secret-safe（秘密保险箱）

Secret-Safe: Secure Credential Handling for Agent Skills

Why this skill exists: Snyk researchers found that 7.1% of all ClawHub skills instruct agents to handle API keys through the LLM context — making every secret an active exfiltration channel. This skill teaches the correct pattern.

The Core Rule

A secret must never appear in:

• The LLM prompt or system context
• Claude's response or reasoning
• Logs, session exports, or .jsonl history files
• File artifacts created by the agent
• Error messages echoed back to the user

A secret must only flow through:

• process.env (injected by OpenClaw before the agent turn)
• The shell environment of a subprocess the agent spawns
• A secrets manager CLI (read at subprocess level, not piped back into context)

Pattern 1: Environment Injection (Preferred)

This is OpenClaw's native, secure path. Use it for any skill that needs an API key.

In SKILL.md frontmatter

---
name: my-service-skill
description: Interact with MyService API.
metadata: {"openclaw": {"requires": {"env": ["MY_SERVICE_API_KEY"]}, "primaryEnv": "MY_SERVICE_API_KEY"}}
---

The requires.env gate ensures the skill will not load if the key isn't present — no silent failures, no prompting the user to paste a key mid-conversation.

The primaryEnv field links to skills.entries.<n>.apiKey in openclaw.json, so the user configures it once in their config file, never in chat.

In skill instructions

## Authentication
The API key is available as `$MY_SERVICE_API_KEY` in the shell environment.
Pass it to CLI tools or curl as an environment variable — never echo it or
include it in any output returned to the user.

Example safe curl invocation (instruct the agent to do this)

# CORRECT — key stays in environment, never in command string visible to LLM
MY_SERVICE_API_KEY="$MY_SERVICE_API_KEY" curl -s \
  -H "Authorization: Bearer $MY_SERVICE_API_KEY" \
  https://api.myservice.com/v1/data

Never instruct the agent to do this:

# WRONG — key is visible in LLM context, command history, and logs
curl -H "Authorization: Bearer sk-abc123realkeyhere" https://api.myservice.com/

Pattern 2: Secrets Manager Integration

For production setups or team environments, read secrets from a manager at subprocess level.

Supported managers

Safe shell wrapper pattern

Create a scripts/run-with-secret.sh in your skill:

#!/usr/bin/env bash
# Fetches the secret at subprocess level — never echoes to stdout
SECRET=$(security find-generic-password -s "my-service-api-key" -w 2>/dev/null)
if [ -z "$SECRET" ]; then
  echo "ERROR: Secret 'my-service-api-key' not found in keychain." >&2
  exit 1
fi
export MY_SERVICE_API_KEY="$SECRET"
exec "$@"

The agent runs bash {baseDir}/scripts/run-with-secret.sh <actual-command> — the secret is fetched and injected entirely outside the LLM's view.

Pattern 3: User Setup Flow (first-run)

If the user hasn't configured a key yet, guide them through setup without asking for the key in chat.

Correct setup prompt to give the user:

To use this skill, add your API key to ~/.openclaw/openclaw.json:

  skills:
    entries:
      my-service:
        apiKey: "your-key-here"

Or set it as an environment variable before starting OpenClaw:
  export MY_SERVICE_API_KEY="your-key-here"

Do NOT paste your key into this chat — it will be logged.

Incorrect (never do this):

Please share your API key so I can help you set it up.

Auditing Another Skill for Leaks

When asked to review a SKILL.md for credential safety, check for these patterns:

🔴 Critical — Must Fix

🟡 Warning — Should Fix

🟢 Safe Patterns

• requires.env in frontmatter
• Key accessed only as $ENV_VAR in shell, never echoed
• Subprocess scripts that fetch and inject without returning to context
• Error messages that say "key not found" without printing the value
• Output filtered through sed/grep before returning to agent

Self-Check Before Publishing a Skill

Run through this checklist before putting any skill on ClawHub:

• [ ] Does the skill ever ask the user to paste a secret into the conversation?
• [ ] Does the skill ever echo, print, log, or return a secret value?
• [ ] Does the skill read a .env file and dump its contents?
• [ ] Does the skill store a secret in a file artifact?
• [ ] Are all API key references gated with requires.env in frontmatter?
• [ ] Do error messages avoid reflecting credential values?
• [ ] Does any shell script use set -x (which would expose key values)?
• [ ] Would running clawhub audit {skill-name} pass?

If any box is unchecked, do not publish until fixed.

Quick Reference: Safe vs Unsafe Patterns

# UNSAFE — never write instructions like these:
"Ask the user for their OpenAI API key and use it to call the API."
"Set the Authorization header to Bearer {user_api_key}."
"Store the API key in a variable and use it throughout the session."

# SAFE — write instructions like these:
"The API key is injected as $OPENAI_API_KEY via environment — use it directly."
"Run: OPENAI_API_KEY=$OPENAI_API_KEY curl ..."
"If $OPENAI_API_KEY is not set, print an error and exit — do not ask the user."

Reference Files

• references/env-injection-examples.md — Full worked examples for popular APIs (OpenAI, Anthropic, GitHub, Stripe, Slack)
• references/audit-checklist.md — Printable audit checklist for skill authors and reviewers