server/sandbox_smoke.py

"""Stand-alone E2B sandbox smoke — boot opencode serve, expose it publicly.

This script isolates "can a sandbox even stand up opencode serve?" from
the rest of the env (no MCP server, no proxy, no primitive, no UI). Good
for when a full rollout fails and you want to rule out the sandbox path.

What it does:
    1. Create a fresh E2B sandbox.
    2. Write ``~/.config/opencode/opencode.json`` pointing at either:
       - the HF Router (default, just needs HF_TOKEN), or
       - a user-provided vLLM URL.
    3. Install opencode via the upstream one-liner.
    4. Start ``opencode serve --port 4096 --hostname 0.0.0.0`` in bg.
    5. ``sandbox.get_host(4096)`` → a public ``https://4096-<sbx>.e2b.app``.
    6. Poll ``{public}/doc`` until it answers 200.
    7. Print the public URL + ``sandbox_id`` and keep the sandbox alive so
       you can hit it manually. Ctrl-C closes the sandbox.

Usage:
    # HF Router (default)
    HF_TOKEN=hf_... uv run python server/sandbox_smoke.py

    # or self-hosted vLLM
    uv run python server/sandbox_smoke.py \\
        --backend vllm \\
        --vllm-url https://my-tunnel.example/v1 \\
        --model Qwen/Qwen3.5-4B

Once it prints the URL you can:

    curl https://4096-<sbx>.e2b.app/global/health
    curl https://4096-<sbx>.e2b.app/config
    # create + send prompt
    SID=$(curl -s -X POST https://4096-<sbx>.e2b.app/session \\
        -H 'content-type: application/json' \\
        -d '{"title":"smoke"}' | python3 -c 'import json,sys;print(json.load(sys.stdin)["id"])')
    curl -X POST https://4096-<sbx>.e2b.app/session/$SID/prompt_async \\
        -H 'content-type: application/json' \\
        -d '{"parts":[{"type":"text","text":"write hello.py"}]}'
    curl -N https://4096-<sbx>.e2b.app/event
"""

from __future__ import annotations

import argparse
import json
import os
import signal
import sys
import time
from pathlib import Path
from typing import Any

# Load the env-server's .env (E2B_API_KEY, HF_TOKEN, etc.) before importing
# anything that needs them. Walks up from this file to find ``openenv/.env``.
try:
    from dotenv import load_dotenv

    _env_path = Path(__file__).resolve().parent.parent / ".env"
    if _env_path.is_file():
        load_dotenv(_env_path, override=False)
        print(f"loaded env from {_env_path}")
except ImportError:
    pass

try:
    from e2b_code_interpreter import Sandbox
except ImportError:
    from e2b import Sandbox  # type: ignore


SERVE_PORT = 4096
CONFIG_DIR = "/home/user/.config/opencode"
CONFIG_PATH = f"{CONFIG_DIR}/opencode.json"
LOG_DIR = "/home/user/logs/agent"
SERVE_LOG = f"{LOG_DIR}/serve.log"


def build_opencode_json(
    *,
    backend: str,
    model_id: str,
    base_url: str,
    api_key: str,
    context_limit: int = 32768,
    output_limit: int = 16384,
) -> str:
    """Emit a minimal, valid opencode.json for the chosen backend."""
    provider_id = "vllm" if backend == "vllm" else "hf-router"
    return json.dumps({
        "$schema": "https://opencode.ai/config.json",
        "model": f"{provider_id}/{model_id}",
        "provider": {
            provider_id: {
                "npm": "@ai-sdk/openai-compatible",
                "name": f"{provider_id} (smoke)",
                "options": {
                    "baseURL": base_url,
                    "apiKey": api_key,
                    "timeout": 600_000,
                },
                "models": {
                    model_id: {
                        "name": model_id,
                        "limit": {"context": context_limit, "output": output_limit},
                    },
                },
            },
        },
        "tools": {"webfetch": False, "question": False},
    }, indent=2)


_START = time.time()


def log(msg: str) -> None:
    """Timestamped progress line, flushed so it appears in real time."""
    t = time.time() - _START
    print(f"[{t:6.1f}s] {msg}", flush=True)


def run_shell(sbx: Any, cmd: str, *, timeout_s: int = 120) -> tuple[int, str, str]:
    """Run a shell command, return (exit_code, stdout, stderr)."""
    out = sbx.commands.run(cmd, timeout=timeout_s)
    return (out.exit_code, out.stdout or "", out.stderr or "")


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--backend", choices=["hf", "vllm"], default="hf")
    ap.add_argument("--model", default="Qwen/Qwen3.5-397B-A17B:together")
    ap.add_argument("--vllm-url", default="")
    ap.add_argument("--hf-token", default=os.environ.get("HF_TOKEN", ""))
    ap.add_argument("--sandbox-timeout-s", type=int, default=900)
    ap.add_argument("--idle-hold-s", type=int, default=1200,
                    help="keep the sandbox alive for this many seconds after boot")
    args = ap.parse_args()

    if args.backend == "hf":
        if not args.hf_token:
            print("ERROR: --backend hf needs --hf-token or $HF_TOKEN", file=sys.stderr)
            return 2
        base_url = "https://router.huggingface.co/v1"
        api_key = args.hf_token
    else:
        if not args.vllm_url:
            print("ERROR: --backend vllm needs --vllm-url", file=sys.stderr)
            return 2
        base_url = args.vllm_url.rstrip("/")
        if not base_url.endswith("/v1"):
            base_url += "/v1"
        api_key = "anything"

    if not os.environ.get("E2B_API_KEY"):
        print("ERROR: E2B_API_KEY not set", file=sys.stderr)
        return 2

    log(f"[1/7] creating sandbox (timeout={args.sandbox_timeout_s}s) …")
    sbx = Sandbox.create(timeout=args.sandbox_timeout_s)
    log(f"      sandbox_id = {sbx.sandbox_id}")

    try:
        log("[2/7] mkdir config + logs …")
        rc, out, err = run_shell(sbx, f"mkdir -p {CONFIG_DIR} {LOG_DIR}")
        if rc != 0:
            log(f"      FAIL rc={rc} stderr={err[:500]}")
            return 1

        log(f"[3/7] writing {CONFIG_PATH} …")
        cfg = build_opencode_json(
            backend=args.backend,
            model_id=args.model,
            base_url=base_url,
            api_key=api_key,
        )
        sbx.files.write(CONFIG_PATH, cfg)
        log(f"      backend={args.backend}  model={args.model}")
        log(f"      baseURL={base_url}")

        log("[4/7] installing opencode via curl opencode.ai/install … (~10-30s cold)")
        rc, out, err = run_shell(
            sbx,
            "curl -fsSL https://opencode.ai/install | bash 2>&1",
            timeout_s=300,
        )
        log(f"      install rc={rc}")
        if out:
            for line in out.strip().splitlines()[-8:]:
                log(f"      │ {line}")
        if rc != 0:
            log("      stderr tail:")
            for line in (err or "").strip().splitlines()[-10:]:
                log(f"      │ {line}")
            return 1

        log("[5/7] verifying opencode binary …")
        rc, out, err = run_shell(sbx, '$HOME/.opencode/bin/opencode --version')
        log(f"      opencode --version rc={rc} out={(out or '').strip()[:120]}")
        if rc != 0:
            log(f"      stderr: {(err or '')[:400]}")
            return 1

        log(f"[6/7] starting opencode serve in bg on :{SERVE_PORT} …")
        serve_cmd = (
            'export PATH="$HOME/.opencode/bin:$PATH" && '
            f"opencode serve --port {SERVE_PORT} --hostname 0.0.0.0 "
            f"> {SERVE_LOG} 2>&1"
        )
        serve_bg = sbx.commands.run(serve_cmd, background=True)
        log(f"      serve pid = {getattr(serve_bg, 'pid', '?')}")

        host = sbx.get_host(SERVE_PORT)
        public_url = f"https://{host}"
        log(f"      public URL = {public_url}")

        log("[7/7] waiting for /doc to answer (polls every 0.5s for 60s) …")
        import httpx
        ok = False
        for i in range(120):
            try:
                r = httpx.get(f"{public_url}/doc", timeout=5)
                if r.status_code == 200:
                    log(f"      /doc ok (poll #{i+1}, {i*0.5:.1f}s)")
                    ok = True
                    break
                elif i % 6 == 5:  # ~every 3s print progress
                    log(f"      /doc → HTTP {r.status_code} (still trying, {i*0.5:.1f}s)")
            except Exception as exc:
                if i % 6 == 5:
                    log(f"      /doc unreachable ({type(exc).__name__}, {i*0.5:.1f}s)")
            time.sleep(0.5)
        if not ok:
            log("      /doc never answered — tailing serve log (last 2KB):")
            try:
                tail = sbx.files.read(SERVE_LOG)[-2000:]
            except Exception as exc:
                tail = f"(could not read log: {exc})"
            for line in tail.splitlines()[-40:]:
                log(f"      │ {line}")
            return 1

        print("\n" + "=" * 70)
        print("sandbox is up — manual probe recipes:")
        print("=" * 70)
        print(f"curl  -s  {public_url}/global/health | jq .")
        print(f"curl  -s  {public_url}/config | jq '.model, .provider'")
        print()
        print(f"SID=$(curl -s -X POST {public_url}/session \\")
        print("        -H 'content-type: application/json' \\")
        print("        -d '{\"title\":\"smoke\"}' | jq -r .id)")
        print(f"curl  -X POST {public_url}/session/$SID/prompt_async \\")
        print("        -H 'content-type: application/json' \\")
        print("        -d '{\"parts\":[{\"type\":\"text\",\"text\":\"write hello.py and run it\"}]}'")
        print(f"curl  -N  {public_url}/event      # SSE stream")
        print()
        print(f"serve log:  sbx.files.read('{SERVE_LOG}')")
        print(f"sandbox_id: {sbx.sandbox_id}")
        print(f"holding for up to {args.idle_hold_s}s — Ctrl-C to close")
        print("=" * 70 + "\n")

        stopper = {"stop": False}
        def _sigh(*_a):
            print("\nsignal — closing sandbox")
            stopper["stop"] = True
        signal.signal(signal.SIGINT, _sigh)
        signal.signal(signal.SIGTERM, _sigh)

        # Periodic /doc ping so we catch opencode-serve crashes in real time.
        # Any non-200 (incl. E2B's 502 "port not open") is a crash signal —
        # dump serve.log and stop the hold.
        import httpx
        last_ok_ts = time.time()
        deadline = time.time() + args.idle_hold_s
        def _dump_serve_log() -> None:
            try:
                tail = sbx.files.read(SERVE_LOG)
                log("    --- serve.log tail (last 4KB) ---")
                for line in tail[-4000:].splitlines()[-60:]:
                    log(f"    │ {line}")
                log("    --- end serve.log ---")
            except Exception as exc2:
                log(f"    could not read serve.log: {exc2}")
            # Also list workdir so we can see if the agent did anything.
            try:
                rc, out, err = run_shell(sbx, "ls -la /home/user/workdir 2>&1 | head -40")
                log("    --- workdir ls ---")
                for line in (out or err).splitlines():
                    log(f"    │ {line}")
            except Exception:
                pass
        while time.time() < deadline and not stopper["stop"]:
            try:
                r = httpx.get(f"{public_url}/doc", timeout=5)
                if r.status_code == 200:
                    last_ok_ts = time.time()
                else:
                    log(f"!!! /doc → HTTP {r.status_code}  "
                        f"(last ok {time.time()-last_ok_ts:.1f}s ago) — "
                        f"opencode serve appears dead, dumping log")
                    _dump_serve_log()
                    break
            except Exception as exc:
                log(f"!!! /doc probe failed: {type(exc).__name__}: {exc}  "
                    f"(last ok {time.time()-last_ok_ts:.1f}s ago)")
                _dump_serve_log()
                break
            time.sleep(10.0)
        return 0

    finally:
        try:
            print("killing sandbox …")
            sbx.kill()
        except Exception as exc:
            print(f"  kill failed (probably already dead): {exc}")


if __name__ == "__main__":
    sys.exit(main())