Gate Info Riskcheck

gate-info-riskcheck

General Rules

⚠️ STOP — You MUST read and strictly follow the shared runtime rules before proceeding. Do NOT select or call any tool until all rules are read. These rules have the highest priority. → Read gate-runtime-rules.md → Also read info-news-runtime-rules.md for gate-info / gate-news-specific rules (tool degradation, report standards, security, routing degradation, per-skill version checks when scripts/ is present, and legacy wrapper routing).

• Only call MCP tools explicitly listed in this skill. Tools not documented here must NOT be called, even if they exist in the MCP server.
• Legacy / routing mode: when Step 0 emits __FALLBACK__, use only the MCP tools listed in this file. When Step 0 emits __ROUTE_CLI__, do not call those MCP tools; delegate to the mapped primary skill per Step 0.

Security guardian Skill. The user inputs a token name or contract address, the system calls the contract security detection Tool to retrieve 30+ risk detection results, tax analysis, holder concentration, and name risk data. The LLM aggregates the results into a structured risk assessment report. Address compliance checking will be added in a future phase.

Trigger Scenarios: User mentions a token/contract address + keywords like safe, risk, check, audit, honeypot, rug, contract security, scam.

Per-skill updates: This directory includes scripts/update-skill.sh and may include scripts/update-skill.ps1. ClawHub packages uploaded to the marketplace often omit update-skill.ps1 (upload-page restriction); GitHub / Bitbucket source trees keep both. Policy: info-news-runtime-rules.md §1. The Trigger update steps below apply in addition to gate-runtime-rules.md; when scripts are present, use this flow for version checks before execution.

Update check — user visibility: Technical failures during version check (missing script, sandbox, network, non-zero exit, no parseable Result= line) must not be explained to the user; continue with Execution per info-news-runtime-rules.md. Only successful check / apply outcomes may be summarized (including update_available / strict exit 3, which is still a success path that requires user confirmation before apply). Do not auto-download update-skill.* from the network. Static reference if .ps1 is missing: canonical scripts live in gate/gate-skills under skills/<name>/scripts/ (same <name> as frontmatter).

Step 0 — Wrapper routing probe

This legacy skill is a compatibility alias for the primary CLI skill gate-info-risk.

Before Trigger update, MCP tool selection, or any legacy Execution Workflow, run a deterministic shell probe:

PRIMARY_SKILL="gate-info-risk"
HAS_PRIMARY=0
for root in \
  "$HOME/.cursor/skills" \
  "$HOME/.codex/skills" \
  "$HOME/.openclaw/skills" \
  "$HOME/.agents/skills" \
  "$HOME/.gemini/antigravity/skills"
do
  if [ -f "$root/$PRIMARY_SKILL/SKILL.md" ]; then
    HAS_PRIMARY=1
    break
  fi
done

if command -v gate-cli >/dev/null 2>&1 && [ "$HAS_PRIMARY" = "1" ]; then
  echo "__ROUTE_CLI__"
else
  echo "__FALLBACK__"
fi

Interpretation:

• __ROUTE_CLI__ → stop here. Do not run Trigger update or the legacy MCP sections below. Follow https://github.com/gate/gate-skills/blob/master/skills/gate-info-risk/SKILL.md and its playbook contract. Carry over the minimum context: token+chain, address+chain, or symbol depending on whether the query is token risk, address risk, or project-level risk.
• __FALLBACK__ → continue with Trigger update and the legacy MCP workflow in this file.

Trigger update (with Execution)

Default (PD / blocking): On each trigger, run check first without GATE_SKILL_UPDATE_MODE=auto. The script only compares local vs remote; if Result=update_available, it does not overwrite files. STOP — ask the user whether to update before any apply, run, or skill Execution. In sandboxes (e.g. Cursor), use full / all permissions when later running apply. This matches blocking update confirmation in the independent-script TEST_REPORT: blocking is enforced by these SKILL instructions plus check-only behavior (not by silent auto-apply).

Strict check (Agent / IDE — recommended on check): Use GATE_SKILL_CHECK_STRICT=1. When Result=update_available, check exits 3 and prints GATE_SKILL_CHECK_EXIT=3 — it stops only that shell step. Hard rule: after you ask the user, end this assistant turn; do not call apply, run, or MCP in the same message. run applies without a prior user step — do not use run in chat for the blocking flow; use check → user confirms → apply.

Step 1 — Check

Install root: check / apply read $DEST/SKILL.md. Typical $SKILL_ROOT: $HOME/.cursor/skills/<name>/, $HOME/.codex/skills/<name>/, $HOME/.openclaw/skills/<name>/, $HOME/.agents/skills/<name>/, $HOME/.gemini/antigravity/skills/<name>/. Single-arg: the script resolves DEST in that order when SKILL.md exists; otherwise DEST is scripts/../. Workspace-only trees need two-arg check / apply with explicit DEST.

Agent selection rules (run at most one of the fenced examples below; choose by OS and files under $SKILL_ROOT/scripts/):

• macOS / Linux / WSL / Git Bash: If update-skill.sh exists → use only the Bash example.
• Windows (native PowerShell): If update-skill.ps1 exists → use only the PowerShell example.
• Windows without .ps1 (typical ClawHub install): If only update-skill.sh exists → try WSL or Git Bash (bash) using the same command shape as the Bash example. If bash is unavailable or execution fails → skip Step 1 silently (no script stderr or “version check failed” message to the user); proceed to Execution.
• Neither script exists: Skip Step 1 silently; proceed to Execution. Do not invent GATE_SKILL_CONFIRM_TOKEN or run apply / revoke-pending.

Bash (blocking — no auto; example Cursor):

GATE_SKILL_CHECK_STRICT=1 bash "$HOME/.cursor/skills/gate-info-riskcheck/scripts/update-skill.sh" check "gate-info-riskcheck"

PowerShell:

$env:GATE_SKILL_CHECK_STRICT = '1'
powershell -ExecutionPolicy Bypass -File "$env:USERPROFILE\.cursor\skills\gate-info-riskcheck\scripts\update-skill.ps1" check "gate-info-riskcheck"

Result semantics: skipped = no action. update_available = remote newer; with blocking flow, do not apply until the user agrees. check_failed = could not compare — proceed with current version per info-news-runtime-rules.md; do not surface technical check failure details to the user.

Agent parse (stdout): GATE_SKILL_UPDATE_AGENT_ACTION=…. BLOCK_UNTIL_USER_CONFIRMS_UPDATE → Step 2 before Execution. CONTINUE_SKILL_EXECUTION → no block from the check script.

Step 2 — Confirm or Reject (blocking)

Runtime: Use the same shell family for Step 2 as for Step 1 (Bash vs PowerShell). If Step 1 was skipped, do not run apply or revoke-pending.

If update_available:

1. STOP — do NOT proceed to Execution yet.

2. Inform the user (e.g. newer version available; summarize if helpful).

3. Wait for the user’s reply — blocking step.

   Hard rule (Cursor / Agent): When check reports update_available, or BLOCK_UNTIL_USER_CONFIRMS_UPDATE, or strict exit 3, end this turn after asking. Only in the user’s next message run apply (if they agree) or revoke-pending (if they decline). Do not chain apply in the same turn as check for this flow.

   • User agrees → run apply with GATE_SKILL_CONFIRM_TOKEN from strict check stdout when required, then Execution.
   • User declines → revoke-pending, then Execution on the current install.

Two-step gate (strict check): apply / run (without GATE_SKILL_UPDATE_MODE=auto) fail until GATE_SKILL_CONFIRM_TOKEN matches .gate-skill-apply-token. User decline → revoke-pending.

GATE_SKILL_CONFIRM_TOKEN="<paste from check stdout>" bash "$HOME/.cursor/skills/gate-info-riskcheck/scripts/update-skill.sh" apply "gate-info-riskcheck"

bash "$HOME/.cursor/skills/gate-info-riskcheck/scripts/update-skill.sh" revoke-pending "gate-info-riskcheck"

$env:GATE_SKILL_CONFIRM_TOKEN = '<paste from check stdout>'
powershell -ExecutionPolicy Bypass -File "$env:USERPROFILE\.cursor\skills\gate-info-riskcheck\scripts\update-skill.ps1" apply "gate-info-riskcheck"

powershell -ExecutionPolicy Bypass -File "$env:USERPROFILE\.cursor\skills\gate-info-riskcheck\scripts\update-skill.ps1" revoke-pending "gate-info-riskcheck"

If Step 1 was not strict (no pending token): apply without GATE_SKILL_CONFIRM_TOKEN is allowed.

If skipped or check_failed: no update step; proceed to Execution.

Optional — GATE_SKILL_UPDATE_MODE=auto

For CI / unattended automation only: setting GATE_SKILL_UPDATE_MODE=auto on check makes the script apply immediately when the remote is newer — no user confirmation and incompatible with blocking update confirmation tests. Do not use auto on check when reproducing the blocking PD flow.

Parameters

• name: Frontmatter name above; must match skills/<name>/ on gate-skills.
• Invoke: Use $SKILL_ROOT/scripts/update-skill.sh (or .ps1) where $SKILL_ROOT/SKILL.md is this skill — e.g. ~/.cursor/skills/<name>, ~/.codex/skills/<name>, ~/.openclaw/skills/<name>, ~/.agents/skills/<name>, ~/.gemini/antigravity/skills/<name>; do not treat ~/.cursor (or any host root without skills/<name>/SKILL.md) as the install. With one arg, the script resolves $SKILL_ROOT in that order before falling back to the script’s directory; workspace installs need explicit DEST. Two-arg check / apply / revoke-pending: canonical order is absolute DEST (skill root) first, then name; update-skill.sh / update-skill.ps1 auto-swap when only one normalized path contains SKILL.md (e.g. agent passes name then path).
• ClawHub vs full tree: Installs without update-skill.ps1 may copy it from gate/gate-skills under skills/<name>/scripts/ (manual only; agents must not auto-download).

Do not dump raw script logs into the user-facing reply except when debugging. On check exit 3 (strict), do not run Execution until Step 2 is resolved. On check_failed or apply failure, still run Execution when appropriate per runtime rules.

MCP Dependencies

Legacy path only — this section applies when Step 0 emitted __FALLBACK__.

Required MCP Servers

MCP Tools Used

Query Operations (Read-only)

• info_coin_get_coin_info
• info_compliance_check_address_risk
• info_compliance_check_token_security
• info_onchain_get_address_info

Authentication

• API Key Required: No
• Credentials Source: None; this skill uses read-only Gate Info / Gate News MCP access only.

Installation Check

• Required: Gate-Info
• Install: Use the local Gate MCP installation flow for the current host IDE before continuing.
• Continue only after the required Gate MCP server is available in the current environment.

Routing Rules

Legacy path only — when Step 0 emitted __ROUTE_CLI__, routing is delegated to gate-info-risk.

Execution Workflow

Legacy path only — this section applies when Step 0 emitted __FALLBACK__.

Step 0: Multi-Dimension Intent Check

Before executing this Skill, check if the user's query involves multiple analysis dimensions:

• If the query is about address safety (e.g., "is this address safe", "check 0x..."), proceed with this Skill (Mode B) — address risk is exclusive to this Skill and NOT covered by gate-info-research.
• If the query is only about token/contract security with no other dimension, proceed with this Skill (Mode A).
• If the query also mentions fundamentals, technicals, news, sentiment, comparison, or any other analysis dimension beyond security, route to gate-info-research — it handles multi-dimension queries with unified tool deduplication and coherent report aggregation.

Mode A: Token Security Check (Core Mode — Ready)

Step 1: Intent Recognition & Parameter Extraction

Extract from user input:

• token: Token symbol (e.g., PEPE, SHIB) — mutually exclusive with address
• address: Contract address (e.g., 0x...) — mutually exclusive with token
• chain: Chain name (eth / bsc / solana / base / arb, etc.) — required

Parameter Completion Strategy:

• If user provides only token without chain: ask "Please specify the chain (e.g., eth, bsc, solana)"
• If user provides a contract address without chain: attempt to infer from address format (0x prefix likely EVM chain, but still confirm specific chain)
• If user asks about major coins (BTC, ETH): inform them "Major coins typically have no contract security risks. If you need to check, please specify the wrapped token or a Meme token on a specific chain"

Step 2: Call 2 MCP Tools in Parallel

Both Tools are called in parallel with no dependencies.

Step 3: LLM Aggregation — Generate Risk Report

Pass the security detection data and fundamentals to the LLM to generate the assessment report using the template below.

Mode B: Address Risk Check (Degraded Mode)

info_compliance_check_address_risk is not yet available (P3 phase). Currently only info_onchain_get_address_info can provide basic address information.

Degradation Handling: Inform the user "Address compliance risk detection is under development. Currently only basic address information is available. For token contract security checks, please provide the token name or contract address."

Report Template (Token Security Mode)

Legacy path only — this section applies when Step 0 emitted __FALLBACK__.

## {token} Contract Security Report

### 1. Risk Overview

| Metric | Result |
|--------|--------|
| Chain | {chain} |
| Contract Address | {address} |
| Overall Risk Level | {risk_level_text} ({highest_risk_level}) |
| High-Risk Items | {high_risk_num} |
| Medium-Risk Items | {middle_risk_num} |
| Low-Risk Items | {low_risk_num} |
| Honeypot Detected | {is_honeypot ? "⛔ Yes" : "✅ No"} |
| Open Source | {is_open_source ? "✅ Yes" : "⚠️ No"} |

### 2. High-Risk Item Details

{If high-risk items exist, list each:}

| Risk Item | Description | Value |
|-----------|------------|-------|
| {risk_name_1} | {risk_desc_1} | {risk_value_1} |
| {risk_name_2} | {risk_desc_2} | {risk_value_2} |
| ... | ... | ... |

{If no high-risk items: "✅ No high-risk items detected"}

### 3. Tax Analysis

| Metric | Value | Status |
|--------|-------|--------|
| Buy Tax | {buy_tax}% | {Normal/Elevated/Extreme} |
| Sell Tax | {sell_tax}% | {Normal/Elevated/Extreme} |
| Transfer Tax | {transfer_tax}% | {Normal/Elevated/Extreme} |

{If multiple DEX pools have different tax rates, list the major pool breakdowns}

### 4. Holder Concentration

| Metric | Value | Status |
|--------|-------|--------|
| Holder Count | {holder_count} | {Many/Normal/Low} |
| Top 10 Holder % | {top10_percent}% | {Normal/High/Extremely Concentrated} |
| Top 100 Holder % | {top100_percent}% | — |
| Developer Holdings | {dev_holding_percent}% | {Normal/High} |
| Insider Holdings | {insider_percent}% | {Normal/High} |
| Largest Single Holder | {max_holder_percent}% | {Normal/High} |

### 5. Name Risk

| Metric | Result |
|--------|--------|
| Domain Token | {is_domain_token ? "⚠️ Yes" : "✅ No"} |
| Contains Sensitive Words | {is_sensitive ? "⚠️ Yes" : "✅ No"} |
| Sensitive Words | {sensitive_words} |

### 6. Project Basic Info (Supplementary)

| Metric | Value |
|--------|-------|
| Project Name | {project_name} |
| Sector | {category} |
| Listed on Major Exchanges | {exchange_list} |

### 7. Overall Assessment

{LLM generates a 3-5 sentence comprehensive risk assessment:}
- Overall contract safety level
- Most critical risk items (if any)
- Whether holder concentration is healthy
- Whether tax rates are reasonable
- Whether further manual audit is recommended

### ⚠️ Risk Warnings

{Auto-generated explicit warnings based on detection results:}
- Honeypot detection (if applicable)
- High tax warning (if applicable)
- Excessive holder concentration (if applicable)
- Contract not open-source (if applicable)

> The above analysis is based on automated on-chain data detection and cannot cover all risk scenarios. Please combine with project due diligence and community research for comprehensive judgment.

Decision Logic

Risk Level Mapping

Error Handling

Cross-Skill Routing

Available Tools & Degradation Notes

Safety Rules

1. Mandatory honeypot warning: When is_honeypot=true is detected, display the "⛔ Critical Risk" warning in the most prominent position — never downplay
2. No investment advice: Risk assessment is based on on-chain data and must include a "not investment advice" disclaimer
3. No absolute safety guarantees: Even if all checks pass, state that "automated detection cannot cover all risks"
4. Data transparency: Label detection data source and timestamp
5. Flag missing data: When any dimension has no data, explicitly inform the user — never fabricate safety conclusions
6. Address privacy: Do not proactively expose address holder identities — only display publicly available on-chain data