# The "extended" image: the delivery image plus the options it leaves out because of their
# licence, the origin of their data, or their size. Each option is one build argument, off by
# default; anything not listed here is not part of what we build or test.
#
#   docker build -f delivery/Dockerfile.extended \
#     --build-arg WITH_CLAUDE=1 --build-arg WITH_CURSOR=1 \
#     --build-arg WITH_NAMES=1  --build-arg WITH_FASTAI=1 \
#     -t parleml-extended .
#
# Option          Why it is not in the standard image                     What it adds
# WITH_CLAUDE=1   Claude Code is "All rights reserved" (Anthropic's       Claude Code CLI (the version the
#                 Commercial Terms): no redistribution grant.             adapter is verified against; override
#                                                                         with CLAUDE_VERSION)
# WITH_CURSOR=1   cursor-sdk is Cursor's proprietary software: no        Cursor sessions (needs CURSOR_API_KEY)
#                 redistribution grant.
# WITH_NAMES=1    names-dataset's corpus was extracted from the 2021      The six name-corpus features (known
#                 Facebook data leak. The code is MIT; we do not ship     first/last name, country of origin...);
#                 the data in a public image. Your call, your risk.       without it they degrade to 0/None/NaN
# WITH_FASTAI=1   Size: fast.ai, spaCy and the AutoGluon fast.ai extra    The "Neural net (fast.ai)" model family
#                 (see delivery/README.md for the measured cost).         in the roster
#
# BASE is the delivery image to build on. The copy inside a released image (and the one on the
# website) defaults to that release, ghcr.io/parleml/parleml:<version>; this source copy defaults to
# the local build, parleml-protected.
#
# Build it yourself and keep it to yourself: never push an extended image to a public registry.
# Publishing it would redistribute exactly what this recipe exists to keep out of our image, under
# licences that do not allow it.
#
# Nothing here signs anything in: a CLI's login or API key stays the user's (delivery/README.md,
# "Agent sessions"). Python packages go into the application's own environment with
# `pip --python /app/.venv/bin/python`: the image's plain `pip` belongs to the system Python,
# whose packages ParleML cannot see, and the application's environment has no pip of its own.
# Build context can be any directory; nothing is copied from it.

ARG BASE=ghcr.io/parleml/parleml:0.8.0
FROM ${BASE}

ARG WITH_CLAUDE=0
# The adapter's argv is verified against this release (parleml_agents/config.py: claude_pinned_version).
ARG CLAUDE_VERSION=2.1.270
ARG WITH_CURSOR=0
ARG WITH_NAMES=0
ARG WITH_FASTAI=0

RUN set -eu; \
    PY=/app/.venv/bin/python; \
    for opt in WITH_CLAUDE WITH_CURSOR WITH_NAMES WITH_FASTAI; do \
        eval "val=\$$opt"; \
        case "$val" in 0|1) ;; *) echo "$opt must be 0 or 1, got '$val'"; exit 1;; esac; \
    done; \
    if [ "$WITH_CLAUDE" = 1 ]; then \
        apt-get update; \
        apt-get install -y --no-install-recommends nodejs npm; \
        rm -rf /var/lib/apt/lists/*; \
        npm install -g "@anthropic-ai/claude-code@${CLAUDE_VERSION}"; \
        claude --version; \
    fi; \
    if [ "$WITH_CURSOR" = 1 ]; then \
        pip --python $PY install --no-cache-dir "cursor-sdk>=1.0.31"; \
        $PY -c "import cursor_sdk"; \
    fi; \
    if [ "$WITH_NAMES" = 1 ]; then \
        pip --python $PY install --no-cache-dir names-dataset; \
        $PY -c "import names_dataset"; \
    fi; \
    if [ "$WITH_FASTAI" = 1 ]; then \
        TORCH="$($PY -c 'import torch; print(torch.__version__)')"; \
        AG="$($PY -c 'import importlib.metadata as m; print(m.version("autogluon.tabular"))')"; \
        pip --python $PY install --no-cache-dir \
            --extra-index-url https://download.pytorch.org/whl/cpu \
            "autogluon.tabular[fastai]==$AG" "torch==$TORCH"; \
        $PY -c "import sys, torch; assert torch.__version__ == sys.argv[1], (torch.__version__, sys.argv[1]); import fastai" "$TORCH"; \
        $PY -c "import importlib.metadata as m, sys; bad = sorted(d.metadata['Name'] for d in m.distributions() if d.metadata['Name'].lower().startswith('nvidia-')); sys.exit('FAIL: CUDA libraries came in: ' + ', '.join(bad)) if bad else print('fast.ai installed, CPU torch unchanged')"; \
    fi