Summary

When switching providers via /provider <name>, if the new provider has no embedding_model configured, semantic tool schema filtering and skill matching silently fall back to all tools/all skills without notifying the user.

Reproduction

Config with two providers — one with embedding_model, one without:

[[llm.providers]]
name = "openai-stt"
model = "gpt-4o-mini"
embedding_model = "text-embedding-3-small"

[[llm.providers]]
name = "openai-quality"
model = "gpt-4o"
# no embedding_model

Session:

/provider openai-quality
What is 23 times 8?

Expected

Either:
a) Keep using the original embedding provider for query embedding regardless of active conversational provider, OR
b) Emit a visible INFO to the user: "Semantic tool filtering unavailable — provider does not support embeddings. Using all tools."

Actual

WARN failed to embed query: embedding not supported by openai
WARN skill matcher returned no results, falling back to all skills
WARN tool filter: query embed failed, using all tools: embedding not supported by openai

Tool selection degrades from semantic top_k=5 to all 23 tools. No user-visible notification.

Impact

• MEDIUM: tool filtering efficiency drops after provider switch
• User has no indication of degraded behaviour
• Discovered in CI-209 live session (2026-03-27)

Fix Direction

Decouple embedding provider from active conversational provider. Tool schema filter and skill matcher should use a dedicated embedding provider (configurable via embedding_provider field or defaulting to whichever provider in the pool has embedding_model set), independent of /provider switch.