diff --git a/.github/workflows/ai-review.yml b/.github/workflows/ai-review.yml
index 471e33c1..eaac91d5 100644
--- a/.github/workflows/ai-review.yml
+++ b/.github/workflows/ai-review.yml
@@ -1,17 +1,15 @@
 # AI Code Review Workflow
-# Uses Claude Code CLI on self-hosted runner with CLIProxyAPI
+# Uses anthropics/claude-code-action with CLIProxy for model flexibility
 # Posts as ccs-agy-reviewer[bot] via GitHub App
 #
 # Triggers:
 # - Automatically on PR open/update
 # - Manually via /review comment on PR
-#
-# Logs stored on runner at: /opt/actions-runner/logs/ai-review/
 
 name: AI Code Review
 
 on:
-  pull_request:
+  pull_request_target:
     types: [opened, synchronize, reopened]
   issue_comment:
     types: [created]
@@ -30,20 +28,32 @@ concurrency:
 jobs:
   review:
     name: Claude Code Review
-    # Only run on self-hosted runner with cliproxy access
     runs-on: [self-hosted, cliproxy]
+    permissions:
+      contents: read
+      pull-requests: write
+      issues: write
+      id-token: write
 
     # Conditions:
-    # - PR event: always run
+    # - PR event: always run (pull_request_target has secrets access for forks)
     # - Comment event: only if it's a PR and contains /review
     if: >
-      github.event_name == 'pull_request' ||
+      github.event_name == 'pull_request_target' ||
       github.event_name == 'workflow_dispatch' ||
       (github.event_name == 'issue_comment' &&
        github.event.issue.pull_request &&
        contains(github.event.comment.body, '/review'))
 
+    # CLIProxy environment for model routing
+    env:
+      ANTHROPIC_BASE_URL: http://localhost:8317
+      REVIEW_MODEL: gemini-claude-opus-4-5-thinking
+
     steps:
+      - name: Clear Claude install locks
+        run: rm -rf ~/.local/state/claude/locks/* 2>/dev/null || true
+
       - name: Generate App Token
         id: app-token
         uses: actions/create-github-app-token@v1
@@ -51,7 +61,7 @@ jobs:
           app-id: ${{ secrets.CCS_REVIEWER_APP_ID }}
           private-key: ${{ secrets.CCS_REVIEWER_PRIVATE_KEY }}
 
-      - name: Checkout PR branch
+      - name: Checkout repository
         uses: actions/checkout@v4
         with:
           fetch-depth: 0
@@ -65,17 +75,6 @@ jobs:
             gh pr checkout $PR_NUM || git checkout ${{ github.event.pull_request.head.sha }} 2>/dev/null || true
           fi
 
-      - name: Determine PR Number
-        id: pr
-        run: |
-          if [ "${{ github.event_name }}" = "pull_request" ]; then
-            echo "number=${{ github.event.pull_request.number }}" >> $GITHUB_OUTPUT
-          elif [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
-            echo "number=${{ github.event.inputs.pr_number }}" >> $GITHUB_OUTPUT
-          else
-            echo "number=${{ github.event.issue.number }}" >> $GITHUB_OUTPUT
-          fi
-
       - name: Add reaction to comment
         if: github.event_name == 'issue_comment'
         run: |
@@ -84,360 +83,51 @@ jobs:
         env:
           GH_TOKEN: ${{ steps.app-token.outputs.token }}
 
-      - name: Ensure Claude Code installed
-        run: |
-          # Force cleanup of any stale lock/install markers
-          rm -rf ~/.claude/.installing ~/.claude/downloads 2>/dev/null || true
-          pkill -f "claude install" 2>/dev/null || true
-
-          if ! command -v claude &> /dev/null; then
-            echo "[i] Installing Claude Code..."
-            # Download and run install script
-            curl -fsSL https://claude.ai/install.sh > /tmp/claude-install.sh
-            chmod +x /tmp/claude-install.sh
-
-            # Run installer; on failure, try running claude install --force directly
-            if ! bash /tmp/claude-install.sh; then
-              echo "[!] Script install failed, trying direct --force..."
-              BINARY=$(find ~/.claude -name "claude" -type f -executable 2>/dev/null | head -1)
-              if [ -n "$BINARY" ]; then
-                "$BINARY" install --force || true
-              fi
-            fi
-            echo "$HOME/.local/bin" >> $GITHUB_PATH
-          else
-            echo "[i] Claude Code already installed: $(claude --version)"
-          fi
-
-      - name: Create review prompt template
-        run: |
-          cat > /tmp/review-instructions.txt << 'INSTRUCTIONS_END'
-          You are a senior staff engineer performing a comprehensive code review.
-
-          ## MANDATORY: Tool Usage Requirements
-          Before writing your review, you MUST:
-
-          1. **Read each changed file in full** using the Read tool
-             - Don't just look at the diff - read the ENTIRE file for context
-             - Understand the surrounding code, imports, and dependencies
-
-          2. **Find related test files** using Glob tool
-             - Pattern: `**/*test*` or `**/*.spec.*` or `**/__tests__/**`
-             - Read the test files to assess coverage
-
-          3. **Search for usages** using Grep tool
-             - Find where modified functions/classes are called
-             - Check for breaking changes impact
-
-          4. **Check project structure** using Glob tool
-             - Understand the codebase architecture
-             - Verify the changes follow existing patterns
-
-          ## Analysis Dimensions
-
-          ### 🔒 Security (CRITICAL)
-          - Auth bypass, privilege escalation, injection (SQL, command, XSS)
-          - Secrets/credentials exposure, insecure defaults
-          - OWASP Top 10 vulnerabilities
-          - Input validation at trust boundaries
-
-          ### ✅ Correctness
-          - Logic errors, off-by-one, null/undefined handling
-          - Race conditions, deadlocks, async issues
-          - Edge cases, boundary conditions
-          - Error handling completeness
-
-          ### 🏗️ Architecture and Design
-          - Coupling, cohesion, separation of concerns
-          - Design patterns usage (appropriate or over-engineered)
-          - API contract changes, backward compatibility
-          - Single responsibility, interface segregation
-
-          ### 🧪 Testing
-          - Test coverage for new/modified code
-          - Edge case coverage, negative tests
-          - Test quality (not just quantity)
-          - Integration vs unit test balance
-
-          ### ⚡ Performance
-          - O(n^2) or worse algorithms, unnecessary iterations
-          - Memory leaks, resource cleanup
-          - Database query efficiency (N+1, missing indexes)
-          - Only flag if measurable impact
-
-          ### 📖 Maintainability
-          - Naming clarity, self-documenting code
-          - Comments for "why" not "what"
-          - Complexity (cyclomatic, cognitive)
-          - Consistency with codebase patterns
-
-          ## Output Format (Markdown)
-
-          ## 🔍 Code Review
-
-          **Verdict**: [Choose one: ✅ Approve | ⚠️ Approve with notes | ❌ Request changes]
-
-          > 🤖 Reviewed by `MODEL_NAME`
-
-          ### Summary
-          [2-3 sentences: what this PR does, the approach taken, and overall assessment]
-
-          ### ✅ Strengths
-          [Specific things done well with file:line-range references]
-          1. **[Category]**: Description
-             - `src/path/file.ts:10-25` - specific code reference
-          2. ...
-
-          ### ⚠️ Issues Found
-          | File:Line | Issue | Severity | Impact |
-          |-----------|-------|----------|--------|
-          | `src/file.ts:123-145` | Description | 🔴/🟡/🟢 | What could go wrong |
-
-          ### 🔍 Code Quality Observations
-          **Positive:**
-          - Observation with `file.ts:line` reference
-
-          **Concerns:**
-          - Concern with `file.ts:line` reference and code example:
-          ```typescript
-          // Current code (problematic):
-          const x = something.slice(0, 60);
-
-          // Suggested improvement:
-          const x = truncateString(something, 60);
-          ```
-
-          ### 🧪 Testing
-          - **Coverage**: [Assessed by reading test files]
-          - **Test files examined**: `tests/path/file.test.ts`
-          - **Missing tests**: [List specific scenarios not covered]
-          - **Test quality**: [Observations about test implementations]
-
-          ### 🔒 Security
-          - [Detailed security analysis or "No concerns identified"]
-
-          ### ⚡ Performance
-          - [Performance analysis or "No concerns identified"]
-
-          ### 📝 Recommendations
-          1. **[Category]**: Specific recommendation
-             ```typescript
-             // Instead of:
-             problematicCode();
-
-             // Consider:
-             betterApproach();
-             ```
-
-          2. ...
-
-          ### 🎯 Alignment with Project Standards
-          - ✅/❌ Follows codebase patterns (cite specific examples)
-          - ✅/❌ Consistent naming/style
-          - ✅/❌ Proper error handling
-
-          ### Verdict: **[Final verdict with emoji]**
-          [One sentence final summary with key takeaway]
-
-          ---
-          **Guidelines:**
-          - Skip empty sections
-          - Only ❌ Request changes for 🔴 High severity issues
-          - Include code snippets for ALL non-trivial suggestions
-          - Reference specific line ranges, not just file names
-          INSTRUCTIONS_END
-
       - name: Run Claude Code Review
-        env:
-          ANTHROPIC_BASE_URL: http://localhost:8317
-          ANTHROPIC_API_KEY: ${{ secrets.CLIPROXY_API_KEY }}
-          ANTHROPIC_MODEL: gemini-3-pro-preview
-          GH_TOKEN: ${{ steps.app-token.outputs.token }}
-          GITHUB_TOKEN: ${{ steps.app-token.outputs.token }}
-        run: |
-          PR_NUM="${{ steps.pr.outputs.number }}"
-          REPO="${{ github.repository }}"
-          TIMESTAMP=$(date +%Y%m%d-%H%M%S)
+        id: claude-review
+        uses: anthropics/claude-code-action@v1
+        with:
+          anthropic_api_key: ${{ secrets.CLIPROXY_API_KEY }}
+          github_token: ${{ steps.app-token.outputs.token }}
+          path_to_claude_code_executable: /home/github-runner/.local/bin/claude
+          track_progress: ${{ github.event_name != 'workflow_dispatch' }}
+          prompt: |
+            ultrathink
 
-          # Setup logging directory on runner
-          LOG_DIR="/opt/actions-runner/logs/ai-review/pr-${PR_NUM}-${TIMESTAMP}"
-          mkdir -p "$LOG_DIR"
+            REPO: ${{ github.repository }}
+            PR NUMBER: ${{ github.event.pull_request.number || github.event.issue.number || github.event.inputs.pr_number }}
 
-          echo "[i] Running Claude Code review for PR #${PR_NUM}"
-          echo "[i] Model: ${ANTHROPIC_MODEL}"
-          echo "[i] Working directory: $(pwd)"
-          echo "[i] Log directory: ${LOG_DIR}"
+            Perform a comprehensive code review. Follow the repository's CLAUDE.md for project-specific guidelines.
 
-          # Get PR metadata
-          PR_TITLE=$(gh pr view $PR_NUM --repo $REPO --json title --jq .title)
-          PR_BODY=$(gh pr view $PR_NUM --repo $REPO --json body --jq .body)
+            ## Review Focus Areas
 
-          # Get list of changed files
-          CHANGED_FILES=$(gh pr diff $PR_NUM --repo $REPO --name-only)
+            1. **Security** - OWASP Top 10, injection, auth bypass, secrets exposure
+            2. **Correctness** - Logic errors, edge cases, error handling
+            3. **Architecture** - Coupling, cohesion, API contracts, backward compatibility
+            4. **Testing** - Coverage, edge cases, test quality
+            5. **Performance** - Algorithm complexity, memory leaks, query efficiency
+            6. **Maintainability** - Naming, complexity, consistency with codebase
 
-          # Get full diff (NO LIMIT - unlimited)
-          DIFF=$(gh pr diff $PR_NUM --repo $REPO)
+            ## Output Structure
 
-          # Replace model name in instructions
-          sed -i "s|MODEL_NAME|${ANTHROPIC_MODEL}|g" /tmp/review-instructions.txt
+            Use visual hierarchy with `---` separators between major sections:
+            - Summary (2-3 sentences)
+            - Strengths (numbered, with file:line references)
+            - Observations & Suggestions (with code examples)
+            - Security Considerations (✅/❌ checklist)
+            - Code Quality Checklist (✅/❌ format)
+            - Recommendations (Priority: High/Medium/Low)
+            - Overall Assessment (APPROVED/APPROVED WITH NOTES/CHANGES REQUESTED)
 
-          # Build the full prompt
-          cat > /tmp/full-prompt.txt << PROMPT_END
-          # Pull Request Review Task
+            Be thorough but concise. Skip empty sections. Include code examples for non-trivial suggestions.
 
-          ## PR Information
-          - **PR**: #${PR_NUM}
-          - **Title**: ${PR_TITLE}
-          - **Repository**: ${REPO}
+            End your review with:
+            > 🤖 Reviewed by `${{ env.REVIEW_MODEL }}`
 
-          ## Description
-          ${PR_BODY}
-
-          ## Changed Files
-          ${CHANGED_FILES}
-
-          ## Diff (for initial understanding only)
-          ${DIFF}
-
-          ## Your Task
-
-          **STEP 1 - MANDATORY EXPLORATION (before writing review):**
-          You MUST use tools to explore the codebase. Do not skip this step.
-
-          a) For EACH file in "Changed Files" above:
-             - Use Read tool to read the FULL file (not just diff)
-             - Examine imports, exports, and dependencies
-
-          b) Find test files:
-             - Use Glob with pattern: tests/**/*.test.* or **/*.spec.*
-             - Read relevant test files to assess coverage
-
-          c) Search for usages:
-             - Use Grep to find where modified functions/classes are used
-             - Assess breaking change impact
-
-          **STEP 2 - WRITE REVIEW:**
-          Only after completing Step 1, write your review following the format below.
-
-          $(cat /tmp/review-instructions.txt)
-
-          ---
-          BEGIN YOUR EXPLORATION NOW. Use Read, Glob, and Grep tools before writing your review.
-          PROMPT_END
-
-          # Save prompt for debugging
-          cp /tmp/full-prompt.txt "$LOG_DIR/prompt.txt"
-
-          # Save metadata
-          cat > "$LOG_DIR/metadata.json" << EOF
-          {
-            "pr_number": ${PR_NUM},
-            "repository": "${REPO}",
-            "model": "${ANTHROPIC_MODEL}",
-            "timestamp": "${TIMESTAMP}",
-            "changed_files_count": $(echo "$CHANGED_FILES" | wc -l),
-            "diff_lines": $(echo "$DIFF" | wc -l),
-            "workflow_run_id": "${{ github.run_id }}"
-          }
-          EOF
-
-          # Run Claude Code with NO LIMITS
-          echo "[i] Starting comprehensive review (unlimited turns)..."
-          START_TIME=$(date +%s)
-
-          # Run with separate stdout/stderr capture
-          set +e
-          cat /tmp/full-prompt.txt | claude -p \
-            --allowed-tools "Read,Glob,Grep" \
-            > "$LOG_DIR/claude-stdout.txt" \
-            2> "$LOG_DIR/claude-stderr.txt"
-          EXIT_CODE=$?
-          set -e
-
-          END_TIME=$(date +%s)
-          DURATION=$((END_TIME - START_TIME))
-
-          # Save timing info
-          cat > "$LOG_DIR/timing.txt" << EOF
-          Start: $(date -d @$START_TIME '+%Y-%m-%d %H:%M:%S')
-          End: $(date -d @$END_TIME '+%Y-%m-%d %H:%M:%S')
-          Duration: ${DURATION} seconds
-          Exit code: ${EXIT_CODE}
-          EOF
-
-          echo "[i] Review completed in ${DURATION}s with exit code ${EXIT_CODE}"
-
-          # Read the review output
-          REVIEW=$(cat "$LOG_DIR/claude-stdout.txt")
-          REVIEW_LEN=${#REVIEW}
-
-          echo "[i] Review output: ${REVIEW_LEN} chars"
-          echo "[i] Stderr output: $(wc -c < "$LOG_DIR/claude-stderr.txt") chars"
-
-          # Log stderr if non-empty (for debugging)
-          if [ -s "$LOG_DIR/claude-stderr.txt" ]; then
-            echo "[!] Claude stderr output:"
-            cat "$LOG_DIR/claude-stderr.txt"
-          fi
-
-          # Validate review output - retry if too short
-          if [ -z "$REVIEW" ] || [ ${REVIEW_LEN} -lt 100 ]; then
-            echo "[!] Review output too short (${REVIEW_LEN} chars), retrying without tool restrictions..."
-
-            cat /tmp/full-prompt.txt | claude -p \
-              > "$LOG_DIR/claude-stdout-retry.txt" \
-              2> "$LOG_DIR/claude-stderr-retry.txt" || true
-
-            REVIEW=$(cat "$LOG_DIR/claude-stdout-retry.txt")
-            REVIEW_LEN=${#REVIEW}
-            echo "[i] Retry output: ${REVIEW_LEN} chars"
-          fi
-
-          # Final fallback with error details
-          if [ -z "$REVIEW" ] || [ ${REVIEW_LEN} -lt 50 ]; then
-            STDERR_CONTENT=$(cat "$LOG_DIR/claude-stderr.txt" 2>/dev/null | head -10 || echo "No stderr")
-            REVIEW="## 🔍 Code Review
-
-          **Verdict**: ⚠️ Review generation failed
-
-          > 🤖 Attempted by \`${ANTHROPIC_MODEL}\`
-
-          The automated review encountered an issue. Details logged to runner.
-
-          **Debug info:**
-          - Exit code: ${EXIT_CODE}
-          - Duration: ${DURATION}s
-          - Output length: ${REVIEW_LEN} chars
-          - Log path: \`${LOG_DIR}\`
-
-          Please request a manual review or check runner logs."
-          fi
-
-          # Save final review
-          echo "$REVIEW" > "$LOG_DIR/final-review.txt"
-
-          # Update metadata with result
-          cat > "$LOG_DIR/metadata.json" << EOF
-          {
-            "pr_number": ${PR_NUM},
-            "repository": "${REPO}",
-            "model": "${ANTHROPIC_MODEL}",
-            "timestamp": "${TIMESTAMP}",
-            "changed_files_count": $(echo "$CHANGED_FILES" | wc -l),
-            "diff_lines": $(echo "$DIFF" | wc -l),
-            "workflow_run_id": "${{ github.run_id }}",
-            "duration_seconds": ${DURATION},
-            "exit_code": ${EXIT_CODE},
-            "review_length": ${REVIEW_LEN},
-            "success": $([ ${REVIEW_LEN} -ge 100 ] && echo "true" || echo "false")
-          }
-          EOF
-
-          # Post review as comment
-          echo "[i] Posting review (${REVIEW_LEN} chars)..."
-          echo "$REVIEW" | gh pr comment $PR_NUM --repo $REPO --body-file -
-          echo "[OK] Review posted"
-          echo "[i] Logs saved to: ${LOG_DIR}"
+          claude_args: |
+            --model ${{ env.REVIEW_MODEL }}
+            --allowedTools "Bash(gh pr comment:*),Bash(gh pr diff:*),Bash(gh pr view:*),Read,Glob,Grep"
+        continue-on-error: true
 
       - name: Add success reaction
         if: success() && github.event_name == 'issue_comment'
diff --git a/package.json b/package.json
index 4f7cdf4e..b64232ac 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@kaitranntt/ccs",
-  "version": "7.16.0",
+  "version": "7.16.0-dev.5",
   "description": "Claude Code Switch - Instant profile switching between Claude Sonnet 4.5 and GLM 4.6",
   "keywords": [
     "cli",