Legion

Legion routes AI-assisted engineering, research, editorial, design and commercial work to the right capability, keeps decisions, changes and independent checks separate, and records evidence before anything is called delivered. It is one doctrine projected into Claude Code, Codex or an AGENTS.md harness.

Set up

Public package-manager installation is not open yet. Current shipping contract is signed native Legion; Node package remains private development tooling. From a source checkout, Node 22.13+, pnpm, & Rust toolchain are required.

pnpm install --frozen-lockfile
cargo build --manifest-path engine/Cargo.toml --release --bin legion
engine/target/release/legion --version

After installing a signed release-bound binary through an approved channel:

legion setup --dry-run
legion setup --confirm
legion setup --check

Setup detects supported clients, previews exact persistent changes, projects canonical skills & MCP integration through registered host adapters, records rollback state, then verifies installed host-visible state. setup --check makes no projection changes, but completes any pending rollback journal. setup remove --confirm removes Legion-owned projections while preserving user-owned config.

Claude Code receives packaged plugin surfaces. Codex receives canonical Agent Skills under ~/.agents/skills plus MCP registration. Portable Agent Plugins receive the portable core, every public canonical skill & MCP declaration, assembled with pinned RightAX tooling.

Harnesses & fidelity

One doctrine & kernel project into harness-native slots. Legion does not pretend every host exposes identical controls.

Run legion doctor to inspect repository, binding, coverage, & host state. legion bind --registrations inventories installed hooks across settings & plugins and detects duplicate registrations.

How work flows

Legion begins with current user intent. Legion selects zero or more capabilities from compact public catalog, deterministic runtime validates IDs, & Legion materializes work graph. Legion attaches authority only where work requires it: Sage (optional) for cross-cutting design, reassessment, & material unresolved meaning/ownership/acceptance, Alchemist for bounded implementation (ambient unless work is locked or explicitly contracted), & Oracle (optional) for independent Completion Validation on explicit request or a concrete outcome or safety risk. Arcane owns cognitive processing & response policy. Guard deterministically gates typed effects, reports enforcement health, & owns effect decisions; persisted receipts are future work. Neither can invent consent.

Work splits into independent units where safe, then reunites at delivery. Capabilities supply expertise, method, workflow, & context; capability never grants authority. Optional domains group catalog entries for discovery only.

Engineering cohort

Oracle is optional independent assurance authority for Completion Validation; it is never a delivery gate.

Catalog grouping

Current catalog uses these optional discovery groups, not routing authorities or a fixed hierarchy:

• Commercial — marketing, advertising, social, SEO
• Research — general, scientific, & market evidence
• Editorial — writing & editorial work
• Design — interface, visual, & brand-identity work
• Engineering — architecture, diagnosis, QA, and platform development

Skills supply reusable methods. Legion owns semantic selection, work graph, evidence, & delivery state across all groups.

Apple development

Use /ios-development for iOS/iPadOS and /macos-development for Mac applications or their Apple-native components. Both are automatically discoverable engineering capabilities with short entrypoints and task-selected references for SwiftUI, concurrency, testing, persistence, performance, system integration, and release preparation. They preserve the project’s deployment targets, architecture, data layer, and any Rust/Tauri/Cargo host.

Each bundle is self-contained for native skill projection. Xcode, SwiftPM, and optional build/UI/documentation/release adapters are checked only for the selected workflow; the skills include official-source setup recipes and a read-only PATH preflight helper. Agents detect and reuse tools, propose missing setup, then apply only authorized changes and verify. CLI setup persists per environment and MCP registration per selected client scope, shared by both skills. Loading a skill grants no installation, persistent-access, account, signing, telemetry, upload, or publishing permission. See packaged skills and third-party notices for provenance and tool boundaries.

Audit engine

Run an audit & verify its evidence with:

legion audit .
legion verify .audit/

Audit freezes revision & file inventory, seals a deterministic provider plan, then runs checks. Security candidates remain separate from adjudication. Output includes JSON, SARIF & Markdown; skipped required checks, stale inventory, missing adjudication, or unproven coverage prevent a clean verdict. See legion --help for current commands.

The flow makes this evidence chain explicit:

flowchart LR
  A["freeze revision + inventory"] --> B["select & seal plan"]
  B --> C["run deterministic providers"]
  C --> D["adjudicate candidates"]
  D --> E["reason over redacted facts"]
  E --> F["JSON · SARIF · Markdown"]

Repository

• Doctrine
• Full manual
• Provider architecture
• Engine interface
• License

Repository truth docs

• Canonical system architecture — ownership boundaries and global invariants
• Role architecture — Sage, Alchemist, Oracle, packaged skills
• Historical provenance — frozen records that cite deleted code; not authoritative

License

Source-available under the Orthic Labs Source Use License v1.0. See LICENSE.

---

Orthic Labs · github.com/Orthic-Labs · More: SellRight · Ember · SwiftRight · Pulse · RightKit · RustRight