51 skills, generated five ways
Amp, Codex, Pi, OpenCode and DeepSeek Harness receive deterministic projections of the canonical Claude skills, including complete resource trees, non-Markdown files and executable modes.
v3.1.0 · full Claude plugin · four generated editions
One Phoenix playbook. Five AI coding runtimes.
Plausible Elixir is not idiomatic Phoenix. phxagents v3 gives Claude Code, Amp, Codex, Pi, OpenCode and DeepSeek Harness the same 51-skill field manual. Claude Code also gets the full 26-agent, hooks and orchestration layer.
51
Canonical skills
26
Claude Code agents
10
Workflows adapted
26
Iron Laws
23
Enforcement hooks
10
Lifecycle events
Every number above is read from the plugin at build time. What measurement said about them afterwards, including which skills have never run in real work, is in the origin story.
What changed in v3.
The release is broader than renamed commands. Canonical skills are projected into native package shapes, core workflows are rewritten around each runtime’s tools, and unsupported Claude-only capabilities stay explicit.
10 workflows adapted, not merely renamed
Each keeps its real contract—artifacts, read-only review, user gates, resumable progress and verification—with optional native workers and a complete sequential fallback.
Claude orchestration stays native
Claude Code remains the full plugin: 26 named agents, 23 lifecycle hooks, permission settings, root instructions and its specialist fan-out.
Client payloads are untrusted
New LiveView guidance treats phx-value-*, form and hook payloads as user-controlled. Exposed IDs are not flaws by themselves; every event must authorize against server-side state.
Runtime-honest safety
Generated targets never pretend Claude hooks or named agents are installed. Codex alone gains an optional, trust-gated destructive-command safeguard; other enforcement remains deferred.
Learned fixes become real skills
/phx:learn-from-fix can now save package-specific lessons as native personal or project skills, with locked-version lookup and safe shadowing checks.
The portable workflow layer.
These are the workflows v3 explicitly adapts across every generated runtime. Other skills still ship as useful domain guidance, but may retain Claude-specific orchestration references.
- Investigate + reviewphx-investigate
Reproduce before fixing; review changed files without mutating them. Run independent tracks with native workers or sequentially.
- Plan + workphx-plan
Keep the canonical plan schema, scratchpad research, task checkboxes and resumable
progress.md. - PR review + fullphx-pr-review
Preserve read-only triage, explicit mutation approval, bounded retries and the complete gated lifecycle.
- Trace + audit + researchphx-audit
Use direct
mix xref, five health tracks, and native web or HTTP tools with local-source fallbacks.
The layer that doesn’t rely on being read.
Skills and agents are instructions a model may or may not follow. Hooks are shell scripts that always run. Session analysis across 400 sessions measured CLAUDE.md prose rules firing ~0% of the time — so anything that genuinely must happen moved out of prose and into 23 auditable bash scripts across 10 lifecycle events.
Every Elixir-specific hook gates on mix.exs, so the plugin is inert in a
non-Elixir repository, and all of them fail open — a broken hook means the guard is off,
never that the session is unusable.
Read how the hooks work →
Choose the runtime you already use.
Every edition ships the same canonical skill knowledge. Packaging, invocation, hooks, custom agents and MCP setup follow each runtime’s real capabilities—not a lowest-common-denominator promise.
Install v3 for your runtime.
Choose the runtime you already use. If hooks or multi-agent orchestration matter to your workflow, compare support first.