Repository navigation
NeckBeard v0.1.0
NeckBeard v0.1.0 — Principal-Engineer Mode for AI Coding Agents
Smallest change that is correct, safe at the boundary it crosses, fast and scalable at a named load, simple for the end user, and finished.
Built on Ponytail's 6-rung minimalist question ladder, with security, performance at a named load, and deep-module user simplicity baked in from line one.
Highlights
- Unified Portable Ruleset (
skills/neckbeard/SKILL.md&AGENTS.md):- 4 Pre-Write Framing Questions: Blast radius & user impact, trust boundary, load as a number (not 100x it), and how necessary complexity is hidden inside the module.
- 6-Rung Question Ladder:
Lean is this ladder, not a fourth gate. - Three Built-In Gates:
- Safe: Edge parsing, fail-closed errors, server-side object authorization (
Authorize the object on the server, every time. Logged in is not allowed.), parameterized SQL, shell argument arrays, path canonicalization + symlink-escape rejection, outbound host allow-lists, platform crypto, and untrusted LLM/tool outputs. - Fast and scalable at the named number: Input limits, idempotent retries, no check-then-write races, no query-per-item (N+1), no unbounded in-memory lists, no quadratic work, no blocking work on the request/UI thread, and no unindexed filters at the named load.
- Done: Fix shared code once + all broken callers/tests, single source of truth per state, backward-compatible interfaces, deep modules that hide complexity behind safe defaults, clear actionable user failure outcomes (never raw stack traces or secrets), and 1 hostile-input test on risky checks.
- Safe: Edge parsing, fail-closed errors, server-side object authorization (
- Four Focused Skills:
/neckbeard [lite|full|paranoid|off]— session coding rules and level switching (/neckbeard,$neckbeard,@neckbeard)./neckbeard-review [--ci]— change review with concrete failure proof, repo-wide variant search on Must fix findings, and--ciJSON gate output ({"verdict":"pass|fail","mustFix":N})./neckbeard-audit— whole-repo audit starting at entry points, checking the three gates, broken doc/agent-instruction promises, least-privilege permissions, supply chain, and offering an opt-inSPEC.mdor Maestro playbook remediation plan./neckbeard-debt— scans tracked and untracked files forrisk:,shortcut:, andponytail:comments and flagsillegal-deferralon trust checks or data-loss writes.
PreToolUseGuard (hooks/guard.js):full(default): Blocks hardcoded secrets (private keys, AWS, GitHub, Slack, Stripe live, Google, Anthropic, OpenAI, npm tokens) and blocks direct agent tampering with immutable VIBES Protocol.ai-audit/provenance and evidence artifacts.paranoid: Also blockscurl … | sh, disabled TLS verification,chmod 777,--no-verify, and literal credentials.
- Multi-Host Adapters: Out-of-the-box manifests for Claude Code, Codex, GitHub Copilot CLI & Editor, Antigravity CLI & IDE, Gemini CLI, Grok Build, and any
AGENTS.md-compatible host. - VIBES Protocol Integration: Zero-config support for
VIBES v1.1,VERIFY v1.0,PRISM v1.0,EVOLVE v1.0, andTRACE v1.0when.ai-audit/is present. - Zero Dependencies: Pure Node.js (
>= 18), 17 unit tests vianode:test.