slopbrick

slopbrick

AI agents forget your architecture. Every session starts fresh.

SlopBrick gives your codebase persistent structure — so agents follow your patterns instead of reinventing them.

The fix is one command: npx slopbrick scan writes .slopbrick/{inventory.json, constitution.json, health.json, structure.md}. The next time your AI agent writes a file — Claude Code, Cursor, Copilot, Aider — it reads .slopbrick/structure.md instead of re-parsing the AST. 100–1000× faster on the agent integration, and the agent's first suggestion matches what the project already uses, not what the LLM trained on.

npm install -D slopbrick
npx slopbrick init        # write .slopbrick/constitution.json
npx slopbrick scan        # write .slopbrick/structure.md
npx slopbrick mcp         # start the MCP server (Claude / Cursor)

For the prevention layer:

slopbrick watch           # re-run scan on every file change
slopbrick install         # install the Git pre-commit hook
slopbrick ci              # CI gate: exit 1 on constitution violation

This isn't CLAUDE.md. CLAUDE.md is a static file the agent reads once per session. .slopbrick/structure.md is a generated artifact that updates on every scan — your repository, encoded for the next agent.

What you get

• Repository Structure — the four .slopbrick/ artifacts (structure, inventory, constitution, health) make your codebase queryable by any AI agent in O(read file) instead of O(parse AST).
• LockBrick prevention — slopbrick watch flags violations as you write, slopbrick install blocks AI-introduced slop at pre-commit, slopbrick ci enforces the same in CI.
• Constitution — declare your canonical stack (state lib, form lib, modal system, API client) once. The agent and the linter enforce it together.

Status: v0.43.0 (current) (current). Engine: 103 rules in 22 categories, 4 headline scores, 9+ language frontends, v10-calibrated against 576,750 real files. See CHANGELOG for the full release notes.

Quick start

# 1. Install
npm install -D slopbrick

# 2. Initialize (8 quick questions about your stack)
npx slopbrick init

# 3. Scan (writes .slopbrick/ artifacts)
npx slopbrick scan

# 4. Optional: start the MCP server so Claude Code / Cursor can
# consume the artifacts
npx slopbrick mcp

That's it. The agent integration is O(read file) for the next session.

For a CI gate, see EXAMPLES.md. For monorepo setup, see EXAMPLES.md. For every other config question, see EXAMPLES.md.

The headlines (4-score model, v0.21.0+)

v0.15.0 introduced the 4-score model. v0.21.0 FLIPPED aiSlopScore to the natural-reading "raw amount of slop" direction (0 = clean, 100 = saturated, lower = cleaner). v0.18.0 added the 4th score (security). The other three scores stay "higher = better". The legacy slopIndex field is kept as optional on ProjectReport for backward compat with existing test fixtures; the v0.14-compat removal is tracked separately.

Score-band messages (v0.21.0+): every score ships with a one-line verdict in the pretty output — e.g. AI Slop Score: 25 → "low amount of AI slop", Security Risk: low. The band mapping for aiSlopScore is 0–9 no slop, 10–29 low, 30–49 medium, 50–69 high, 70–100 saturated (v0.21.0 lower-is-better direction). See src/report/pretty.ts.

The same numbers live in .slopbrick/health.json.

For the full math, the 4-score quadrant, and which one to focus on, see docs/scoring-explained.md.

For per-rule precision/recall/FPR (auditable), see src/rules/signal-strength.json.

Telemetry (opt-in)

Starting in v0.24.0, slopbrick can send a single one-shot usage ping after slopbrick scan completes. This is opt-in — the default is OFF — and is intended for self-hosted CI telemetry.

How to opt in

# 1. Set the endpoint env var
export SLOPBRICK_TELEMETRY_ENDPOINT="https://your-host.example/ingest"

# 2. Pass the flag on the CLI
slopbrick scan --report-usage

Both conditions are required. If either is missing, no request is sent, no warning is printed, and exit code is unaffected.

What is sent

A single POST with Content-Type: application/json and exactly 8 fields:

Privacy promise

The payload is frozen at exactly 8 fields. We will never send:

• file paths, file hashes, or file contents
• rule ids, rule violations, or rule categories
• user identifiers, IP addresses, or environment variables
• timestamps other than what process.version provides indirectly

Failure mode

The beacon is fire-and-forget with a 5-second socket timeout. Network errors, DNS failures, 4xx/5xx responses, and timeouts are all silent — slopbrick scan exit code is never affected. The request is also unidirectional: no retries, no follow-up calls.

Scope

Only slopbrick scan fires the beacon. slopbrick watch, slopbrick ci, and programmatic scanProject calls are unaffected regardless of the flag or env var.

Local flywheel

This is separate from the local flywheel. The local flywheel writes detailed scan results to .slopbrick/flywheel/scans.jsonl and is gated by --no-telemetry (default ON, opt-out per-run or via config.telemetry = false). The new beacon is gated by --report-usage + SLOPBRICK_TELEMETRY_ENDPOINT (default OFF).

See docs/research/beacon-design.md for the full design doc, threat model, and OPSEC requirements for the receiver.

Example output

$ npx slopbrick scan --brief
[v0.43.0] auto-suppressed 184 INVERTED/NOISY issue(s) from 18 default-off rule(s).
Memory persisted to .slopbrick/ (0 patterns, 0 components, 537 bytes of structure.md).

Repo is low (25/100). The biggest problem is AI patterns — worst file is packages/slopbrick/src/engine/parser-rust.ts.

  AI Slop Score         25   low  (aiSlopScore)
  Engineering Hygiene  100   excellent  (engineeringHygiene)
  Security             100   excellent  (security)
  Repository Health     57   needs work  (repositoryHealth)

  CI gate: AI Slop Score <= 15 -> fail

  Scanned 593 files, 346 issues. Run with --all for the full report.
1 threshold failed: meanSlop (score 25 > 15)

--brief (CI/scripts): same headline + threshold + delta in 4 lines. --why-failing: top 5 rules ranked by weighted impact. --suggest: per-rule auto-fix advice. --human-only / --ai-only: filter issues by category.

Documentation

The 19 subcommands are auto-generated from commander and run slopbrick --help to see them.

Installation

npm install -D slopbrick

Requires Node 20+ (verified by slopbrick doctor). The package ships ESM + CJS dual builds, TypeScript types, and is published to npm as slopbrick.

For the MCP server, add to your AI agent's config:

{
  "mcpServers": {
    "slopbrick": { "command": "npx", "args": ["slopbrick", "mcp"] }
  }
}

See docs/MCP.md for Cursor, Continue, and other clients.

Contributing

See CONTRIBUTING.md — tl;dr: copy src/rules/visual/naturalness-anomaly.ts, edit the analyze() body, add a test in tests/rules/, then add a defaultOff: true entry to src/rules/signal-strength.json. v0.14.5k's calibration will validate the rule on the next corpus run.

We follow the Contributor Covenant.

License

MIT © 2026 Brick.dev