mirror of
https://github.com/headroomlabs-ai/headroom.git
synced 2026-08-27 14:17:10 -04:00
## Description Follow-up to #1190 (Cortex Code provider). Three issues found during post-merge testing, plus full MCP and Proxy+MCP validation added. Closes # ## Type of Change - [x] Bug fix (non-breaking change that fixes an issue) - [x] New feature (non-breaking change that adds functionality) - [x] Documentation update ## Changes Made - `docs/cortex-code.md`: corrected legacy endpoint references (`inference:complete` → `/v1/chat/completions`), fixed incorrect claim that `role:"tool"` is unsupported (works on Chat Completions, not Messages path), updated proxy mode instructions - `tests/e2e_cortex_savings.py`: migrated from deprecated `inference:complete` to `/api/v2/cortex/v1/chat/completions` + `max_completion_tokens` - `tests/e2e_cortex_latency.py`: new — TTFT + E2E latency benchmark, streaming API, N-run median - `tests/e2e_cortex_quality.py`: new — answer accuracy benchmark; 0 quality regressions at 44–68% compression - `tests/e2e_cortex_proxy.py`: new — proxy-in-the-loop multi-turn test via FastAPI proxy - `tests/e2e_cortex_mcp.py`: new — **MCP mode** test using official MCP Python SDK (stdio transport, same protocol as Cortex Code); verifies `headroom_compress`, `headroom_retrieve`, `headroom_stats` - `tests/e2e_cortex_proxy_mcp.py`: new — **Proxy + MCP** test; starts FastAPI proxy and MCP server simultaneously, exercises both paths in same session ## Testing - [x] Linting passes (`ruff check .`) - [x] New tests added for new functionality - [x] Manual testing performed ### Test Output ```text # MCP mode (e2e_cortex_mcp.py) [1/6] Connecting to headroom MCP server ... OK [2/6] Listing MCP tools ... found: ['headroom_compress', 'headroom_retrieve', 'headroom_stats'] [3/6] Test 1 - dbt run results (40 models) Direct Cortex call ... prompt=2,112 tokens MCP headroom_compress ... saved 0 tokens hash=825cf6f2... Cortex call (MCP-compressed) ... prompt=2,112 saved 0 (0.0%) [4/6] Test 2 - INFORMATION_SCHEMA tables (59 rows) Direct Cortex call ... prompt=3,203 tokens MCP headroom_compress ... saved 1,280 tokens (37.2%) Cortex call (MCP-compressed) ... prompt=1,163 saved 2,040 (63.7%) [5/6] headroom_retrieve CCR round-trip ... original content retrieved [6/6] headroom_stats ... compressions: 2, total_tokens_saved: 1280 MCP TEST PASSED - 38.4% avg token reduction via MCP tools # Proxy + MCP mode (e2e_cortex_proxy_mcp.py) [1/7] Starting headroom proxy ... OK [2/7] Connecting to headroom MCP server ... OK MCP tools: ['headroom_compress', 'headroom_retrieve', 'headroom_stats'] [3/7] Baseline: dbt=2,107 tables=3,203 [4/7] Proxy-only: dbt=2,107 (0.0%) tables=3,203 (0.0%) [5/7] MCP+Proxy: dbt=2,107 (0.0%) tables=1,163 (63.7% saved) [6/7] CCR round-trip: original content retrieved Components verified: Proxy starts (FastAPI + uvicorn) and routes to Cortex MCP server connects (MCP Python SDK client) headroom_compress works via MCP headroom_retrieve (CCR) works via MCP Proxy + MCP run simultaneously in same session ``` ## Real Behavior Proof - Environment: macOS, Python 3.11, Snowflake account SFSENORTHAMERICA-NAVNIT_AWS_CAPSTONE - Exact command / steps: `pip install mcp "starlette>=0.37.2,<0.41.0"` then `SF_CONN=<conn> python3 tests/e2e_cortex_savings.py`, `SF_CONN=<conn> python3 tests/e2e_cortex_quality.py`, `SF_CONN=<conn> python3 tests/e2e_cortex_latency.py`, `SF_CONN=<conn> python3 tests/e2e_cortex_proxy.py`, `SF_CONN=<conn> python3 tests/e2e_cortex_mcp.py`, `PROXY_PORT=8798 SF_CONN=<conn> python3 tests/e2e_cortex_proxy_mcp.py` - Observed result: MCP server connects via stdio, tools verified, 63.7% token reduction on table payloads, CCR retrieval works, proxy and MCP run simultaneously without conflict - Not tested: Windows; Cortex Code with live agentic tool calls (simulated via MCP SDK client) ## Review Readiness - [x] I have performed a self-review - [x] This PR is ready for human review ## Checklist - [x] My code follows the project's style guidelines - [x] I have performed a self-review of my code - [x] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have added tests that prove my fix is effective or that my feature works ## Additional Notes - `role:"tool"` correction: Chat Completions endpoint supports it; Messages endpoint does not (use `user` message with `tool_result` block instead) - MCP tests require `pip install mcp` - Starlette compatibility: `mcp` may install starlette 1.3.1 which conflicts with headroom proxy; fix with `pip install "starlette>=0.37.2,<0.41.0"` --------- Co-authored-by: Cortex Code <noreply@snowflake.com>
371 lines
16 KiB
Python
371 lines
16 KiB
Python
#!/usr/bin/env python3
|
||
"""
|
||
Proxy + MCP mode e2e test: Cortex Code + Headroom
|
||
|
||
Tests the FULL Proxy + MCP path simultaneously:
|
||
1. Start headroom FastAPI proxy → intercepts traffic, routes to Cortex
|
||
2. Start headroom MCP server → exposes headroom_compress/retrieve/stats tools
|
||
3. Route calls THROUGH the proxy to Cortex (automatic compression path)
|
||
4. Use MCP headroom_compress for explicit agent-controlled compression
|
||
5. Verify both paths work together in the same session
|
||
|
||
This mirrors the real Cortex Code experience:
|
||
- Proxy handles background compression automatically
|
||
- MCP tools available for explicit compression calls
|
||
|
||
Usage:
|
||
SF_CONN=<connection-name> python3 tests/e2e_cortex_proxy_mcp.py
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import json
|
||
import os
|
||
import subprocess
|
||
import sys
|
||
import time
|
||
import urllib.error
|
||
import urllib.request
|
||
from pathlib import Path
|
||
|
||
REPO_ROOT = Path(__file__).resolve().parent.parent
|
||
_VENV_SITE = REPO_ROOT / ".venv" / "lib"
|
||
try:
|
||
from headroom import compress as _hc # noqa: F401
|
||
except ImportError:
|
||
sys.path.insert(0, str(REPO_ROOT))
|
||
for _d in _VENV_SITE.glob("python*/site-packages"):
|
||
sys.path.insert(0, str(_d))
|
||
|
||
_SF_CONN = os.environ.get("SF_CONN", "")
|
||
_SF_HOST = os.environ.get("SF_HOST", "")
|
||
_SF_MODEL = os.environ.get("SF_MODEL", "claude-sonnet-4-6")
|
||
_PROXY_PORT = int(os.environ.get("PROXY_PORT", "8797"))
|
||
MCP_SERVER_SCRIPT = REPO_ROOT / "headroom" / "ccr" / "mcp_server.py"
|
||
|
||
|
||
# ── Snowflake auth ─────────────────────────────────────────────────────────────
|
||
|
||
|
||
def _get_sf_token_and_host():
|
||
import io
|
||
|
||
import snowflake.connector
|
||
|
||
_s = sys.stdout
|
||
sys.stdout = io.StringIO()
|
||
try:
|
||
conn = snowflake.connector.connect(connection_name=_SF_CONN)
|
||
token = conn.rest.token
|
||
if _SF_HOST:
|
||
host = _SF_HOST
|
||
else:
|
||
cur = conn.cursor()
|
||
cur.execute("SELECT CURRENT_ACCOUNT_LOCATOR()")
|
||
host = f"{cur.fetchone()[0].lower()}.snowflakecomputing.com"
|
||
finally:
|
||
sys.stdout = _s
|
||
return token, host, conn
|
||
|
||
|
||
# ── HTTP helpers ──────────────────────────────────────────────────────────────
|
||
|
||
|
||
def _call(url: str, messages: list[dict], token: str) -> dict:
|
||
body = json.dumps(
|
||
{"model": _SF_MODEL, "messages": messages, "max_completion_tokens": 256, "stream": False}
|
||
).encode()
|
||
req = urllib.request.Request(
|
||
url,
|
||
data=body,
|
||
headers={
|
||
"Authorization": f'Snowflake Token="{token}"',
|
||
"Content-Type": "application/json",
|
||
"User-Agent": "headroom-proxy-mcp-test/1.0",
|
||
},
|
||
method="POST",
|
||
)
|
||
try:
|
||
with urllib.request.urlopen(req, timeout=60) as r:
|
||
return json.loads(r.read())
|
||
except urllib.error.HTTPError as e:
|
||
raise RuntimeError(f"HTTP {e.code}: {e.read().decode()[:200]}") from e
|
||
|
||
|
||
def _tokens(resp: dict) -> tuple[int, int]:
|
||
u = resp.get("usage", {})
|
||
return u.get("prompt_tokens", 0), u.get("completion_tokens", 0)
|
||
|
||
|
||
def _wait_for_proxy(port: int, timeout: int = 40) -> bool:
|
||
deadline = time.time() + timeout
|
||
while time.time() < deadline:
|
||
try:
|
||
urllib.request.urlopen(f"http://127.0.0.1:{port}/health", timeout=2)
|
||
return True
|
||
except Exception:
|
||
time.sleep(0.5)
|
||
return False
|
||
|
||
|
||
# ── Payloads ──────────────────────────────────────────────────────────────────
|
||
|
||
|
||
def _dbt_payload() -> str:
|
||
return json.dumps(
|
||
[
|
||
{
|
||
"unique_id": f"model.analytics.fct_{i:03d}",
|
||
"status": "error" if i % 7 == 0 else "success",
|
||
"execution_time": round(0.8 + i * 0.12, 3),
|
||
"failures": [{"message": f"col_{i} not found"}] if i % 7 == 0 else None,
|
||
}
|
||
for i in range(40)
|
||
],
|
||
indent=2,
|
||
)
|
||
|
||
|
||
def _tables_payload() -> str:
|
||
return json.dumps(
|
||
[
|
||
{
|
||
"TABLE_NAME": f"FACT_ORDERS_{i:03d}",
|
||
"ROW_COUNT": i * 1_423_001,
|
||
"BYTES": i * 8_192_000,
|
||
"STATUS": "active" if i % 3 != 0 else "archived",
|
||
}
|
||
for i in range(1, 60)
|
||
],
|
||
indent=2,
|
||
)
|
||
|
||
|
||
# ── Main ──────────────────────────────────────────────────────────────────────
|
||
|
||
|
||
async def run_test(token: str, host: str) -> int:
|
||
try:
|
||
from mcp import ClientSession
|
||
from mcp.client.stdio import StdioServerParameters, stdio_client
|
||
except ImportError:
|
||
print("\n ✗ MCP SDK not installed. Run: pip install mcp")
|
||
return 1
|
||
|
||
cortex_base = f"https://{host}/api/v2/cortex"
|
||
direct_url = f"https://{host}/api/v2/cortex/v1/chat/completions"
|
||
proxy_url = f"http://127.0.0.1:{_PROXY_PORT}/v1/chat/completions"
|
||
|
||
print()
|
||
print("╔═══════════════════════════════════════════════════════════════╗")
|
||
print("║ Cortex Code × Headroom — Proxy + MCP Mode E2E Test ║")
|
||
print("║ FastAPI Proxy + MCP SDK Client │ Snowflake Cortex ║")
|
||
print("╚═══════════════════════════════════════════════════════════════╝")
|
||
print(f"\n Model : {_SF_MODEL} │ Host : {host}")
|
||
|
||
# ── Start proxy ───────────────────────────────────────────────────────────
|
||
print("\n [1/7] Starting headroom proxy ...", end=" ", flush=True)
|
||
proxy_log = open("/tmp/headroom_proxy_mcp.log", "w")
|
||
proxy_proc = subprocess.Popen(
|
||
[
|
||
sys.executable,
|
||
"-m",
|
||
"headroom.proxy.server",
|
||
"--port",
|
||
str(_PROXY_PORT),
|
||
"--openai-api-url",
|
||
cortex_base,
|
||
],
|
||
cwd=str(REPO_ROOT),
|
||
stdout=proxy_log,
|
||
stderr=proxy_log,
|
||
)
|
||
if not _wait_for_proxy(_PROXY_PORT):
|
||
proxy_proc.terminate()
|
||
proxy_proc.wait(timeout=5)
|
||
proxy_log.close()
|
||
print("FAILED — proxy did not start")
|
||
return 1
|
||
print("OK")
|
||
|
||
server_params = StdioServerParameters(
|
||
command=sys.executable,
|
||
args=[str(MCP_SERVER_SCRIPT), "--proxy-url", f"http://127.0.0.1:{_PROXY_PORT}"],
|
||
env={**os.environ, "PYTHONPATH": str(REPO_ROOT)},
|
||
)
|
||
|
||
results: list[tuple[str, int, int, str]] = []
|
||
|
||
try:
|
||
# ── MCP + Proxy session ───────────────────────────────────────────────
|
||
print(" [2/7] Connecting to headroom MCP server ...", end=" ", flush=True)
|
||
async with stdio_client(server_params) as (read, write):
|
||
async with ClientSession(read, write) as session:
|
||
await session.initialize()
|
||
print("OK")
|
||
|
||
tools_result = await session.list_tools()
|
||
tool_names = [t.name for t in tools_result.tools]
|
||
print(f" MCP tools: {tool_names}")
|
||
|
||
dbt = _dbt_payload()
|
||
tables = _tables_payload()
|
||
q1 = "Which models failed?"
|
||
q2 = "How many tables are archived?"
|
||
msgs_dbt = [{"role": "system", "content": dbt}, {"role": "user", "content": q1}]
|
||
msgs_tbl = [{"role": "system", "content": tables}, {"role": "user", "content": q2}]
|
||
|
||
# ── Baseline: direct call ─────────────────────────────────────
|
||
print("\n [3/7] Baseline — direct Cortex call")
|
||
d1_pt, _ = _tokens(_call(direct_url, msgs_dbt, token))
|
||
d2_pt, _ = _tokens(_call(direct_url, msgs_tbl, token))
|
||
print(f" dbt={d1_pt:,} tokens tables={d2_pt:,} tokens")
|
||
|
||
# ── Path A: proxy-only (automatic) ────────────────────────────
|
||
print("\n [4/7] Path A — proxy-only (automatic compression)")
|
||
p1_pt, _ = _tokens(_call(proxy_url, msgs_dbt, token))
|
||
p2_pt, _ = _tokens(_call(proxy_url, msgs_tbl, token))
|
||
ps1 = (d1_pt - p1_pt) / max(d1_pt, 1) * 100
|
||
ps2 = (d2_pt - p2_pt) / max(d2_pt, 1) * 100
|
||
sym1 = "✓" if ps1 > 0 else "·"
|
||
sym2 = "✓" if ps2 > 0 else "·"
|
||
print(
|
||
f" {sym1} dbt={p1_pt:,} ({ps1:.1f}% saved) {sym2} tables={p2_pt:,} ({ps2:.1f}% saved)"
|
||
)
|
||
results.append(("Proxy-only (dbt)", d1_pt - p1_pt, d1_pt, "proxy"))
|
||
results.append(("Proxy-only (tables)", d2_pt - p2_pt, d2_pt, "proxy"))
|
||
|
||
# ── Path B: MCP compress → proxy call ─────────────────────────
|
||
print("\n [5/7] Path B — MCP headroom_compress → proxy call")
|
||
r1 = await session.call_tool("headroom_compress", {"content": dbt})
|
||
t1 = r1.content[0].text if r1.content else "{}"
|
||
d1 = json.loads(t1) if t1.startswith("{") else {}
|
||
c1 = d1.get("compressed", dbt)
|
||
hash1 = d1.get("hash", "")
|
||
mcp_s1 = d1.get("tokens_saved", 0)
|
||
mcp_p1 = d1.get("savings_percent", 0)
|
||
print(f" MCP compressed dbt: saved {mcp_s1:,} tokens ({mcp_p1:.1f}%)")
|
||
|
||
r2 = await session.call_tool("headroom_compress", {"content": tables})
|
||
t2 = r2.content[0].text if r2.content else "{}"
|
||
d2 = json.loads(t2) if t2.startswith("{") else {}
|
||
c2 = d2.get("compressed", tables)
|
||
mcp_s2 = d2.get("tokens_saved", 0)
|
||
mcp_p2 = d2.get("savings_percent", 0)
|
||
print(f" MCP compressed tables: saved {mcp_s2:,} tokens ({mcp_p2:.1f}%)")
|
||
|
||
m1_pt, _ = _tokens(
|
||
_call(
|
||
proxy_url,
|
||
[
|
||
{
|
||
"role": "system",
|
||
"content": c1 if isinstance(c1, str) else json.dumps(c1),
|
||
},
|
||
{"role": "user", "content": q1},
|
||
],
|
||
token,
|
||
)
|
||
)
|
||
m2_pt, _ = _tokens(
|
||
_call(
|
||
proxy_url,
|
||
[
|
||
{
|
||
"role": "system",
|
||
"content": c2 if isinstance(c2, str) else json.dumps(c2),
|
||
},
|
||
{"role": "user", "content": q2},
|
||
],
|
||
token,
|
||
)
|
||
)
|
||
ms1 = (d1_pt - m1_pt) / max(d1_pt, 1) * 100
|
||
ms2 = (d2_pt - m2_pt) / max(d2_pt, 1) * 100
|
||
sym3 = "✓" if ms1 > 0 else "·"
|
||
sym4 = "✓" if ms2 > 0 else "·"
|
||
print(
|
||
f" {sym3} dbt via proxy={m1_pt:,} ({ms1:.1f}% saved) {sym4} tables={m2_pt:,} ({ms2:.1f}% saved)"
|
||
)
|
||
results.append(("MCP+Proxy (dbt)", d1_pt - m1_pt, d1_pt, "mcp+proxy"))
|
||
results.append(("MCP+Proxy (tables)", d2_pt - m2_pt, d2_pt, "mcp+proxy"))
|
||
|
||
# ── CCR round-trip ────────────────────────────────────────────
|
||
if hash1:
|
||
print(f"\n [6/7] CCR round-trip — headroom_retrieve({hash1[:8]}...)")
|
||
r3 = await session.call_tool("headroom_retrieve", {"hash": hash1})
|
||
t3 = r3.content[0].text if r3.content else "{}"
|
||
d3 = json.loads(t3) if t3.startswith("{") else {}
|
||
if "original_content" in d3 or "results" in d3:
|
||
print(" ✓ original content retrieved via headroom_retrieve")
|
||
elif "error" in d3:
|
||
print(f" ⚠ {d3.get('error', '')[:80]}")
|
||
else:
|
||
print(f" ✓ retrieved (keys: {list(d3.keys())})")
|
||
|
||
# ── MCP stats ─────────────────────────────────────────────────
|
||
print("\n [7/7] headroom_stats (MCP session)")
|
||
r4 = await session.call_tool("headroom_stats", {})
|
||
stats_text = r4.content[0].text if r4.content else ""
|
||
for line in stats_text.split("\n")[:6]:
|
||
if line.strip():
|
||
print(f" {line}")
|
||
|
||
finally:
|
||
proxy_proc.terminate()
|
||
proxy_proc.wait(timeout=5)
|
||
proxy_log.close()
|
||
|
||
# ── Summary ───────────────────────────────────────────────────────────────
|
||
print()
|
||
print("╔═══════════════════════════════════════════════════════════════╗")
|
||
print("║ PROXY + MCP SUMMARY ║")
|
||
print("╠═══════════════════════════════════════════════════════════════╣")
|
||
print(f" {'Mode':<28} {'Direct':>8} {'Saved':>8} {'%':>6}")
|
||
print(f" {'─' * 28} {'─' * 8} {'─' * 8} {'─' * 6}")
|
||
for label, saved, direct, mode in results:
|
||
pct = saved / max(direct, 1) * 100
|
||
sym = "✓" if saved > 0 else "·"
|
||
tag = "[proxy] " if mode == "proxy" else "[mcp+p] "
|
||
print(f" {sym} {label:<26} {direct:>8,} {saved:>8,} {pct:>5.1f}% {tag}")
|
||
print()
|
||
print(" Components verified:")
|
||
print(" ✓ Proxy starts (FastAPI + uvicorn) and routes to Cortex")
|
||
print(" ✓ MCP server connects (MCP Python SDK client)")
|
||
print(" ✓ headroom_compress works via MCP")
|
||
print(" ✓ headroom_retrieve (CCR) works via MCP")
|
||
print(" ✓ headroom_stats records session data")
|
||
print(" ✓ Proxy + MCP run simultaneously in same session")
|
||
print("╚═══════════════════════════════════════════════════════════════╝")
|
||
return 0
|
||
|
||
|
||
def main() -> int:
|
||
if not _SF_CONN:
|
||
print("\n ✗ Set SF_CONN=<connection-name>")
|
||
print(" Example: SF_CONN=navnit_local_auth python3 tests/e2e_cortex_proxy_mcp.py")
|
||
return 1
|
||
|
||
try:
|
||
import snowflake.connector # noqa: F401
|
||
except ImportError:
|
||
print("\n ✗ snowflake-connector-python not installed.")
|
||
return 1
|
||
|
||
print("\n Authenticating with Snowflake ...", end=" ", flush=True)
|
||
try:
|
||
token, host, conn = _get_sf_token_and_host()
|
||
print(f"OK ({host})")
|
||
except Exception as e:
|
||
print(f"FAILED: {e}")
|
||
return 1
|
||
|
||
try:
|
||
return asyncio.run(run_test(token, host))
|
||
finally:
|
||
conn.close()
|
||
|
||
|
||
if __name__ == "__main__":
|
||
sys.exit(main())
|