core-llm-routing-and-finops

# LLM Routing and FinOps ## Opening commitment Cost is a quality dimension, not a separate concern. A plan that "works" but burns ten times the necessary budget has shipped a defect — the defect happens to live in the dispatch graph rather than in the code, but it produces the same kind of regret (post-hoc surprise, hard to reverse, accountable to the same Owner). This skill teaches the LLM FinOps Architect how to reason about routing choices so cost is treated as a first-class invariant alongside correctness, security, and performance. The rule that anchors everything below: **a wrong-tier dispatch is a finding with severity equal to the worst case it enables.** Routing `code-reviewer` to Sonnet on a security-touching diff is not a cost optimization — it is a CRITICAL governance violation that bypasses ADR-052's VETO floor. Routing a one-shot fixture generator to Opus is not safety — it is a MINOR cost finding that, repeated across 30 spawns per session, becomes a MAJOR budget regression. ## What This Skill Is (and isn't) This skill is the **cognitive layer** for routing and cost decisions. It complements but does not replace the mechanical enforcement. | Concern | Mechanism | Layer | Reference | |---|---|---|---| | VETO floor for `code-reviewer` / `security-engineer` is hard-pinned to Opus | Canonical agent files (`.claude/agents/<slug>.md`) carry `model: claude-opus-4-8` frontmatter; `check_agent_spawn.py` PreToolUse Agent matcher validates that file at spawn; agent ...