LLM Pipeline

Editorial enrichment (editorial._llm_complete) with a free-first provider chain and
per-feature cost accounting.

Provider order

Groq → OpenRouter → DeepSeek → Ollama

Free tiers first, paid DeepSeek third, local Ollama last. (Racing + RSSSPORTS editorial
must stay free-first; see deepseek-token-harness for DeepSeek-Pro usage.)

Cost tracking

• Every successful completion appends to /opt/sportstlm/data/llm_usage.jsonl via
  app/services/llm_log.py with: provider, model, feature label, tokens, estimated cost.
• Feature labels in use: paraphrase:sportsmole, preview:scraper, debrief, draft.
• Any new tool that calls _llm_complete(..., feature="x") appears automatically.

Reporting

• /opt/ops/llm_cost_report.py [--days N] [--telegram] → console table +
  /opt/ops/llm-cost.json.
• llm-cost-report.timer refreshes every 15 min.
• llm-cost-digest.timer sends daily summary at 21:00 SGT.
• The 08:00 health digest also carries the cost line.

Pricing

llm_log.PRICES holds the per-1M-token table. Unpriced models are counted and
flagged, not silently zeroed.

Provider health / balance

• provider_balance.py / provider_guard.py + provider-balance.timer track provider
  health, thresholds and balance (see /opt/ops/provider-state/).

Guardrail

Public site copy must never reveal AI/LLM/paraphrasing usage — neutral wording only.