soe

soe

Disciplined intent → plan → TDD → verify — driven autonomously by a crash-safe, human-debuggable engine that uses the best installed specialist for every job.

---

soe fuses three things into one Claude Code plugin:

• the Superpowers 6.1.1 discipline pipeline (brainstorm → plan → TDD → verify),
• a simplified, tested multi-agent orchestration engine — the Evaluate-Loop, adapted from Conductor (parallel workers, Board of Directors, quality gates, bounded loops) with every integrity-critical decision as real code, not prose in a prompt,
• and the best of Everything Claude Code (instincts, logging, self-audit) — plus session-model-led multi-model orchestration, token-efficient cross-plugin discovery, and a minimal-code discipline.

45 skills · 22 agents · 16 commands · 15 tested lib/ modules · 340 passing tests.

Features

• One engine, three front doors — /go (human brainstorm), /go-auto (autonomous spec), /go-all (dual derivation + reconciliation) all feed one autonomous Evaluate-Loop.
• Real multi-agent orchestration — the session model is the orchestrator; it dispatches tier-pinned leaf agents (planner, board, workers, evaluator, fixer) asynchronously and collects each native completion (no polling), applying results serially as the sole state writer.
• Session-model-led multi-model tiering — agents pinned to latest full model IDs (claude-opus-4-8 reasoning · claude-sonnet-5 mechanical · claude-fable-5 top-stakes · claude-haiku-4-5 cheap), with a config Fable gate (fable_enabled:false → falls back to Opus).
• Board of Directors + devil's-advocate — a cheap collapsed 5-lens board by default, a full 5-director vote for high-stakes, and an adversarial red-team gate before any plan executes.
• Guarantees in code, not prompts — 15 tested lib/ modules: atomic single-writer state, crash-safe idempotent resume, bounded fix/plan loops, a deterministic risk matrix + fail-safe scrutiny (only ever raises), a context-firewall validator, and the model resolver.
• Context firewall — delegated agents return only path + 3-line summary + confidence; full output stays out of the orchestrator's context.
• Worktree-isolated workers, mandatory TDD — parallel workers can't trample; every change is RED → GREEN → REFACTOR.
• Token-frugal by design — lean core (language depth from installed packs), collapsed-vs-full board by stakes, and an always-on minimal-code discipline (shortest working diff; code only).
• Cross-plugin capability discovery — auto-detects and reuses installed skills / agents / MCP servers (graphify, chrome-devtools, Codex, Figma) by role, degrading gracefully when absent.
• Escalation that learns — escalates only genuine judgment calls or irreversible actions, and learns from each so routine interruptions fade.
• Ambient, no ceremony — /simplify, /critique, /self-audit, /logging and the discipline skills all work standalone.

What This Project Provides

One engine, three front doors. A single autonomous loop (PLAN → EVALUATE_PLAN → EXECUTE → EVALUATE_EXEC → (FIX↺ | COMPLETE)) fronted by three entry commands that differ only in how the spec is derived:

Command	How the spec is derived
/go <goal>	Human-in-the-loop brainstorm first (aligns intent, cuts hallucination), then the loop
/go-auto <goal>	Autonomous spec from goal + codebase, then the loop
/go-all <goal>	Brainstorm ∥ an independent background auto-spec → 3-way reconciliation (you decide) → the loop

Real, not fictional. The reliability-critical machinery is tested Node code: atomic + single-writer state store, crash-safe resume with idempotency, bounded fix/plan loops, a deterministic risk matrix + fail-safe scrutiny, a context-firewall validator. Prompt-following handles the reasoning; code handles the guarantees.

Token-frugal by design. Lean core (language depth comes from installed plugins, not a monolith), context firewall on delegated work, collapsed-vs-full Board by stakes, and an always-on minimal-code discipline (write the shortest working diff — code only, never docs, never on high-stakes shortcuts).

How It Works

You state a goal with /go. Instead of guessing, the agent first sits down with you and brainstorms the intent — teasing a design out of the conversation, showing it back in digestible chunks, and binding an approved design doc to the track. (Already brainstormed? It skips straight to planning. Want it fully autonomous? Use /go-auto. Want both derivations cross-checked? /go-all.)