v0.3.0 — CLAUDE.md → BECs, new checks, agent tooling
becwright turns the deterministic parts of a CLAUDE.md into commit-blocking BECs — automatically, gradually, and with an AI agent — while judgment-based guidance stays as prose. CLAUDE.md asks an agent to behave; becwright verifies the result on every commit.
Highlights
CLAUDE.md → enforceable rules
becwright init --from-claude-mdreads yourCLAUDE.mdand turns the prohibitions it recognizes (secrets,eval,debugger,console.log, breakpoints, wildcard imports, a per-file line cap, …) into real rules, reporting which phrase matched each. It ignores what it can't verify (architecture, naming quality, function length) — that stays inCLAUDE.md.
Adopt on any codebase
becwright init --baselinestarts rules that already have violations aswarningand clean ones asblocking, so a legacy repo is guarded from day one without blocking commits on existing debt. Graduate each rule toblockingas you clean it.- Per-rule
exclude:globs carve files out of a rule (vendored/generated code, or the check's own implementation) — no more false positives forcing you to delete a useful rule.
New language-agnostic checks
require— a regex that must be present (e.g. a license header); the inverse offorbid.max_lines— a file-length cap (--max N).filename— file-name conventions (--forbid/--require).
AI agent / MCP authoring loop
The MCP server (becwright mcp) now lets an agent build rules from your CLAUDE.md, safely:
propose_rules_from_claude_md— the deterministic starting point.preview_rule— dry-run a candidate rule before writing it.add_rule— persist it behind a trust gate (previews unlessconfirm=true; built-in checks only).
The Claude Code skill/plugin teaches this whole flow.
Robustness & onboarding
- The pre-commit path checks staged content (matches exactly what the commit records).
- Checks have a timeout; a malformed
.bec/rules.yamlor a rule pointing at an unknown check now fails with a clear message instead of a confusing block. - Offline catalog:
becwright search/addinstall ready-made BECs with one command. - Plug into the pre-commit framework or Husky.
See the full changelog.