ci: close attribution-gate gaps from Ox-Agent review

.github/workflows/ci.yml

@@ -2,6 +2,9 @@ name: ci
 
 on:
   pull_request:
+    # edited re-fires the gate on title/body changes — otherwise a
+    # post-green metadata edit could land attribution unchecked.
+    types: [opened, synchronize, reopened, edited]
   push:
     branches: [main]
 
@@ -16,6 +19,8 @@ jobs:
   commit-lint:
     name: commit-lint
     runs-on: ubuntu-latest
+    permissions:
+      contents: read
     steps:
       - uses: actions/checkout@v4
         with:
@@ -33,7 +38,11 @@ jobs:
           PR_TITLE: ${{ github.event.pull_request.title }}
         run: scripts/commit-check.sh --pr-title "$PR_TITLE"
       - name: scan PR body for attribution markers
-        if: github.event_name == 'pull_request'
+        # Bot-authored bodies (dependabot, renovate) are platform-
+        # generated boilerplate — GitHub injects zero-width spaces into
+        # @-mentions for anti-ping, which this scan exists to catch.
+        # Not a human attribution surface; commits/title still lint.
+        if: github.event_name == 'pull_request' && github.event.pull_request.user.type != 'Bot'
         env:
           PR_BODY: ${{ github.event.pull_request.body }}
         run: |
@@ -45,8 +54,10 @@ jobs:
           # a stray start marker could hide attribution after it.
           open_m='<!-- This is an auto-generated comment: release notes by coderabbit.ai -->'
           close_m='<!-- end of auto-generated comment: release notes by coderabbit.ai -->'
-          opens=$(grep -cF "$open_m" pr-body.txt || :)
-          closes=$(grep -cF "$close_m" pr-body.txt || :)
+          # -oF counts occurrences, not matching lines — same-line
+          # pairs would read as one open and one close to -c.
+          opens=$(grep -oF "$open_m" pr-body.txt | wc -l || :)
+          closes=$(grep -oF "$close_m" pr-body.txt | wc -l || :)
           if [ "$opens" != "$closes" ]; then
             echo "commit-check: unbalanced CodeRabbit release-notes markers in PR body" >&2
             exit 1
@@ -67,6 +78,12 @@ jobs:
               }
               printf "%s", out rest
             }' pr-body.txt > pr-body-clean.txt
+          # equal counts but wrong order leave markers behind in the
+          # cleaned body — a stray open/close is not a managed block.
+          if grep -qF "$open_m" pr-body-clean.txt || grep -qF "$close_m" pr-body-clean.txt; then
+            echo "commit-check: misordered CodeRabbit release-notes markers in PR body" >&2
+            exit 1
+          fi
           scripts/commit-check.sh --scan-file pr-body-clean.txt
       - name: lint push range
         if: github.event_name == 'push'
@@ -78,6 +95,14 @@ jobs:
             echo "new branch push — no commit range to lint"
             exit 0
           fi
+          # a force-pushed before-SHA can be unreachable from every ref;
+          # fetch it best-effort — if it stays unresolvable the script
+          # fails the check rather than silently linting zero commits.
+          # Never --depth: a shallow boundary hides PUSH_BEFORE's
+          # ancestry, and rev-list would then lint all of history. The
+          # fetch-depth:0 checkout already has it for normal pushes —
+          # fetch only when absent, without a depth limit.
+          git cat-file -e "$PUSH_BEFORE^{commit}" 2>/dev/null             || git fetch --no-tags origin "$PUSH_BEFORE" 2>/dev/null             || true
           scripts/commit-check.sh --range "$PUSH_BEFORE..$PUSH_AFTER"
 
   # Site gate: type-aware oxlint + oxfmt/prettier check + full static build.

scripts/commit-check.sh

@@ -36,32 +36,58 @@ warn() { printf '%s\n' "commit-check: warning: $*" >&2; }
 
 check_attribution() {
     _msg=$1
+    # One vendor alternation shared by every pattern below — extend it
+    # here, never in a single rule, or the rules drift apart.
+    _VENDOR='claude|cursor|copilot|chatgpt|codex|gemini|devin|openai|anthropic|codeium|windsurf|github[[:space:]]+copilot|claude[[:space:]]+code|(claude|gemini)[[:space:]]+(opus|sonnet|pro|ultra)|swe-[0-9a-z.-]*' 
     # AI co-author / bot trailers. A human Co-authored-by is fine; the
     # ban is on AI vendors, bots, and *-ai-* / [bot] addresses.
-    if printf '%s\n' "$_msg" | grep -iE '^[[:space:]]*co-?authored-by:' \
-        | grep -qiE 'claude|cursor|copilot|chatgpt|codex|gemini|devin-ai|cognition\.ai|devin\.ai[[:space:]>]|openai|anthropic|codeium|windsurf|swe-|\[bot\]|[a-z0-9-]*bot@|ai-integration'; then
+    _co=$(printf '%s\n' "$_msg" | grep -iE '^[[:space:]]*co-?authored-by:' || :)
+    # Unambiguous machine signals anywhere in the trailer: [bot] tags,
+    # bot@ addresses, *-ai-integration logins, AI-vendor domains (literal
+    # list of vendor-owned domains — never github.com, or humans on the
+    # privacy relay get flagged), and a vendor token inside a bare email
+    # (no <display name>) — bot forms like
+    # "Co-authored-by: 175728472+Copilot@users.noreply.github.com".
+    if printf '%s\n' "$_co" | grep -qiE '\[bot\]|(^|[<.[:space:]_+-])bot@|ai-integration|noreply@(anthropic|openai)\.|@(anthropic|openai|cognition|devin|codeium|windsurf|claude-code|cursor|gemini|chatgpt|codex|copilot|google)\.(com|ai|dev|io)([[:space:]>]|$)|:[[:space:]]*[^<[:space:]]*('"$_VENDOR"')[^<[:space:]]*@'; then
+        err "AI Co-authored-by trailer"
+    # A bare vendor identity as the display name — "Co-authored-by:
+    # Claude <noreply@anthropic.com>" hits the domain rule above; this
+    # catches no-email and generic-mailbox forms. Human names that merely
+    # CONTAIN a vendor token ("Claude Martin <claude.martin@…>") stay
+    # legal per .cursor/rules/no-ai-attribution.mdc.
+    elif printf '%s\n' "$_co" | sed 's/^[^:]*:[[:space:]]*//; s/<.*//' \
+        | grep -qiE '^\[?('"$_VENDOR"')\]?(\([^)]*\))?([[:space:]-]?ai)?[[:space:]]*$'; then
         err "AI Co-authored-by trailer"
     fi
     # generator footers: "Generated with <tool>", "Made with: X",
     # "Authored via X" — the tool names that sign output, plus the
     # trailer forms regardless of vendor.
-    if printf '%s\n' "$_msg" | grep -iqE '(generated|made|created|authored|written|produced)[[:space:]]+(with|by|via)[[:space:]]*\[?[[:space:]]*(claude|cursor|copilot|chatgpt|codex|gemini|devin|openai|anthropic|codeium|windsurf|swe-)'; then
+    # connector may carry a colon ("Created with: Cursor"); the vendor
+    # token needs a non-alnum tail so "generated by cursors" stays clean.
+    if printf '%s\n' "$_msg" | grep -iqE '(generated|made|created|authored|written|produced)[[:space:]]+(with|by|via):?[[:space:]]*\[?[[:space:]]*('"$_VENDOR"')([^a-z0-9]|$)'; then
         err "AI generator footer"
     fi
     if printf '%s\n' "$_msg" | grep -iqE '^[[:space:]]*(generated|made|authored)[- ](with|by|via)[[:space:]]*:'; then
         err "generator trailer"
     fi
-    # bot-emoji generator signature
-    if printf '%s\n' "$_msg" | grep -q '🤖'; then
+    # standalone "via <tool>" / "by <tool>" / "with <tool>" lines — the
+    # footer regex above needs a leading verb, so a lone via-line slips.
+    if printf '%s\n' "$_msg" | grep -qiE '^[[:space:]]*(via|by|using|with)[[:space:]]+\[?('"$_VENDOR"')\]?(\([^)]*\))?([[:space:]-]?ai)?([^a-z0-9]|$)'; then
+        err "attribution via-line"
+    fi
+    # bot-emoji signature: a bare 🤖 line, or 🤖 sharing a line with
+    # attribution vocabulary — a mid-sentence robot mention stays clean.
+    _bot=$(printf '%s\n' "$_msg" | grep '🤖' || :)
+    if printf '%s\n' "$_bot" | grep -qiE '^[[:space:]]*🤖[[:space:]]*$|generated|made|created|authored|written|produced|built|co-?authored|via[[:space:]]|with[[:space:]]|by[[:space:]]'; then
         err "bot-emoji attribution marker"
     fi
     # AI named as the author/agent — not as the subject of the change.
     # (kept narrow on purpose: "add Copilot webhook" must stay legal)
-    if printf '%s\n' "$_msg" | grep -iqE '\b(as an ai|i generated this|authored by (claude|cursor|copilot|chatgpt|codex|gemini|devin|openai|swe-))\b'; then
+    if printf '%s\n' "$_msg" | grep -iqE '\b(as an ai|i generated this|authored by \[?('"$_VENDOR"'))\b'; then
         err "AI self-attribution"
     fi
     # vendor-stamped HTML comments smuggled into the message
-    if printf '%s\n' "$_msg" | grep -iqE '<!--[^>]*(claude|cursor|copilot|chatgpt|codex|gemini|devin|openai|anthropic|generated|🤖)'; then
+    if printf '%s\n' "$_msg" | grep -iqE '<!--[^>]*('"$_VENDOR"'|generated|🤖)'; then
         err "attribution HTML comment"
     fi
     # zero-width / invisible markers: U+200B..D, U+2060, U+FEFF as raw