ypollak2/llm-router

Added

• llm-router test "<prompt>" dry-run CLI (src/llm_router/cli.py)

  Simulates a routing decision for any prompt without making an API call. Uses the existing 5-layer classifier to determine task type, complexity, and confidence, then maps to the cheapest appropriate model and shows an estimated cost vs Sonnet baseline.

  llm-router test "refactor the auth module to use JWT"
  → Task: analyze / moderate / 85% confidence (via gemini-flash-lite)
  → Chosen: claude-sonnet-4-6  Baseline: claude-sonnet-4-5
  → Saved: $0.00465 (100% cheaper)
  

• llm_savings MCP tool (src/llm_router/tools/admin.py)

  Text-based savings dashboard with time-bucketed aggregates: today / this week / this month / all-time. Shows actual spend, Sonnet baseline, savings, and the efficiency multiplier (Nx) — the "wow" metric that makes routing value tangible.

• DB schema migration v2.1 (src/llm_router/cost.py)

  Four new columns added to usage table via idempotent ALTER TABLE (safe for existing DBs): baseline_model, potential_cost_usd, saved_usd, is_simulated.

• get_savings_by_period() (src/llm_router/cost.py) — async savings query used by status bar and llm_savings. Falls back to Sonnet estimation for pre-v2.1 rows.

• Enhanced status bar v3 (src/llm_router/hooks/status-bar.py) — D/W savings, provider health icons (auto-hidden until health.json active), enforcement mode badge, Nx efficiency multiplier. Full mode via LLM_ROUTER_STATUS_MODE=full.

Fixed

• Ruff F541 (src/llm_router/cli.py:1275) — spurious f prefix on string with no placeholders; broke CI lint on both Python 3.11 and 3.12.