alexrs's picture
gradio server (#1)
c312d97
"""Gradio Space that showcases execution-grounded code generation."""
from __future__ import annotations
import html
import json
import logging
import os
import re
import subprocess
import sys
import tempfile
import textwrap
from collections.abc import Iterator
from dataclasses import dataclass
from pathlib import Path
from typing import Any
from gradio import Server
from fastapi.responses import HTMLResponse, FileResponse
from cohere import ClientV2
from cohere.core.api_error import ApiError
APP_TITLE = "North Mini Code 1.0 Demo"
CLIENT_NAME = "hf-space-north-mini-code"
MODEL_ID = "north-mini-code-1-0"
MODEL_URL = "https://huggingface.co/CohereLabs/North-Mini-Code-1.0"
OPENCODE_URL = "https://opencode.ai"
DEFAULT_TEMPERATURE = 0.2
PY_TIMEOUT_S = 12
PY_MEM_LIMIT_MB = 1024
MAX_STDIO_CHARS = 16_000
OUTPUT_PNG = "output.png"
THINKING_BLOCK_RE = re.compile(r"<\s*think\s*>.*?<\s*/\s*think\s*>", re.IGNORECASE | re.DOTALL)
CODE_BLOCK_RE = re.compile(r"```([a-zA-Z0-9_+.#-]*)\s*\n(.*?)```", re.DOTALL)
logger = logging.getLogger(__name__)
logging.basicConfig(level=logging.INFO)
SYSTEM_PROMPT = """You are a coding model in a demo where generated code can be run.
Only respond to coding-related requests: code generation, debugging, code review,
software design, developer tooling, programming concepts, or reasoning about code.
If the user asks for something unrelated to coding, briefly say you can only help
with coding-related requests.
If the user asks a coding question that does not require runnable code, answer it
directly and do not force a code block.
If the user asks you to generate, modify, or fix runnable code, return exactly one
fenced code block and no extra prose. Use a correct language tag: ```python for
Python, or ```html for Web.
For Python, prefer standard library or common packages such as matplotlib.
For Python, do not use network calls, subprocesses, shell commands, or long-running loops.
For Web, return a single self-contained HTML document with any CSS and JavaScript inline.
For Web, make the page fully responsive so it fills the area it is given: set html and body
to margin:0 and 100% width/height, prefer relative sizes (100%, 100vw/100vh, flexbox) over
fixed pixel dimensions, and size any <canvas> to its container and re-fit it on window resize
so nothing is clipped or scrolled.
"""
# Curated starter prompts. Each entry is (chip label, prompt, target language).
EXAMPLE_PROMPTS: list[tuple[str, str, str]] = [
(
"🌀 Spiral plot",
"Create a Python script that plots a colorful spiral with matplotlib, prints a short "
"description of what it drew, and does not require any external files.",
"Python",
),
(
"📊 Sine waves",
"Plot three sine waves with different frequencies and amplitudes on one matplotlib "
"figure with a legend and grid, and print the equation of each wave.",
"Python",
),
(
"✨ Particles",
"Create a self-contained HTML/CSS/JavaScript demo with an animated particle field "
"that reacts to the mouse and includes a small title.",
"Web",
),
(
"🖌️ Blackboard",
"Create a self-contained HTML/CSS/JavaScript blackboard drawing app: a dark chalkboard "
"canvas you can draw on with the mouse (and touch), a small palette of chalk colors, an "
"adjustable brush size, and a button to clear the board.",
"Web",
),
(
"🎲 Monte Carlo π",
"Estimate π with a Monte Carlo simulation: sample random points in a unit square, count "
"how many fall inside the quarter circle, and print the estimate, the true value, and the "
"error. Plot the sampled points with matplotlib, colored by whether they land inside the "
"circle.",
"Python",
),
(
"✅ Todo app",
"Create a self-contained HTML/CSS/JavaScript todo app: add tasks, mark them complete, "
"delete them, filter by all/active/completed, and show a live count of remaining tasks, "
"with a clean, modern, responsive UI.",
"Web",
),
]
@dataclass
class PythonExecutionResult:
stdout: str
stderr: str
image_path: str | None
returncode: int | None
timed_out: bool = False
def _disable_parent_proc_inspection() -> None:
"""Best-effort hardening against same-UID reads of the parent process env.
The subprocess receives a scrubbed env, but Linux /proc can sometimes expose a
same-user process's environment. Marking the Gradio process non-dumpable helps
prevent generated code from reading `/proc/<parent>/environ`.
"""
if sys.platform != "linux":
return
try:
import ctypes
pr_set_dumpable = 4
libc = ctypes.CDLL(None)
libc.prctl(pr_set_dumpable, 0, 0, 0, 0)
except Exception:
logger.warning("Could not disable parent /proc inspection", exc_info=True)
def _build_client(api_key: str) -> ClientV2 | None:
if api_key:
return ClientV2(api_key=api_key, client_name=CLIENT_NAME)
else:
logger.warning("COHERE_API_KEY is not set; inference is disabled until configured.")
return None
_raw_api_key = os.getenv("COHERE_API_KEY", "").strip()
API_KEY_CONFIGURED = bool(_raw_api_key)
CLIENT = _build_client(_raw_api_key)
# Do not keep the secret in the parent environment longer than needed. The Cohere
# client has already been constructed, and subprocesses are passed explicit envs.
if _raw_api_key:
os.environ.pop("COHERE_API_KEY", None)
_disable_parent_proc_inspection()
_raw_api_key = ""
def _extract_content_parts(content: object) -> tuple[str, str]:
"""Extract visible text and reasoning text from Cohere content shapes."""
if content is None:
return "", ""
if isinstance(content, str):
return content, ""
if isinstance(content, list):
parts = [_extract_content_parts(block) for block in content]
return "".join(text for text, _ in parts), "".join(thinking for _, thinking in parts)
if isinstance(content, dict):
text = str(content.get("text") or "")
thinking = str(content.get("thinking") or "")
if not text and not thinking and "content" in content:
return _extract_content_parts(content.get("content"))
return text, thinking
text = getattr(content, "text", None)
thinking = getattr(content, "thinking", None)
return (str(text) if text is not None else ""), (str(thinking) if thinking is not None else "")
def _strip_thinking_blocks(text: str) -> str:
return THINKING_BLOCK_RE.sub("", text).strip()
def _format_response(output: str, thinking: str) -> str:
thinking = thinking.strip()
if not thinking:
return output
if not output:
return f"<think>{thinking}</think>"
return f"<think>{thinking}</think>\n\n{output}"
def _no_output_note(finish_reason: str) -> str:
if finish_reason == "MAX_TOKENS":
return "_The model hit its output-token cap before producing visible code._"
if finish_reason == "ERROR":
return "_The model returned an error before producing code. Please try again._"
return f"_The model finished without visible output (finish_reason={finish_reason})._"
def _format_api_error(exc: ApiError) -> str:
body = exc.body
if isinstance(body, dict):
message = body.get("message") or body.get("error") or ""
body_text = str(message) if message else str(body)
else:
body_text = str(body or "").strip()
if exc.status_code == 404:
return f"Model `{MODEL_ID}` was not found on the configured Cohere endpoint."
if exc.status_code in (401, 403):
return "The `COHERE_API_KEY` secret was rejected. Check the Space secret."
if exc.status_code == 429:
return "The Cohere API rate limit was reached. Please wait and try again."
return body_text[:240] or f"HTTP {exc.status_code}"
def call_model(messages: list[dict[str, Any]]) -> Iterator[str]:
"""Stream cumulative model text.
All Cohere-specific details are intentionally isolated here: model name,
client method, streaming event shape, and reasoning handling.
"""
if CLIENT is None:
if not API_KEY_CONFIGURED:
yield "This Space needs a `COHERE_API_KEY` secret before it can call Cohere."
else:
yield "Cohere client is not configured."
return
output = ""
thinking_output = ""
finish_reason: str | None = None
event_counts: dict[str, int] = {}
try:
stream = CLIENT.chat_stream(
model=MODEL_ID,
messages=messages,
temperature=DEFAULT_TEMPERATURE,
thinking={"type": "enabled"},
)
for event in stream:
event_type = getattr(event, "type", None) or "unknown"
event_counts[event_type] = event_counts.get(event_type, 0) + 1
delta = getattr(event, "delta", None)
if event_type in ("content-delta", "content-start"):
msg = getattr(delta, "message", None) if delta is not None else None
if msg is None:
continue
text, thinking = _extract_content_parts(getattr(msg, "content", None))
if thinking:
thinking_output += thinking
yield _format_response(output, thinking_output)
if text:
output += text
yield _format_response(output, thinking_output)
elif event_type == "message-end":
finish_reason = getattr(delta, "finish_reason", None)
if finish_reason is None and isinstance(delta, dict):
finish_reason = delta.get("finish_reason")
logger.info(
"Cohere stream ended: finish_reason=%s, output_len=%d, thinking_len=%d, events=%s",
finish_reason,
len(output),
len(thinking_output),
event_counts,
)
if not output:
yield _format_response(_no_output_note((finish_reason or "unknown").upper()), thinking_output)
except ApiError as exc:
logger.exception("Cohere API error (status=%s)", exc.status_code)
yield _format_response(f"_Cohere API error_: {_format_api_error(exc)}", thinking_output)
except Exception as exc:
logger.exception("Unexpected error calling Cohere API")
yield _format_response(f"_Unexpected error calling Cohere_: {exc}", thinking_output)
def _chat_history_to_messages(history: list[dict[str, str]]) -> list[dict[str, Any]]:
messages: list[dict[str, Any]] = [{"role": "system", "content": SYSTEM_PROMPT}]
for item in history:
role = item.get("role")
content = str(item.get("content") or "").strip()
if role not in {"user", "assistant"} or not content:
continue
if role == "assistant":
content = _strip_thinking_blocks(content)
messages.append({"role": role, "content": content})
return messages
def _clip_context(text: str, limit: int = 6_000) -> str:
if len(text) <= limit:
return text
return text[:limit] + f"\n... truncated {len(text) - limit} characters ..."
def _iteration_context(execution_context: dict[str, Any] | None) -> str:
if not execution_context or not execution_context.get("code"):
return ""
code = _clip_context(str(execution_context.get("code") or ""), 8_000)
target = str(execution_context.get("target") or "code")
fence_lang = str(execution_context.get("fence_lang") or target)
status = str(execution_context.get("status") or "")
stdout = _clip_context(str(execution_context.get("stdout") or ""), 2_000)
stderr = _clip_context(str(execution_context.get("stderr") or ""), 2_000)
parts = [
"Previous generated code and run result are available for iteration.",
f"Previous target: {target}",
f"Previous status: {status}",
f"Previous code:\n```{fence_lang}\n{code}\n```",
]
if stdout:
parts.append(f"Previous stdout:\n{stdout}")
if stderr:
parts.append(f"Previous stderr / traceback:\n{stderr}")
parts.append("If the user asks to revise, debug, extend, or explain the prior code, use this context.")
return "\n\n".join(parts)
def _targeted_prompt(
prompt: str,
target_language: str,
execution_context: dict[str, Any] | None = None,
) -> str:
target = "Python" if target_language == "Python" else "Web"
iteration_context = _iteration_context(execution_context)
context_block = f"\n\n{iteration_context}" if iteration_context else ""
if target == "Python":
return (
"Target: Python. Stay within coding-related requests only. "
"If the user asks a coding question or wants reasoning that does not require running code, "
"answer directly without a fenced block. If they ask to generate, revise, or fix runnable "
"code, return one ```python fenced block only. The code will be executed in a short-lived "
"subprocess."
f"{context_block}\n\n"
f"User request:\n{prompt}"
)
return (
"Target: Web. Stay within coding-related requests only. "
"If the user asks a coding question or wants reasoning that does not require running code, "
"answer directly without a fenced block. If they ask to generate, revise, or fix runnable "
"web code, return one ```html fenced block only. The code is rendered inside a sandboxed "
"iframe that fills the preview panel, so design the page to fill that iframe responsively: "
"html/body at margin:0 and 100% width/height, "
"avoid fixed widths larger than the iframe, and resize any <canvas> to its container "
"(including on window resize) so the whole app is visible without horizontal scrolling."
f"{context_block}\n\n"
f"User request:\n{prompt}"
)
def extract_code(response: str) -> tuple[str, str | None]:
"""Return the first fenced code block and its language tag."""
visible_response = _strip_thinking_blocks(response)
match = CODE_BLOCK_RE.search(visible_response)
if not match:
return "", None
return match.group(2).strip(), (match.group(1).strip().lower() or None)
def _normalize_language(target_language: str | None, fence_lang: str | None) -> str:
if fence_lang in {"python", "py"}:
return "python"
if fence_lang in {"html", "web", "javascript", "js", "css"}:
return "web"
if target_language in {"Python", "Web"}:
return target_language.lower()
return "python"
def _truncate_output(text: str) -> str:
if len(text) <= MAX_STDIO_CHARS:
return text
remaining = len(text) - MAX_STDIO_CHARS
return text[:MAX_STDIO_CHARS] + f"\n\n... truncated {remaining} characters ..."
def _decode_timeout_output(value: str | bytes | None) -> str:
if value is None:
return ""
if isinstance(value, bytes):
return value.decode("utf-8", errors="replace")
return value
def _apply_subprocess_limits() -> None:
"""Apply child-only CPU and memory caps before Python user code starts."""
import resource
mem_bytes = PY_MEM_LIMIT_MB * 1024 * 1024
resource.setrlimit(resource.RLIMIT_AS, (mem_bytes, mem_bytes))
resource.setrlimit(resource.RLIMIT_CPU, (PY_TIMEOUT_S, PY_TIMEOUT_S))
def _python_runner_source() -> str:
return textwrap.dedent(
f"""
import os
import runpy
import sys
import traceback
os.environ.setdefault("MPLBACKEND", "Agg")
exit_code = 0
try:
runpy.run_path(os.path.join(os.getcwd(), "user_code.py"), run_name="__main__")
except SystemExit as exc:
code = exc.code
exit_code = code if isinstance(code, int) else 1
except Exception:
traceback.print_exc()
exit_code = 1
finally:
try:
import matplotlib
matplotlib.use("Agg", force=True)
import matplotlib.pyplot as plt
if plt.get_fignums():
plt.savefig(os.environ["OUTPUT_PNG"], bbox_inches="tight")
except ModuleNotFoundError as exc:
if exc.name != "matplotlib":
traceback.print_exc()
except Exception:
traceback.print_exc()
raise SystemExit(exit_code)
"""
).strip()
def run_python(code: str) -> PythonExecutionResult:
"""Execute generated Python in a subprocess with baseline containment.
Security boundary for v1:
- Scrubbed env: never pass os.environ, so COHERE_API_KEY is absent.
- Hard timeout: kill code that hangs the Space.
- Memory/CPU caps: reduce the blast radius of runaway code.
Accepted limitation: a standard non-privileged Gradio Space cannot reliably
block network egress or filesystem reads from this subprocess. Full isolation
would require a privileged Docker Space with nsjail/gVisor, or an external
executor such as E2B/Modal. Do not refactor these comments away; they define
the risk boundary of this demo.
"""
with tempfile.TemporaryDirectory(prefix="coding_model_run_") as tmp:
workdir = Path(tmp)
runner_path = workdir / "runner.py"
user_path = workdir / "user_code.py"
image_path = workdir / OUTPUT_PNG
runner_path.write_text(_python_runner_source(), encoding="utf-8")
user_path.write_text(code, encoding="utf-8")
env = {
# SECURITY: explicit scrubbed env. Never pass parent os.environ, which
# may contain COHERE_API_KEY or other HF Space secrets.
"PATH": "/usr/bin:/bin",
"HOME": str(workdir),
"TMPDIR": str(workdir),
"MPLBACKEND": "Agg",
"MPLCONFIGDIR": str(workdir / ".matplotlib"),
"OUTPUT_PNG": str(image_path),
"PYTHONIOENCODING": "utf-8",
"PYTHONNOUSERSITE": "1",
"PYTHONUNBUFFERED": "1",
"LANG": "C.UTF-8",
# Keep numeric libraries from spawning many workers inside the capped process.
"OPENBLAS_NUM_THREADS": "1",
"OMP_NUM_THREADS": "1",
"MKL_NUM_THREADS": "1",
"NUMEXPR_NUM_THREADS": "1",
}
try:
completed = subprocess.run(
[sys.executable, "-I", str(runner_path)],
cwd=workdir,
env=env,
capture_output=True,
text=True,
encoding="utf-8",
errors="replace",
timeout=PY_TIMEOUT_S,
preexec_fn=_apply_subprocess_limits if sys.platform == "linux" else None,
check=False,
)
stdout = _truncate_output(completed.stdout)
stderr = _truncate_output(completed.stderr)
if completed.returncode and not stderr:
stderr = f"Process exited with status {completed.returncode}."
saved_image: str | None = None
if image_path.exists() and image_path.stat().st_size > 0:
saved = tempfile.NamedTemporaryFile(
prefix="coding_model_plot_", suffix=".png", delete=False
)
saved.close()
Path(saved.name).write_bytes(image_path.read_bytes())
saved_image = saved.name
return PythonExecutionResult(
stdout=stdout,
stderr=stderr,
image_path=saved_image,
returncode=completed.returncode,
)
except subprocess.TimeoutExpired as exc:
stdout = _truncate_output(_decode_timeout_output(exc.stdout))
stderr = _truncate_output(_decode_timeout_output(exc.stderr))
timeout_note = f"Timed out after {PY_TIMEOUT_S} seconds; the process was killed."
stderr = f"{stderr}\n{timeout_note}".strip()
return PythonExecutionResult(
stdout=stdout,
stderr=stderr,
image_path=None,
returncode=None,
timed_out=True,
)
def _web_document(code: str, fence_lang: str | None) -> str:
lang = (fence_lang or "").lower()
if lang in {"javascript", "js"}:
return f"<!doctype html><html><body><script>\n{code}\n</script></body></html>"
if lang == "css":
return f"<!doctype html><html><head><style>\n{code}\n</style></head><body></body></html>"
if re.search(r"<!doctype|<html[\s>]", code, flags=re.IGNORECASE):
return code
return f"<!doctype html><html><head><meta charset='utf-8'></head><body>\n{code}\n</body></html>"
def build_iframe(code: str, fence_lang: str | None = None) -> str:
"""Render web code in a sandboxed iframe.
SECURITY: sandbox allows scripts so demos can run, but deliberately omits
allow-same-origin. Without same-origin, generated code cannot share the
parent origin, cookies, or storage.
"""
document = _web_document(code, fence_lang)
srcdoc = html.escape(document, quote=True)
return (
'<iframe class="web-frame" '
'sandbox="allow-scripts" '
'allow="fullscreen" '
"allowfullscreen "
f'srcdoc="{srcdoc}" '
'style="width:100%; min-height:680px; border:0; border-radius:14px; '
'background:white;"></iframe>'
)
# ---------- gradio.Server application ----------
def _run_extracted_code(
code: str,
target: str,
) -> tuple[str, str, str | None, str, str]:
"""Execute code and return (stdout, stderr, image_path, status_text, status_state)."""
if target == "python":
result = run_python(code)
if result.timed_out:
return result.stdout, result.stderr, result.image_path, f"Timed out after {PY_TIMEOUT_S}s", "error"
if result.returncode:
return result.stdout, result.stderr, result.image_path, "Finished with errors", "error"
return result.stdout, result.stderr, result.image_path, "Ran successfully", "success"
return "", "", None, "Preview ready", "success"
def _updated_execution_context(
*,
code: str,
target: str,
fence_lang: str | None,
stdout: str,
stderr: str,
image_path: str | None,
status: str,
download_path: str | None,
) -> dict[str, Any]:
return {
"code": code,
"target": target,
"fence_lang": fence_lang or target,
"stdout": stdout,
"stderr": stderr,
"image_path": image_path,
"status": status,
"download_path": download_path,
}
# In-memory registry of temp files to serve (images, downloads)
_served_files: dict[str, str] = {}
app = Server()
@app.get("/", response_class=HTMLResponse)
async def homepage():
html_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "index.html")
with open(html_path, "r", encoding="utf-8") as f:
content = f.read()
# Inject runtime config
config = json.dumps({
"api_key_configured": API_KEY_CONFIGURED,
"app_title": APP_TITLE,
"model_id": MODEL_ID,
"model_url": MODEL_URL,
"opencode_url": OPENCODE_URL,
"examples": [{"label": label, "prompt": prompt, "target": target} for label, prompt, target in EXAMPLE_PROMPTS],
})
content = content.replace("__RUNTIME_CONFIG__", config)
return content
@app.get("/images/{filename}")
async def serve_image(filename: str):
path = _served_files.get(f"img:{filename}")
if path and os.path.exists(path):
return FileResponse(path, media_type="image/png")
return HTMLResponse("Not found", status_code=404)
@app.get("/download/{filename}")
async def serve_download(filename: str):
path = _served_files.get(f"dl:{filename}")
if path and os.path.exists(path):
return FileResponse(path, filename=filename, media_type="application/octet-stream")
return HTMLResponse("Not found", status_code=404)
@app.api(name="chat", concurrency_limit=2)
def handle_chat(prompt: str, target_language: str, history_json: str, exec_context_json: str) -> str:
"""Stream chat responses with code execution. Yields JSON strings."""
history = json.loads(history_json) if history_json else []
execution_context = json.loads(exec_context_json) if exec_context_json else {}
prompt = (prompt or "").strip()
if not prompt:
yield json.dumps({"type": "error", "status_text": "Enter a prompt to get started.", "status_state": "info", "history": history, "execution": execution_context})
return
# Add user message and placeholder assistant message
history = list(history) + [
{"role": "user", "content": prompt},
{"role": "assistant", "content": ""},
]
yield json.dumps({"type": "status", "status_text": "Thinking…", "status_state": "working", "history": history, "execution": execution_context})
# Build messages for Cohere — use targeted prompt
cohere_history = list(history[:-1]) # everything except empty assistant
# Replace the last user message with the targeted version
cohere_history[-1] = {"role": "user", "content": _targeted_prompt(prompt, target_language, execution_context)}
messages = _chat_history_to_messages(cohere_history)
final_response = ""
for partial in call_model(messages):
final_response = partial
history[-1]["content"] = partial
yield json.dumps({"type": "streaming", "status_text": "Generating…", "status_state": "working", "history": history, "execution": execution_context})
if not final_response:
history[-1]["content"] = "The model did not return a response."
yield json.dumps({"type": "error", "status_text": "No model response.", "status_state": "error", "history": history, "execution": execution_context})
return
code, fence_lang = extract_code(final_response)
target = _normalize_language(target_language, fence_lang)
if not code:
yield json.dumps({"type": "complete", "status_text": "Answered without running code.", "status_state": "info", "history": history, "execution": execution_context})
return
yield json.dumps({"type": "status", "status_text": "Running…", "status_state": "working", "history": history, "execution": execution_context})
stdout, stderr, image_path, status_text, status_state = _run_extracted_code(code, target)
# Register image for serving
image_url = None
if image_path:
filename = os.path.basename(image_path)
_served_files[f"img:{filename}"] = image_path
image_url = f"/images/{filename}"
# Register code for download
download_url = None
if code:
ext = "py" if target == "python" else "html"
dl_filename = f"generated.{ext}"
dl_dir = tempfile.mkdtemp(prefix="coding_model_dl_")
dl_path = os.path.join(dl_dir, dl_filename)
Path(dl_path).write_text(code, encoding="utf-8")
_served_files[f"dl:{dl_filename}"] = dl_path
download_url = f"/download/{dl_filename}"
execution_context = {
"code": code,
"target": target,
"fence_lang": fence_lang or target,
"stdout": stdout,
"stderr": stderr,
"image_url": image_url,
"image_path": image_path,
"status": status_text,
"language": "python" if target == "python" else "html",
"suggested_tab": "preview" if (image_path or target == "web") else "console",
"download_url": download_url,
}
yield json.dumps({"type": "complete", "status_text": status_text, "status_state": status_state, "history": history, "execution": execution_context})
app.launch(show_error=True)