diff --git a/.agents/skills/autoreview/SKILL.md b/.agents/skills/autoreview/SKILL.md
index c132173d429..6da7daee3e5 100644
--- a/.agents/skills/autoreview/SKILL.md
+++ b/.agents/skills/autoreview/SKILL.md
@@ -28,6 +28,7 @@ Use when:
 - Treat the helper's successful exit plus absence of actionable findings as the clean review result, even if the underlying Codex CLI output is terse.
 - If rejecting a finding as intentional/not worth fixing, add a brief inline code comment only when it explains a real invariant or ownership decision that future reviewers should know.
 - Do not push just to review. Push only when the user requested push/ship/PR update.
+- In OpenClaw, keep autoreview validation Crabbox/Testbox-aware. A review pass may inspect files and run cheap non-Node probes, but it must not start local `pnpm`, Vitest, `tsgo`, `npm test`, or `node scripts/run-vitest.mjs` from a Codex/worktree review unless the operator explicitly requested local proof. For runtime proof, use existing evidence or route through `OPENCLAW_TESTBOX=1` / Crabbox/Testbox and report the id.
 
 ## Pick Target
 
@@ -50,7 +51,7 @@ git fetch origin
 codex review --base origin/main
 ```
 
-Do not pass an inline prompt with `--base`; current CLI rejects `--base` + `[PROMPT]` even though help text is ambiguous. If custom instructions are needed, run the plain base review first, then do a local/manual follow-up pass.
+Do not pass an inline prompt with `--base`; some CLI versions reject `--base` + `[PROMPT]` even though help text is ambiguous. If custom instructions are needed, prefer stdin prompt form (`codex review --base <ref> -`) as used by the helper.
 
 If an open PR exists, use its actual base:
 
@@ -115,6 +116,7 @@ The helper:
 - writes only to stdout unless `--output` or `AUTOREVIEW_OUTPUT` is set
 - supports `--dry-run`, `--parallel-tests`, and commit refs
 - runs nested review with `--dangerously-bypass-approvals-and-sandbox --sandbox danger-full-access` by default
+- injects OpenClaw validation policy into native Codex review so local memory-heavy Node/Vitest checks are avoided in favor of Crabbox/Testbox proof
 - keeps accepting `--full-access`; use `--no-yolo` or `AUTOREVIEW_YOLO=0` to opt out
 - still accepts legacy `CODEX_REVIEW_*` env vars when the matching `AUTOREVIEW_*` var is unset
 - prints `autoreview clean: no accepted/actionable findings reported` when the selected review command exits 0
diff --git a/.agents/skills/autoreview/scripts/autoreview b/.agents/skills/autoreview/scripts/autoreview
index c442edec774..0931c07cebd 100755
--- a/.agents/skills/autoreview/scripts/autoreview
+++ b/.agents/skills/autoreview/scripts/autoreview
@@ -51,6 +51,7 @@ yolo=${AUTOREVIEW_YOLO:-${CODEX_REVIEW_YOLO:-1}}
 output=${AUTOREVIEW_OUTPUT:-${CODEX_REVIEW_OUTPUT:-}}
 parallel_tests=
 dry_run=false
+codex_review_prompt=
 
 while [[ $# -gt 0 ]]; do
   case "$1" in
@@ -201,6 +202,19 @@ else
   review_cmd=("$codex_bin" "${codex_args[@]}" review --base "$base_ref")
 fi
 
+repo_url=$(git -C "$repo_root" config --get remote.origin.url 2>/dev/null || true)
+if [[ "$repo_url" == *"openclaw/openclaw"* ]]; then
+  codex_review_prompt=$(cat <<'EOF'
+OpenClaw autoreview validation policy:
+- Review the diff by reading code, tests, and dependency contracts.
+- Do not run local memory-heavy Node validation from review mode. This includes local pnpm checks/tests, Vitest, tsgo, npm test, and node scripts/run-vitest.mjs.
+- If runtime proof is needed, use existing proof or route validation through OPENCLAW_TESTBOX=1 / Crabbox / Blacksmith Testbox and report the exact provider and id.
+- If remote validation is not necessary for the finding, state the targeted proof that should be run instead of starting local tests.
+EOF
+)
+  review_cmd+=(-)
+fi
+
 printf 'autoreview target: %s\n' "$review_kind"
 printf 'branch: %s\n' "${current_branch:-detached}"
 if [[ -n "$pr_url" ]]; then
@@ -221,6 +235,9 @@ if [[ "$reviewer" == auto || "$reviewer" == codex ]]; then
   printf 'review:'
   printf ' %q' "${review_cmd[@]}"
   printf '\n'
+  if [[ -n "$codex_review_prompt" ]]; then
+    printf 'review policy: OpenClaw Crabbox/Testbox-aware validation prompt injected\n'
+  fi
 else
   printf 'review: %s prompt review\n' "$reviewer"
 fi
@@ -265,7 +282,11 @@ trap cleanup EXIT
 
 run_review() {
   mkdir -p "$(dirname "$review_output")"
-  "${review_cmd[@]}" 2>&1 | tee "$review_output"
+  if [[ -n "$codex_review_prompt" ]]; then
+    printf '%s\n' "$codex_review_prompt" | "${review_cmd[@]}" 2>&1 | tee "$review_output"
+  else
+    "${review_cmd[@]}" 2>&1 | tee "$review_output"
+  fi
 }
 
 diff_for_review() {
@@ -306,6 +327,7 @@ Rules:
 - Review the proposed code change as a closeout reviewer.
 - Focus on the diff below. If your CLI exposes read-only repository tools, inspect surrounding code and tests to verify findings; never modify files.
 - Do not modify files.
+- In OpenClaw, do not run local memory-heavy Node validation from review mode. Avoid local pnpm checks/tests, Vitest, tsgo, npm test, and node scripts/run-vitest.mjs. If runtime proof is needed, use existing proof or route validation through OPENCLAW_TESTBOX=1 / Crabbox / Blacksmith Testbox and report the exact provider and id.
 - Report only discrete, actionable issues introduced by this change.
 - Prioritize correctness, regressions, security, data loss, performance cliffs, and missing tests that would catch a real bug.
 - Do not report pre-existing issues, speculative risks, broad rewrites, style nits, changelog gaps, or findings that depend on unstated assumptions.