Model tiering convention for Orryx Group agents and subagents. Triggers when authoring a new subagent, configuring an agent's model, or deciding which model tier to use for a task. Defines haiku (explorer with permissionMode plan, doc-writer), sonnet (code-reviewer, test-writer, migration-writer, default delivery), and opus (security-reviewer, pr-investigator, deep-research, high-stakes review). Canonical Orryx convention; lifted from Triora.
How this skill is triggered — by the user, by Claude, or both
Slash command
/model-tiering-convention:model-tiering-conventionThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Three tiers, picked by agent role:
Three tiers, picked by agent role:
explorer (with permissionMode: plan), doc-writer, anything that reads/lists/summarisescode-reviewer, test-writer, migration-writer, default for delivery agentssecurity-reviewer, pr-investigator, deep-research, high-stakes architectural decisionsPicked from Triora's mature convention; promoted to cross-cutting in Wave 1.
| Agent role | Tier | Why |
|---|---|---|
explorer | haiku | Read-only with permissionMode: plan; speed > sophistication |
doc-writer | haiku | Format-heavy work where consistency > insight |
summariser | haiku | Lossy compression; sonnet is overkill |
code-reviewer | sonnet | Catches most issues; cost-efficient at scale |
test-writer | sonnet | Pattern-matching test generation |
migration-writer | sonnet | Mechanical transformation work |
qa-runner | sonnet | Executing test suites + interpreting results |
ops-auto | sonnet | Routine deploy / sync / status operations |
security-reviewer | opus | Stakes are high; misses are expensive |
pr-investigator | opus | Synthesises across many files |
deep-researcher | opus | Open-ended exploration |
architect | opus | Design decisions with long-term impact |
takeoff-estimator (Pillarworks BOQ) | sonnet | Domain-specific, well-structured |
handoff-protocol mediator | sonnet | Coordinates between agents |
| Default if unsure | sonnet | Workhorse default |
---
name: code-reviewer
description: Reviews code changes for security, performance, and correctness.
tools: [Read, Grep, Glob]
model: sonnet
permissionMode: plan
color: yellow
---
| Tier | Default permissionMode | Why |
|---|---|---|
| haiku | plan (read-only) | Speed agents shouldn't be allowed to write |
| sonnet | acceptEdits (writes OK, no bash without confirm) | Workhorse delivery default |
| opus | acceptEdits (or default) | High-stakes; trust the more capable model |
The explorer agent uniquely has model: haiku + permissionMode: plan — combining speed-tier model with read-only mode. This is the smart pattern Triora pioneered.
| Tier | ~ Cost relative to opus | When NOT worth it |
|---|---|---|
| haiku | 1/15× | Anything requiring multi-file synthesis |
| sonnet | 1/5× | Open-ended exploration; ambiguous decisions |
| opus | 1× | High-volume routine work (deploy checks, lint suggestions) |
A typical Orryx Group session: 1 opus orchestrator + 3-5 sonnet workers + 1-2 haiku explorers. Cost dominated by opus, but its calls are sparse.
D:\Clinical.Trials\.claude\agents\ — Triora's mature tier assignments (the canonical reference)references/triora-agent-inventory.md — full Triora agent → tier mapProvides CDSS development patterns for drug interaction checking, dose validation, clinical scoring (NEWS2, qSOFA), and alert classification integrated into EMR workflows.
npx claudepluginhub alexmclaren/orryx-knowledge --plugin model-tiering-convention