- Home
- Skills
- Agents & Orchestration
- ⚖ Agent Referee
⚖ Agent Referee
Repository-level governance and lifecycle management for AI agents using structured hooks, audits, and specs.
Free
Works with the AI Agents you already use
⚖ Agent Referee
Example session with this skill installed
Run an audit on the last 20 sessions to see why the router keeps missing the 'ship' skill.
- Read your context and instructions
- Compiled the agent referee
I have audited the last 20 sessions and the trigger evaluations.
Session Audit & Postmortem Report
| Metric | Value |
|---|---|
| Sessions Scanned | 20 |
| Router Precision | 85% |
| 'ship' Skill Misses | 3 |
| Most Frequent Failure | routed-skill-loaded ✗ |
Postmortem: 'ship' routing failure
What happened: The user asked to "push these changes to a new branch," but the agent performed a standard git commit without loading the ship skill.
Why: The router (.referee/routing.json) lacks a pattern for "push these changes." The request was handled by the default agent behavior instead of the structured ship workflow.
Fix (Router Layer): Added "push these changes" and "new branch" to the trigger patterns in .referee/routing.json.
Trigger Evaluation Results
node .referee/scripts/eval-triggers.mjs --live
[PASS] "full send" -> ship
[PASS] "fix this bug" -> ship
[FAIL] "push these changes" -> (none) [Expected: ship]
[FIXED] Updated .referee/routing.json with 2 new patterns.
Next steps
- Run
node .referee/scripts/eval-triggers.mjsto verify the new routing patterns. - Review
.referee/audits/session-bf23.mdfor a detailed trace of the last missed routing event. - Delete the redundant
deploy-helperskill; it has had 0 loads in 20 sessions.
Connects securely to your tools. The creator never sees your data.
About this skill
Stop guessing why your AI agents are failing or hallucinating. Agent Referee is a suite of 13 specialized skills designed to enforce guardrails, audit performance, and manage the entire development lifecycle within your repository. It moves agent governance from vague prompts into verifiable system hooks and structured knowledge.
What it does
- Enforces Guardrails via
.referee/scripts/hooks/guard.mjsto block unauthorized edits to protected paths, secrets, or main branch merges. - Audits Performance by running postmortems on agent failures, tracking skill load frequency, and identifying redundant or weak instructions.
- Standardizes Shipping with a structured "Ship" workflow that covers branching, planning, testing, and PR generation with human-readable evidence.
- Captures Knowledge in a
kb/directory to store technical decisions, task history, and lessons learned that code alone cannot express. - Validates Output using a "Verification Before Completion" protocol that requires fresh command output and test results before any claim of success.
How it works
- Onboard: Run
referee-onboardto scan the codebase and configure repository-specific protections in.referee/config.json. - Spec: Use
spec-firstto agree on test cases in plain English with the user before a single line of code is written. - Execute: The agent follows the
shiporwriting-plansskill to implement changes, strictly adhering to the "Explore before change" rule. - Review: Invoke the
review-loopto trigger an independent subagent or CLI reviewer for a second opinion on the diff. - Audit: Periodically run
method-upkeepto clean up stale skills, fix routing misses, and review session logs.
Frameworks & tools
Built for environments using Claude Code, Cursor, or custom agent harnesses. It utilizes a Node.js runtime for evaluation scripts and integrates with the GitHub CLI for PR management.
Why this beats prompting it yourself
Prompting "don't touch my .env file" fails when context gets long. This skill uses a hard-coded script hook (guard.mjs) to physically block the shell command, making safety a system property rather than a suggestion.
Use cases
- Preventing agents from editing critical infrastructure or legacy migrations.
- Automatically generating high-quality PR descriptions with evidence and agent tags.
- Reducing agent loop costs by enforcing plan-first and spec-first workflows.
- Onboarding new agents to complex repos with existing architectural decisions.
Known limitations
Requires a Node.js environment to run the referee scripts and hooks. Protections only work if the agent harness respects the provided hook scripts.
Requires Node.js runtime for script-based hooks and audits. Protective guards require the agent harness to execute hook scripts before shell/file operations.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 11 days ago
- Free with an Agensi account
Requires