// the find
oliver-kriska/claude-elixir-phoenix
Claude Code plugin for Elixir/Phoenix/LiveView — 26 specialist agents, Iron Laws enforcement, and Tidewave MCP integration. Plan features with parallel research agents, execute with automatic verification, review with 4-agent parallel audits, and capture learnings as reusable knowledge.
A Claude Code plugin that bundles 26 specialist subagents, 51 skills, and 23 deterministic hooks to bake Elixir/Phoenix/LiveView/Ecto/Oban conventions (float money fields, assign_new on reconnect, Oban idempotency) directly into Claude's plan/work/review loop. For teams already running Claude Code against a real Phoenix codebase who want enforced house rules instead of a CLAUDE.md nobody reads.
The hooks are actual shell scripts that scan diffs for known antipatterns, so the guardrails run every time instead of being instructions a model can forget three turns into a session. The context-supervisor pattern solves a real problem: fanning out 4-8 research or review agents can return 50k+ tokens that would otherwise blow out the orchestrator's context, so it compresses and dedupes before the parent ever sees it. There's an actual eval harness (lab/eval, lab/plugin_eval) with per-skill graders, so skill triggering and plan quality are checked against something, not just vibes. Hooks self-gate on mix.exs, so installing it globally doesn't leak Elixir-specific behavior into unrelated repos.
The surface area is large and still expanding: 26 agents, 51 skills, 23 hooks, plus generated editions for five other runtimes (Amp, Codex, Pi, OpenCode, dsh), and the docs already admit those editions aren't at parity — that's a lot to keep synchronized as Claude Code's plugin API changes under it. The upgrade path is visibly fragile: the README spells out a multi-step, order-dependent recovery procedure for the case where a plain `claude plugin update` silently drops all 36 /phx:* commands because of an undeclared dependency. /phx:plan and /phx:review spawn several parallel subagents, some on Opus, per invocation, with no cost guidance anywhere, so routine feature planning could get expensive fast without anyone noticing. The actual enforcement mechanism for anything not covered by the hardcoded hook patterns is still just a model following written rules, so the Iron Laws are only as reliable as the handful of antipatterns someone thought to grep for.