name: integration-slow

# Weekly gate for slow integration tests: real training, export, ollama
# round trip, pack / unpack / prompt identity. Not run per-PR because
# the end-to-end path is ~15-20 min on ubuntu-latest; opt in by labeling
# a PR with `slow-run-please`.
#
# Sprint 14.5 owns this workflow. The bodies live under tests/integration/
# and skip gracefully when `ollama` / `vendor/llama.cpp` / tiny-model
# cache aren't present, so the job is also safe to dispatch manually
# from the Actions tab.

on:
  schedule:
    # Sundays at 07:00 UTC — offset from the template-drift workflow at
    # 06:00 UTC so the two jobs don't contend for runner capacity.
    - cron: "0 7 * * 0"
  workflow_dispatch: {}
  pull_request:
    types: [labeled, synchronize]

concurrency:
  group: integration-slow-${{ github.ref }}
  cancel-in-progress: false

env:
  UV_VERSION: "0.11.6"
  PYTHON_VERSION: "3.11"
  # Same SHA as ci.yml / weekly-template-drift.yml — keep in sync.
  TINY_MODEL_REVISION: "12fd25f77366fa6b3b4b768ec3050bf629380bac"

jobs:
  slow:
    name: slow integration suite
    # On pull_request runs, only fire when the `slow-run-please` label is present.
    # Scheduled / manual dispatch always run.
    if: >-
      github.event_name != 'pull_request'
      || contains(github.event.pull_request.labels.*.name, 'slow-run-please')
    runs-on: ubuntu-latest
    timeout-minutes: 45
    steps:
      - name: Checkout with submodules
        uses: actions/checkout@v4
        with:
          submodules: recursive

      - name: Install uv
        uses: astral-sh/setup-uv@v4
        with:
          version: ${{ env.UV_VERSION }}

      - name: Sync dependencies
        run: uv sync --all-extras --dev

      - name: Restore HF cache
        id: hf-cache
        uses: actions/cache@v4
        with:
          path: ${{ github.workspace }}/.hf-cache
          key: hf-tiny-${{ env.TINY_MODEL_REVISION }}-${{ hashFiles('pyproject.toml') }}
          restore-keys: |
            hf-tiny-${{ env.TINY_MODEL_REVISION }}-

      - name: Pre-warm tiny model
        env:
          HF_HOME: ${{ github.workspace }}/.hf-cache
          DLM_TINY_MODEL_REVISION: ${{ env.TINY_MODEL_REVISION }}
        run: |
          uv run python - <<'PY'
          from tests.fixtures.tiny_model import tiny_model_path
          print("tiny model at:", tiny_model_path())
          PY

      - name: Restore llama.cpp build cache
        id: llama-cpp-cache
        uses: actions/cache@v4
        with:
          path: vendor/llama.cpp/build
          key: llama-cpp-build-${{ hashFiles('.gitmodules', 'vendor/llama.cpp/VERSION') }}

      - name: Build llama-quantize + llama-imatrix (if not cached)
        if: steps.llama-cpp-cache.outputs.cache-hit != 'true'
        run: |
          set -euxo pipefail
          command -v cmake >/dev/null 2>&1 || sudo apt-get install -y cmake
          scripts/bump-llama-cpp.sh build

      - name: Install Ollama
        run: |
          set -euxo pipefail
          curl -fsSL https://ollama.com/install.sh | sh
          # The ollama_daemon fixture starts its own `ollama serve` when
          # nothing is listening — we don't need to launch one here.
          ollama --version

      - name: Run slow integration suite
        env:
          HF_HOME: ${{ github.workspace }}/.hf-cache
          DLM_TINY_MODEL_REVISION: ${{ env.TINY_MODEL_REVISION }}
        run: uv run pytest -m slow -v