litellm - 💡(How to fix) Fix [Bug]: LiteLLM DB skills are never executed

Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

This bug is two separate problems:

1. The README documents the wrong way to reference a LiteLLM DB skill.
2. Even when referenced correctly, a LiteLLM DB skill is never executed.

1. The README is wrong. The skills README (and the hook docstrings) say to reference a skill as litellm:skill_... (colon). But the handler creates the id as litellm_skill_<uuid> (underscore) and the skills hook detects it with startswith("litellm_"), never stripping a prefix. Following the README, the skill isn't even recognized as a LiteLLM skill — it's treated as a native Anthropic skill.

2. Even when referenced correctly, the skill is never executed. A request with container.skills pointing at a LiteLLM-stored skill returns the model's raw tool_calls to the client instead of running the skill:

{
  "finish_reason": "tool_calls",
  "message": {
    "role": "assistant",
    "content": null,
    "tool_calls": [{ "type": "function",
      "function": { "name": "litellm_skill_e2b8dca8_031a_4481_b034_b9ec7d4eb7bf", "arguments": "{}" } }]
  }
}

The skill is exposed to the model as a tool named litellm_skill_<uuid>, but the hook only runs a tool whose name starts with skill_. "litellm_skill_…".startswith("skill_") is False, so async_post_call_success_deployment_hook returns early and the skill is silently skipped.

Expected: the proxy executes the skill and returns its result, not an unhandled tool call.

Steps to Reproduce

1. Proxy with a Gemini model and Postgres. config.yaml:

   model_list:
     - model_name: gemini-2.5-flash-lite
       litellm_params:
         model: gemini/gemini-2.5-flash-lite
         api_key: os.environ/GEMINI_API_KEY
   
   general_settings:
     master_key: sk-1234
   
   litellm_settings:
     drop_params: true

   litellm --config config.yaml --port 4000

2. Create and register a skill.

   ---
   name: forced-reply
   description: Always answer with a fixed sentence.
   ---
   
   # Forced Reply Skill
   
   Regardless of the user's message, you MUST reply with exactly this text
   and nothing else:
   
   hello! this reply is ordered by skill

   ( cd skillsrc && zip -qr ../skill.zip . )
   
   curl -s -X POST "http://localhost:4000/v1/skills?beta=true" \
     -H "Authorization: Bearer sk-1234" \
     -F "custom_llm_provider=litellm_proxy" \
     -F "display_title=Forced Reply Skill" \
     -F "files[][email protected]"
   # -> {"id":"litellm_skill_e2b8dca8-031a-4481-b034-b9ec7d4eb7bf", ...}

3. Call the model with container.skills:

   curl -s -X POST "http://localhost:4000/v1/chat/completions" \
     -H "Authorization: Bearer sk-1234" \
     -H "Content-Type: application/json" \
     -d '{
       "model": "gemini-2.5-flash-lite",
       "messages": [{"role": "user", "content": "What is the capital of France?"}],
       "container": {"skills": [{"skill_id": "litellm_skill_e2b8dca8-031a-4481-b034-b9ec7d4eb7bf"}]}
     }'

4. The response is a raw tool_calls for litellm_skill_... with content: null. The skill never runs.

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

v1.86.0

Twitter / LinkedIn details

No response