starlight · vibeguard v1.1

Stop the slop.
Before the commit.

Native rules + real-time hooks + static guards to catch what AI coding agents get wrong — before it reaches your codebase. Duplicate files, invented APIs, dangerous shell commands, unverified "I'm done" claims. Blocked at the source.

$ bash ~/vibeguard/setup.sh
how it works →
109+
native rules
7
constraint layers
6
hooks active
2
agents supported
why vibeguard

Prompt guidelines don't stop bad actions.

without vibeguard

  • Agent creates utils_v2.py next to utils.py — duplicate files pile up.
  • Agent invents user.fullName that never existed on the model.
  • Agent runs git push --force on main.
  • Agent claims "tests pass" without running them.
  • Every mistake is a lesson you pay for once per session.

with vibeguard

  • L1 search-before-create blocks duplicates at write time.
  • L4 data-integrity rejects undeclared APIs and fields.
  • pre-bash-guard suggests --force-with-lease.
  • W-03 verification gate demands fresh command output.
  • /vibeguard:learn turns each mistake into a new guard.

Mechanical enforcement

Hooks block bad actions at the source. Not prompt guidelines the model can forget.

Every error is a fix

Interceptions return OBSERVATION → FIX → WHY. The agent self-corrects in one turn.

Mistake → defense

Repeated errors become new guards via /vibeguard:learn. The system gets stricter over time.

three-layer defense

Rules bias. Hooks block. Guards scan.

layerwhat it does
Native rules109+ constraints auto-loaded into CLAUDE.md / AGENTS.md. Bias the model away from bad decisions before it acts.
Runtime hooksPreToolUse / PostToolUse intercept writes, bash, edits. Block dangerous or low-quality actions in real time.
Static guardsProject scans for duplicate files, slop patterns, structural issues. Run via /vibeguard:check.
Slash commands/vibeguard:preflight, :review, :check, :learn — workflows for each phase of the loop.
Learning loopRepeated mistakes surface as candidate rules. Reviewed, then compiled into the native rule set.
ObservabilityMetrics, logs, and health checks for every interception. You see what was blocked and why.
a real interception

Blocked with a fix, not just a no.

hook · pre-bash-guard 
┌─ VibeGuard ─────────────────────────────────────────┐
✗ blocked  pre-bash-guard

  → command:  git push --force origin main
  → rule:     L7 · commit discipline · W-force-destructive
  → reason:   destructive overwrite on protected ref

  ⚠  fix
    use git push --force-with-lease instead
    or resolve upstream locally before pushing

└─────────────────────────────────────────────────────┘
agent: retrying with --force-with-lease
7-layer constraint index

Seven rules. Injected at session start.

L1
Search before create
No similar-file pattern. Extend existing code first.
L2
Naming conventions
snake_case internally, camelCase at API boundary. No aliases.
L3
Quality baseline
No silent exception swallowing. No Any-typed public methods.
L4
Data integrity
No data = blank. No undeclared API or field exists.
L5
Minimal changes
Do what was asked. No "easy improvements" on the side.
L6
Process gates
1–2 files direct · 3–5 preflight · 6+ interview → spec.
L7
Commit discipline
No AI tags. No force push. No secrets in commits.
+
U · RS · TS · SEC · W
Language and security rule families extend the seven core layers.
get started

Install in 30 seconds.

$ git clone https://github.com/majiayu000/vibeguard.git ~/vibeguard
$ bash ~/vibeguard/setup.sh

Works with Claude Code and Codex CLI · MIT licensed · no telemetry