diff --git a/qwen3_6_scripts/patch_ops.sh b/qwen3_6_scripts/patch_ops.sh
index 1035b2f..82d23c1 100755
--- a/qwen3_6_scripts/patch_ops.sh
+++ b/qwen3_6_scripts/patch_ops.sh
@@ -42,7 +42,8 @@ cp ./paged_attn.py /usr/local/corex/lib/python3/dist-packages/vllm/attention/ops
 python3 ./patch_model_runner.py
 
 # --- transformers: Qwen3_5 tokenizer / model files --------------------------
-pip install transformers==4.55.3 -i https://pypi.tuna.tsinghua.edu.cn/simple
+pip install transformers==4.55.3 -i https://pypi.tuna.tsinghua.edu.cn/simple || \
+  echo "[patch_ops] WARN: transformers==4.55.3 install failed; continuing with installed transformers."
 cp -r ./qwen3_5 /usr/local/lib/python3.10/site-packages/transformers/models/
 cp -r ./qwen3_5_moe /usr/local/lib/python3.10/site-packages/transformers/models/
 python3 ./patch_transformers_qwen3_5.py
diff --git a/worklogs/2026-07-13-initial-run.md b/worklogs/2026-07-13-initial-run.md
new file mode 100644
index 0000000..73ba5c5
--- /dev/null
+++ b/worklogs/2026-07-13-initial-run.md
@@ -0,0 +1,149 @@
+# 2026-07-13 Initial Run Worklog
+
+## Working Rules
+
+- Local repository is the source of truth for code changes.
+- Remote Phanthy GPU server is used for build, runtime, and validation.
+- Every meaningful experiment records: code version, command, environment, result, issue, and next action.
+- Changes should be committed with git after a coherent milestone or before risky experiments.
+
+## Remote Target
+
+- Host: `ssh-55c3b0b3.default.gpu.phanthy.com`
+- SSH requires TLS ProxyCommand on port `32222`.
+- User: `root`
+
+## Status
+
+- Local repository inspected.
+- SSH connectivity confirmed with `whoami` and `hostname`.
+- Remote environment inspected.
+
+## Remote Environment Findings
+
+- Remote shell user: `root`
+- Remote hostname: `cc-55c3b0b3-8c03-4fe2-8ef8-7109a6aff0d6-0`
+- Remote appears to already be inside a container.
+- `docker` is not installed in the remote runtime container.
+- CoreX is installed under `/usr/local/corex -> /usr/local/corex-3.2.3`.
+- Iluvatar devices are visible as `/dev/iluvatar0` through `/dev/iluvatar3`.
+- PyTorch is available only when `PYTHONPATH` and `LD_LIBRARY_PATH` include CoreX paths.
+- PyTorch reports `torch.cuda.is_available() == True` and `torch.cuda.device_count() == 4`.
+- Model path found: `/root/public-storage/models/Qwen/Qwen3.6-35B-A3B`.
+- Model `config.json` already has `architectures: ["Qwen3_5MoeForCausalLM"]`.
+- Because remote has no Docker, first run will patch the current runtime directly instead of building an image.
+
+## First Run Plan
+
+1. Sync only lightweight submission files to `/root/work/enginex-vllm-bi100-qwen36`.
+2. Backup target runtime files before applying `qwen3_6_scripts/patch_ops.sh`.
+3. Apply patches in the remote CoreX container.
+4. Start OpenAI-compatible API server against `/root/public-storage/models/Qwen/Qwen3.6-35B-A3B`.
+5. Run minimal smoke tests before deeper compatibility tests.
+
+## 2026-07-14 First Patch Attempt Diagnosis
+
+- User-provided remote log showed repeated `$'\r': command not found` in `patch_ops.sh`.
+- Root cause: the shell script, and likely copied `.py` patch helpers, arrived on Linux with Windows CRLF line endings.
+- Consequence: `python3 ./patch_model_runner.py\r` and similar patch commands did not execute correctly.
+- Later server startup failed with `unrecognized arguments: --reasoning-parser qwen3`.
+- Interpretation: OpenAI API runtime files such as `cli_args.py` and `api_server.py` were not patched into `/usr/local/corex/lib/python3/dist-packages/vllm/...`.
+- Next action: normalize CRLF to LF on the remote copy, rerun `patch_ops.sh`, verify `--reasoning-parser` exists in the installed vLLM CLI, then restart the API server.
+
+## 2026-07-14 First Server Start After CRLF Fix
+
+- User verified `python3 -m vllm.entrypoints.openai.api_server --help | grep reasoning-parser` now prints `--reasoning-parser`.
+- Startup log shows `reasoning_parser='qwen3'`, so OpenAI API argument patch is active.
+- `ixsmi` shows no vLLM GPU process and `/health` returns `Connection refused`, so the API server process exited before serving.
+- Need inspect the tail of `/root/work/logs/server_first_run.log` after `Downcasting torch.float32 to torch.float16`.
+- Additional suspicion to verify: some patches target `/usr/local/corex/lib/python3/dist-packages` while the Qwen3_5 registry patch targets `/usr/local/corex/lib64/python3/dist-packages`; confirm the actual imported registry path and whether both `lib` and `lib64` contain `Qwen3_5MoeForCausalLM`.
+
+## 2026-07-14 Server Still Initializing
+
+- User checked `jobs -l` and `ps`; master process PID 2052 is still running.
+- Log has no `Traceback`, `ERROR`, or common fatal exceptions.
+- Log reached `Worker ready; awaiting tasks` for three worker processes, so multiprocess executor startup is progressing.
+- Both `/usr/local/corex/lib/.../vllm` and `/usr/local/corex/lib64/.../vllm` contain `registry.py` patched with `Qwen3_5MoeForCausalLM`, and `qwen3_5.py` exists in both.
+- Earlier `/health` failure was likely checked before API server finished engine initialization and started listening.
+- Next action: wait and monitor model loading/GPU memory, then retry `/health`; only diagnose hang if no new log/GPU memory movement for several minutes.
+
+## 2026-07-14 Port Occupied By Stale First Process
+
+- User started a second server while PID 2052 from the first run was still alive.
+- Second run exited with `OSError: [Errno 98] Address already in use` at `sock.bind(("", args.port))`.
+- `ps` still shows PID 2052 holding the API server command, but `ixsmi` shows no model GPU memory/processes.
+- Interpretation: PID 2052 is a stale or stuck master process occupying port 1111, not a healthy loaded model service.
+- Next action: stop PID 2052 and any VllmWorkerProcess children, verify port 1111 is free, restart once with a fresh log filename, then wait for explicit `Uvicorn running` / startup-complete logs before testing `/health`.
+
+## 2026-07-14 New Instance Startup Fix
+
+- New host: `ssh-8d2ae743.default.gpu.phanthy.com`.
+- Initial failure: `api_server.py: error: unrecognized arguments: --reasoning-parser qwen3`.
+- Cause 1: new instance runtime had not yet applied `qwen3_6_scripts/patch_ops.sh`.
+- Applied `patch_ops.sh`; `cli_args.py` in both `/usr/local/corex/lib/...` and `/usr/local/corex/lib64/...` then contained `--reasoning-parser`.
+- Cause 2: service was started from repository root `/root/data-disk-1/enginex-vllm-bi100-qwen36`, whose local `./vllm` package shadowed the patched CoreX runtime vLLM.
+- Fix: start service from `/root` and set CoreX `PYTHONPATH`/`LD_LIBRARY_PATH`, using `python3 -B` to avoid stale bytecode.
+- Working log: `/root/work/logs/server_from_root_B.log`.
+- Health check reached `GET /health HTTP/1.1" 200 OK`.
+- Minimal `/v1/chat/completions` request reached `POST /v1/chat/completions HTTP/1.1" 200 OK`.
+- `ixsmi` shows model loaded at roughly 29GB per BI-V100 card.
+
+## 2026-07-14 Baseline Smoke And Mini Benchmark
+
+- Remote benchmark script copied to `/root/work/remote_smoke_bench.py`.
+- Result file: `/root/work/logs/baseline_smoke_20260713_195201.json`.
+- `/health` status: `200`.
+- Non-stream smoke request: `200`, 16 completion tokens in `2.41s`, about `6.63 tok/s`.
+- Streaming 3-run summary:
+  - TTFT average: `1.03s`.
+  - TTFT P90 from 3 samples: about `0.80s` using the simple small-sample estimator.
+  - Output TPS after TTFT: about `8.91 tok/s`.
+- Prefix-cache probe:
+  - First long-prefix request: `cached_tokens=0`, 16 completion tokens in `6.59s`.
+  - Second related long-prefix request: `cached_tokens=1056 / 1069 prompt tokens`, 16 completion tokens in `3.61s`.
+  - Server metrics also reported GPU prefix cache hit rate around `50.36%`.
+- Current status: model is fully runnable and API-compatible enough for smoke tests, but generation throughput is far below the contest target of Output TPS P10 >= 20.
+- First performance direction: reduce thinking-token waste, profile decode path, verify XFormers fallback cost, then tune serving parameters after a larger benchmark.
+
+## 2026-07-14 Official-Like Benchmark Dataset Preparation
+
+- Current repository dataset `chat_dataset_v0.json` is only a 4-conversation smoke dataset.
+- It has short multi-turn chat/role-play prompts and no realistic `tools`, assistant `tool_calls`, or `tool` role messages.
+- It is not representative of the official benchmark, which is an 881-request long-context Agent coding workload.
+- Added `worklogs/formal_perf_bench.py` support for OpenAI-style request items with `messages`, `tools`, `tool_choice`, `stream`, and per-request `max_tokens`.
+- Added `worklogs/generate_official_like_dataset.py` to synthesize official-like long-context, stream=true, tool-heavy requests.
+- Uploaded updated scripts to:
+  - `/root/work/formal_perf_bench.py`
+  - `/root/work/generate_official_like_dataset.py`
+- Server smoke generation succeeded:
+  - `/root/work/logs/synthetic_agent_perf_16.jsonl`
+  - 16 requests, 13 sessions, approx prompt average `26611.5`, approximate weighted cache ratio `0.638`.
+  - File size `4.5M`.
+
+## 2026-07-14 Official-Like Synthetic Benchmark Round 1
+
+- Fixed synthetic request format issues:
+  - Assistant tool-call messages now use `content: ""` instead of `content: null`.
+  - Synthetic system/context content is merged into a single leading system message because the current runtime rejects multiple system messages with `System message must be at the beginning`.
+  - Added `--max-prompt-tokens` to cap generated prompt sizes under the current server `--max-model-len 100000`.
+- Generated dataset:
+  - `/root/work/logs/synthetic_agent_perf_16_fixed2_cap90k.jsonl`
+  - 16 requests, approximate average prompt tokens `25712`, approximate intended cache ratio `0.637`.
+- Benchmark:
+  - Command output file: `/root/work/logs/perf_synth16_fixed2_cap90k_c1_r4.json`.
+  - Concurrency `1`, requests `4`, max output tokens `64`.
+- Results:
+  - Success rate: `100%` (`4/4`).
+  - Prompt tokens total: `166080`.
+  - Completion tokens total: `256`.
+  - Cached tokens total: `96`.
+  - Actual cache hit rate: `0.058%`, far below official-like target; generator does not yet create exact cumulative session-prefix reuse.
+  - TTFT P90: `129.07s`.
+  - Output TPS P10 per request: `5.23 tok/s`.
+  - Aggregate output TPS: `0.607 tok/s`.
+  - Aggregate uncached input TPS: `393.52 tok/s`.
+  - Weighted token throughput: `1111.78`, far below `8000`.
+- Interpretation:
+  - Long-context uncached prefill is the immediate bottleneck.
+  - Current synthetic generator must be improved to make same-session adjacent requests share exact prefix, otherwise cache behavior is not representative of the official 65.6% cached-token workload.
+  - Current runtime also needs template compatibility work for multiple/irregular system messages if official requests include them unmodified.
diff --git a/worklogs/formal_perf_bench.py b/worklogs/formal_perf_bench.py
new file mode 100644
index 0000000..dc4f631
--- /dev/null
+++ b/worklogs/formal_perf_bench.py
@@ -0,0 +1,255 @@
+import argparse
+import concurrent.futures
+import json
+import math
+import statistics
+import time
+import urllib.error
+import urllib.request
+from dataclasses import dataclass, asdict
+from datetime import datetime
+from pathlib import Path
+
+
+@dataclass
+class RequestResult:
+    ok: bool
+    elapsed_sec: float
+    ttft_sec: float | None
+    prompt_tokens: int
+    cached_tokens: int
+    completion_tokens: int
+    reasoning_tokens: int
+    output_tps: float | None
+    error: str | None = None
+
+
+def percentile(values, pct):
+    if not values:
+        return None
+    values = sorted(values)
+    if len(values) == 1:
+        return values[0]
+    pos = (len(values) - 1) * pct / 100.0
+    lo = math.floor(pos)
+    hi = math.ceil(pos)
+    if lo == hi:
+        return values[lo]
+    return values[lo] * (hi - pos) + values[hi] * (pos - lo)
+
+
+def load_dataset(path):
+    text = Path(path).read_text(encoding="utf-8")
+    if path.endswith(".jsonl"):
+        data = [json.loads(line) for line in text.splitlines() if line.strip()]
+    else:
+        data = json.loads(text)
+
+    # Official-like format: each item is already an OpenAI chat completion
+    # request with messages/tools/tool_choice/stream/etc.
+    if data and isinstance(data[0], dict) and "messages" in data[0]:
+        return data
+
+    requests = []
+    for item in data:
+        system_prompt = item.get("system_prompt") or "You are a helpful assistant."
+        history = [{"role": "system", "content": system_prompt}]
+        for question in item.get("user_questions", []):
+            messages = history + [{"role": "user", "content": question}]
+            requests.append({"messages": messages})
+            # Synthetic assistant placeholder keeps later prompts multi-turn.
+            history = messages + [{"role": "assistant", "content": "好的，我们继续。"}]
+    return requests
+
+
+def post_stream(url, model, request_item, max_tokens, timeout):
+    allowed_fields = {
+        "messages",
+        "tools",
+        "tool_choice",
+        "model",
+        "max_tokens",
+        "temperature",
+        "top_p",
+        "stop",
+        "presence_penalty",
+        "frequency_penalty",
+        "n",
+        "response_format",
+    }
+    payload = {k: v for k, v in dict(request_item).items() if k in allowed_fields}
+    payload["model"] = payload.get("model") or model
+    # The benchmark CLI controls output length, even if the synthetic dataset
+    # stores a larger official-like max_tokens value.
+    payload["max_tokens"] = max_tokens
+    payload["temperature"] = payload.get("temperature", 0)
+    payload["stream"] = True
+    payload["stream_options"] = {"include_usage": True}
+    req = urllib.request.Request(
+        url.rstrip("/") + "/v1/chat/completions",
+        data=json.dumps(payload, ensure_ascii=False).encode("utf-8"),
+        headers={"Content-Type": "application/json"},
+        method="POST",
+    )
+
+    start = time.perf_counter()
+    first_token_at = None
+    usage = {}
+    try:
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            for raw in resp:
+                line = raw.decode("utf-8", errors="replace").strip()
+                if not line or not line.startswith("data:"):
+                    continue
+                data = line[5:].strip()
+                if data == "[DONE]":
+                    break
+                obj = json.loads(data)
+                if obj.get("usage"):
+                    usage = obj["usage"]
+                for choice in obj.get("choices") or []:
+                    delta = choice.get("delta") or {}
+                    text = delta.get("content") or delta.get("reasoning_content") or ""
+                    if text and first_token_at is None:
+                        first_token_at = time.perf_counter()
+        elapsed = time.perf_counter() - start
+        ttft = first_token_at - start if first_token_at else None
+        completion_tokens = int(usage.get("completion_tokens") or 0)
+        prompt_tokens = int(usage.get("prompt_tokens") or 0)
+        reasoning_tokens = int(usage.get("reasoning_tokens") or 0)
+        details = usage.get("prompt_tokens_details") or {}
+        cached_tokens = int(details.get("cached_tokens") or 0)
+        decode_sec = elapsed - ttft if ttft is not None else elapsed
+        output_tps = completion_tokens / decode_sec if completion_tokens and decode_sec > 0 else None
+        return RequestResult(
+            ok=True,
+            elapsed_sec=elapsed,
+            ttft_sec=ttft,
+            prompt_tokens=prompt_tokens,
+            cached_tokens=cached_tokens,
+            completion_tokens=completion_tokens,
+            reasoning_tokens=reasoning_tokens,
+            output_tps=output_tps,
+        )
+    except urllib.error.HTTPError as exc:
+        body = ""
+        try:
+            body = exc.read().decode("utf-8", errors="replace")
+        except Exception:
+            pass
+        elapsed = time.perf_counter() - start
+        return RequestResult(
+            ok=False,
+            elapsed_sec=elapsed,
+            ttft_sec=None,
+            prompt_tokens=0,
+            cached_tokens=0,
+            completion_tokens=0,
+            reasoning_tokens=0,
+            output_tps=None,
+            error=f"HTTPError {exc.code}: {body[:1000]}",
+        )
+    except Exception as exc:
+        elapsed = time.perf_counter() - start
+        return RequestResult(
+            ok=False,
+            elapsed_sec=elapsed,
+            ttft_sec=None,
+            prompt_tokens=0,
+            cached_tokens=0,
+            completion_tokens=0,
+            reasoning_tokens=0,
+            output_tps=None,
+            error=f"{type(exc).__name__}: {exc}",
+        )
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--url", default="http://127.0.0.1:1111")
+    parser.add_argument("--model", default="llm")
+    parser.add_argument("--dataset", required=True)
+    parser.add_argument("--concurrency", type=int, default=1)
+    parser.add_argument("--max-requests", type=int, default=20)
+    parser.add_argument("--max-tokens", type=int, default=128)
+    parser.add_argument("--timeout", type=int, default=600)
+    parser.add_argument("--out", default="/root/work/logs/formal_perf_result.json")
+    args = parser.parse_args()
+
+    reqs = load_dataset(args.dataset)
+    if not reqs:
+        raise SystemExit("empty dataset")
+    scheduled = [reqs[i % len(reqs)] for i in range(args.max_requests)]
+
+    wall_start = time.perf_counter()
+    results = []
+    with concurrent.futures.ThreadPoolExecutor(max_workers=args.concurrency) as pool:
+        futures = [
+            pool.submit(post_stream, args.url, args.model, request_item, args.max_tokens, args.timeout)
+            for request_item in scheduled
+        ]
+        for fut in concurrent.futures.as_completed(futures):
+            results.append(fut.result())
+            print(f"done {len(results)}/{len(scheduled)} ok={results[-1].ok}", flush=True)
+    wall_sec = time.perf_counter() - wall_start
+
+    ok_results = [r for r in results if r.ok]
+    success_rate = len(ok_results) / len(results) if results else 0.0
+    ttfts = [r.ttft_sec for r in ok_results if r.ttft_sec is not None]
+    output_tps_values = [r.output_tps for r in ok_results if r.output_tps is not None]
+
+    prompt_tokens = sum(r.prompt_tokens for r in ok_results)
+    cached_tokens = sum(r.cached_tokens for r in ok_results)
+    input_tokens_uncached = max(prompt_tokens - cached_tokens, 0)
+    output_tokens = sum(r.completion_tokens for r in ok_results)
+
+    aggregate_output_tps = output_tokens / wall_sec if wall_sec > 0 else 0.0
+    aggregate_input_tps = input_tokens_uncached / wall_sec if wall_sec > 0 else 0.0
+    aggregate_cache_tps = cached_tokens / wall_sec if wall_sec > 0 else 0.0
+    weighted = aggregate_output_tps * 16.796 + aggregate_input_tps * 2.799 + aggregate_cache_tps * 0.56
+    cache_hit_rate = cached_tokens / prompt_tokens if prompt_tokens else 0.0
+
+    summary = {
+        "created_at": datetime.now().isoformat(timespec="seconds"),
+        "url": args.url,
+        "model": args.model,
+        "dataset": args.dataset,
+        "concurrency": args.concurrency,
+        "max_requests": args.max_requests,
+        "max_tokens": args.max_tokens,
+        "wall_sec": wall_sec,
+        "success_rate": success_rate,
+        "ttft_p90_sec": percentile(ttfts, 90),
+        "output_tps_p10_per_request": percentile(output_tps_values, 10),
+        "aggregate_output_tps": aggregate_output_tps,
+        "aggregate_input_tps_uncached": aggregate_input_tps,
+        "aggregate_cache_tps": aggregate_cache_tps,
+        "cache_hit_rate": cache_hit_rate,
+        "weighted_token_throughput": weighted,
+        "totals": {
+            "requests": len(results),
+            "success": len(ok_results),
+            "prompt_tokens": prompt_tokens,
+            "input_tokens_uncached": input_tokens_uncached,
+            "cached_tokens": cached_tokens,
+            "completion_tokens": output_tokens,
+            "reasoning_tokens": sum(r.reasoning_tokens for r in ok_results),
+        },
+        "targets": {
+            "output_tps_p10_per_request_gte_20": (percentile(output_tps_values, 10) or 0) >= 20,
+            "ttft_p90_lte_5": (percentile(ttfts, 90) or 999) <= 5,
+            "cache_hit_rate_gte_50pct": cache_hit_rate >= 0.5,
+            "success_rate_gte_99pct": success_rate >= 0.99,
+            "weighted_token_throughput_gte_8000": weighted >= 8000,
+        },
+        "results": [asdict(r) for r in results],
+    }
+
+    Path(args.out).parent.mkdir(parents=True, exist_ok=True)
+    Path(args.out).write_text(json.dumps(summary, ensure_ascii=False, indent=2), encoding="utf-8")
+    print(json.dumps({k: v for k, v in summary.items() if k != "results"}, ensure_ascii=False, indent=2))
+    print("RESULT_FILE", args.out)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/worklogs/generate_official_like_dataset.py b/worklogs/generate_official_like_dataset.py
new file mode 100644
index 0000000..5db4de9
--- /dev/null
+++ b/worklogs/generate_official_like_dataset.py
@@ -0,0 +1,296 @@
+import argparse
+import json
+import random
+from pathlib import Path
+
+
+TOOL_NAMES = [
+    "terminal_exec",
+    "process_list",
+    "read_file",
+    "edit_file",
+    "write_file",
+    "glob_files",
+    "grep_code",
+    "web_search",
+    "web_fetch",
+    "browser_open",
+    "mcp_codebase_memory_search",
+    "mcp_codebase_memory_graph",
+    "sessions_spawn",
+    "sessions_send",
+    "subagents_run",
+    "memory_search",
+]
+
+
+SYSTEM_PROMPT = """You are Codex, a coding agent. You help the user modify, test, and reason about code.
+Follow the available tools carefully. Prefer precise shell commands, inspect files before editing,
+and return concise progress updates. When tool calls are needed, produce valid tool call arguments."""
+
+
+COMMON_PROJECT_CONTEXT = """
+Repository: /home/user/workspace/project
+Task: implement features, inspect logs, run tests, patch code, summarize results.
+Conventions:
+- Use rg for search.
+- Avoid destructive commands.
+- Keep a worklog.
+- Preserve unrelated user changes.
+"""
+
+
+CODE_SNIPPET = """
+def fetch_prices(symbols, start_date, end_date, retries=3):
+    results = []
+    for symbol in symbols:
+        payload = {
+            "symbol": symbol,
+            "start": start_date,
+            "end": end_date,
+            "adjust": "qfq",
+        }
+        # retry network request with backoff
+        for attempt in range(retries):
+            try:
+                results.append(client.get("/prices", params=payload))
+                break
+            except TimeoutError:
+                time.sleep(1.5 * (attempt + 1))
+    return results
+"""
+
+
+def make_tool(name):
+    if "terminal" in name or "process" in name:
+        props = {"command": {"type": "string", "description": "Shell command to execute."}}
+        required = ["command"]
+    elif "read" in name:
+        props = {"path": {"type": "string"}}
+        required = ["path"]
+    elif "edit" in name:
+        props = {
+            "file_path": {"type": "string"},
+            "old_string": {"type": "string"},
+            "new_string": {"type": "string"},
+        }
+        required = ["file_path", "old_string", "new_string"]
+    elif "web" in name:
+        props = {"query": {"type": "string"}, "count": {"type": "integer"}}
+        required = ["query"]
+    else:
+        props = {"input": {"type": "string"}, "limit": {"type": "integer"}}
+        required = ["input"]
+    return {
+        "type": "function",
+        "function": {
+            "name": name,
+            "description": f"Synthetic benchmark tool: {name}",
+            "parameters": {
+                "type": "object",
+                "properties": props,
+                "required": required,
+            },
+        },
+    }
+
+
+def rand_text(rng, approx_tokens):
+    # Mostly ASCII/code-like text keeps the chars/token ratio simple enough for
+    # synthetic load generation. This is approximate; server usage is authoritative.
+    unit = (
+        "tool output line: status=ok path=/home/user/project/src/module.py "
+        "grep result includes function names, stack frames, JSON fields, "
+        "中文说明：这里包含工具返回、日志片段、代码上下文和转义字符。 "
+        + CODE_SNIPPET.replace("\n", "\\n")
+        + "\n"
+    )
+    chars = max(1, approx_tokens * 4)
+    return (unit * (chars // len(unit) + 1))[:chars]
+
+
+def make_tool_call(rng, idx):
+    name = rng.choice(TOOL_NAMES)
+    if "terminal" in name:
+        args = {
+            "command": "cd /home/user/project && rg -n \"TODO|FIXME|error\" src tests 2>/dev/null | head -50"
+        }
+    elif "read" in name:
+        args = {"path": "/home/user/workspace/.agents/skills/data-analysis/SKILL.md"}
+    elif "edit" in name:
+        args = {
+            "file_path": "/home/user/scripts/fetch_prices.py",
+            "old_string": "# 全局频率限制器实例\n_rate_limiter = RateLimiter(min_interval=2.0, max_per_minute=20)",
+            "new_string": "# 全局频率限制器实例 - 根据配额上调\n_rate_limiter = RateLimiter(min_interval=1.5, max_per_minute=40)",
+        }
+    elif "web" in name:
+        args = {"query": "深圳 周末活动 推荐 2026年7月", "count": 10}
+    else:
+        args = {"input": "inspect project memory and summarize related files", "limit": 20}
+    return {
+        "id": f"call_{idx}_{rng.randrange(10**8)}",
+        "type": "function",
+        "function": {"name": name, "arguments": json.dumps(args, ensure_ascii=False)},
+    }
+
+
+def target_prompt_tokens(rng, max_prompt_tokens=None):
+    # Approximate official distribution: many 4K-16K, meaningful 32K+ tail,
+    # small number of 128K+ requests.
+    x = rng.random()
+    if x < 0.08:
+        value = rng.randint(1000, 4000)
+        return min(value, max_prompt_tokens) if max_prompt_tokens else value
+    if x < 0.50:
+        value = rng.randint(4000, 16000)
+        return min(value, max_prompt_tokens) if max_prompt_tokens else value
+    if x < 0.74:
+        value = rng.randint(16000, 32000)
+        return min(value, max_prompt_tokens) if max_prompt_tokens else value
+    if x < 0.95:
+        high = min(128000, max_prompt_tokens) if max_prompt_tokens else 128000
+        value = rng.randint(32000, max(32000, high))
+        return value
+    high = min(235000, max_prompt_tokens) if max_prompt_tokens else 235000
+    return rng.randint(128000, max(128000, high)) if high >= 128000 else high
+
+
+def target_output_tokens(rng):
+    x = rng.random()
+    if x < 0.66:
+        return rng.randint(64, 256)
+    if x < 0.96:
+        return rng.randint(256, 2000)
+    return rng.randint(2000, 8192)
+
+
+def build_request(rng, req_id, session_id, shared_prefix_tokens, total_prompt_tokens):
+    tool_count = max(1, int(rng.lognormvariate(2.8, 0.8)))
+    tool_count = min(tool_count, 92)
+    tools = [make_tool(rng.choice(TOOL_NAMES)) for _ in range(tool_count)]
+
+    messages = [
+        {
+            "role": "system",
+            "content": SYSTEM_PROMPT + "\n" + COMMON_PROJECT_CONTEXT + rand_text(rng, shared_prefix_tokens),
+        },
+    ]
+
+    remaining = max(512, total_prompt_tokens - shared_prefix_tokens)
+    turns = rng.randint(12, 65)
+    per_turn = max(32, remaining // max(1, turns))
+    call_idx = 0
+    for turn in range(turns):
+        messages.append({
+            "role": "user",
+            "content": (
+                f"Session {session_id}, turn {turn}: inspect the codebase and continue the task. "
+                "Need exact commands, possible edits, and concise next action."
+            ),
+        })
+        if rng.random() < 0.78:
+            calls = [make_tool_call(rng, call_idx)]
+            call_idx += 1
+            messages.append({
+                "role": "assistant",
+                # This runtime validates that every message has content or
+                # reasoning_content, even when assistant emits tool_calls.
+                "content": "",
+                "tool_calls": calls,
+            })
+            tool_len = int(per_turn * rng.uniform(0.5, 2.0))
+            if rng.random() > 0.99:
+                tool_len = rng.randint(12000, 50000)
+            elif rng.random() > 0.90:
+                tool_len = rng.randint(2500, 16000)
+            messages.append({
+                "role": "tool",
+                "tool_call_id": calls[0]["id"],
+                "name": calls[0]["function"]["name"],
+                "content": rand_text(rng, tool_len),
+            })
+        else:
+            messages.append({
+                "role": "assistant",
+                "content": "我会先检查相关文件和日志，再给出下一步修改建议。",
+            })
+
+    messages.append({
+        "role": "user",
+        "content": "基于以上工具结果，继续完成当前任务。需要时发起下一步 tool_call。",
+    })
+
+    return {
+        "id": f"synthetic_{req_id:04d}",
+        "model": "llm",
+        "messages": messages,
+        "tools": tools,
+        "tool_choice": "auto",
+        "stream": True,
+        "temperature": 0,
+        "max_tokens": target_output_tokens(rng),
+        "metadata": {
+            "session_id": session_id,
+            "target_prompt_tokens_approx": total_prompt_tokens,
+            "shared_prefix_tokens_approx": shared_prefix_tokens,
+        },
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--out", required=True)
+    parser.add_argument("--requests", type=int, default=881)
+    parser.add_argument("--sessions", type=int, default=768)
+    parser.add_argument("--seed", type=int, default=20260714)
+    parser.add_argument("--jsonl", action="store_true")
+    parser.add_argument(
+        "--max-prompt-tokens",
+        type=int,
+        default=None,
+        help="Cap synthetic prompt-token target. Use 90000 for a server started with --max-model-len 100000.",
+    )
+    args = parser.parse_args()
+
+    rng = random.Random(args.seed)
+    session_ids = [f"sess_{i:04d}" for i in range(args.sessions)]
+    # Bias toward near-neighbor session reuse by reusing some recent sessions.
+    requests = []
+    recent = []
+    for i in range(args.requests):
+        if recent and rng.random() < 0.22:
+            session_id = rng.choice(recent[-32:])
+        else:
+            session_id = session_ids[i % len(session_ids)]
+        recent.append(session_id)
+        total = target_prompt_tokens(rng, args.max_prompt_tokens)
+        # Official weighted cache hit is about 65.6%; request-level varies.
+        if total >= 128000:
+            cache_ratio = rng.uniform(0.1, 0.35)
+        elif total >= 16000:
+            cache_ratio = rng.uniform(0.55, 0.72)
+        else:
+            cache_ratio = rng.uniform(0.75, 0.92)
+        shared = int(total * cache_ratio)
+        requests.append(build_request(rng, i, session_id, shared, total))
+
+    out = Path(args.out)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    if args.jsonl:
+        out.write_text("\n".join(json.dumps(x, ensure_ascii=False) for x in requests) + "\n", encoding="utf-8")
+    else:
+        out.write_text(json.dumps(requests, ensure_ascii=False, indent=2), encoding="utf-8")
+
+    total_prompt = sum(x["metadata"]["target_prompt_tokens_approx"] for x in requests)
+    total_shared = sum(x["metadata"]["shared_prefix_tokens_approx"] for x in requests)
+    print(json.dumps({
+        "out": str(out),
+        "requests": len(requests),
+        "sessions": len({x["metadata"]["session_id"] for x in requests}),
+        "approx_prompt_tokens_avg": total_prompt / len(requests),
+        "approx_cache_ratio_weighted": total_shared / total_prompt,
+    }, ensure_ascii=False, indent=2))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/worklogs/remote_smoke_bench.py b/worklogs/remote_smoke_bench.py
new file mode 100644
index 0000000..e53315c
--- /dev/null
+++ b/worklogs/remote_smoke_bench.py
@@ -0,0 +1,135 @@
+import json
+import statistics
+import sys
+import time
+import urllib.request
+from datetime import datetime
+from pathlib import Path
+
+
+BASE_URL = sys.argv[1] if len(sys.argv) > 1 else "http://127.0.0.1:1111"
+OUT_DIR = Path(sys.argv[2]) if len(sys.argv) > 2 else Path("/root/work/logs")
+MODEL = sys.argv[3] if len(sys.argv) > 3 else "llm"
+
+
+def post_json(path, payload, timeout=300):
+    req = urllib.request.Request(
+        BASE_URL + path,
+        data=json.dumps(payload, ensure_ascii=False).encode("utf-8"),
+        headers={"Content-Type": "application/json"},
+        method="POST",
+    )
+    with urllib.request.urlopen(req, timeout=timeout) as resp:
+        return resp.status, resp.read().decode("utf-8")
+
+
+def health():
+    with urllib.request.urlopen(BASE_URL + "/health", timeout=10) as resp:
+        return resp.status
+
+
+def chat_once(prompt, max_tokens=32):
+    payload = {
+        "model": MODEL,
+        "messages": [{"role": "user", "content": prompt}],
+        "max_tokens": max_tokens,
+        "temperature": 0,
+    }
+    start = time.perf_counter()
+    status, body = post_json("/v1/chat/completions", payload)
+    elapsed = time.perf_counter() - start
+    obj = json.loads(body)
+    usage = obj.get("usage") or {}
+    completion_tokens = usage.get("completion_tokens") or 0
+    return {
+        "status": status,
+        "elapsed_sec": elapsed,
+        "usage": usage,
+        "output_tps": completion_tokens / elapsed if elapsed and completion_tokens else None,
+        "content_preview": (obj["choices"][0]["message"].get("content") or "")[:200],
+    }
+
+
+def chat_stream(prompt, max_tokens=64):
+    payload = {
+        "model": MODEL,
+        "messages": [{"role": "user", "content": prompt}],
+        "max_tokens": max_tokens,
+        "temperature": 0,
+        "stream": True,
+        "stream_options": {"include_usage": True},
+    }
+    req = urllib.request.Request(
+        BASE_URL + "/v1/chat/completions",
+        data=json.dumps(payload, ensure_ascii=False).encode("utf-8"),
+        headers={"Content-Type": "application/json"},
+        method="POST",
+    )
+    start = time.perf_counter()
+    first_token_at = None
+    usage = None
+    pieces = []
+    with urllib.request.urlopen(req, timeout=300) as resp:
+        for raw in resp:
+            line = raw.decode("utf-8", errors="replace").strip()
+            if not line or not line.startswith("data:"):
+                continue
+            data = line[5:].strip()
+            if data == "[DONE]":
+                break
+            obj = json.loads(data)
+            if obj.get("usage"):
+                usage = obj["usage"]
+            choices = obj.get("choices") or []
+            if choices:
+                delta = choices[0].get("delta") or {}
+                text = delta.get("content") or delta.get("reasoning_content") or ""
+                if text:
+                    if first_token_at is None:
+                        first_token_at = time.perf_counter()
+                    pieces.append(text)
+    elapsed = time.perf_counter() - start
+    completion_tokens = (usage or {}).get("completion_tokens") or 0
+    gen_time = elapsed - (first_token_at - start) if first_token_at else elapsed
+    return {
+        "elapsed_sec": elapsed,
+        "ttft_sec": (first_token_at - start) if first_token_at else None,
+        "usage": usage,
+        "output_tps_after_ttft": completion_tokens / gen_time if gen_time and completion_tokens else None,
+        "content_preview": "".join(pieces)[:200],
+    }
+
+
+def main():
+    OUT_DIR.mkdir(parents=True, exist_ok=True)
+    results = {
+        "base_url": BASE_URL,
+        "model": MODEL,
+        "created_at": datetime.now().isoformat(timespec="seconds"),
+    }
+    results["health_status"] = health()
+    results["smoke_nonstream"] = chat_once("你好，请用一句话介绍你自己。", max_tokens=16)
+
+    short_prompt = "请用中文简要说明什么是模型推理服务。"
+    stream_runs = [chat_stream(short_prompt, max_tokens=64) for _ in range(3)]
+    results["stream_runs"] = stream_runs
+    ttfts = [x["ttft_sec"] for x in stream_runs if x["ttft_sec"] is not None]
+    tps = [x["output_tps_after_ttft"] for x in stream_runs if x["output_tps_after_ttft"] is not None]
+    results["stream_summary"] = {
+        "ttft_avg_sec": statistics.mean(ttfts) if ttfts else None,
+        "ttft_p90_sec": sorted(ttfts)[int(0.9 * (len(ttfts) - 1))] if ttfts else None,
+        "output_tps_avg": statistics.mean(tps) if tps else None,
+    }
+
+    repeated = "以下是一段用于测试前缀缓存的公共上下文：" + ("模型部署竞赛关注吞吐、延迟、缓存命中和稳定性。" * 80)
+    results["prefix_cache_probe_1"] = chat_once(repeated + "\n请总结一句话。", max_tokens=16)
+    results["prefix_cache_probe_2"] = chat_once(repeated + "\n请换一种说法总结一句话。", max_tokens=16)
+
+    out = OUT_DIR / f"baseline_smoke_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json"
+    out.write_text(json.dumps(results, ensure_ascii=False, indent=2), encoding="utf-8")
+    print(json.dumps(results, ensure_ascii=False, indent=2), flush=True)
+    print("RESULT_FILE", out, flush=True)
+
+
+if __name__ == "__main__":
+    main()