diff --git a/claude-dev/Dockerfile b/claude-dev/Dockerfile
index 5fc7d46..36d15f1 100644
--- a/claude-dev/Dockerfile
+++ b/claude-dev/Dockerfile
@@ -6,9 +6,11 @@ COPY required-tools.txt /opt/claude-dev/required-tools.txt
 COPY scripts /opt/claude-dev/scripts
 
 # LiteLLM translation proxy config
+RUN mkdir -p /etc/claude-dev
 COPY config/litellm.yaml /etc/claude-dev/litellm.yaml
 COPY scripts/start-litellm.sh /opt/claude-dev/scripts/start-litellm.sh
 
 RUN chmod +x /opt/claude-dev/scripts/*.sh
 
-CMD ["bash"]
+# Start LiteLLM proxy, then bash
+CMD ["sh", "-c", "DEEPSEEK_API_KEY=${DEEPSEEK_API_KEY} litellm --config /etc/claude-dev/litellm.yaml --port 4008 --host 0.0.0.0 > /tmp/litellm.log 2>&1 & exec bash"]
diff --git a/claude-dev/Dockerfile.base b/claude-dev/Dockerfile.base
index 9358645..8bf218a 100644
--- a/claude-dev/Dockerfile.base
+++ b/claude-dev/Dockerfile.base
@@ -37,56 +37,38 @@ RUN wget -q -O /usr/local/bin/tea "https://gitea.com/gitea/tea/releases/download
 
 RUN npm install -g @anthropic-ai/claude-code
 
-# Bake LiteLLM auto-start script into the image (used by Dockerfile runtime)
-COPY scripts/start-litellm.sh /opt/claude-dev/scripts/start-litellm.sh
-RUN chmod +x /opt/claude-dev/scripts/start-litellm.sh
-
-# Preflight bypass: Claude Code tries to validate connectivity against api.anthropic.com.
-# Since we use a custom proxy (LiteLLM → DeepSeek), we disable this check.
-# Strategy: replace the preflight function body with a no-op return.
+# Preflight bypass: replace all Anthropic URLs in the bundled binary with localhost:4008
+# (LiteLLM proxy). This is a binary-level patch since the bundled cli.js is minified.
 RUN python3 - <<'PY'
-from pathlib import Path
-import re
+path = "/usr/local/lib/node_modules/@anthropic-ai/claude-code/cli.js"
+with open(path, "rb") as f:
+    d = bytearray(f.read())
 
-path = Path("/usr/local/lib/node_modules/@anthropic-ai/claude-code/cli.js")
-if not path.exists():
-    raise SystemExit(f"claude cli not found: {path}")
+replacements = [
+    (b"https://api.anthropic.com", b"http://localhost:4008"),
+    (b"https://platform.claude.com", b"http://localhost:4008"),
+    (b"https://claude.ai", b"http://localhost:4008"),
+    (b"api.anthropic.com", b"localhost:4008"),
+    (b"platform.claude.com", b"localhost:4008"),
+]
+total = 0
+for old, new in replacements:
+    c = d.count(old)
+    if c:
+        d = d.replace(old, new)
+        total += c
+d = d.replace(b"https://localhost:4008", b"http://localhost:4008")
 
-content = path.read_text()
+with open(path, "wb") as f:
+    f.write(d)
 
-# For v2.1.x bundled binary: find and remove the preflight connectivity check
-# Pattern: the function that GETs api/hello and oauth/hello endpoints
-# Strategy 1: try the known old regex pattern
-old_pattern = r'let A=U7\(\),q=new URL\(A\.TOKEN_URL\),K=\[`\$\{A\.BASE_API_URL\}/api/hello`,`\$\{q\.origin\}/v1/oauth/hello`\],Y=async\(_\)=>\{try\{let \$=await I8\.get\(_,\{headers:\{"User-Agent":ey\(\)\}\}\);if\(\$\.status!==200\)return\{success:!1,error:`Failed to connect to \$\{new URL\(_\)\.hostname\}: Status \$\{\$\.status\}`\};return\{success:!0\}\}catch\(\$\)\{'
-replacement = 'return'
-patched, count = re.subn(old_pattern, replacement, content, count=1)
-
-if count == 1:
-    path.write_text(patched)
-    print("Applied preflight bypass (old pattern)")
-else:
-    # Strategy 2: more general - remove any function calling api/hello for connectivity check
-    print("Old pattern not found, trying broader patch...")
-    # Find the connectivity check function and replace it
-    broad_pattern = r'Failed to connect to \$\{new URL\(_\)\.hostname\}: Status'
-    if re.search(broad_pattern, content):
-        # Replace whole preflight function - match from "let A=U7" up to the catch block
-        content = re.sub(
-            r'let \w+=U7\(\)[^;]+;[^}]+Failed to connect to[^}]+}\}catch\([^}]+\)\{',
-            'return',
-            content,
-            count=1
-        )
-        path.write_text(content)
-        print("Applied preflight bypass (broad pattern)")
-    else:
-        # Strategy 3: replace api/hello references with empty health endpoint
-        content = content.replace(
-            'api/hello',
-            'x-no-op'
-        )
-        path.write_text(content)
-        print("Applied preflight bypass (url-replace pattern)")
+# Verify
+with open(path, "rb") as f:
+    d = f.read()
+remaining = sum(d.count(p) for p in [b"api.anthropic.com", b"platform.claude.com"])
+if remaining:
+    raise SystemExit(f"Patch incomplete: {remaining} anthropic refs remain")
+print(f"Patched {total} Anthropic URL refs → localhost:4008")
 PY
 
 CMD ["bash"]