Trade-Audit — Mandatory Audit Gate for Trading & Transfer Decisions

When to use

This skill is designed for auditing financial decisions — buy, sell, swap, transfer, liquidity pool entry/exit, or any on-chain value movement. The user may ask you to run it before executing a trade or transfer, or to always use it as a pre-check for financial actions.

What this skill does

Takes agent-prepared decision material and sends it to the Apus deterministic inference API running on an NVIDIA H100 TEE. Returns a structured, hardware-attested decision packet with:

• - Bundle Hash — SHA-256 of the normalized decision material
• INLINECODE1 — SHA-256 of the model's structured decision packet
• INLINECODE2 — hardware attestation for that specific run
• INLINECODE3 — APPROVE / REJECT / WAIT
• INLINECODE4 — 1-100 integer, gated by --min-confidence (default 60)

Every run is logged to ~/.trade-audit/audit.jsonl.

No wallet or API key required. This skill only reads public data and calls the Apus inference API. It does not execute any transactions.

Important boundary:

The script is at {baseDir}/analyze.py.

• - The agent collects the page contents, address information, pool details, rules, and relevant facts.
• The agent organizes that material into either a text/markdown file or a JSON decision bundle.
• This script does not fetch pages or explorer data itself.
• Reuse the bundled templates when preparing inputs:

Step 1 — Prepare the decision material

The audit model (gemma-3-27b-it) performs best with concise, focused inputs. The agent MUST distill raw data into core decision points before submitting.

Data preparation rules:

• - Extract only: prices, thresholds, numeric values, rules/conditions, addresses, risk factors
• Strip out: page chrome, disclaimers, marketing text, navigation, repeated boilerplate
• Keep material under 4,000 characters when possible (warning at 4k, hard truncation at 12k)
• Each fact should be one short bullet — no paragraphs
• If a page has 50 data points, pick the 5-10 that directly affect the decision

Create one of these:

1. 1. A text or markdown file containing the organized facts.
2. A JSON bundle containing the organized facts plus decision_goal.

For example, a prepared text file might contain:

CODEBLOCK0

Common data sources (no auth required)

When preparing decision material, prefer public APIs over scraping JS-rendered pages.

Polymarket

Use the CLOB API to get market data — no wallet or login needed:

CODEBLOCK1

Key fields to extract: question, tokens[].outcome (YES/NO), tokens[].price, end_date_iso, description (resolution rules).

Crypto prices

CODEBLOCK2

On-chain data

CODEBLOCK3

The agent should fetch data from these APIs, extract the core numbers, and organize them into the decision material template. Do not pass raw API responses directly — distill to key facts first.

Step 2 — Run the audit

No external dependencies required — the script uses only Python stdlib. Just run with python3:

Standard mode (always returns exit 0 on success)

CODEBLOCK4

Gate mode (exit code reflects verdict)

CODEBLOCK5

Exit codes in gate mode:

• - 0 = APPROVE — proceed with the action
• INLINECODE19 = REJECT — do NOT proceed
• INLINECODE20 = WAIT — insufficient information, gather more data first

If the agent already prepared a JSON bundle:

CODEBLOCK6

Print the full terminal output to the user exactly as it appears.

After the audit completes, the agent MUST include these TEE attestation fields in the response to the user:

CODEBLOCK7

These fields prove the decision was made inside a Trusted Execution Environment. Omitting them defeats the purpose of using this skill. Always show them alongside the verdict summary.

Step 3 — Integration pattern for other skills

Any skill or agent workflow that involves a trade or transfer must use this pattern:

CODEBLOCK8

The --min-confidence flag (default 60) auto-rejects any APPROVE verdict below the threshold. Set higher for high-value transactions.

Step 4 — Agent preparation templates

For --input-file, use this structure:

CODEBLOCK9

Use the bundled file for a copyable version:

INLINECODE23

For --bundle-file, use:

INLINECODE25

Step 5 — Audit log

Every run automatically appends a record to ~/.trade-audit/audit.jsonl. Each line is a JSON object:

CODEBLOCK10

Step 6 — Explain the attestation

After the report, add this note:

To reproduce the decision exactly, rerun the skill on the same saved bundle with the same decision goal. If the bundle is identical, the Output Hash should match. The TEE Nonce changes on each run because it is bound to that specific execution.