diff --git a/skills/ci-stabilization/SKILL.md b/skills/ci-stabilization/SKILL.md
new file mode 100644
index 00000000000..86c788fb9c4
--- /dev/null
+++ b/skills/ci-stabilization/SKILL.md
@@ -0,0 +1,246 @@
+# Skill: CI/CD Stabilization for LiteLLM
+
+## When to Use
+Use this skill when CircleCI tests are failing on the `main` branch and you need to get them all passing. This applies to any CI system but is tailored for LiteLLM's CircleCI setup.
+
+## Prerequisites
+- CircleCI API token (for checking pipeline status)
+- Write access to the repository
+- Branch name must start with `litellm_` to trigger CircleCI workflows
+
+## Overview
+
+The process follows a systematic loop: **Analyze → Fix → Push → Wait → Check → Repeat** until all tests pass.
+
+---
+
+## Phase 1: Reconnaissance
+
+### Step 1.1: Check Current CI Status
+```bash
+# Get the latest pipeline on main
+curl -s -H "Circle-Token: $CIRCLECI_TOKEN" \
+  "https://circleci.com/api/v2/project/gh/BerriAI/litellm/pipeline?branch=main" | \
+  python3 -c "import sys,json; d=json.load(sys.stdin); print(d['items'][0]['id'])"
+```
+
+### Step 1.2: Get All Job Statuses
+```bash
+# List all jobs and their status for a workflow
+curl -s -H "Circle-Token: $CIRCLECI_TOKEN" \
+  "https://circleci.com/api/v2/workflow/$WORKFLOW_ID/job" | \
+  python3 -c "
+import sys,json
+data=json.load(sys.stdin)
+for j in data['items']:
+    if j['status'] == 'failed':
+        print(f\"FAILED: {j['name']:55s} #{j['job_number']}\")
+"
+```
+
+### Step 1.3: Get Failure Details
+Use the v1.1 API to get step-level output:
+```bash
+curl -s "https://circleci.com/api/v1.1/project/github/BerriAI/litellm/$JOB_NUMBER?circle-token=$CIRCLECI_TOKEN" | \
+  python3 -c "
+import sys,json,subprocess
+d=json.load(sys.stdin)
+for step in d.get('steps',[]):
+    for a in step.get('actions',[]):
+        if a.get('failed'):
+            url = a.get('output_url','')
+            print(f'FAILED: {a.get(\"name\",\"\")}')
+            if url:
+                r = subprocess.run(['curl','-s',url], capture_output=True, text=True)
+                data = json.loads(r.stdout)
+                print(data[0]['message'][-3000:])
+"
+```
+
+### Step 1.4: Categorize Failures
+
+Sort every failure into one of these buckets:
+
+| Category | Description | Action |
+|---|---|---|
+| **Code Bug** | Linting, type errors, wrong assertions | Fix the code |
+| **Config Bug** | Version mismatch, missing migration, wrong CI params | Fix config |
+| **Auth Bug** | Tests missing auth mock/override | Add `app.dependency_overrides` |
+| **Parallel Bug** | Worker crashes, timeouts in `-n 8` | Add `xdist_group`, reduce `-n` |
+| **External API** | Live API returns errors | Add retries, graceful skip |
+| **Timing Bug** | Async operations not complete when checked | Add polling/retry |
+
+---
+
+## Phase 2: Fix Categories (Priority Order)
+
+### 2.1: Linting & Type Errors
+These block ALL other jobs. Fix first.
+
+```bash
+# Check ruff
+poetry run ruff check litellm/
+
+# Common fixes:
+# - Add `# noqa: PLR0915` for too-many-statements
+# - Add `# type: ignore[error-code]` for mypy false positives
+# - Fix actual type annotations
+```
+
+### 2.2: Version & Schema Issues
+```bash
+# Check proxy-extras version
+cat litellm-proxy-extras/pyproject.toml | grep version
+
+# IMPORTANT: Only bump version in source package pyproject.toml
+# Do NOT bump in requirements.txt (package isn't published yet)
+
+# Sync schemas
+cp litellm/proxy/schema.prisma litellm-proxy-extras/litellm_proxy_extras/schema.prisma
+cp litellm/proxy/schema.prisma schema.prisma
+
+# Create migrations for schema changes
+mkdir -p litellm-proxy-extras/litellm_proxy_extras/migrations/YYYYMMDDHHMMSS_description/
+# Write migration.sql with the SQL diff
+```
+
+### 2.3: Security Vulnerabilities
+```bash
+# Check npm vulnerabilities
+cd docs/my-website && npm audit
+cd ui/litellm-dashboard && npm audit
+
+# Fix with overrides in package.json or npm audit fix
+# ALSO check Dockerfile for pinned vulnerable versions
+grep 'npm install -g' Dockerfile
+```
+
+### 2.4: Test Auth Issues
+Many proxy tests fail with 401 because they don't override FastAPI auth:
+
+```python
+# PATTERN: Add this to tests that call proxy endpoints
+from litellm.proxy._types import LitellmUserRoles, UserAPIKeyAuth
+import litellm.proxy.proxy_server as ps
+
+app.dependency_overrides[ps.user_api_key_auth] = lambda: UserAPIKeyAuth(
+    user_role=LitellmUserRoles.PROXY_ADMIN, user_id="test-user"
+)
+try:
+    # ... test code ...
+finally:
+    app.dependency_overrides.pop(ps.user_api_key_auth, None)
+```
+
+### 2.5: Parallel Test Isolation
+Worker crashes (OOM) in parallel pytest:
+
+```python
+# Add to heavy test modules/classes
+import pytest
+pytestmark = pytest.mark.xdist_group("heavy_group_name")
+
+# Or reduce parallelism in .circleci/config.yml
+# Change: -n 8  →  -n 4
+```
+
+### 2.6: External API Resilience
+For tests that call live APIs:
+
+```python
+# Add retry decorator
+@pytest.mark.flaky(retries=3, delay=5)
+
+# For Jest/Node.js tests - add graceful skip
+if (!spendData || !spendData.length || !spendData[0]) {
+    console.warn('Data not available - skipping assertions');
+    return;
+}
+```
+
+### 2.7: UI Test Timeouts
+Vitest tests with `userEvent.type()` are slow:
+
+```typescript
+// Fix: Remove typing delay
+const user = userEvent.setup({ delay: null });
+
+// Increase timeout for heavy tests
+it("should ...", async () => { ... }, { timeout: 15000 });
+```
+
+---
+
+## Phase 3: Push & Monitor Loop
+
+### Step 3.1: Create CI Branch
+```bash
+# Branch MUST match /litellm_.*/ for CircleCI to run
+git checkout -b litellm_<descriptive_name>
+git push -u origin litellm_<descriptive_name>
+```
+
+### Step 3.2: Monitor Pipeline
+```bash
+# Poll pipeline status
+curl -s -H "Circle-Token: $CIRCLECI_TOKEN" \
+  "https://circleci.com/api/v2/project/gh/BerriAI/litellm/pipeline?branch=litellm_<branch>" | \
+  python3 -c "
+import sys,json,subprocess
+data=json.load(sys.stdin)
+pid=data['items'][0]['id']
+wf=json.loads(subprocess.run(['curl','-s','-H',f'Circle-Token: {TOKEN}',
+  f'https://circleci.com/api/v2/pipeline/{pid}/workflow'],
+  capture_output=True, text=True).stdout)
+# ... check job statuses
+"
+```
+
+### Step 3.3: Iterate
+After each CI run:
+1. Check which jobs failed
+2. Get the exact failing test and error message
+3. Categorize the failure
+4. Apply the fix pattern from Phase 2
+5. Commit, push, wait for next run
+6. Repeat until 0 failures
+
+---
+
+## Common Pitfalls
+
+### ❌ Don't bump litellm-proxy-extras in requirements.txt
+The package isn't published to PyPI yet. Only bump the source `pyproject.toml`. Bumping `requirements.txt` will break ALL jobs.
+
+### ❌ Don't use `-n 8` for heavy proxy tests
+8 parallel workers each loading the full FastAPI proxy causes OOM. Use `-n 4`.
+
+### ❌ Don't assume test failures are code bugs
+Check if the same test passes in other runs. External API failures SHUFFLE between runs randomly.
+
+### ❌ Don't patch auth with `unittest.mock.patch()`
+FastAPI dependency injection requires `app.dependency_overrides`, not `patch()`.
+
+### ❌ Don't create migrations without checking the actual diff
+Run `prisma migrate diff` output from CI to see exactly what SQL is needed.
+
+---
+
+## Key Files Reference
+
+| File | Purpose |
+|---|---|
+| `.circleci/config.yml` | CI job definitions, parallelism, timeouts |
+| `litellm-proxy-extras/pyproject.toml` | Proxy extras version |
+| `schema.prisma` (3 copies) | DB schema - must stay in sync |
+| `litellm-proxy-extras/.../migrations/` | DB migrations |
+| `Dockerfile` | Docker image deps (check for pinned vulnerable versions) |
+| `docs/my-website/package.json` | npm overrides for security |
+| `ui/litellm-dashboard/package-lock.json` | UI dependencies |
+
+## Success Criteria
+- All CircleCI jobs pass (green)
+- No worker crashes
+- No auth 401 errors in tests
+- External API tests have retries/graceful handling
+- Security scans pass (no HIGH/CRITICAL CVEs)