Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/eqtylab/cupcake/pr-reviewgit clone --depth 1 https://github.com/eqtylab/cupcakeWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.05605 | $0.05605 |
| Opus 5 | $0.02802 | $0.02802 |
| Sonnet 5 | $0.01121 | $0.01121 |
| Haiku 4.5 | $0.00560 | $0.00560 |
Grade C, and why
cupcake pr-review.instructions.md scanned grade C with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Recursive force deletehighDestructive command
rm -rf with a variable or a broad path is one typo away from removing the wrong tree.
contains(input.tool_input.command, "rm -rf") # Spacing bypass! How it starts
The opening of the file, as written. The whole thing — 858 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Cupcake PR Review Guide
Project Overview
Cupcake is a policy enforcement and governance layer for AI coding agents. It intercepts tool calls from AI agents (Claude Code, Cursor) and evaluates them against user-defined policies written in OPA Rego, returning Allow/Block/Warn/Ask decisions before actions execute.
Core Value Proposition:
- Deterministic rule-following for AI agents
- Early-warning alerts when agents violate rules repeatedly
- Performance boost by moving rules out of context into guaranteed enforcement layer
Architecture Pattern: Hybrid Model
- Rego (WASM): Declares policies, evaluates rules, returns decision verbs
- Rust (Engine): Routes events, gathers signals, synthesizes decisions
High-Level Architecture
Core Flow
AI Agent → Hook Event → Cupcake Engine → Policy Evaluation → Decision → AI Agent
↓
1. Preprocess (normalize input)
2. Route (O(1) metadata lookup for signal gating)
3. Early Exit if no policies match
4. Gather Signals (shell commands)
5. Evaluate (ALL policies via WASM)
6. Synthesize (apply priority)
7. Format Response
Key Components
Engine (cupcake-core/src/engine/)
- Scanner: Discovers policy files
- Metadata Parser: Extracts routing requirements
- Router: O(1) signal gating and early exit (does NOT control which Rego rules execute)
- Compiler: Converts Rego → WASM (single entrypoint, all policies)
- Evaluator: Executes WASM in sandbox (all compiled policies run via
walk()) - Synthesis: Applies decision priority hierarchy
Preprocessing (cupcake-core/src/preprocessing/)
- Whitespace normalization (defense against spacing bypasses)
- Symlink resolution (defense against path traversal)
- Script inspection (defense against hidden commands in scripts)
- Automatic, always-on defense-in-depth
Harnesses (cupcake-core/src/harness/)
- Claude Code: PreToolUse, PostToolUse, UserPromptSubmit events
- Cursor: beforeShellExecution, beforeFileEdit, etc.
- Native event formats (no normalization layer)
- Harness-specific response builders
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 858 lines · 5,605 tokens per session scan C 7e27dc00920f
cupcake pr-review.instructions.md is an instructions file published in the GitHub repository eqtylab/cupcake (288 stars, last pushed 6mo ago), licensed Apache-2.0. It adds 5,605 tokens to every session, about $0.0280 per session on Opus 5. A static security scan graded it C with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other instructions, from other repositories
kagi-skills AGENTS.md
Instructions for joelazar/kagi-skills: Multi-module monorepo. Each skill lives in its own directory with its own go.mod.
agent-hook-schemas CLAUDE.md
Claude Code instructions for mherod/agent-hook-schemas, covering claude.md, build & test commands, npm publish, architecture and key patterns.
claude-codex-settings CLAUDE.md
Claude Code instructions for fcakyon/claude-codex-settings, covering claude-settings, repo structure, cross-tool plugin format reference, plugin manifests and root marketplace files.
awesome-reviewers CLAUDE.md
Instructions for baz-scm/awesome-reviewers, covering awesome reviewers — repository guidelines, what this project is, source of truth, site layer and machine interface.
coral CLAUDE.md
Instructions for cdknorow/coral, covering claude.md - coral go, mission, testing, go unit tests and legacy parity tools (historical reference).
cli-continues parsers.instructions.md
Instructions for yigitkonur/cli-continues, covering parser review guidelines, crash safety (critical), required exports, jsonl streaming and tool summarizer.