session-e2e.yaml

name: Session E2E

# Drives the real `lk agent daemon` start/say/stop lifecycle against the minimal
# one-file echo agent in cmd/lk/testdata/echo-agent, on Linux, macOS, and
# Windows. This exercises the detached daemon, the readiness handshake, the
# console IPC transport, and the model round-trip end to end -- runtime behavior
# that `go test` alone never covers.
#
# Runs on manual dispatch and on pushes to any repo branch (forks can't trigger
# `push`, so secrets are only exposed to trusted collaborators). It needs live
# LiveKit credentials -- set these repo secrets first: LIVEKIT_API_KEY,
# LIVEKIT_API_SECRET, LIVEKIT_URL. The echo agent drives its LLM through LiveKit
# Inference, so no other provider keys are needed.
#
# The echo agent depends on plain PyPI livekit-agents (synced by `uv sync` from
# its pyproject.toml). Note: on current releases/main, `cli.run_app()` routes
# `console` through the legacy click CLI, which has no --connect-addr (that
# lives behind `python -m livekit.agents`). The fixture's __main__ dispatches
# console mode to the TCP console directly to bridge the daemon's
# `python <entry> console --connect-addr` launch.
#
# Windows is split into two jobs: pkg/apm/webrtc's C++ uses MSVC SEH
# (__try/__except) that the runner's mingw GCC can't compile, so we build with
# zig (clang) exactly as .goreleaser.yaml does. zig-as-CC must run on Linux: on
# native Windows the cgo link of ~560 webrtc/portaudio objects overflows the
# ~32KB command-line limit. So `cross-build-windows` cross-compiles lk.exe AND
# the e2e test binary on Linux and uploads them; `e2e-windows` downloads them
# and runs the test natively (LK_SESSION_E2E_BIN points the test at the
# prebuilt lk, so nothing is rebuilt on the Windows runner).
#
# Node is intentionally not in the matrix yet: this branch's session daemon only
# supports Python agents (`detectProject` rejects non-Python), and Node console
# support depends on the brian/agent-session-node-support CLI line (#868/#878)
# plus agents-js #1804. Add a node arm once those land.

on:
  workflow_dispatch:
  push:
    branches: ['**']

concurrency:
  group: session-e2e-${{ github.ref }}
  cancel-in-progress: true

jobs:
  # Linux and macOS build + run natively in one job -- their linkers have no
  # command-line-length limit, so `go test` building lk in-process is fine.
  e2e:
    strategy:
      fail-fast: false
      matrix:
        os: [ubuntu-latest, macos-latest]

    runs-on: ${{ matrix.os }}
    name: Agent Session with Python Agent on ${{ matrix.os }}

    permissions:
      contents: read

    steps:
      - name: Checkout livekit-cli
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6
        with:
          # pkg/portaudio/pa_src is a submodule with the vendored PortAudio C
          # source; needed now that console (cgo) builds unconditionally.
          submodules: true

      - name: Install ALSA headers (Linux)
        if: matrix.os == 'ubuntu-latest'
        run: sudo apt-get update && sudo apt-get install -y libasound2-dev

      - name: Set up Go
        uses: actions/setup-go@4a3601121dd01d1626a1e23e37211e3254c1c06c # v6
        with:
          go-version-file: go.mod
          cache: true

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: "3.12"

      - name: Set up uv
        uses: astral-sh/setup-uv@v5
        with:
          enable-cache: true

      - name: Sync echo agent deps
        working-directory: cmd/lk/testdata/echo-agent
        run: uv sync

      - name: Run session e2e
        env:
          LIVEKIT_API_KEY: ${{ secrets.LIVEKIT_API_KEY }}
          LIVEKIT_API_SECRET: ${{ secrets.LIVEKIT_API_SECRET }}
          LIVEKIT_URL: ${{ secrets.LIVEKIT_URL }}
        run: go test ./cmd/lk -run TestSessionE2E -count=1 -v -timeout 600s

  # Cross-compile the Windows artifacts on Linux with zig, mirroring
  # .goreleaser.yaml's lk-windows-amd64 build. Produces lk.exe (the binary under
  # test) and the compiled e2e test binary, both shipped to e2e-windows.
  cross-build-windows:
    runs-on: ubuntu-latest
    name: Cross-build Windows artifacts (zig)

    permissions:
      contents: read

    steps:
      - name: Checkout livekit-cli
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6
        with:
          submodules: true

      - name: Set up Zig
        uses: mlugg/setup-zig@d1434d08867e3ee9daa34448df10607b98908d29 # v2
        with:
          version: 0.14.1

      - name: Set up Go
        uses: actions/setup-go@4a3601121dd01d1626a1e23e37211e3254c1c06c # v6
        with:
          go-version-file: go.mod
          cache: true

      # Generate the MinGW import libs lld needs for Go's /DEFAULTLIB references
      # (dbghelp, bcrypt, ...), exactly as goreleaser's before-hook does.
      - name: Generate MinGW import libs
        run: scripts/setup-cross.sh windows/amd64

      - name: Cross-compile lk.exe and the e2e test binary
        env:
          GOOS: windows
          GOARCH: amd64
          CGO_ENABLED: "1"
          # zig cc/c++ as the cgo toolchain, matching .goreleaser.yaml's Windows
          # build. -fms-extensions (SEH) isn't on cgo's allowlist; -fno-sanitize
          # is a zig UBSan-under-lld workaround -- both copied from goreleaser.
          CC: zig cc -target x86_64-windows-gnu
          CXX: zig c++ -target x86_64-windows-gnu
          CGO_CXXFLAGS_ALLOW: -fms-extensions
          CGO_CXXFLAGS: -fno-sanitize=all
        run: |
          ldflags="-extldflags=-L$PWD/.cross/windows_amd64/mingw_lib"
          mkdir -p dist
          go build -ldflags="$ldflags" -o dist/lk.exe ./cmd/lk
          go test -c -ldflags="$ldflags" -o dist/session-e2e.test.exe ./cmd/lk

      - name: Upload Windows artifacts
        uses: actions/upload-artifact@v4
        with:
          name: windows-e2e
          path: dist/*.exe
          retention-days: 1
          if-no-files-found: error

  # Run the prebuilt Windows artifacts natively. No Go/zig/cgo here: the test
  # binary just drives lk.exe (via LK_SESSION_E2E_BIN) against the echo agent.
  e2e-windows:
    needs: cross-build-windows
    runs-on: windows-latest
    name: Agent Session with Python Agent on windows-latest

    permissions:
      contents: read

    steps:
      - name: Checkout livekit-cli
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6
        # No submodules: the prebuilt binaries already contain the cgo code;
        # only the echo-agent fixture (plain repo files) is needed at runtime.

      - name: Download Windows artifacts
        uses: actions/download-artifact@v4
        with:
          name: windows-e2e
          path: dist

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: "3.12"

      - name: Set up uv
        uses: astral-sh/setup-uv@v5
        with:
          enable-cache: true

      - name: Sync echo agent deps
        working-directory: cmd/lk/testdata/echo-agent
        run: uv sync

      - name: Run session e2e
        # Run from cmd/lk so the test's default testdata/echo-agent/agent.py
        # path resolves; point it at the cross-built lk.exe so nothing rebuilds.
        shell: bash
        working-directory: cmd/lk
        env:
          # Relative to cmd/lk; Go's filepath.Abs resolves it. Forward slashes
          # keep bash happy (GITHUB_WORKSPACE would carry Windows backslashes).
          LK_SESSION_E2E_BIN: ../../dist/lk.exe
          LIVEKIT_API_KEY: ${{ secrets.LIVEKIT_API_KEY }}
          LIVEKIT_API_SECRET: ${{ secrets.LIVEKIT_API_SECRET }}
          LIVEKIT_URL: ${{ secrets.LIVEKIT_URL }}
        run: |
          ../../dist/session-e2e.test.exe \
            -test.run TestSessionE2E -test.count=1 -test.v -test.timeout 600s