pi-allium-port/allium-port-plan.md

# Allium → Pi.dev Port Plan

## Goal
Port allium (https://github.com/juxt/allium) to pi.dev, keeping upstream files by reference where possible.
Test with the turn-limit extension (https://gitea.apps.sustainabledelivery.com/QWAN/monotonic-pi-extensions).

## Directory layout

```
pi-allium/                          ← this repo
├── allium-main/                    ← git clone of juxt/allium (submodule or checkout)
├── turn-limit/                     ← checkout of monotonic-pi-extensions/packages/pi-turn-limit
├── .pi/
│   └── skills/
│       ├── allium/                 ← root skill (entry point, routing table)
│       │   ├── SKILL.md            ← adapted frontmatter + symlink or include of upstream content
│       │   └── references/         ← symlinks to allium-main/references/*
│       ├── elicit/
│       │   ├── SKILL.md
│       │   └── references/         ← symlinks to allium-main/skills/elicit/references/*
│       ├── distill/
│       │   ├── SKILL.md
│       │   └── references/         ← symlinks to allium-main/skills/distill/references/*
│       └── propagate/
│           ├── SKILL.md
│           └── references/         ← (upstream has none currently)
├── allium-port-plan.md             ← this file
└── learning-goal.md                ← goal card
```

## Key differences: Claude Code plugin vs Pi skill

| Aspect | Claude Code | Pi.dev |
|--------|-------------|--------|
| Skill location | `.claude-plugin/` + `skills/` | `.pi/skills/{name}/SKILL.md` |
| Frontmatter | `name`, `description`, `version`, `auto_trigger` | `name`, `description`, `disable-model-invocation`, `license`, `metadata` |
| References dir | `resources/` or `references/` | `references/` |
| Agents | `.claude/agents/{name}.md` with `model:`, `tools:` | No native agent support — needs extension or workaround |
| Hooks | `.claude/settings.json` → PostToolUse hooks | Not directly supported — needs extension |
| Rules | `.claude/rules/*.md` with `globs:` | No equivalent — content goes into SKILL.md or references |
| Auto-trigger | `auto_trigger: file_patterns, keywords` | `disable-model-invocation: true/false` |
| Invocation | `/allium`, `/allium:elicit` | `/skill:allium`, `/skill:elicit` |

## Phases

### Phase 1: Setup and smoke test
_Goal: get the root `/skill:allium` working in pi with a local model._

- [x] Clone allium upstream into `allium-main/`
  ```
  cd ~/research/pi-allium
  git clone https://github.com/juxt/allium.git allium-main
  ```
- [x] Create `.pi/skills/allium/SKILL.md` with pi-compatible frontmatter
  - Source: `allium-main/skills/allium/SKILL.md`
  - Change: add `disable-model-invocation: true`, `license: MIT`, `metadata` block
  - Change: update routing table to use `/skill:elicit`, `/skill:distill`, `/skill:propagate` instead of Claude skill/agent references
  - Change: update reference paths from `../../references/` to `references/`
  - Keep: the full language reference body
- [x] Symlink references: `ln -s ../../../allium-main/references .pi/skills/allium/references`
- [x] Smoke test: `pi -p` in this directory, invoke `/skill:allium`, verify it loads and responds with the routing table
- [x] Verify references load: ask allium about language syntax, confirm it can read `references/language-reference.md`

### Phase 2: Port elicit, distill, propagate skills
_Goal: all three sub-skills work via `/skill:elicit` etc._

- [x] Create `.pi/skills/elicit/SKILL.md` — adapt frontmatter from `allium-main/skills/elicit/SKILL.md`
- [x] Symlink references: individual symlinks in `.pi/skills/elicit/references/` (language-reference.md + library-spec-signals.md)
- [x] Smoke test `/skill:elicit` — start a mini elicitation session, verify it follows the methodology
- [x] Create `.pi/skills/distill/SKILL.md` — adapt frontmatter from `allium-main/skills/distill/SKILL.md`
- [x] Symlink references: individual symlinks in `.pi/skills/distill/references/` (language-reference.md + worked-examples.md)
- [x] Smoke test `/skill:distill`
- [x] Create `.pi/skills/propagate/SKILL.md` — adapt frontmatter from `allium-main/skills/propagate/SKILL.md`
- [x] Smoke test `/skill:propagate`

### Phase 3: Test with turn-limit extension (TDD)
_Goal: use distill → propagate on real code, verify allium produces useful output._

- [ ] Checkout turn-limit into this workspace
  ```
  git clone https://gitea.apps.sustainabledelivery.com/QWAN/monotonic-pi-extensions.git turn-limit-repo
  ln -s turn-limit-repo/packages/pi-turn-limit turn-limit
  ```
- [ ] Run `/skill:distill` against `turn-limit/` — extract a `.allium` spec from existing code
- [ ] Review the generated spec: does it capture turn-limit constraints, enable/disable, UI separation?
- [ ] Run `/skill:propagate` against the generated spec — generate test suggestions
- [ ] Write at least one test based on propagate output (TDD red step)
- [ ] Implement to make the test pass (green step)
- [ ] Iterate: use allium to identify more test obligations, write them, make them pass
- [ ] Run `/skill:elicit` to explore the "disable turn limit" feature requirement
- [ ] Generate spec + tests for the new feature via propagate

### Phase 4: Fold in allium rules content
_Goal: pi agent knows allium syntax rules when editing .allium files._

- [ ] Merge content from `allium-main/.claude/rules/allium.md` into `SKILL.md` or a reference file
  - This contains syntax gotchas and anti-patterns that the model needs when writing .allium files
  - In Claude Code this auto-triggers on `**/*.allium` globs — in pi we need it in the skill context
- [ ] Test: create/edit a `.allium` file via pi, verify the model follows naming conventions and avoids anti-patterns

### Phase 5: Agent support (tend & weed)
_Goal: find a way to support tend/weed agent functionality in pi._

Tend and weed are Claude Code agents with `model: opus` and specific tool permissions (Read, Glob, Grep, Edit, Write, Bash). Pi doesn't have native agent support.

Options to investigate:
- [ ] **Option A: Skills as agents** — port tend.md and weed.md as skills with `disable-model-invocation: true`. The model won't have enforced tool restrictions but the skill prompt still guides behavior. Simplest path.
- [ ] **Option B: Pi extension (TypeScript)** — write a TypeScript extension that registers `/tend` and `/weed` commands, controlling context injection. More control, more work.
- [ ] **Option C: npm agents package** — check if pi.dev has an agents extension or if one exists on npm (`npx skills` ecosystem). Check https://pi.dev/docs and https://github.com/nicepkg/agent-skills.
- [ ] Decide on approach and implement
- [ ] Test: use tend to modify a `.allium` spec, verify it follows allium conventions
- [ ] Test: use weed to check spec-code alignment on the turn-limit extension

### Phase 6: Hook support (allium-check validation)
_Goal: .allium files get validated on write._

The Claude Code version runs `allium-check.mjs` as a PostToolUse hook on Edit/Write.

- [ ] Investigate pi.dev hook/extension support for post-write validation
- [ ] If supported: configure the hook pointing to `allium-main/.claude/hooks/allium-check.mjs`
- [ ] If not: add validation instructions to the skill prompt ("after writing .allium files, run `node allium-check.mjs <file>`")
- [ ] Test: write a `.allium` file with deliberate errors, verify validation catches them

### Phase 7: Documentation and upstream
_Goal: publishable result._

- [ ] Add README.md with attribution (MIT license from JUXT)
- [ ] Add LICENSE
- [ ] Document the symlink-based approach for staying in sync with upstream
- [ ] Document known limitations vs Claude Code version
- [ ] Write blog post draft (separate file)
- [ ] Share with allium team and pi.dev community

## Key URLs
- Allium upstream: https://github.com/juxt/allium
- Allium docs: https://juxt.github.io/allium/
- Allium installation: https://juxt.github.io/allium/installation
- Turn-limit extension: https://gitea.apps.sustainabledelivery.com/QWAN/monotonic-pi-extensions
- Pi.dev: https://pi.dev
- Local allium cache (current): `~/.claude/plugins/cache/juxt-plugins/allium/f6fb08dd301f/`

## Notes for the agent on the other machine
- Pi is invoked with `pi -p` to load project skills from `.pi/skills/`
- Default model: `qwen3.5:35b-a3b` via Ollama on `192.168.0.1:11434`
- Skills are invoked with `/skill:name` (requires `enableSkillCommands: true` in pi settings if not already set)
- Symlinks must resolve on the target machine — clone allium-main in the same relative position
- The `.allium` language version is 3 (check `allium-main/VERSION`)