Run any Skill in Manus with one click

artifex-model-tiering

Stars0

Forks0

UpdatedJuly 23, 2026 at 06:48

Provider-neutral Model Router + 4-tier model policy rules for Artifex. Use when editing packages/model-router/**, the Model Catalog, the Tier Policy engine, the admission gate, or anything choosing which model an agent runs. Triggers on "model router", "model catalog", "tier", "admission gate", "Ollama", "vLLM", "Qwen", "provider", "which model".

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

mhylle

mhylle/artifex

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

name	artifex-model-tiering
description	Provider-neutral Model Router + 4-tier model policy rules for Artifex. Use when editing packages/model-router/**, the Model Catalog, the Tier Policy engine, the admission gate, or anything choosing which model an agent runs. Triggers on "model router", "model catalog", "tier", "admission gate", "Ollama", "vLLM", "Qwen", "provider", "which model".

Model tiering & routing (ADR-0002/0003)

Router is the spine; the Claude Agent SDK is one backend behind it. Never bind the runtime hard to Claude.
Tier is computed, not hardcoded. The Tier Policy engine derives a logical tier per staffing decision from blast radius, fan-in, reversibility, task class, budget, and clade score. A tier-bump is an escalation-ladder rung.
4-tier ladder: 0 = no LLM (mechanical checks) · 1 = local small (Qwen2.5) · 2 = local ~32B quantized · 3 = frontier (Claude).
Catalog resolves tier → concrete model and is swappable data (principle #3). A missing entry is a typed error, never a silent default.
Admission gate: a model enters the catalog only after producing schema-valid output on the real shared-types schemas.
Tier-2 fallback: attempt the local 32B; if it fails the gate, logical Tier-2 resolves to Claude — explicitly and logged (ADR-0003).
Budget vs blast radius is governed by the per-mission autonomy dial (supervised = hard floor + escalate; high-autonomy = may downgrade, loudly logged).
Every tier choice + its inputs is a ledger event.

Consult the claude-api skill before touching the Claude backend. Install deps via npm install.

More from this repository

same repository

artifex-invariants

mhylle/artifex

The 7 non-negotiable Artifex design invariants. Use BEFORE writing or reviewing any code in packages/worker/** or anything touching the meta-agents (Orchestrator, Agent Creator, Reviewer, Learning Agent), Constitution, Context Broker, Worker Swarm, the mission loop, verification gates, or the audit ledger. Triggers on "invariant", "constitution", "gate A/B", "fold-up", "context broker", "learning agent", "ledger", or editing worker/meta-agent code.

2026-07-230

artifex-ledger

mhylle/artifex

Append-only Audit Ledger rules for Artifex. Use when writing DB migrations or any code that reads/writes the ledger, the Model Catalog, or the memory-fabric stores (packages that touch Postgres, migrations, LISTEN/NOTIFY, replay). Triggers on "ledger", "append-only", "migration", "LISTEN/NOTIFY", "event store", "time-travel", "replay", "audit event".

2026-07-230

artifex-schemas

mhylle/artifex

TypeBox schema conventions for Artifex's shared-types. Use when creating or editing schemas/types in packages/shared-types/** — the contract, ledger event, evidence bundle, verdict, or capability-manifest schemas — or anywhere a schema is handed to an LLM for structured output. Triggers on "schema", "TypeBox", "contract type", "ledger event type", "JSON schema", "structured output", "validate contract".

2026-07-230

artifex-tdd

mhylle/artifex

Artifex test & phase-implementation discipline. Use when implementing any v0 phase (P0–P13), writing tests, or marking work done. Triggers on "implement phase", "write tests", "TDD", "acceptance criteria", "is this done", "RED/GREEN", "verify".

2026-07-230

name	artifex-model-tiering
description	Provider-neutral Model Router + 4-tier model policy rules for Artifex. Use when editing packages/model-router/**, the Model Catalog, the Tier Policy engine, the admission gate, or anything choosing which model an agent runs. Triggers on "model router", "model catalog", "tier", "admission gate", "Ollama", "vLLM", "Qwen", "provider", "which model".

Model tiering & routing (ADR-0002/0003)

Router is the spine; the Claude Agent SDK is one backend behind it. Never bind the runtime hard to Claude.
Tier is computed, not hardcoded. The Tier Policy engine derives a logical tier per staffing decision from blast radius, fan-in, reversibility, task class, budget, and clade score. A tier-bump is an escalation-ladder rung.
4-tier ladder: 0 = no LLM (mechanical checks) · 1 = local small (Qwen2.5) · 2 = local ~32B quantized · 3 = frontier (Claude).
Catalog resolves tier → concrete model and is swappable data (principle #3). A missing entry is a typed error, never a silent default.
Admission gate: a model enters the catalog only after producing schema-valid output on the real shared-types schemas.
Tier-2 fallback: attempt the local 32B; if it fails the gate, logical Tier-2 resolves to Claude — explicitly and logged (ADR-0003).
Budget vs blast radius is governed by the per-mission autonomy dial (supervised = hard floor + escalate; high-autonomy = may downgrade, loudly logged).
Every tier choice + its inputs is a ledger event.

Consult the claude-api skill before touching the Claude backend. Install deps via npm install.