_files": 1, "blocked": true,
  "results": [ { "id": "no-debugger-js", "severity": "blocking", "passed": false, "intent": "...", "why_it_matters": "...", "output": "app.js:1\n      > ..." } ] }

becwright mcp — universal MCP server

A stdio MCP server exposing check and list_checks to any MCP-capable agent (Claude, Cursor, Windsurf). Lazily imported and shipped behind the optional mcp extra, so the core stays pyyaml-only; a missing extra prints a helpful install hint.

Client config:

{ "mcpServers": { "becwright": { "command": "becwright", "args": ["mcp"] } } }

Notes

• git.repo_root() gains an optional cwd so MCP tools can target a given path.
• The MCP server ships only with the Python package (not the npm binary); CLI/hook use without Python keeps working via npm.

Test plan

• 121 tests pass; coverage 91% (threshold 80%)
• check --json verified (blocking + clean repos), valid JSON, correct exit codes
• MCP tools registered (check, list_checks); check returns correct payload; list_checks returns all built-ins (tests use importorskip("mcp"); mcp added to the dev extra so CI runs them)
• Missing-extra path returns exit 2 with install hint
• claude plugin validate still passes after the skill note
• Dogfood becwright check passes