Merge pull request #117 from logos-co/ci/validate-without-fork-checkout

ci: fetch submission files instead of checking out the fork

.github/scripts/validate-submission.sh

@@ -4,8 +4,11 @@
 #
 # Runtime contract (from validate-submission.yml):
 #   - Runs on ubuntu-latest (GNU grep / sed / coreutils assumed).
-#   - `base/` contains the trusted base-branch checkout.
-#   - `pr/`   contains the PR-head checkout, treated as untrusted data only.
+#   - `base/` contains the trusted base-branch checkout, and is the only
+#     source for repo-owned files such as the prize specs.
+#   - `pr/solutions/` contains just the submission markdown the PR changed,
+#     fetched by path from the API and treated as untrusted data only. The
+#     fork's working tree is never checked out.
 #   - Env: PR_TITLE, PR_REPO, BASE_REPO, CHANGED_FILES.
 set -euo pipefail
 
@@ -230,7 +233,9 @@ if $IS_SOLUTION; then
   # -------------------------------------------------------------------------
   # 3h. Prize exists and is open
   # -------------------------------------------------------------------------
-  PRIZE_FILE="pr/prizes/${PRIZE_ID}.md"
+  # Read from base/: the prize spec is repo-owned, so the authoritative copy
+  # is the base branch's, not whatever stale revision the fork branched from.
+  PRIZE_FILE="base/prizes/${PRIZE_ID}.md"
   if [[ ! -f "$PRIZE_FILE" ]]; then
     err "Prize \`${PRIZE_ID}\` not found in \`prizes/\`. Check the ID."
   else
@@ -261,7 +266,12 @@ if $IS_SOLUTION; then
 
     CLONE_DIR="/tmp/submission-repo"
     rm -rf "$CLONE_DIR"
-    if git clone --depth=1 "$REPO_URL" "$CLONE_DIR" 2>/dev/null; then
+    # `REPO_URL` is submitter-controlled, so the clone is bounded: no auth
+    # prompts, no submodules, one shallow branch, and a wall-clock ceiling so
+    # an oversized or slow repo cannot consume the whole job.
+    if GIT_TERMINAL_PROMPT=0 timeout "${LP_CLONE_TIMEOUT:-300}" \
+         git clone --depth=1 --single-branch --no-tags \
+                   --no-recurse-submodules "$REPO_URL" "$CLONE_DIR" 2>/dev/null; then
 
       # 4a. AI workspace artifacts in the external repo
       AI_ARTIFACTS=()
@@ -336,7 +346,9 @@ if $IS_SOLUTION; then
 
         # Logos Messaging / Waku integration
         if echo "$PRIZE_CONTENT" | grep -qi 'Logos Messaging\|Logos Chat\|Waku'; then
-          waku_ref=$(grep -ril 'waku\|logos.messaging\|logos.chat' "$CLONE_DIR" \
+          # Time-bounded: this is the one full-content scan of the clone.
+          waku_ref=$(timeout "${LP_SCAN_TIMEOUT:-60}" \
+            grep -ril 'waku\|logos.messaging\|logos.chat' "$CLONE_DIR" \
             --include='*.rs' --include='*.go' --include='*.ts' --include='*.js' \
             --include='*.toml' --include='*.json' 2>/dev/null | head -1 || true)
           if [[ -z "$waku_ref" ]]; then

.github/workflows/validate-submission.yml

@@ -2,8 +2,8 @@ name: Validate Submission
 
 # `pull_request_target` is used so the workflow can comment on PRs from forks
 # (which is how solutions are submitted). It runs on the base branch with
-# write perms; the PR-head checkout is treated as untrusted data only and
-# never executed.
+# write perms, so the fork's tree is never checked out here: the only
+# submission content the validator reads is fetched by path as data.
 on:
   pull_request_target:
     types: [opened, synchronize, reopened, edited]
@@ -26,13 +26,8 @@ jobs:
         with:
           ref: ${{ github.event.pull_request.base.sha }}
           path: base
-
-      - name: Checkout PR head (untrusted data, never executed)
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event.pull_request.head.sha }}
-          repository: ${{ github.event.pull_request.head.repo.full_name }}
-          path: pr
+          # Nothing in this job pushes, so the token has no reason to sit in
+          # base/.git/config for the life of the run.
           persist-credentials: false
 
       - name: Get changed files
@@ -52,6 +47,31 @@ jobs:
             echo "__LP_EOF__"
           } >> "$GITHUB_OUTPUT"
 
+      - name: Fetch changed solution files (untrusted data, never executed)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          REPO: ${{ github.repository }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+          CHANGED_FILES: ${{ steps.changed.outputs.files }}
+        run: |
+          # The validator reads exactly one thing from the PR: the submission
+          # markdown. Fetching those blobs by path keeps the fork's working
+          # tree off the runner entirely, so there is nothing here to execute
+          # and no `allow-unsafe-pr-checkout` opt-in to carry. Paths come from
+          # the PR files API and are constrained to a single markdown file
+          # directly under `solutions/`.
+          mkdir -p pr/solutions
+          solution_files=$(printf '%s\n' "$CHANGED_FILES" \
+            | grep -E '^solutions/[A-Za-z0-9._-]+\.md$' || true)
+          # Unquoted on purpose: the pattern above admits no spaces or globs.
+          for f in $solution_files; do
+            if ! gh api -H "Accept: application/vnd.github.raw" \
+                 "/repos/${REPO}/contents/${f}?ref=${HEAD_SHA}" > "pr/${f}"; then
+              rm -f "pr/${f}"
+              echo "::warning::Could not fetch ${f} at ${HEAD_SHA}."
+            fi
+          done
+
       - name: Run validation
         id: validate
         env: