Orchestrator

The Lamina orchestrator coordinates design → ready_to_build → external build → verify. It routes slash commands to workflows, selects capability skills via audit profiles, and manages subagent patterns during verify.

Quick path

Mode	When	Loads
Workflow	Slash commands (`/lamina-design`, etc.)	Orchestrator → workflow → audit profiles → skills
Direct	`/lamina <focused question>`	`lamina-core` → one skill

Architecture


/lamina* command
    ↓
lamina-orchestrator
    ↓
workflow (router | init | design | verify)
    ↓
audit-profiles.yaml → capability skills
    ↓
.lamina/ artifacts

Mode B guardrail: the orchestrator writes only under .lamina/. Never app source.

Workflows

Workflow	File	Purpose
Router	`workflows/router.md`	Intent parsing for `/lamina`
Init	`workflows/init.md`	Domain charter bootstrap
Design	`workflows/design.md`	Contract → `ready_to_build`
Verify	`workflows/verify.md`	Persona walks → `findings[]`

Orchestrator steps

Select — pick skills from audit profiles for the current workflow section
Apply — load each skill; write run.json incrementally
Deliver — output contract; design ends at ready_to_build + implement.md; verify ends at findings[] + fix.md

Audit profiles

audit-profiles.yaml maps workflow sections to capability skills. Examples:

Profile section	Skills loaded
`design-domain`	system-structure, invariants, dependencies
`design-scenarios`	edge-cases, error-handling
`design-intake`	problem-framing, feature-discovery
`verify-contract`	usability-evaluation, research-synthesis

During design, the orchestrator loads profiles sequentially or in parallel (when the host supports it) to build the contract section by section.

Subagent patterns

Used during verify and brownfield design:

Pattern	Purpose	When
persona-panel	One subagent per actor	`/lamina-verify`
visual-walkthrough	Browser capture at `base_url`	Verify and brownfield
parallel-review	Multiple expert lenses in parallel	Design review, a11y
fresh-context	Bounded brownfield repo scan	Init and design on existing code

Advanced: Subagents return fragments to the orchestrator — they do not write files directly. This enforces the .lamina/ write boundary.

Persona panel

One subagent per primary: true actor in personas.json. Each walks the live app independently, checking their workflows and permissions.

Visual walkthrough

Captures screenshots and step annotations at base_url. Evidence lands in walkthrough/ and informs findings.

Parallel review

During design, loads accessibility and heuristic review skills in parallel when supported — catching issues before implementation.

Fresh context

Bounded repo scan for brownfield projects. Reads source files and may walk live UI to ground contracts in existing behavior.

Init gate

Before design or verify workflows run, the orchestrator checks prerequisites/init-required.md. Without valid business-context.md, it returns the init-blocked output and stops.

Merge rules

merge-rules.md defines how skill outputs combine into run.json — grounding requirements, conflict resolution via decisions.md, and incremental writes.

Key files in the repo

File	Purpose
`skills/lamina-orchestrator/SKILL.md`	Orchestrator entry point
`skills/lamina-orchestrator/artifacts.md`	Schema and lifecycle
`skills/lamina-orchestrator/audit-profiles.yaml`	Section → skill mapping
`skills/lamina-orchestrator/workflows/*.md`	Command behavior
`skills/lamina-orchestrator/patterns/*.md`	Subagent patterns
`skills/lamina-core/guardrails.md`	Write boundaries

Next steps

Skill map — all capability skills
Validation — validate contracts
LaminaBench — benchmark methodology