diagnose-llm-advisor-health
Verify a native LLM voting advisor before trusting advisor-sensitive evaluations.
SKILL.mdtemplates/optimizers/242026/crewrift-specific/skills/diagnose-llm-advisor-health/SKILL.md
GitHubDiagnose LLM advisor health
The advisor reads a meeting snapshot and returns a vote/chat decision. A completed episode does not prove that it ran; scripted fallback can hide an inactive advisor.
- Confirm the exact uploaded version was built for native Anthropic Messages or OpenAI Chat and opted in with
--use-llm. - Check
COWORLD_LLM_ENDPOINT,COWORLD_LLM_ENABLED, andCOWORLD_LLM_MODELnames/presence without printing secrets. - Verify a real call/result in player traces and distinguish request rejection, timeout, malformed vote, and fallback.
- Validate the returned vote against offered options. Check elapsed time against the game's vote deadline.
- Re-run deterministic fallback alone before comparing advisor-sensitive scores.
Use canonical model IDs, such as anthropic/claude-haiku-4.5, only when permitted by the current platform policy.
Hosted SDK credentials are placeholders; only the sidecar holds the provider key. Local calls require a configured
native endpoint or host OpenRouter key. Rebuild and republish old advisor images; no automatic protocol migration exists.
These templates describe the required native contract, not proof that an external notsus image has migrated.