orchestration-routing

name	orchestration-routing
description	Routing table, classification protocol, and ask protocol for the Sando Orchestrator. Extracted from the orchestrator to reduce its context size. Contains the keyword → agent mapping, multi-agent workflow definitions, skill injection reference, and the SDD architectural gate. Trigger: loaded once per session by the orchestrator before any delegation.
license	MIT
metadata	{"version":"1.0.0","author":"Sando Design System"}

Orchestration Routing

Routing intelligence for the Sando Orchestrator. This skill replaces the inline classification_system, ask_protocol, delegation_protocol, and multi_agent_workflows sections.

---

Automatic Routing Table

Scan the user request for keywords and route immediately:

Keywords Detected	Classification	Route To
token, color, spacing, typography, flavor, ingredient, recipe, theme, --sando-	TOKEN_WORK	sando-tokens
implement, component, fix bug, add feature, refactor, Lit, render, @property	COMPONENT_CODE	sando-developer
test, coverage, a11y, accessibility, axe, WCAG, audit, quality, validate	QUALITY_WORK	sando-quality
story, stories, storybook, argTypes, controls, addon, CSF, preview.ts	STORYBOOK_WORK	sando-storybook
document, docs, JSDoc, MDX, README, API reference, guide	DOCUMENTATION	sando-documenter
architecture, pattern, structure, breaking change, decision, RFC, ADR, build system	ARCHITECTURE	sando-architect
UX, user experience, states, behavior, flow, empty state, error message, motion	UX_DESIGN	sando-ux-designer
copy, microcopy, write text, tagline, changelog entry, content audit, tone	UX_WRITING	sando-ux-writer
create component, new component, scaffold, build [component-name]	FULL_WORKFLOW	Multi-agent

Classification Protocol

1. READ the user's request
2. SCAN for keywords above
3. MATCH to a classification
4. IF single classification → DELEGATE IMMEDIATELY (Task tool)
5. IF multiple classifications → plan multi-agent workflow
6. IF no match or unclear → ASK THE USER

---

SDD Architectural Gate

Before delegating ARCHITECTURE requests, check if the change is architectural:

IS THIS ARCHITECTURAL?
├── Modifies token system layers (new layer, renames, structural changes)? → YES
├── Introduces a new component pattern (new mixin, new composition approach)? → YES
├── Changes a guideline that affects 2+ agents or 3+ components? → YES
├── Is a breaking change (removes/renames public API, changes token naming)? → YES
├── Adds or modifies a Flavor (theme)? → YES (if structural, not just values)
└── Everything else (new component, bug fix, new token values, new story) → NO

If YES → Run SDD Workflow FIRST, then delegate implementation:

Load skill: sdd-architectural-workflow
Execute: SDD pipeline (explore → propose → spec → design → tasks)
ONLY AFTER all SDD phases complete → delegate implementation to specialists

If NO → Route normally using the routing table above.

---

Ask Protocol

Use this template when the request is unclear:

No puedo determinar automáticamente qué agente debe manejar esta tarea.

**Tu solicitud:** "[quote]"

**Mi duda:** [why unclear]

**Opciones:**
1. Si quieres [A] → delegaré a `[agent]`
2. Si quieres [B] → delegaré a `[agent]`

¿Cuál es tu intención?

When to ask instead of delegate:

• No keyword matches
• 2+ agents could handle it equally
• Scope is ambiguous ("fix the button" — fix what?)
• New territory not covered by any agent

Cardinal rule:

IF IN DOUBT → ASK
IF NOT IN DOUBT → DELEGATE
IF TEMPTED TO DO IT YOURSELF → YOU ARE IN DOUBT → ASK

---

Skill Injection Reference

When delegating, inject skills from .opencode/skills/ (all skills are local to the project):

If delegating to...	Inject these skills
Any agent (always)	agent-guidelines-compact (SHARED + role block), verification-protocol, engram-protocol
sando-developer for a new component	+ component-creator, component-development-workflow
sando-developer for a loading state	+ skeleton-creator
sando-developer for any component work	+ component-creator (structural rules always useful)
sando-architect for an architectural task	+ sdd-architectural-workflow (if SDD gate fires) OR advisor mode (no extra skill needed)
Any agent creating a PR	+ branch-pr
Any agent creating a GitHub issue	+ issue-creation

Judgment Day: inject compact rules for ALL skills relevant to the files being reviewed.

All skills live at .opencode/skills/{skill-name}/SKILL.md. Never reference .atl/ — that path does not exist in this project.

---

Standard Delegation Prompt Format

## Task for [agent-name]

### Context
[Relevant background — what exists, what's needed]

### Requirements
[Specific list of what to accomplish]

### Deliverables
- [ ] File or outcome 1
- [ ] File or outcome 2

### Constraints
[Any rules or limitations]

---

Multi-Agent Workflows

Full Component Creation

USER: "Create a Checkbox component"

1. Create TODO list with phases
2. Load skill registry (Step 0) — inject component-creator + component-development-workflow
3. DELEGATE sando-tokens → Create Recipe tokens
4. WAIT — check Return Envelope STATUS
5. DELEGATE sando-developer → Implement component
   Inject: component-creator + component-development-workflow
6. WAIT — check Return Envelope STATUS
7. PARALLEL DELEGATE (delegate tool — true async):
   - sando-quality → Write tests
   - sando-storybook → Write stories
8. WAIT for all — check both Return Envelopes
9. IF any STATUS = partial/blocked → fix before continuing
10. DELEGATE sando-quality → Final validation
11. Report summary to user

Component Modification

USER: "Add a loading state to Button"

1. Check if token changes needed (quick read)
2. If tokens needed → DELEGATE sando-tokens first
3. WAIT — check Return Envelope STATUS
4. DELEGATE sando-developer → Add feature (inject skeleton-creator if loading state)
5. WAIT — check Return Envelope STATUS
6. PARALLEL DELEGATE: sando-quality + sando-storybook
7. WAIT for all — check Return Envelopes
8. Report summary

Token + Component Change

USER: "Add a new primary-alt color and use it in Card"

1. DELEGATE sando-tokens → Create token
2. WAIT for completion
3. DELEGATE sando-developer → Update component
4. Report summary

---

Parallelization Rules

CAN run in parallel: Tests + Stories (after component exists), multiple independent components, documentation for different components.

MUST run sequentially: Tokens → Component, Component → Tests, Architecture decision → Implementation.

---

Compact Rules (for skill-registry)

• Load this skill once per session BEFORE any delegation
• Check SDD Architectural Gate for every ARCHITECTURE-classified request
• Inject agent-guidelines-compact + verification-protocol + engram-protocol into EVERY sub-agent prompt
• Use Ask Protocol template verbatim when unclear — no improvising
• Use delegate tool (async) for parallel work, task tool (sync) for sequential
• NEVER route implementation tasks to sdd-apply — use the correct Sando specialist (sando-developer, sando-tokens, etc.)
• All skills are local: .opencode/skills/{name}/SKILL.md — no global fallback needed