Agent Receipts Gate
Deterministic pre-send linting to catch agent hallucinations, missing receipts, and filler chatter.
$9
Works with the AI Agents you already use
Agent Receipts Gate
Example session with this skill installed
Analyze this draft: 'I have finished the code review for the auth module. Everything looks good, consider it handled.' Context: { 'hasTools': true, 'lastSentAgoSec': 300 }
- Read your context and instructions
- Compiled the agent receipts gate
The agent reply was analyzed. The gate flagged a fabrication because the claim of a completed review lacked a receipt (file path, PR #, or link) and used a premature closure phrase.
Validation Results
| Metric | Score | Status | Details |
|---|---|---|---|
| Grounding | 10/100 | FAIL | No file path, PR #, or URL provided for 'auth module'. |
| Action | 20/100 | FAIL | Claimed 'handled' without specific tool output. |
| Honesty | 90/100 | PASS | No false tool denials detected. |
| Noise | 100/100 | PASS | Not a duplicate; sent 300s after last message. |
| Verdict | REPAIR | 2 | Block sending until evidence is added. |
Suggested Repair
"I have finished the code review for src/auth/middleware.ts in PR #42. The changes meet our security standards."
Next steps
- Append the specific file path or PR number to the draft.
- Rerun the gate to confirm the
sendaction (Exit Code 0). - Check the
reasonfield in the JSON output to identify other missing receipts.
Connects securely to your tools. The creator never sees your data.
About this skill
Agent communication often suffers from "lazy hallucinations" where models claim a task is complete without performing the action or deny access to tools they actually possess. This skill provides a deterministic, pattern-based linting gate to validate agent replies before they reach the end user. It enforces the inclusion of "receipts"—tangible evidence like file paths, commit SHAs, or URLs—while suppressing redundant duplicates and filler chatter.
What it does
- Evidence verification identifies claims of completion or review that lack specific identifiers like URLs, file paths, or PR numbers.
- Action validation flags instances where an agent claims a task is "handled" before any tool execution has occurred.
- Honesty checking detects "self-denials" where an agent incorrectly claims it cannot perform a task despite having the necessary tools.
- Noise suppression blocks duplicate messages sent within short windows or messages containing only filler chatter.
- Reply scoring generates a 0-100 metric across grounding, action, honesty, voice, noise, and closure.
How it works
- Pass the draft reply along with JSON context containing tool availability and message history.
- Evaluate the action code returned by the gate:
send,repair, orsuppress. - Parse the verdict to identify specific failure modes like
fabricationorpromised-done. - Automate the response by using exit codes (0, 2, or 3) to trigger pipeline blocks or rewrite logic.
Frameworks & tools
This is a deterministic, offline Node.js script (gate.cjs) that uses regex pattern matching rather than LLM calls for low-latency validation.
Why this beats prompting it yourself
Prompting an LLM to check its own honesty often results in recursive hallucinations. This skill uses hard-coded rules and pattern matching to provide an objective, offline second opinion that doesn't add to your token costs.
Use cases
- CI/CD bots that must provide a commit hash or log path when claiming a build passed.
- Support agents to prevent "I'll look into that" filler when no ticket was actually opened.
- Coding assistants to ensure file modifications are accompanied by specific line numbers or file names.
Known limitations
The tool is pattern-based and supports English only. It may flag benign conversational sentences as fabrications if they lack hard links, and it currently lacks a comprehensive list of all external websites.
Pattern-based, English only. May flag benign claims lacking hard links as fabrications. Does not recognize all external URLs. Requires node environment.
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
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed today
Requires