If you've been watching AI coding assistants stumble through Phoenix projects, phxagents.dev is the intervention you've been waiting for. This plugin teaches Claude Code โ€” stable today โ€” how Phoenix actually works by loading 45 skills, 22 agents, and 22 Iron Laws into your development environment in about 30 seconds. The project just landed on Hacker News with a stable release targeting Claude Code, while support for Codex, OpenCode, and Pi drops together in v3.0. The core problem here is one every Elixir developer knows too well: generic AI editors hallucinate Repo.preload calls in the wrong places, swallow exceptions with bare rescues instead of proper error tuples, and skip Ecto.Multi when transactions actually matter. These tools don't understand Phoenix-shaped patterns โ€” they just pattern-match on code they scraped from GitHub. phxagents flips that by giving your AI editor actual domain knowledge about contexts, LiveView async patterns, Oban idempotency, and the production safety rails that keep generated Elixir code from becoming an incident at 2 AM. Enter the Iron Laws: 22 non-negotiable rules enforced at both generation time and during pull request review. No Repo in loops. No bare rescues. No untyped Oban args. No missing preload calls before association access. The judge agent literally blocks PRs that violate these constraints, which means your AI pair programmer has actual taste โ€” or at least better brakes than the baseline models. The plugin also auto-installs 128 references into CLAUDE.md and registers activation rules so context loads intelligently without manual prompting. The workflow isn't just autocomplete with extra steps โ€” it's a four-phase cycle designed for real engineering work: /phx:plan โ†’ /phx:work โ†’ /phx:review โ†’ /phx:compound. Each phase produces artifacts the next phase reads, and every step has compile, format, and test gates built in. Generated code that doesn't pass mix test simply doesn't get committed. The PlanSpecialist agents draft multi-step plans with specific files, schemas, and Oban jobs to touch, while the verification-runner agent handles context-aware reads of your mix.exs to discover Credo, Dialyzer, Sobelow, and ExCheck configurations โ€” honoring custom aliases and knowing whether you're on Phoenix 1.7 versus 1.8.

Four-Track Parallel Review

The /phx:review command runs four specialists in parallel against every diff: elixir-reviewer, security-analyzer, testing-reviewer, and verification-runner each read the changes in fresh context before findings are deduplicated and triaged interactively. This isn't a single-pass lint โ€” it's structured peer review from agents that actually understand your stack. The compound docs system captures every fixed bug as a searchable solution document with YAML frontmatter, so next time you hit a similar Ecto constraint error, the answer is one query away instead of an hour of debugging. Supply chain security gets attention too with /phx:deps-audit and /phx:deps-vet commands that audit Hex dependencies for bidirectional character tricks, compile-time exec vulnerabilities, maintainer changes, typosquats, and CVEs. Vetted versions get recorded in hex_vet.exs so you don't re-review the same package twice โ€” a genuinely useful feature when your dependency tree is deeper than a few top-level packages.

Current Status and Roadmap

Shipping today on Claude Code with full Phase 1 + Phase 2 parity from a single source codebase. The v3.0 release targets Codex, OpenCode, and Pi adapters simultaneously with complete feature parity โ€” tracked in PR #46 for the curious. MIT licensed means you can fork it, audit it, and extend it without licensing headaches. Install takes 30 seconds: one plugin manifest in your .claude folder, then Iron Laws and all 45 skills auto-load.

Key Takeaways

  • phxagents brings Phoenix-specific context to AI editors that currently hallucinate Repo calls and skip Ecto.Multi
  • 22 Iron Laws block PRs violating production safety rules โ€” no bare rescues, no Repo in loops, proper Oban typing
  • Four-phase workflow (/phx:plan โ†’ /phx:work โ†’ /phx:review โ†’ /phx:compound) with test gates at each step
  • Stable on Claude Code today; Codex, OpenCode, and Pi support lands together in v3.0

The Bottom Line

This is exactly the kind of tooling the Elixir ecosystem needs right now โ€” not another prompt template, but actual enforcement mechanisms that make AI-assisted development safer. If you've been burned by AI-generated code that passed all tests but broke at scale, phxagents' Iron Laws approach looks like a real step toward trustworthy pair programming with LLMs.