made claude.md

agents.md is not what I wanted. but it is a step.

AGENTS.md

@@ -0,0 +1,31 @@
+# AGENTS.md
+
+Subagent instructions for this repository. Keep prompts minimal.
+
+## Context
+
+Porting allium (Claude Code plugin) to pi.dev skills. Upstream lives in `allium-main/`. Port output goes in `.pi/skills/`. See `allium-port-plan.md` for full plan.
+
+## Agent: skill-porter
+
+Adapt a Claude Code skill to a pi.dev skill.
+
+**Input:** upstream skill path (e.g. `allium-main/skills/elicit/SKILL.md`)
+**Output:** `.pi/skills/{name}/SKILL.md` + symlinked `references/`
+
+Steps:
+1. Read upstream SKILL.md
+2. Replace frontmatter: add `disable-model-invocation: true`, `license: MIT`, `metadata` block
+3. Replace Claude invocation references (`/allium:X`) with pi references (`/skill:X`)
+4. Preserve body content verbatim
+5. Create `references/` symlink to upstream references dir (relative path)
+
+## Agent: smoke-tester
+
+Verify a ported skill loads and responds correctly.
+
+**Input:** skill name
+**Steps:**
+1. Check `.pi/skills/{name}/SKILL.md` exists and has valid frontmatter
+2. Check `references/` symlink resolves
+3. Report ready for manual `pi -p` test with `/skill:{name}`

CLAUDE.md

@@ -0,0 +1,45 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Project
+
+Port of [allium](https://github.com/juxt/allium) (a Claude Code plugin by JUXT) to [pi.dev](https://pi.dev), keeping upstream files by reference (symlinks) rather than forking. Test target: the [turn-limit extension](https://gitea.apps.sustainabledelivery.com/QWAN/monotonic-pi-extensions).
+
+See `allium-port-plan.md` for the full phased plan and task checklist.
+
+## Architecture
+
+```
+pi-allium-port/
+├── allium-main/              ← git clone of juxt/allium (upstream, read-only)
+├── turn-limit/               ← symlink to turn-limit-repo/packages/pi-turn-limit
+├── .pi/skills/               ← pi.dev skill definitions (the port output)
+│   ├── allium/SKILL.md       ← root skill, routing table
+│   ├── elicit/SKILL.md       ← elicitation sub-skill
+│   ├── distill/SKILL.md      ← distillation sub-skill
+│   └── propagate/SKILL.md    ← propagation sub-skill
+└── allium-port-plan.md       ← plan and checklist
+```
+
+Each skill's `references/` directory symlinks to the corresponding upstream directory in `allium-main/`.
+
+## Key platform differences (Claude Code → Pi.dev)
+
+- Skill location: `.pi/skills/{name}/SKILL.md` (not `.claude-plugin/`)
+- Invocation: `/skill:name` (not `/name` or `/name:sub`)
+- Frontmatter: uses `disable-model-invocation`, `license`, `metadata` (not `auto_trigger`, `version`)
+- No native agent or hook support in pi.dev — workarounds needed for tend/weed agents and allium-check validation
+- Rules content must be inlined into SKILL.md or references (no glob-triggered rules)
+
+## Commands
+
+Pi.dev is invoked with `pi -p` to load project skills from `.pi/skills/`.
+
+Default model: `Qwen3.6-35B-A3B-MXFP4_MOE.gguf` via llama.cpp at `http://127.0.0.1:8000/v1`.
+
+## Conventions
+
+- Symlinks must use relative paths so they resolve on any machine with `allium-main/` cloned in the same position
+- Allium language version is 3 (check `allium-main/VERSION`)
+- When adapting upstream SKILL.md files, preserve the body content and only change frontmatter and invocation references