# Model Menu + Persistent Daemon Implementation Plan

> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.

**Goal:** Add a model menu (3 permissive SDXL models), a persistent HTTP daemon that holds one model resident in XPU VRAM for fast prompt iteration, and docker compose lifecycle, to the local Intel Arc image generator.

**Architecture:** `generate.py` is refactored to expose a model registry and reusable `resolve_model`/`build_pipe`/`generate` functions, keeping its one-shot CLI. A new `server.py` imports that core and serves `/status`, `/generate`, `/model` over stdlib HTTP on localhost. A `compose.yaml` runs the daemon as the host user; the `imggen` wrapper gains `start`/`stop`/`status`/`list` subcommands and routes prompts to the daemon (with a one-shot fallback).

**Tech Stack:** Python 3 stdlib (`http.server`, `json`, `argparse`), diffusers `StableDiffusionXLPipeline` + `AutoencoderKL`, torch XPU, docker compose v2, bash.

---

## File Structure

- `generate.py` (modify) — model registry, `resolve_model`, refactored `build_pipe(key)` with variant fallback, unchanged `generate`, `-m/--model` CLI flag.
- `server.py` (create) — stdlib HTTP daemon holding one resident pipe; imports from `generate.py`.
- `compose.yaml` (create) — daemon service, runs as host user, GPU + mounts + port.
- `.env.example` (create) — documents `.env` shape (real `.env` is gitignored, already handled).
- `imggen` (modify, in `~/bin/`) — subcommand dispatch + daemon client + one-shot fallback + interactive menu.
- `test_generate.py` (create) — pure-logic asserts (registry, resolve_model).
- `README.md` (modify) — new usage, lifecycle, VRAM warning.
- `CLAUDE.md` (modify) — new architecture, corrected defaults.

Note: `.gitignore` already committed (protects `.env`). HF cache chown + `.env` creation are done by the user, not tasks here.

---

## Task 1: Model registry + resolve_model (pure logic, TDD)

**Files:**
- Modify: `generate.py:27-30` (replace `MODEL`/`VAE` constants)
- Test: `test_generate.py` (create)

- [ ] **Step 1: Write the failing test**

Create `test_generate.py`:

```python
import generate


def test_registry_integrity():
    assert generate.DEFAULT_MODEL in generate.MODELS
    for key, repo in generate.MODELS.items():
        assert isinstance(key, str) and key
        assert isinstance(repo, str) and "/" in repo


def test_resolve_known_key():
    assert generate.resolve_model("sdxl") == generate.MODELS["sdxl"]
    assert generate.resolve_model("realvis") == generate.MODELS["realvis"]


def test_resolve_unknown_key_raises_listing_valid():
    try:
        generate.resolve_model("nope")
    except SystemExit as e:
        assert "sdxl" in str(e) and "realvis" in str(e) and "pony" in str(e)
    else:
        assert False, "expected SystemExit for unknown key"


if __name__ == "__main__":
    test_registry_integrity()
    test_resolve_known_key()
    test_resolve_unknown_key_raises_listing_valid()
    print("ok")
```

- [ ] **Step 2: Run test to verify it fails**

Run: `python3 test_generate.py`
Expected: FAIL — `AttributeError: module 'generate' has no attribute 'MODELS'` (or `resolve_model`).

- [ ] **Step 3: Write minimal implementation**

In `generate.py`, replace lines 27-29 (the `MODEL` and `VAE` constants) with:

```python
MODELS = {
    "sdxl":    "stabilityai/stable-diffusion-xl-base-1.0",  # default, general
    "realvis": "SG161222/RealVisXL_V5.0",                   # photoreal, permissive
    "pony":    "AstraliteHeart/pony-diffusion-v6-xl",       # unrestricted, versatile
}
DEFAULT_MODEL = "sdxl"
# fp16-fix VAE avoids the fp32 upcast that OOMs a 12GB card at decode time.
VAE = "madebyollin/sdxl-vae-fp16-fix"


def resolve_model(key):
    if key not in MODELS:
        valid = ", ".join(sorted(MODELS))
        raise SystemExit(f"unknown model '{key}'. valid: {valid}")
    return MODELS[key]
```

Note: importing `generate.py` currently runs nothing at import time (all logic is under `main()` / `if __name__`), so `import generate` is safe for the test. Confirm no top-level executable code was added.

- [ ] **Step 4: Run test to verify it passes**

Run: `python3 test_generate.py`
Expected: `ok`

- [ ] **Step 5: Commit**

```bash
git add generate.py test_generate.py
git commit -S -m "Add model registry and resolve_model with valid-key errors

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>"
```

---

## Task 2: Refactor build_pipe to take a model key with variant fallback

**Files:**
- Modify: `generate.py:33-53` (`build_pipe`)

This is GPU/download code; no cheap unit test. Verified in Task 7 (integration on real Arc). TDD does not apply to the XPU load itself.

- [ ] **Step 1: Rewrite build_pipe**

Replace `build_pipe` (currently lines 33-53) with:

```python
def build_pipe(model_key):
    if not torch.xpu.is_available():
        # Fail loud and specific. A silent CPU fallback would run but take
        # minutes per image, which is worse than a clear error.
        print(
            "ERROR: torch.xpu is not available. The Level Zero runtime is not "
            "visible inside this container. Check that you passed --device "
            "/dev/dri and that the base image ships the Intel GPU runtime.",
            file=sys.stderr,
        )
        sys.exit(2)

    repo = resolve_model(model_key)
    vae = AutoencoderKL.from_pretrained(VAE, torch_dtype=torch.float16)
    try:
        # SDXL base publishes an fp16 variant (smaller download).
        pipe = StableDiffusionXLPipeline.from_pretrained(
            repo, vae=vae, torch_dtype=torch.float16,
            use_safetensors=True, variant="fp16",
        )
    except (ValueError, OSError, EnvironmentError):
        # Many fine-tunes do not publish a separate fp16 variant; load the
        # default files at fp16 dtype instead.
        pipe = StableDiffusionXLPipeline.from_pretrained(
            repo, vae=vae, torch_dtype=torch.float16, use_safetensors=True,
        )
    return pipe.to("xpu")
```

- [ ] **Step 2: Verify it still imports (no GPU needed)**

Run: `python3 -c "import generate; print('import ok')"`
Expected: `import ok` (import does not touch XPU).

- [ ] **Step 3: Commit**

```bash
git add generate.py
git commit -S -m "Refactor build_pipe to take a model key with fp16-variant fallback

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>"
```

---

## Task 3: Add -m/--model flag to the one-shot CLI

**Files:**
- Modify: `generate.py` (argparse in `main`, the `build_pipe()` call, the load message)

- [ ] **Step 1: Add the argument**

In `main()`, after the `prompt` positional (currently around line 80), add:

```python
    ap.add_argument(
        "-m",
        "--model",
        default=DEFAULT_MODEL,
        help=f"Model key: {', '.join(sorted(MODELS))} (default: {DEFAULT_MODEL}).",
    )
```

- [ ] **Step 2: Pass the key into build_pipe and the load message**

Replace the load block (currently lines 133-136):

```python
    print(f"Loading {args.model} + fp16-fix VAE on XPU ...", file=sys.stderr)
    t0 = time.time()
    pipe = build_pipe(args.model)
    print(f"Loaded in {time.time() - t0:.1f}s", file=sys.stderr)
```

- [ ] **Step 3: Verify the CLI parses (no GPU)**

Run: `python3 generate.py --help`
Expected: help text shows `-m/--model` with `sdxl, realvis, pony` and default `sdxl`.

Run: `python3 generate.py "x" -m nope 2>&1 | head -1` is NOT valid here (would hit XPU/build first only after arg parse). Instead verify unknown-key handling stays in `resolve_model` (covered by Task 1). Skip runtime call.

- [ ] **Step 4: Commit**

```bash
git add generate.py
git commit -S -m "Add -m/--model flag to one-shot CLI

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>"
```

---

## Task 4: HTTP daemon (server.py)

**Files:**
- Create: `server.py`

Holds one resident pipe. Reuses `build_pipe`/`generate`/`resolve_model` from `generate.py`. Serial by construction (`http.server` default), which is required (one GPU).

- [ ] **Step 1: Write server.py**

Create `server.py`:

```python
#!/usr/bin/env python3
# Copyright (C) 2026 Danilo M. <danix@danix.xyz>
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License version 2 as
# published by the Free Software Foundation.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.

# Long-lived daemon that keeps ONE SDXL model resident in XPU VRAM so prompt
# iteration skips the model-reload tax. Stdlib HTTP only. Serial by design:
# one GPU, so concurrent generation would OOM. See README.md.

import io
import json
import os
import sys
from http.server import BaseHTTPRequestHandler, HTTPServer

import torch

import generate

PORT = int(os.environ.get("IMGGEN_PORT", "8765"))

STATE = {"model": None, "pipe": None}


def load(model_key):
    if STATE["model"] == model_key and STATE["pipe"] is not None:
        return
    if STATE["pipe"] is not None:
        del STATE["pipe"]
        STATE["pipe"] = None
        torch.xpu.empty_cache()
    STATE["pipe"] = generate.build_pipe(model_key)
    STATE["model"] = model_key


class Handler(BaseHTTPRequestHandler):
    def _json(self, code, obj):
        body = json.dumps(obj).encode()
        self.send_response(code)
        self.send_header("Content-Type", "application/json")
        self.send_header("Content-Length", str(len(body)))
        self.end_headers()
        self.wfile.write(body)

    def _read_json(self):
        length = int(self.headers.get("Content-Length", "0"))
        raw = self.rfile.read(length) if length else b"{}"
        return json.loads(raw)  # raises on bad JSON -> caught by caller

    def do_GET(self):
        if self.path == "/status":
            self._json(200, {"model": STATE["model"], "ready": STATE["pipe"] is not None})
        else:
            self._json(404, {"error": "not found"})

    def do_POST(self):
        try:
            data = self._read_json()
        except (ValueError, json.JSONDecodeError):
            self._json(400, {"error": "bad JSON"})
            return

        if self.path == "/model":
            name = data.get("name")
            try:
                generate.resolve_model(name)
            except SystemExit as e:
                self._json(400, {"error": str(e)})
                return
            try:
                load(name)
            except Exception as e:  # download / load failure
                self._json(500, {"error": f"load failed: {e}"})
                return
            self._json(200, {"model": STATE["model"], "ready": True})
            return

        if self.path == "/generate":
            try:
                image, elapsed = generate.generate(
                    STATE["pipe"],
                    data["prompt"],
                    data.get("negative", ""),
                    int(data.get("steps", 25)),
                    float(data.get("guidance", 7.0)),
                    data.get("seed"),
                    int(data.get("width", 1024)),
                    int(data.get("height", 1024)),
                )
            except torch.OutOfMemoryError:
                torch.xpu.empty_cache()
                self._json(500, {"error": "OOM, try fewer steps or smaller size"})
                return
            except Exception as e:
                self._json(500, {"error": str(e)})
                return
            buf = io.BytesIO()
            image.save(buf, format="PNG")
            body = buf.getvalue()
            self.send_response(200)
            self.send_header("Content-Type", "image/png")
            self.send_header("X-Elapsed", f"{elapsed:.2f}")
            self.send_header("Content-Length", str(len(body)))
            self.end_headers()
            self.wfile.write(body)
            return

        self._json(404, {"error": "not found"})

    def log_message(self, fmt, *args):
        sys.stderr.write("%s - %s\n" % (self.address_string(), fmt % args))


def main():
    startup = os.environ.get("IMGGEN_MODEL", generate.DEFAULT_MODEL)
    print(f"Loading {startup} on XPU ...", file=sys.stderr)
    load(startup)
    print(f"Ready on 127.0.0.1:{PORT} with model {startup}", file=sys.stderr)
    HTTPServer(("0.0.0.0", PORT), Handler).serve_forever()


if __name__ == "__main__":
    main()
```

Note: bind `0.0.0.0` inside the container; compose publishes only to
`127.0.0.1` on the host, so it is not exposed to the network.

- [ ] **Step 2: Verify it imports and parses (no GPU, no serve)**

Run: `python3 -c "import ast; ast.parse(open('server.py').read()); print('parse ok')"`
Expected: `parse ok`

(Cannot import-run `main()` without XPU; syntax check is the cheap gate. Full run is Task 7.)

- [ ] **Step 3: Commit**

```bash
git add server.py
git commit -S -m "Add persistent HTTP daemon holding one model resident in VRAM

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>"
```

---

## Task 5: compose.yaml + .env.example + Dockerfile CMD check

**Files:**
- Create: `compose.yaml`
- Create: `.env.example`
- Read (verify): `Dockerfile` (ensure `server.py` gets copied in; add if missing)

- [ ] **Step 1: Verify the Dockerfile copies server.py**

Run: `grep -nE 'COPY|ADD' Dockerfile`
Expected: a `COPY` that brings in the app files. If it copies `generate.py`
explicitly (e.g. `COPY generate.py .`), add a sibling line `COPY server.py .`
right after it. If it copies the whole context (`COPY . .`), no change needed.
Make the minimal edit to ensure `server.py` lands in the image working dir.

- [ ] **Step 2: Write compose.yaml**

Create `compose.yaml`:

```yaml
services:
  imggen:
    build: .
    image: imggen:local
    user: "${IMGGEN_UID:-1000}:${IMGGEN_GID:-100}"
    group_add:
      - video          # gid for /dev/dri/card* access as non-root
    devices:
      - /dev/dri
    volumes:
      - ${IMGGEN_OUT:-./out}:/out
      - ${IMGGEN_CACHE:-./models}:/models
    environment:
      - HF_HOME=/models
      - IMGGEN_MODEL=${IMGGEN_MODEL:-sdxl}
      - IMGGEN_PORT=${IMGGEN_PORT:-8765}
    ports:
      - "127.0.0.1:${IMGGEN_PORT:-8765}:8765"
    command: ["python3", "server.py"]
```

- [ ] **Step 3: Write .env.example**

Create `.env.example`:

```
# Copy to .env (gitignored) and set your real host paths.
IMGGEN_OUT=/path/to/output
IMGGEN_CACHE=/path/to/model-cache
# IMGGEN_PORT=8765
# IMGGEN_MODEL=sdxl
# IMGGEN_UID=1000
# IMGGEN_GID=100
```

- [ ] **Step 4: Validate compose config (needs .env present)**

Run: `docker compose config >/dev/null && echo "compose ok"`
Expected: `compose ok`. If it errors on missing `.env` variables, confirm the
user's `.env` exists (they created it); the `:-` defaults cover the rest.

- [ ] **Step 5: Commit**

```bash
git add compose.yaml .env.example Dockerfile
git commit -S -m "Add compose service (runs as host user) and .env.example

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>"
```

---

## Task 6: Rewrite the imggen wrapper (subcommands + daemon client + fallback)

**Files:**
- Modify: `~/bin/imggen` (and keep the repo copy `./imggen` in sync — copy after)

The wrapper must `cd` to the repo dir so `docker compose` finds `compose.yaml`
and `.env`. Set `IMGGEN_DIR` to the repo path.

- [ ] **Step 1: Write the new wrapper**

Replace `~/bin/imggen` with:

```bash
#!/bin/bash
# Copyright (C) 2026 Danilo
# Licensed under the GNU General Public License version 2.
#
# Host-side wrapper for local SDXL image generation on Intel Arc.
# Subcommands manage a persistent daemon (model resident in VRAM); a bare
# prompt talks to the daemon if up, else falls back to a one-shot run.
#
# Usage:
#   imggen start [-m realvis]     # spin up daemon, load model, hold VRAM
#   imggen stop                   # kill daemon, free VRAM
#   imggen status                 # is it up? which model?
#   imggen list                   # show model keys
#   imggen "a red bicycle" [-m realvis] [--steps 30] [--seed 42] [-o out.png] [-n "blurry"]

set -euo pipefail

IMGGEN_DIR="${IMGGEN_DIR:-$HOME/Programming/GIT/imggen}"
cd "$IMGGEN_DIR"

# Load .env for OUT dir + port on the host side.
if [ -f .env ]; then set -a; . ./.env; set +a; fi

OUT_DIR="${IMGGEN_OUT:-$HOME/Pictures/AI-gen}"
PORT="${IMGGEN_PORT:-8765}"
BASE="http://127.0.0.1:${PORT}"
mkdir -p "$OUT_DIR"

usage() {
    echo "usage:" >&2
    echo "  imggen start [-m KEY] | stop | status | list" >&2
    echo "  imggen \"<prompt>\" [-m KEY] [--steps N] [--guidance G] [--seed S] [-n NEG] [-o name.png]" >&2
    exit 1
}

daemon_up() { curl -sf "${BASE}/status" >/dev/null 2>&1; }

case "${1:-}" in
    start)
        shift
        model="${IMGGEN_MODEL:-sdxl}"
        if [ "${1:-}" = "-m" ] || [ "${1:-}" = "--model" ]; then model="$2"; fi
        IMGGEN_MODEL="$model" docker compose up -d
        echo "daemon starting with model '$model' (VRAM ~9GB held until 'imggen stop')"
        exit 0
        ;;
    stop)
        docker compose down
        echo "daemon stopped, VRAM freed"
        exit 0
        ;;
    status)
        if daemon_up; then curl -s "${BASE}/status"; echo; else echo "daemon down"; fi
        exit 0
        ;;
    list)
        echo "sdxl     general (default)"
        echo "realvis  photoreal, permissive"
        echo "pony     unrestricted, versatile"
        exit 0
        ;;
    "" ) usage ;;
    -h|--help) usage ;;
esac

# --- prompt path ---
PROMPT="$1"; shift
MODEL=""; NEG=""; STEPS=""; GUID=""; SEED=""; OUT=""; W=""; H=""
while [ "$#" -gt 0 ]; do
    case "$1" in
        -m|--model)    MODEL="$2"; shift 2 ;;
        -n|--negative) NEG="$2"; shift 2 ;;
        -s|--steps)    STEPS="$2"; shift 2 ;;
        -g|--guidance) GUID="$2"; shift 2 ;;
        --seed)        SEED="$2"; shift 2 ;;
        -W|--width)    W="$2"; shift 2 ;;
        -H|--height)   H="$2"; shift 2 ;;
        -o|--out)      OUT="$2"; shift 2 ;;
        *) echo "unknown arg: $1" >&2; usage ;;
    esac
done

OUT_NAME="${OUT:-sdxl-$(date +%Y%m%d-%H%M%S).png}"
OUT_PATH="${OUT_DIR}/${OUT_NAME}"

if daemon_up; then
    # Interactive menu only when no -m and a TTY is present.
    if [ -z "$MODEL" ] && [ -t 0 ] && [ -t 1 ]; then
        echo "which model? 1) sdxl  2) realvis  3) pony" >&2
        read -rp "choice [1]: " choice
        case "$choice" in 2) MODEL=realvis ;; 3) MODEL=pony ;; *) MODEL=sdxl ;; esac
    fi
    MODEL="${MODEL:-sdxl}"

    # Swap resident model if needed.
    current="$(curl -s "${BASE}/status" | sed -n 's/.*"model": *"\([^"]*\)".*/\1/p')"
    if [ "$current" != "$MODEL" ]; then
        echo "swapping daemon model $current -> $MODEL ..." >&2
        curl -sf -X POST "${BASE}/model" -H 'Content-Type: application/json' \
            -d "{\"name\":\"$MODEL\"}" >/dev/null
    fi

    # Build JSON request. jq not assumed; hand-build (values are ours/simple).
    req="{\"prompt\":$(printf '%s' "$PROMPT" | python3 -c 'import json,sys; print(json.dumps(sys.stdin.read()))')"
    [ -n "$NEG" ]   && req="$req,\"negative\":$(printf '%s' "$NEG" | python3 -c 'import json,sys; print(json.dumps(sys.stdin.read()))')"
    [ -n "$STEPS" ] && req="$req,\"steps\":$STEPS"
    [ -n "$GUID" ]  && req="$req,\"guidance\":$GUID"
    [ -n "$SEED" ]  && req="$req,\"seed\":$SEED"
    [ -n "$W" ]     && req="$req,\"width\":$W"
    [ -n "$H" ]     && req="$req,\"height\":$H"
    req="$req}"

    if curl -sf -X POST "${BASE}/generate" -H 'Content-Type: application/json' \
        -d "$req" -o "$OUT_PATH"; then
        echo "Saved ${OUT_PATH}"
    else
        echo "generate failed (see daemon logs: docker compose logs)" >&2
        rm -f "$OUT_PATH"
        exit 1
    fi
else
    # One-shot fallback: non-interactive, model defaults to sdxl or -m KEY.
    echo "daemon down, one-shot run (slower: model loads then exits)" >&2
    set -- "$PROMPT"
    [ -n "$MODEL" ] && set -- "$@" -m "$MODEL"
    [ -n "$NEG" ]   && set -- "$@" -n "$NEG"
    [ -n "$STEPS" ] && set -- "$@" -s "$STEPS"
    [ -n "$GUID" ]  && set -- "$@" -g "$GUID"
    [ -n "$SEED" ]  && set -- "$@" --seed "$SEED"
    [ -n "$W" ]     && set -- "$@" -W "$W"
    [ -n "$H" ]     && set -- "$@" -H "$H"
    set -- "$@" -o "$OUT_NAME"
    docker compose run --rm imggen python3 generate.py "$@"
fi
```

- [ ] **Step 2: Syntax check**

Run: `bash -n ~/bin/imggen && echo "syntax ok"`
Expected: `syntax ok`

- [ ] **Step 3: Test subcommand routing without a daemon (no GPU)**

Run: `~/bin/imggen list`
Expected: three lines (sdxl / realvis / pony).

Run: `~/bin/imggen status`
Expected: `daemon down` (daemon not started yet).

- [ ] **Step 4: Sync repo copy and commit**

```bash
cp ~/bin/imggen ./imggen
git add imggen
git commit -S -m "Rewrite wrapper: daemon lifecycle, client, and one-shot fallback

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>"
```

---

## Task 7: Integration on real Arc (agent runs on this machine)

**Files:** none (verification task). Prereqs done by user: HF cache chown to
`danix:users`, `.env` created with real paths.

- [ ] **Step 1: Rebuild the image with server.py**

Run: `cd "$IMGGEN_DIR" && docker compose build`
Expected: builds `imggen:local`, no error, `server.py` present in image.

- [ ] **Step 2: One-shot SDXL still works (cached weights)**

Run: `~/bin/imggen "a red bicycle against a stone wall" -o test-oneshot.png`
Expected: daemon down → one-shot path → `Saved .../test-oneshot.png`.
Verify: `file "$HOME/Pictures/AI-gen/test-oneshot.png"` shows PNG 1024x1024,
and `ls -l` shows owner `danix:users` (permissions fix works).

- [ ] **Step 3: Daemon lifecycle on SDXL**

Run: `~/bin/imggen start`
Then: `~/bin/imggen status`
Expected: `{"model": "sdxl", "ready": true}`.

Run: `~/bin/imggen "a foggy harbour at dawn" -m sdxl -o test-daemon.png`
Expected: no reload message, `Saved .../test-daemon.png`, owner `danix:users`.

- [ ] **Step 4: Model download + swap (realvis, then pony) — approved pulls**

Run: `~/bin/imggen "a portrait of a fox, photoreal" -m realvis -o test-realvis.png`
Expected: `swapping daemon model sdxl -> realvis`, ~7GB downloads once to
`/data/LLM-models/imggen-models`, then `Saved`. If load fails with a
diffusers-format / variant error, the fp16-variant fallback (Task 2) should
handle the missing-variant case; a single-file-only repo would need
`from_single_file` — record the exact error and stop for a fix decision.

Run: `~/bin/imggen "anime character" -m pony -o test-pony.png`
Expected: swaps to pony, downloads once, `Saved`. Same fallback caveat.

- [ ] **Step 5: Stop frees VRAM**

Run: `~/bin/imggen stop`
Then: `~/bin/imggen status`
Expected: `daemon down`. (Optionally confirm VRAM freed via `xpu-smi` if
available; otherwise container-gone is the signal.)

- [ ] **Step 6: Commit any fixes made during integration**

If Task 2/4 needed adjustment (e.g. single-file load path for pony), commit:

```bash
git add -A
git commit -S -m "Fix model load path discovered during Arc integration

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>"
```

If nothing changed, skip.

---

## Task 8: Update README.md and CLAUDE.md

**Files:**
- Modify: `README.md`
- Modify: `CLAUDE.md`

- [ ] **Step 1: Update README.md**

Add a "Models" section (the three keys + descriptions), a "Daemon / lifecycle"
section documenting `start`/`stop`/`status`/`list`, the VRAM-locked-while-up
warning (~9GB held until `stop`), the one-shot fallback behavior, and `.env`
setup (copy `.env.example` → `.env`, set real paths, `IMGGEN_UID`/`GID`).
Keep the existing "Development Approach" section at the bottom verbatim.

- [ ] **Step 2: Update CLAUDE.md**

Update to reflect: 3 files → 5 (`imggen`, `generate.py`, `server.py`,
`compose.yaml`, `.env.example`); daemon is in scope as local burst-iteration
(still no refiner/Flux/batching); compose replaces raw `docker run`; the
registry is the source of truth for models; correct the container-boundary
defaults to note `.env` overrides and that the real paths live there, not in
the repo. Keep the 12GB / fp16-fix VAE / XPU-only constraints intact.

- [ ] **Step 3: Commit**

```bash
git add README.md CLAUDE.md
git commit -S -m "Document model menu, daemon lifecycle, and compose setup

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>"
```

---

## Self-Review notes

- **Spec coverage:** registry (T1), load robustness/variant fallback (T2),
  `-m` CLI (T3), daemon API status/generate/model (T4), compose + user perms +
  .env.example (T5), wrapper subcommands/menu/fallback (T6), integration incl.
  approved pulls + perms check (T7), README + CLAUDE.md (T8). All spec sections
  mapped.
- **Type consistency:** `build_pipe(model_key)`, `resolve_model(key)`,
  `generate(pipe, prompt, negative, steps, guidance, seed, width, height)`
  used identically in `generate.py`, `server.py`, and the wrapper's JSON keys
  (`prompt/negative/steps/guidance/seed/width/height`).
- **Placeholders:** none; every code step has full code. Docs tasks (T8)
  describe exact sections to write, not "TODO".
- **Known open risk:** pony/realvis single-file-vs-diffusers load path is only
  resolvable at download time (T7 step 4), with a defined fallback and a
  stop-for-decision if it is single-file-only.