Skip to main content

🧩 LLMFeed Extensions

This document lists optional fields, experimental structures, and advanced practices that extend the core LLMFeed protocol.

They are not mandatory — but agents may support and react to them.

✅ Core extensions​

Extension FieldPurpose
flags[]Public warnings, reported risks or revocation status
llm_trust_level_requiredLimits execution to high-trust agents
recommended_toolchainSuggest compatible agent runtimes
agent_behaviorDeclares expected agent UX mode
session_behaviorHints whether the feed is stateless or transient
risk_score, safety_tierFeed-level caution or reputation estimate
sandbox_policy.llmfeed.jsonRuntime constraints for execution (experimental)
attachments[]Files or assets referenced in feed
examples[]Sample use cases, usage strings
tags[]Categorical labels for filtering or display

⚠️ flags[]​

Flags can be added by agents or moderators to warn about:

  • mismatched capabilities
  • deprecated versions
  • suspicious or malicious usage
"flags": [
  {
    "type": "risk",
    "submitted_by": "agent://previewbot",
    "reason": "Declared actions not matching real API",
    "date": "2025-05-19T12:00:00Z",
    "status": "pending",
    "source": "https://llmca.org/flag/234"
  }
]

🔐 Trust extensions​

Agents may use:

  • trust.scope = certified
  • certifier = https://llmca.org
  • signed_blocks = [ "metadata", "prompts", "trust" ]

Verification interfaces (like /verify) will display results clearly.

🧠 Agent behavior hints​

Declare how the LLM should behave:

"agent_behavior": {
  "mode": "suggest-only",
  "requires_user_confirmation": true
}

🧪 Runtime policies (experimental)​

You can expose:

"sandbox": {
  "max_calls": 10,
  "device_scope": "local-only",
  "runtime_constraints": "No background tasks"
}

Or create a separate sandbox-policy.llmfeed.json.

📚 Related​