| name | Swarm Architect |
| description | Orchestrate multi-agent delivery with phase→wave→swarm planning, worker bootstrap packets for fresh CLI sessions, GitHub synchronization, and OpenViking-ready memory capture. |
| globs | ["DesignSpec.md","ProjectArchitecture.md",".context/**/*.md","docs/architecture/**/*.md"] |
| requiredSources | ["github"] |
Swarm Architect
Use this skill to design and orchestrate multi-agent execution-ready delivery plans from specs, architecture docs, and repository context.
This skill is not a lightweight checklist generator. It is the orchestration layer for:
- discovery,
- plan shaping,
- phase → wave → swarm decomposition,
- contract-first parallelization,
- GitHub issue synchronization,
- and collision-resistant multi-agent delivery.
When to Use
Activate this skill when the work involves any of the following:
- architecture or implementation planning with 3+ meaningful steps,
- multiple independent execution tracks,
- multiple coding agents working in the same repository,
- GitHub issue creation/update from a delivery plan,
- phase/wave-based rollout or hardening,
- or a need for explicit validation and dependency management.
If the task is trivial, use a lighter planning flow instead.
Mandatory Companion Files
When this skill is invoked, load these files before producing the plan.
Operating rules
playbooks/multi-agent-boundaries.md
playbooks/worktree-strategy.md
playbooks/verification-gates.md
playbooks/github-sync.md
playbooks/claude-codex-copilot-gemini-operating-model.md when the active stack matches that workflow
playbooks/openviking-memory-ops.md when memory-aware swarm execution is requested
Output scaffolds
templates/discovery-template.md
templates/phase-wave-swarm-template.md
templates/github-issue-template.md
templates/agent-handoff-template.md
templates/openviking-memory-capture-template.md when memory records need to be written explicitly
templates/shared-contract-packet-template.md when fresh worker CLI sessions must share frozen contracts
templates/cli-session-bootstrap-template.md when launching a fresh Codex, Copilot, Gemini, or similar worker session
templates/validation-brief-template.md when launching a dedicated validation session
Structured defaults
schemas/task-schema.json
schemas/issue-mapping-schema.json
schemas/agent-role-matrix.yaml
schemas/runtime-role-catalog.yaml when runtime responsibilities or owner role clarity are needed
Lightweight runbooks
runbooks/spec-to-plan.md when starting from a feature brief, spec, or architecture request
runbooks/plan-to-github.md when the plan must become GitHub-tracked execution work
runbooks/wave-close.md when a wave needs integration closeout and next-wave readiness review
runbooks/validation-gate.md when validation requirements need to be explicit before execution
runbooks/memory-capture.md when OpenViking-compatible memory records must be shaped
runbooks/launch-worker-session.md when a fresh worker CLI session must be launched from a scoped handoff packet
runbooks/superset-workspace-bootstrap.md when Superset setup/teardown automation should provision worker workspaces from generated packets
Usage references
examples/sample-plan.md when the user asks for a sample full plan
examples/sample-wave.md when the user asks for a detailed wave example
examples/sample-agent-assignment.md when the user asks how agents should be split
docs/openviking-memory-mapping.md when memory-aware swarm execution or retrieval design is in scope
docs/bootstrap.md when first-run environment setup, upstream cloning, or OpenViking preparation is requested
machine-readable/reference.yaml when a compact package map is needed
machine-readable/workflows.yaml when workflow selection is ambiguous
machine-readable/profiles.yaml when role/profile overlays need a compact summary
machine-readable/session-bootstrap-schema.yaml when generating launch manifests or worker startup packets
Adjacent Skills to Leverage
When available in the active runtime, use these as companion protocols:
using-superpowers → skill discovery and ordering discipline
dispatching-parallel-agents → independent parallel swarm execution
coding-agent → launching supported coding agents safely when implementation is delegated
gemini → test design, regression analysis, and adversarial validation
Workflow
Before planning, choose the closest operating path:
- use
runbooks/spec-to-plan.md for normal planning from specs or architecture briefs
- use
runbooks/plan-to-github.md when GitHub execution tracking is a first-class output
- use
runbooks/validation-gate.md when the quality bar or release gate must be made explicit early
- use
runbooks/wave-close.md when assessing integration readiness at a wave boundary
- use
runbooks/memory-capture.md when OpenViking-compatible records or retrieval paths must be shaped
- use
runbooks/launch-worker-session.md when a fresh Codex, Copilot, Gemini, or similar CLI session must be started from a scoped handoff packet
- use
runbooks/superset-workspace-bootstrap.md when Superset should provision and clean up the worker workspace automatically
If workflow selection is ambiguous, consult machine-readable/workflows.yaml before proceeding.
1. Discovery is mandatory
Run a short discovery pass unless the user already supplied the answers.
Capture at minimum:
- planning depth:
lean | standard | deeply detailed
- delivery mode:
prototype | production | hardening
- CI/CD expectations:
none | basic | production-grade
- release model:
single milestone | phased rollout
- quality bar: testing, observability, performance, security
- team topology:
solo | small squad | multi-squad
- coding agents available and intended role split
- external constraints: deadline, compliance, platform, data, or regulatory constraints
Use templates/discovery-template.md as the canonical prompt/output frame.
2. Load planning context
Prioritize these inputs when present:
DesignSpec.md
ProjectArchitecture.md
.context/architecture/overview.md
.context/architecture/patterns.md
.context/auth/overview.md
.context/testing.md
.context/workflows.md
.context/errors.md
.context/api/headers.md
.context/feature-flags.md
.context/performance.md
.context/monitoring.md
.context/ui/patterns.md
If the user asks to process all context docs, enumerate relevant .context/** files instead of silently omitting them.
3. Build the plan with strict structure
Default target: 80 tasks.
Minimum granularity: 70 tasks unless the user explicitly wants smaller scope.
Required hierarchy:
- top-level Phases,
- Phase 1 split into at least 3 Waves by default,
- each Wave containing at least 2 Swarms,
- each Swarm focused on one primary concern.
Use the schema in schemas/task-schema.json for every task.
4. Enforce contract-first parallelism
Before assigning parallel build work:
- freeze API contracts,
- freeze shared schema/type expectations,
- freeze config/env expectations,
- freeze UI integration boundaries,
- and define validation evidence up front.
Do not send multiple agents into overlapping implementation zones without an explicit integration contract.
5. Apply multi-agent safety rules
The default operating model is:
- one issue → one owner → one branch/worktree,
- shared-file lock zones are serialized,
- contract changes force re-planning or an integration swarm,
- and merges happen at wave boundaries, not continuously.
Use playbooks/multi-agent-boundaries.md and playbooks/worktree-strategy.md to define safe ownership.
6. Synchronize with GitHub
When the user wants execution tracking:
- map tasks to GitHub issues,
- preserve dependencies in issue text/checklists/comments,
- align phase/wave/swarm IDs with issue metadata,
- and post concise wave summaries as work advances.
Use playbooks/github-sync.md and templates/github-issue-template.md as the canonical mapping rules.
7. Verify before declaring done
Every wave must include dedicated validation tasks.
Never mark work complete without evidence such as:
- tests,
- logs,
- status checks,
- screenshots,
- metrics,
- or diff validation.
Use playbooks/verification-gates.md for the completion standard.
Output Contract
A Swarm Architect response should present, in order:
- Discovery summary
- Assumptions and constraints
- Agent ownership model
- Phase map
- Detailed Phase 1 Wave/Swarm layout
- Full task list using the defined schema
- Dependency rationale
- Verification strategy
- GitHub sync and dispatch strategy
- Worker bootstrap packet strategy when fresh external CLI workers will be launched
- Risks and fallback plan
Default Agent Model
Unless the user overrides it, use this baseline split:
- Planner / orchestrator agent → orchestration, planning, issue graph, review
- UI / app implementation agent → architecture implementation, UI, app integration
- Cloud / backend agent → cloud, infra, deployment, backend integration
- Validation agent → tests, regressions, adversarial validation
Example mapping when available:
- Claude can fill the planner / orchestrator role
- Codex can fill the UI / app implementation role
- Copilot can fill the cloud / backend role
- Gemini can fill the validation role
Refine ownership using schemas/agent-role-matrix.yaml and schemas/runtime-role-catalog.yaml.
Definition of Done
A Swarm Architect plan is done only when:
- discovery is complete,
- required context has been loaded,
- structure includes phases, waves, and swarms,
- tasks are schema-complete and dependency-aware,
- agent ownership and collision boundaries are explicit,
- verification strategy is explicit,
- and GitHub synchronization is defined when requested.