[pull] main from fern-api:main

.github/actions/auto-update-seed/action.yaml

@@ -25,6 +25,14 @@ inputs:
     description: The matrix strategy index of the job (to differentiate when parallelizing jobs)
     required: false
     default: "0"
+  dockerhub-username:
+    description: Docker Hub username for authenticated pulls (avoids rate limits)
+    required: false
+    default: ""
+  dockerhub-token:
+    description: Docker Hub token for authenticated pulls (avoids rate limits)
+    required: false
+    default: ""
 
 runs:
   using: "composite"
@@ -40,6 +48,8 @@ runs:
         allow-unexpected-failures: ${{ inputs.allow-unexpected-failures }}
         skip-scripts: ${{ inputs.skip-scripts }}
         skip-git-diff-check: true
+        dockerhub-username: ${{ inputs.dockerhub-username }}
+        dockerhub-token: ${{ inputs.dockerhub-token }}
 
     # Add changes first to simplify git diff checks, no changes will NOT error here
     - name: Check for changes

.github/actions/cached-seed/action.yaml

@@ -20,13 +20,52 @@ inputs:
     description: Whether to allow unexpected failures during seed generation
     required: false
     default: "false"
+  dockerhub-username:
+    description: Docker Hub username for authenticated pulls (avoids rate limits)
+    required: false
+    default: ""
+  dockerhub-token:
+    description: Docker Hub token for authenticated pulls (avoids rate limits)
+    required: false
+    default: ""
 
 runs:
   using: "composite"
   steps:
     - name: Install
       uses: ./.github/actions/install
 
+    - name: Log in to Docker Hub
+      if: ${{ inputs.dockerhub-username != '' }}
+      uses: docker/login-action@v3
+      with:
+        username: ${{ inputs.dockerhub-username }}
+        password: ${{ inputs.dockerhub-token }}
+
+    - name: Clear stale Docker image cache
+      shell: bash
+      run: rm -rf /tmp/docker-cache
+
+    - name: Cache Docker images
+      id: docker-cache
+      uses: actions/cache@v4
+      with:
+        path: /tmp/docker-cache
+        key: ${{ runner.os }}-docker-${{ inputs.generator-name }}-${{ hashFiles('**/Dockerfile*') }}
+        restore-keys: |
+          ${{ runner.os }}-docker-${{ inputs.generator-name }}-
+
+    - name: Load cached Docker images
+      shell: bash
+      run: |
+        for f in /tmp/docker-cache/*.tar; do
+          [ -f "$f" ] && docker load -i "$f" || true
+        done
+
+    - name: Record pre-existing Docker images
+      shell: bash
+      run: docker images -q | sort -u > /tmp/docker-images-before.txt
+
     - name: Seed Test
       shell: bash
       env:
@@ -55,3 +94,19 @@ runs:
           echo "Seed test failed (attempt $attempt/3), retrying in 10s..."
           sleep 10
         done
+
+    - name: Save Docker images for cache
+      if: always()
+      shell: bash
+      run: |
+        if [ ! -f /tmp/docker-images-before.txt ]; then
+          echo "Skipping cache save: docker-images-before.txt not found"
+          exit 0
+        fi
+        mkdir -p /tmp/docker-cache
+        comm -13 /tmp/docker-images-before.txt <(docker images -q | sort -u) | while read -r id; do
+          img=$(docker inspect --format '{{index .RepoTags 0}}' "$id" 2>/dev/null || true)
+          [ -z "$img" ] && continue
+          fname=$(echo "$img" | tr '/:' '_')
+          docker save "$img" -o "/tmp/docker-cache/${fname}.tar" 2>/dev/null || true
+        done

.github/actions/run-seed-process/action.yaml

@@ -34,6 +34,14 @@ inputs:
     description: "Whether to skip the git diff check (used by auto-update-seed where changes are expected)"
     required: false
     default: "false"
+  dockerhub-username:
+    description: Docker Hub username for authenticated pulls (avoids rate limits)
+    required: false
+    default: ""
+  dockerhub-token:
+    description: Docker Hub token for authenticated pulls (avoids rate limits)
+    required: false
+    default: ""
 
 outputs:
   start_time:
@@ -78,6 +86,8 @@ runs:
         fixtures-to-run: ${{ steps.json-to-string.outputs.as-string }}
         skip-scripts: ${{ inputs.skip-scripts }}
         allow-unexpected-failures: ${{ inputs.allow-unexpected-failures }}
+        dockerhub-username: ${{ inputs.dockerhub-username }}
+        dockerhub-token: ${{ inputs.dockerhub-token }}
 
     - name: Summarize Test Results
       if: always() && !cancelled()

.github/scripts/format-benchmark-report.sh

@@ -6,7 +6,8 @@
 #   {"generator":"ts-sdk","spec":"square","duration_seconds":45}
 #
 # The main-results-dir may contain either:
-#   - Flat .jsonl files (single baseline, legacy format)
+#   - Flat .jsonl files (single baseline) with optional e2e/ subdirectory
+#     for E2E results (artifacts API format)
 #   - A history/ subdirectory with dated run folders, each containing .jsonl files.
 #     When history exists, the median of all historical runs is used as the baseline.
 #     E2E results are stored in history/<run>/e2e/ subdirectories.
@@ -26,7 +27,8 @@ compute_median() {
 }
 
 # Look up E2E baseline duration for a given generator+spec.
-# Reads from history/<run>/e2e/ subdirectories.
+# Checks flat e2e/ subdirectory first (artifacts API format),
+# then falls back to history/<run>/e2e/ subdirectories.
 # Sets: E2E_BASELINE_VAL, E2E_BASELINE_RUNS
 lookup_e2e_baseline() {
   local generator="$1"
@@ -51,6 +53,17 @@ lookup_e2e_baseline() {
     if [ "$E2E_BASELINE_RUNS" -gt 0 ]; then
       E2E_BASELINE_VAL=$(printf '%s\n' "${durations[@]}" | compute_median)
     fi
+  else
+    # Flat E2E layout: MAIN_DIR/e2e/{generator}.jsonl (artifacts API format)
+    local e2e_file="${MAIN_DIR}/e2e/${generator}.jsonl"
+    if [ -f "$e2e_file" ]; then
+      local dur
+      dur=$(jq -r --arg spec "$spec" 'select(.spec == $spec) | .duration_seconds' "$e2e_file" 2>/dev/null || true)
+      if [ -n "$dur" ] && [ "$dur" != "null" ] && [ "$dur" != "0" ]; then
+        E2E_BASELINE_VAL="$dur"
+        E2E_BASELINE_RUNS=1
+      fi
+    fi
   fi
 }
 
@@ -105,7 +118,7 @@ if [ -n "${BASELINE_TIMESTAMP:-}" ]; then
     HISTORY_COUNT=$(ls -d "${MAIN_DIR}/history"/*/ 2>/dev/null | wc -l)
     echo "Comparing PR branch against **median of ${HISTORY_COUNT} nightly run(s)** on \`main\` (latest: ${BASELINE_TIMESTAMP})."
   else
-    echo "Comparing PR branch against cached \`main\` baseline (generated ${BASELINE_TIMESTAMP})."
+    echo "Comparing PR branch against latest nightly baseline on \`main\` (${BASELINE_TIMESTAMP})."
   fi
 else
   echo "Comparing PR branch against \`main\` baseline."

.github/scripts/test-format-benchmark-report.sh

@@ -153,7 +153,7 @@ test_baseline_timestamp() {
 
   OUTPUT=$(BASELINE_TIMESTAMP="2026-03-30T04:00:00Z" bash "$REPORT_SCRIPT" "$PR_DIR" "$MAIN_DIR")
 
-  assert_contains "$OUTPUT" "cached" "Shows cached baseline label"
+  assert_contains "$OUTPUT" "latest nightly baseline" "Shows nightly baseline label"
   assert_contains "$OUTPUT" "2026-03-30T04:00:00Z" "Shows baseline timestamp"
   assert_contains "$OUTPUT" "benchmark-baseline" "Shows link to refresh workflow"
 }
@@ -522,6 +522,25 @@ test_e2e_column_with_history
 test_e2e_column_missing
 test_e2e_partial_coverage
 test_e2e_median_odd
+
+# Test 22: E2E column from flat e2e/ directory (artifacts API layout)
+test_e2e_flat_layout() {
+  echo "Test: E2E column populated from flat e2e/ directory (artifacts API)"
+  setup_dirs
+  echo '{"generator":"ts-sdk","spec":"square","duration_seconds":50,"exit_code":0}' > "$PR_DIR/ts-sdk.jsonl"
+
+  # Flat layout: generator-only at root, E2E in e2e/ subdir
+  echo '{"generator":"ts-sdk","spec":"square","duration_seconds":45,"exit_code":0}' > "$MAIN_DIR/ts-sdk.jsonl"
+  mkdir -p "$MAIN_DIR/e2e"
+  echo '{"generator":"ts-sdk","spec":"square","duration_seconds":200,"exit_code":0}' > "$MAIN_DIR/e2e/ts-sdk.jsonl"
+
+  OUTPUT=$(BASELINE_TIMESTAMP="2026-04-06T04:00:00Z" bash "$REPORT_SCRIPT" "$PR_DIR" "$MAIN_DIR")
+
+  assert_contains "$OUTPUT" "| ts-sdk | square | 45s | 200s | 50s |" "Flat layout: both gen and E2E shown"
+  assert_contains "$OUTPUT" "latest nightly baseline" "Flat layout: shows nightly baseline label"
+}
+
+test_e2e_flat_layout
 echo ""
 echo "=== PostHog JSON validation tests ==="
 echo ""

.github/workflows/benchmark-baseline.yml

@@ -78,7 +78,7 @@ jobs:
         with:
           name: benchmark-baseline-${{ matrix.generator }}
           path: benchmark-results/
-          retention-days: 7
+          retention-days: 30
 
   # Run full E2E benchmarks (with build/test scripts) for informational purposes.
   # These capture the complete customer-observable generation time including
@@ -116,100 +116,7 @@ jobs:
         with:
           name: benchmark-e2e-${{ matrix.generator }}
           path: benchmark-results/
-          retention-days: 7
-
-  # Aggregate results into rolling 7-day history cache
-  save-baseline-cache:
-    runs-on: ubuntu-latest
-    timeout-minutes: 10
-    needs: [baseline, baseline-e2e]
-    if: ${{ always() && !cancelled() && needs.baseline.result != 'skipped' }}
-    steps:
-      - name: Download generator-only baseline artifacts
-        uses: actions/download-artifact@v4
-        with:
-          pattern: benchmark-baseline-*
-          path: todays-results
-          merge-multiple: true
-
-      - name: Download E2E baseline artifacts
-        uses: actions/download-artifact@v4
-        with:
-          pattern: benchmark-e2e-*
-          path: todays-e2e-results
-          merge-multiple: true
-
-      - name: Restore existing history cache
-        id: history-cache
-        uses: actions/cache/restore@v4
-        with:
-          path: benchmark-baseline-results
-          # Prefix fallback finds the most recent history cache entry
-          key: benchmark-baseline-never-matches
-          restore-keys: benchmark-baseline-
-
-      - name: Append today's results and prune old entries
-        shell: bash
-        run: |
-          mkdir -p benchmark-baseline-results/history
-          TODAY=$(date -u +%Y-%m-%d)
-          RUN_DIR="benchmark-baseline-results/history/${TODAY}_${{ github.run_id }}"
-          mkdir -p "$RUN_DIR"
-
-          # Copy generator-only results (used for PR comparisons)
-          if [ -d todays-results ] && ls todays-results/*.jsonl 1>/dev/null 2>&1; then
-            for f in todays-results/*.jsonl; do
-              [ -f "$f" ] && cp "$f" "$RUN_DIR/"
-            done
-          else
-            echo "::warning::No generator-only baseline results found. Removing empty history entry."
-            rm -rf "$RUN_DIR"
-            echo "SKIP_CACHE_SAVE=true" >> "$GITHUB_ENV"
-          fi
-
-          # Only proceed with E2E + metadata if generator-only results were found
-          if [ -d "$RUN_DIR" ]; then
-            # Copy full E2E results into a separate subdirectory (informational only)
-            if [ -d todays-e2e-results ] && ls todays-e2e-results/*.jsonl 1>/dev/null 2>&1; then
-              mkdir -p "$RUN_DIR/e2e"
-              for f in todays-e2e-results/*.jsonl; do
-                [ -f "$f" ] && cp "$f" "$RUN_DIR/e2e/"
-              done
-            fi
-
-            # Store metadata for this run
-            echo "{\"timestamp\":\"$(date -u +%Y-%m-%dT%H:%M:%SZ)\",\"sha\":\"${{ github.sha }}\"}" \
-              > "$RUN_DIR/_metadata.json"
-          fi
-
-          # Prune history older than 7 days
-          # Note: 'date -d' is GNU coreutils syntax (CI runs on Ubuntu; not portable to macOS)
-          CUTOFF=$(date -u -d '7 days ago' +%Y-%m-%d)
-          for dir in benchmark-baseline-results/history/*/; do
-            [ -d "$dir" ] || continue
-            DIR_DATE=$(basename "$dir" | cut -d_ -f1)
-            if [[ "$DIR_DATE" < "$CUTOFF" ]]; then
-              echo "Pruning old baseline: $dir (date: $DIR_DATE, cutoff: $CUTOFF)"
-              rm -rf "$dir"
-            fi
-          done
-
-          # Write top-level metadata summarizing the history
-          RUNS=$(ls -d benchmark-baseline-results/history/*/ 2>/dev/null | wc -l)
-          echo "{\"latest_timestamp\":\"$(date -u +%Y-%m-%dT%H:%M:%SZ)\",\"sha\":\"${{ github.sha }}\",\"total_runs\":${RUNS}}" \
-            > benchmark-baseline-results/_metadata.json
-
-          echo "Baseline history summary:"
-          cat benchmark-baseline-results/_metadata.json
-          echo "History entries:"
-          ls -la benchmark-baseline-results/history/
-
-      - name: Save baseline to cache
-        if: ${{ env.SKIP_CACHE_SAVE != 'true' }}
-        uses: actions/cache/save@v4
-        with:
-          path: benchmark-baseline-results
-          key: benchmark-baseline-${{ github.run_id }}
+          retention-days: 30
 
   # Send benchmark results to PostHog for long-term trend analysis and dashboards.
   # Uses the same PostHog project + API key as seed test metrics (seed.yml).

.github/workflows/ci.yml

@@ -179,6 +179,36 @@ jobs:
       - name: Install
         uses: ./.github/actions/install
 
+      - name: Log in to Docker Hub
+        if: ${{ env.DOCKERHUB_USERNAME != '' }}
+        uses: docker/login-action@v3
+        env:
+          DOCKERHUB_USERNAME: ${{ secrets.DOCKER_USERNAME_PUBLIC_READONLY }}
+        with:
+          username: ${{ secrets.DOCKER_USERNAME_PUBLIC_READONLY }}
+          password: ${{ secrets.DOCKER_PASSWORD_PUBLIC_READONLY }}
+
+      - name: Clear stale Docker image cache
+        run: rm -rf /tmp/docker-cache
+
+      - name: Cache Docker images
+        id: docker-cache
+        uses: actions/cache@v4
+        with:
+          path: /tmp/docker-cache
+          key: ${{ runner.os }}-docker-ete-${{ hashFiles('**/Dockerfile*') }}
+          restore-keys: |
+            ${{ runner.os }}-docker-ete-
+
+      - name: Load cached Docker images
+        run: |
+          for f in /tmp/docker-cache/*.tar; do
+            [ -f "$f" ] && docker load -i "$f" || true
+          done
+
+      - name: Record pre-existing Docker images
+        run: docker images -q | sort -u > /tmp/docker-images-before.txt
+
       - name: Cache Fern bin dependencies
         uses: actions/cache@v4
         with:
@@ -232,6 +262,21 @@ jobs:
             echo "No snapshot changes to commit"
           fi
 
+      - name: Save Docker images for cache
+        if: always()
+        run: |
+          if [ ! -f /tmp/docker-images-before.txt ]; then
+            echo "Skipping cache save: docker-images-before.txt not found"
+            exit 0
+          fi
+          mkdir -p /tmp/docker-cache
+          comm -13 /tmp/docker-images-before.txt <(docker images -q | sort -u) | while read -r id; do
+            img=$(docker inspect --format '{{index .RepoTags 0}}' "$id" 2>/dev/null || true)
+            [ -z "$img" ] && continue
+            fname=$(echo "$img" | tr '/:' '_')
+            docker save "$img" -o "/tmp/docker-cache/${fname}.tar" 2>/dev/null || true
+          done
+
       - name: Ensure no changes to git-tracked files
         if: ${{ !github.event.inputs.update_snapshots }}
         run: git --no-pager diff --exit-code