ypollak2/llm-router

Added

• LLM_ROUTER_EXPLAIN=1 response prefix (src/llm_router/tools/text.py)

  When set, every routed response (llm_query, llm_research, llm_generate, llm_analyze, llm_code) is prefixed with a compact routing summary:

  [→ gemini-2.5-flash · query · $0.00003 · 42.9x cheaper than Sonnet]
  

  Shows: model used, task type, per-call cost, and cost ratio vs Sonnet baseline — the "why this model?" answer right in the response stream.

• llm_classify cost comparison table (src/llm_router/tools/routing.py)

  The classification output now includes a "Why not a more expensive model?" breakdown showing Opus/Sonnet/Haiku costs side-by-side with the chosen tier, including a multiplier for each skipped tier (e.g. "↑ 60x more expensive — unnecessary for simple task"). Always shown; no env var required.

• reason_code DB column (src/llm_router/cost.py)

  New column in routing_decisions table for storing classification reasoning codes (idempotent migration). log_routing_decision() updated with reason_code: str | None = None parameter.

• router.py reason_code propagation — passes reason_code from classification metadata to log_routing_decision().

Technical

• _explain_prefix() helper: pure function, zero overhead when env var not set.
• Cost table uses per-1k-output-token pricing — representative of real-world savings signal.
• Routing tip injected into llm_classify output when LLM_ROUTER_EXPLAIN is not set.