@skathio/somi

SoMi

An opinionated, reusable multi-agent engineering workflow system for Claude Code and GitHub Copilot.

The latest published version is on npm and in the GitHub Releases; see CHANGELOG.md for release notes.

SoMi gives engineering teams a shared, version-controlled "operating system" for working with Claude: three first-class build workflows — plan → code → review — plus an upstream discovery workflow that turns a raw product idea into a research-grounded requirements & design foundation. All backed by specialised subagents, deterministic guardrail hooks, composable skills, and a global ruleset that enforces SOLID, clean code, and OWASP defenses.

It is designed to be:

• Reusable across many repositories and teams
• Installable as a Claude Code plugin (via marketplace or npm) or as a GitHub Copilot chat extension
• Opinionated but extensible — strong defaults, clean overrides
• Deterministic where possible (hooks) and judgment-heavy where needed (agents)

---

The workflows

Command	Tier	Agent	Purpose
/discover	MAX	discovery-analyst	Research the competition, then author the requirements & design foundation (BRD/SRS/FRD/SDD/TDD) + a brief.md for a new product
/design	MAX	designer	Settle a brownfield feature's architecture against the codebase; compile the brief.md the cheap tier executes against
/plan	ECO	planner	Sequence the design (brief) into phases, risks, slices, DoD, test & rollout strategy
/code	ECO	coder	Execute against an approved plan + brief with senior-level design judgment
/review	MAX	reviewer	Strict, skeptical, fresh-context review of code / plans / designs with severity-graded findings
/ship	both	planner+coder+reviewer	Full (optional MAX front-load →) plan → code → review pipeline, gated at every stage

Two economic tiers. The MAX tier (opus) front-loads expensive reasoning — research, design, decisions, complexity mapping, fresh-eyes review — into a dense, bounded brief.md. The ECO tier (sonnet) executes against that brief without re-researching, so the high-volume work (plan detail, iterative coding) runs cheaply. /discover (new product) and /design (brownfield feature) are the MAX front-loads that feed /plan; /ship-loop runs the whole pipeline continuously, gating once at the MAX→ECO model switch. Supporting MAX agents (by handoff): security-reviewer, architecture-reviewer, test-strategist, refactorer. See docs/AGENTS.md.

---

Install

Option 1 — Claude Code plugin (marketplace)

# Add the SoMi marketplace and install the plugin:
/plugin marketplace add https://github.com/skathio/somi
/plugin install somi@somi

Updates flow through /plugin update somi.

Option 2 — npm (Claude Code)

npm install -g @skathio/somi

Then in Claude Code: /plugin install somi.

Option 3 — GitHub Copilot (extension marketplace)

SoMi is also a GitHub Copilot extension, installable the same way as the Claude Code plugin:

copilot plugin marketplace add https://github.com/skathio/somi
copilot plugin install somi@somi

Once installed, use @somi in GitHub Copilot chat:

@somi /plan  Add per-team rate limiting to the public webhook endpoint
@somi /code  rate-limiting-webhooks phase 1, iteration 1
@somi /review  rate-limiting-webhooks

Parity caveat — Copilot is not feature-equivalent to Claude Code. Two things do not carry over, because they depend on Claude Code host capabilities Copilot doesn't expose:

• The deterministic guardrail hooks do not fire. Blocking dangerous bash, secret-writes, protected-path writes, dep-install gating, and the audit log are Claude Code hooks — on Copilot they are absent, so those safety nets are not enforced. The agent/skill judgment still applies, but the hard stops don't.
• Multi-agent orchestration degrades to sequential. The loops (/code-loop, /plan-loop, /ship) and the parallel commands (/review-panel, /code-parallel) drive Claude Code sub-agents via the Task tool. Where the host can't spawn sub-agents concurrently, these run one lens / one iteration at a time — same result, no parallelism.
• The MAX/ECO model split is a Claude Code feature. The economy depends on per-agent model tiering (opus for MAX, sonnet for ECO) and the cache-correct subagent-model split. Where the host runs a single model, the workflow shape (MAX front-load → dense brief.md → ECO execution) still holds and still helps — but the cost split does not. Priority is Claude Code; quality is not sacrificed for Copilot parity — Copilot gets the portable subset.

The commands, agents, skills, rules, and templates are shared; the enforcement, concurrency, and model-tiering layers are Claude Code features. Treat Copilot as the portable subset, not a drop-in equal. See docs/HOOKS.md and docs/PLUGIN.md.

---

What's in the box

.claude-plugin/   Plugin + marketplace manifests (Claude Code plugin distribution)
agents/           Subagent definitions (planner, coder, reviewer, + support)
commands/         Slash-command entrypoints (/plan, /code, /review, /ship, ...)
skills/           On-demand expert knowledge packs (OWASP, SOLID, test strategy, ...)
rules/            Global ruleset composed into CLAUDE.md
hooks/            Deterministic guardrails (block dangerous bash, secret writes, ...)
templates/        Artifact templates (BRIEF [MAX→ECO handoff], DESIGN, CONTEXT, SPEC, DECISIONS, PHASE, PROGRESS, DIARY, REVIEW, ADR, DOD; R&D: RD-README, RESEARCH, BRD, SRS, FRD, SDD, TDD)
.copilot-extension/ Copilot extension + marketplace manifests (mirrors .claude-plugin/)
examples/         Worked examples + a minimal consuming project
docs/             Full documentation

When you use SoMi in a project, workflows write their artifacts into a .somi/ directory at the project root. Discovery foundations live under .somi/rd/<slug>/ (research report, BRD, SRS, FRD, SDD, TDD); plans live under .somi/plans/<slug>/ (context, spec, decisions, progress, diary, phases); reviews live under .somi/reviews/<slug>/ — separate directories, no clutter. See docs/WORKFLOWS.md for the full layout.

---

Quick start (after install)

For a brand-new product, start one step earlier with discovery:

> /discover  A self-hosted alternative to Calendly for clinics, with HIPAA-aware
            scheduling, SMS reminders, and no per-seat pricing.

# Claude proposes a slug ("clinic-scheduler"), researches competitors and real user
# complaints (cited), pauses on each crossroads (persona, scope, build-vs-integrate,
# expensive-to-reverse architecture), and writes a traceable requirements & design set
# to .somi/rd/clinic-scheduler/ (research report, BRD, SRS, FRD, SDD, TDD). Then run
# /plan clinic-scheduler — the planner consumes that foundation.

For a design-heavy feature on an existing repo, front-load the design (MAX) so planning and coding run cheaply against it:

> /design  Add per-team rate limiting to the public webhook endpoint, sharing budget
          across replicas, with an emergency kill switch.

# Claude (opus) reads the codebase, ingests the repo's own CLAUDE.md/AGENTS.md once,
# resolves the expensive-to-reverse calls with you (storage backend, where the limiter
# lives), maps the complexity hotspots, and compiles .somi/plans/rate-limiting-webhooks/
# brief.md — the dense MAX→ECO handoff. Then /plan rate-limiting-webhooks sequences it
# into phases on the cheaper (sonnet) tier, /code-loop implements against the brief.

For an incremental change with a settled design (the daily loop), start at planning:

> /plan  Add per-team rate limiting to the public webhook ingestion endpoint
        with audit logging and an emergency kill switch.

# Claude proposes a slug ("rate-limiting-webhooks"), reads the repo, drafts context.md,
# then pauses inline on each architectural decision — presenting options with concrete
# pros/cons, plus "Other" and "Discover" escape hatches. You decide. Verified decisions
# land in decisions.md; the spec, phases, progress, and diary fill in. Review the
# artifacts at .somi/plans/rate-limiting-webhooks/, edit if needed, approve.

> /code  rate-limiting-webhooks phase 1, iteration 1

# Claude implements with senior-engineer judgment, writes tests, updates docs, and
# keeps the plan in sync — if implementation reveals the plan needs to change, it
# updates spec/decisions/phases in place and appends a diary entry.

> /review  rate-limiting-webhooks

# Claude returns severity-graded findings (written under reviews/), rejects weak
# solutions, flags plan-vs-code divergence.

For the all-in-one pipeline: /ship <problem statement>.

---

Why a shared OS, not per-project setups

• Consistency — every repo gets the same review bar, the same security posture, the same plan shape.
• Upgrade once — update the plugin; every project benefits.
• Override locally — projects keep their own CLAUDE.md and rules/99-overrides.md; SoMi never silently overrides them.
• Auditable — hooks log denied actions; reviewers can see what the system blocked vs. what humans approved.

---

Documentation

• Installation — Claude Code plugin, npm, or Copilot extension
• Usage — running each workflow with examples
• Workflows — plan / code / review semantics and handoffs
• Agents — full agent catalogue, escalation rules
• Hooks — guardrails and how to add your own
• Skills — on-demand expertise packs
• Rules — global ruleset philosophy and conflict resolution
• Commands — slash-command reference
• Extending — adding workflows, agents, skills
• Versioning — SemVer policy, breaking-change rules
• Governance — how teams adopt updates safely
• Plugin distribution — marketplace and VS Code setup
• Architecture — how the pieces fit together

---

Versioning

SoMi follows Semantic Versioning. Releases are automated from Conventional Commits: merging to main runs the publish workflow, which derives the next version, publishes to npm (with a signed provenance attestation via OIDC trusted publishing), pushes the v<version> git tag, and creates the matching GitHub Release.

The published git tags and the npm registry are the source of truth for the released version — the in-repo VERSION / package.json are not committed back on each release, so don't rely on them for "what's published." Check the npm page or Releases instead.

See docs/VERSIONING.md for the breaking-change policy and migration guide template.

---

License

MIT — see LICENSE.