zed/crates/eval_cli/zed_eval/judge_proxy.py
Anant Goel 10f501d700
eval_cli: Add remote benchmark orchestration (#59802)
Summary:

- Add the `zed-eval` Python CLI for Modal/Harbor/Pier benchmark
orchestration, including content-addressed remote builds, run/suite
management, reporting, rejudge, baseline, and cleanup workflows.
- Extend `eval-cli` for remote evals with provider/model overrides and
step/tool-call metrics in `result.json`.
- Add install/source-run helper scripts so `zed-eval` can be installed
or run from the checkout without manually setting `PYTHONPATH`.
- Harden the remote harness wrappers around exit-code preservation,
archive extraction, custom secret wiring, and Harbor/Pier option parity,
with regression coverage.

Testing:

- Using the CLI for two weeks
- `PYTHONPATH=crates/eval_cli python3 -m compileall -q
crates/eval_cli/zed_eval`
- `uv run --project crates/eval_cli/zed_eval python -m unittest discover
-s crates/eval_cli/zed_eval/tests`
- `bash -n crates/eval_cli/script/install-zed-eval
crates/eval_cli/script/zed-eval`
- `cargo check -p eval_cli`
- `cargo fmt --package eval_cli -- --check`
- `cargo test -p eval_cli --no-run`
- `./script/clippy -p eval_cli`

Release Notes:

- N/A
2026-06-24 15:32:41 +00:00

104 lines
4.8 KiB
Python

"""The in-sandbox judge proxy shim, shared by the Harbor verifier and the
offline rejudge path so judge calls go through one identical code path
everywhere.
Stdlib-only and free of any `harbor`/`modal` imports on purpose: `verifier.py`
imports these constants for the live in-sandbox install, and `rejudge.py` runs
the very same proxy in a plain container to re-grade stored runs. Keeping the
shim here means there is exactly one implementation of the upstream fixups
(Anthropic `response_format` stripping, `max_tokens` -> `max_completion_tokens`,
the `ZED_JUDGE_MAX_TOKENS` floor) rather than a copy per call site.
"""
from __future__ import annotations
# Stdlib-only OpenAI-protocol shim. The SWE-Atlas RF judge hardcodes
# `response_format={"type": "json_object"}`, which Anthropic's OpenAI-compatible
# endpoint rejects (400: "response_format.type: Input should be 'json_schema'").
# The shim listens on 127.0.0.1:PORT and forwards to ZED_JUDGE_UPSTREAM, applying
# upstream-specific fixups:
# - `response_format` is stripped ONLY for Anthropic (Baseten/OpenAI accept
# json_object, and the Kimi/DeepSeek judges rely on it);
# - `max_tokens` is renamed to `max_completion_tokens` (accepted by both
# Anthropic-compat and OpenAI, including GPT-5.x reasoning models that reject
# `max_tokens`) and raised to the ZED_JUDGE_MAX_TOKENS floor so the runtime
# judge matches the offline-calibrated token budget.
# When ZED_JUDGE_AUTH_ENV is set, it replaces verifier auth with the named
# sandbox env var; otherwise it passes auth headers through unchanged.
JUDGE_PROXY_SCRIPT = """\
import json, os, urllib.request, urllib.error
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
UPSTREAM = os.environ.get("ZED_JUDGE_UPSTREAM", "https://api.anthropic.com/v1").rstrip("/")
PORT = int(os.environ.get("ZED_JUDGE_PROXY_PORT", "8089"))
AUTH_ENV = os.environ.get("ZED_JUDGE_AUTH_ENV")
MAX_TOKENS_FLOOR = int(os.environ.get("ZED_JUDGE_MAX_TOKENS", "0"))
class Handler(BaseHTTPRequestHandler):
def do_POST(self):
length = int(self.headers.get("Content-Length", 0))
body = self.rfile.read(length)
try:
payload = json.loads(body)
if "anthropic" in UPSTREAM:
payload.pop("response_format", None)
sent = payload.pop("max_tokens", None)
current = sent if sent is not None else payload.get("max_completion_tokens")
if current is not None or MAX_TOKENS_FLOOR:
payload["max_completion_tokens"] = max(current or 0, MAX_TOKENS_FLOOR)
body = json.dumps(payload).encode()
except Exception as error:
print(f"proxy: passing body through unmodified: {error}", flush=True)
path = self.path
if path.startswith("/v1/"):
path = path[len("/v1"):]
request = urllib.request.Request(UPSTREAM + path, data=body, method="POST")
request.add_header("Content-Type", "application/json")
auth_value = os.environ.get(AUTH_ENV, "") if AUTH_ENV else ""
if auth_value:
request.add_header("Authorization", f"Bearer {auth_value}")
else:
for header in ("Authorization", "x-api-key"):
value = self.headers.get(header)
if value:
request.add_header(header, value)
for header in ("anthropic-version",):
value = self.headers.get(header)
if value:
request.add_header(header, value)
try:
with urllib.request.urlopen(request, timeout=900) as response:
data = response.read()
status = response.status
except urllib.error.HTTPError as error:
data = error.read()
status = error.code
except Exception as error:
data = json.dumps({"error": {"message": str(error)}}).encode()
status = 502
self.send_response(status)
self.send_header("Content-Type", "application/json")
self.send_header("Content-Length", str(len(data)))
self.end_headers()
self.wfile.write(data)
def log_message(self, format, *args):
print("proxy: " + format % args, flush=True)
if __name__ == "__main__":
print(f"judge proxy listening on 127.0.0.1:{PORT} -> {UPSTREAM}", flush=True)
ThreadingHTTPServer(("127.0.0.1", PORT), Handler).serve_forever()
"""
# Idempotent starter. Uses bash's /dev/tcp (pgrep isn't in every image) to check
# the port; the daemon inherits ZED_JUDGE_UPSTREAM from the verifier setup exec
# env. stdout/stderr/stdin are fully detached so the spawning exec doesn't hang
# on open pipes.
JUDGE_PROXY_ENSURE_SCRIPT = """\
#!/bin/bash
if ! (exec 3<>/dev/tcp/127.0.0.1/${ZED_JUDGE_PROXY_PORT:-8089}) 2>/dev/null; then
nohup python3 /usr/local/lib/zed_judge_proxy.py \\
>>/tmp/zed-judge-proxy.log 2>&1 </dev/null &
sleep 1
fi
"""