Status: Accepted (shipped in v2.x — evolved from the original v2.4.0 estimate) Date: 2026-05-20 Amended: 2026-07-03 (update status and clarify enforcement model)

Using the most capable model for every skill wastes cost on tasks where lighter models are sufficient. Using the cheapest model everywhere degrades quality on complex reasoning tasks (planning, architecture decisions, assumption surfacing) by 15–23%.

Skills are assigned a model tier based on task complexity:

Tier	Model	Tasks	Token budget
Light	Haiku-equivalent	Code review, verification, commit messages	100K
Standard	Sonnet-equivalent	Code writing, research, planning	200K
Deep	Opus-equivalent	Strategic planning, grill-me, ADR decisions	250K

The assignment lives in docs/references/model-profiles.md. Skills declare their tier in SKILL.md frontmatter via the model: field.

Routing is declared in frontmatter but enforcement is harness-dependent.

• Declarative: Every SKILL.md carries a model: field (see model-profiles.md for the canonical tier list). This is the single source of truth.
• Harness-dependent: The host agent harness (Claude Code, Cursor, Gemini CLI, pi) reads the model: field but each enforces routing differently — or not at all. Claude Code and pi respect tier hints natively; Cursor and Gemini CLI currently ignore them.
• No central orchestrator: The original ADR assumed orchestrate-project would enforce routing. Instead, each skill’s model: declaration is read directly by the harness at launch time — simpler, no single point of failure, but with per-harness variance.

• ~35% cost reduction vs. all-premium baseline (harness-dependent — varies by tool).
• Quality maintained on complex tasks (premium models where it matters).
• The model: field in frontmatter is declarative and non-breaking — harnesses that don’t support routing silently ignore it. This is intentional: bigpowers can’t control host behavior, only declare intent.
• model-profiles.md must stay in sync with SKILL.md updates (same consequence as original).