ci.yml

name: CI
on:
  push:
    branches: [main, develop]
  pull_request:
    branches: [main, develop]
jobs:
  build-and-test:
    strategy:
      # One platform's failure says nothing about the other's; let both finish.
      fail-fast: false
      matrix:
        os: [ubuntu-latest, macos-latest]
    runs-on: ${{ matrix.os }}
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"
      - uses: dtolnay/rust-toolchain@stable
        with:
          targets: wasm32-wasip1
      - uses: actions/cache@v4
        with:
          # ~/.almide holds the build cache for test files that take Almide's
          # native path; without it every run rebuilds them from scratch.
          path: |
            ~/.cargo
            ~/.almide
            target
          # The cached path list is not part of the key, so adding a path to an
          # existing key would restore the old entry and never save the new one.
          # Bump this token whenever the paths change.
          key: ${{ runner.os }}-porta-v2-${{ hashFiles('almide.toml', '.almide-version') }}
          restore-keys: ${{ runner.os }}-porta-v2-
      - name: Install pinned Almide
        run: bash scripts/install-almide.sh
      - name: Install pinned Wasmtime
        # Almide runs a test file through wasmtime when it is on PATH and
        # otherwise builds it natively, which dominated the job time. This is an
        # accelerator, not a dependency: a failed download must leave the job
        # slow rather than red, so the tests still decide whether CI passes.
        continue-on-error: true
        run: |
          bash scripts/install-wasmtime.sh
          echo "$PWD/.tools/wasmtime" >> "$GITHUB_PATH"
      - name: Code quality grade
        # The score depends on the analyzer's own thresholds and on how much of
        # this repository's Almide its grammar can parse, so the version is
        # pinned and the analysis runs on one platform. The floor is the grade-A
        # boundary: this gate exists to catch the repository dropping below it,
        # and moving the floor is a deliberate edit, not a passing fix.
        if: matrix.os == 'ubuntu-latest'
        run: |
          bash scripts/install-codopsy.sh
          .tools/codopsy/codopsy analyze . -o "$RUNNER_TEMP/codopsy.json"
          python3 scripts/check_grade.py "$RUNNER_TEMP/codopsy.json" --min-score 90
      - name: Evaluation grader tests
        run: |
          python3 -m unittest discover -s scripts/evals -p 'test_*.py'
          python3 scripts/audit_eval.py docs/benchmarks/2026-09-19-real-model.json.gz
          python3 scripts/audit_eval.py docs/benchmarks/2026-09-19-verification-fixed.json.gz
          python3 scripts/audit_eval.py docs/benchmarks/2026-09-19-compact-results.json.gz
          python3 scripts/audit_eval.py docs/benchmarks/2026-09-19-pre-tool-policy.json.gz
          python3 scripts/audit_eval.py docs/benchmarks/2026-09-20-loop-notice.json.gz
          python3 scripts/audit_eval.py docs/benchmarks/2026-09-20-loop-control.json.gz
          python3 scripts/audit_eval.py docs/benchmarks/2026-09-20-shared-compute.json.gz
          python3 scripts/audit_containment.py docs/benchmarks/2026-09-20-containment.json.gz
          python3 scripts/audit_containment.py docs/benchmarks/2026-09-20-containment-fixed.json.gz
          python3 scripts/audit_compute_smoke.py docs/benchmarks/2026-09-20-compute-smoke.json.gz
      - name: Audit saved startup and memory measurements
        run: |
          python3 scripts/audit_benchmark.py docs/benchmarks/2026-09-20-startup.json.gz
          python3 scripts/audit_benchmark.py docs/benchmarks/2026-09-20-startup-recorded.json.gz --porta-record
      - name: Check
        run: .tools/almide/almide check src/main.almd
      - name: Build
        run: .tools/almide/almide build src/main.almd -o target/porta
      # TLS is rustls only (reqwest without its default features), so the Linux
      # binary needs no libssl.so.3 from whatever image it lands in (#35). A
      # dependency that turns native-tls back on shows up here.
      - name: No OpenSSL linked
        if: runner.os == 'Linux'
        run: |
          ldd target/porta
          ! ldd target/porta | grep -E 'libssl|libcrypto'
      - name: Test
        run: .tools/almide/almide test --ci
      # The memory ceiling is a cgroup v2 scope asked of the systemd user
      # manager, and a runner job has no login session, so no user manager
      # runs for it. Lingering starts one, and the suite then proves the
      # ceiling on Linux (an OOM kill at 64 MiB) instead of only its refusal.
      - name: Start a systemd user manager for the memory ceiling
        if: matrix.os == 'ubuntu-latest'
        run: |
          sudo loginctl enable-linger "$USER"
          for _ in $(seq 1 20); do [ -S "/run/user/$(id -u)/bus" ] && break; sleep 1; done
          ls -la "/run/user/$(id -u)/bus"
          echo "XDG_RUNTIME_DIR=/run/user/$(id -u)" >> "$GITHUB_ENV"
          echo "DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/$(id -u)/bus" >> "$GITHUB_ENV"
      # --why traces the run with strace on Linux.
      - name: Install strace for --why
        if: matrix.os == 'ubuntu-latest'
        run: command -v strace || (sudo apt-get update -qq && sudo apt-get install -y -qq strace)
      - name: Integration
        run: python3 scripts/integration.py target/porta
      - name: Escape corpus
        run: python3 scripts/escapes.py target/porta
      # The runner restricts unprivileged user namespaces through AppArmor, as
      # Ubuntu does from 23.10, so the two steps above prove the run without
      # them. The profile Ubuntu documents for such a program lets porta have
      # them, and the same two suites then prove the namespaces on x86_64.
      # The .deb's postinst loads the AppArmor profile for /usr/bin/porta on a
      # runner that restricts user namespaces, as this one does; removing the
      # package must take the profile with it.
      - name: The .deb gives porta its namespaces and removes cleanly
        if: matrix.os == 'ubuntu-latest'
        run: |
          cp target/porta "$RUNNER_TEMP/porta" && strip "$RUNNER_TEMP/porta"
          bash scripts/build-deb.sh "$RUNNER_TEMP/porta" 0.0.0 amd64 "$RUNNER_TEMP/deb"
          sudo apt-get install -y "$RUNNER_TEMP/deb/porta_0.0.0_amd64.deb"
          /usr/bin/porta check | grep -E '^ok +own PID, mount and network namespace'
          sudo apt-get remove -y porta
          test ! -e /etc/apparmor.d/usr.bin.porta
      - name: Integration and escape corpus with user namespaces
        if: matrix.os == 'ubuntu-latest'
        run: |
          # The build sits in the checkout, which the runner's user can write;
          # the helper script refuses it, and porta setup makes a root-owned
          # copy with the profile for that copy alone.
          ! sudo bash scripts/apparmor-userns.sh target/porta
          # The runner's /usr/local/bin is not root's alone, so setup puts the
          # copy in /usr/libexec/porta and links /usr/local/bin/porta to it.
          sudo target/porta setup
          set_up=$(readlink -f /usr/local/bin/porta)
          "$set_up" check | grep -E '^ok +own PID, mount and network namespace'
          python3 scripts/integration.py "$set_up"
          python3 scripts/escapes.py "$set_up"
          sudo "$set_up" setup --undo
          ! ls /etc/apparmor.d/*.porta 2>/dev/null
      # The seeds that found each fixed bug, so the fix stays fixed, then one
      # fresh seed per push, printed, so a new find can be replayed.
      - name: Fuzz
        run: |
          python3 scripts/fuzz.py target/porta policy --iterations 250 --seed 2
          python3 scripts/fuzz.py target/porta proxy --iterations 4 --seed 7
          python3 scripts/fuzz.py target/porta cli --iterations 300 --seed 1
          python3 scripts/fuzz.py target/porta policy --iterations 150
          python3 scripts/fuzz.py target/porta proxy --iterations 4
      # Operations strung together at random: runs with random grants and
      # misbehaviour, snapshots and rollbacks, explain, init and up, runs
      # killed part way. Seed 13 found two snapshots in one second rolled
      # back to the wrong one; it stays, beside one fresh seed per push.
      - name: Monkey
        run: |
          python3 scripts/monkey.py target/porta --steps 150 --seed 13
          python3 scripts/monkey.py target/porta --steps 100
      # Not a gate: the numbers move with the runner. Printed so every push
      # leaves what porta cost on that commit in its log.
      - name: Overhead
        run: python3 scripts/overhead.py target/porta --runs 20
      - name: Build WASM agents and probes
        run: |
          .tools/almide/almide build examples/chat-agent/src/mod.almd --target wasm -o examples/chat-agent/agent.wasm
          .tools/almide/almide build examples/demo-agent/src/mod.almd --target wasm -o examples/demo-agent/agent.wasm
          .tools/almide/almide build scripts/fixtures/guest_isolation.almd --target wasm -o target/guest-isolation.wasm
          .tools/almide/almide build scripts/fixtures/verify_readonly.almd --target wasm -o target/verify-readonly.wasm
          cargo test --locked --manifest-path examples/verify-json/Cargo.toml --target-dir target/verify-json
          cargo build --locked --manifest-path examples/verify-json/Cargo.toml --target wasm32-wasip1 --release --target-dir target/verify-json
          cp target/verify-json/wasm32-wasip1/release/porta-verify-json.wasm examples/verify-json/check.wasm
          cp target/verify-json/wasm32-wasip1/release/verify-tool.wasm examples/verify-json/verify-tool.wasm
          cp target/verify-json/wasm32-wasip1/release/before-tool.wasm examples/verify-json/before-tool.wasm
          cargo test --locked --manifest-path examples/compute/Cargo.toml --target-dir target/compute
          cargo build --locked --manifest-path examples/compute/Cargo.toml --target wasm32-wasip1 --release --target-dir target/compute
          cp target/compute/wasm32-wasip1/release/porta-compute.wasm examples/compute/compute.wasm
      - name: WASM computation and exact JSON transport
        run: python3 scripts/compute_integration.py target/porta examples/chat-agent/agent.wasm examples/compute/compute.wasm examples/demo-agent/agent.wasm
      - name: WASM agents and teams
        run: python3 scripts/agent_integration.py target/porta examples/chat-agent/agent.wasm examples/demo-agent/agent.wasm target/guest-isolation.wasm
      - name: Durable resume and replay
        run: python3 scripts/journal_integration.py target/porta examples/chat-agent/agent.wasm examples/demo-agent/agent.wasm
      - name: Tool argument validation
        run: python3 scripts/schema_integration.py target/porta examples/chat-agent/agent.wasm examples/demo-agent/agent.wasm
      - name: MCP agent security and recovery
        run: python3 scripts/mcp_agent_integration.py target/porta examples/chat-agent/agent.wasm examples/verify-json/verify-tool.wasm
      - name: Official MCP SDK interoperability
        run: |
          python3 -m venv "$RUNNER_TEMP/porta-mcp-sdk"
          "$RUNNER_TEMP/porta-mcp-sdk/bin/python" -m pip install mcp==1.30.0
          "$RUNNER_TEMP/porta-mcp-sdk/bin/python" scripts/mcp_sdk_integration.py target/porta examples/chat-agent/agent.wasm
          "$RUNNER_TEMP/porta-mcp-sdk/bin/python" scripts/compute_mcp_integration.py target/porta examples/compute/compute.wasm examples/demo-agent/agent.wasm
      - name: Verified completion and recovery
        run: python3 scripts/verification_integration.py target/porta examples/chat-agent/agent.wasm examples/demo-agent/agent.wasm examples/verify-json/check.wasm target/verify-readonly.wasm
      - name: Pre-tool policies and recovery
        run: python3 scripts/before_tool_integration.py target/porta examples/chat-agent/agent.wasm examples/demo-agent/agent.wasm examples/verify-json/before-tool.wasm examples/verify-json/check.wasm
      - name: Model batch repair and recovery
        run: python3 scripts/batch_recovery_integration.py target/porta examples/chat-agent/agent.wasm examples/demo-agent/agent.wasm
      - name: Artifact pins and strict team loading
        run: python3 scripts/artifact_pins_integration.py target/porta examples/chat-agent/agent.wasm examples/demo-agent/agent.wasm
      - name: Offline team configuration inspection
        run: python3 scripts/agent_check_integration.py target/porta examples/chat-agent/agent.wasm examples/demo-agent/agent.wasm

  # The action a workflow uses to confine its own steps: install a released
  # porta against its signature, and on Ubuntu give it the namespaces the
  # runner's AppArmor otherwise withholds. Tried as a user would use it.
  action:
    strategy:
      matrix:
        os: [ubuntu-latest, macos-latest]
    runs-on: ${{ matrix.os }}
    steps:
      - uses: actions/checkout@v4
      - id: porta
        uses: ./
        with:
          version: v0.6.11
          why: "true"
      - name: Confine a step
        run: |
          porta run /bin/sh -v . -- -c 'echo inside > granted.txt && cat granted.txt'
          ! porta run /bin/sh -v . -- -c "echo outside > \"$HOME/porta-action-escape\"" 
          test ! -e "$HOME/porta-action-escape"
      - name: The namespaces are there
        if: runner.os == 'Linux'
        run: porta check | grep -E '^ok +own PID, mount and network namespace'