Persist full GSI payloads to per-match JSONL for later analysis.

Default-on dump keeps every POST body under samples/raw/<matchid>/gsi.jsonl, with CLI overrides to disable when disk use is a concern. Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-26 12:16:47 +08:00
parent 2fb298ebf1
commit 7dc3e515e9
7 changed files with 61 additions and 4 deletions
@@ -30,7 +30,7 @@ Windows 上的 Dota 2 **天梯选将识别**工具：用 GSI 感知选将阶段

 运行时产物（**勿提交**，见 `.gitignore`）：

- `samples/raw/<matchid>/` — GSI 会话截图；手动 `capture.py` 可写在 `raw/` 根下
+- `samples/raw/<matchid>/` — GSI 会话截图与 `gsi.jsonl`；手动 `capture.py` 可写在 `raw/` 根下
 - `preview/` — 标定 / sheet 预览
 - `results/` — 每局 JSON
 - `failures/` — `--truth` 调试用错识裁切
@@ -39,6 +39,7 @@ Windows 上的 Dota 2 **天梯选将识别**工具：用 GSI 感知选将阶段

 ```
 Dota 2 GSI → gsi_watch.py (:3223)
+              → 全量 payload → samples/raw/<matchid>/gsi.jsonl（可关）
              → DraftSession 轮询截屏 → samples/raw/<matchid>/
              → recognize_image (CDN 模板 + 可选天梯遮罩)
              → roles / grid / modes（辅助）
@@ -4,6 +4,11 @@

 ## [Unreleased]

+### Added
+
+- GSI 全量落盘：默认把每包 POST 追加到 `samples/raw/<matchid>/gsi.jsonl`
+ （`gsi.dump_payloads` / `--dump-gsi` / `--no-dump-gsi`）。
+
 ### Changed

 - 项目定名：**上分帝** / **Climperor**（目录 `climperor`），由 `dota2-draft-vision` 迁出独立仓库。
@@ -76,6 +76,12 @@ Valve 已明确关闭普通玩家视角的实时 draft
 GSI 会话写入 `samples/raw/{matchid}/`；手动 `capture.py` 仍写 `samples/raw/` 根目录。
 结果 JSON 仍为 `results/draft_<时间戳>.json`（内含 `match_id`）。

+**⑥ GSI 全量落盘**
+
+`gsi.dump_payloads`（默认开）把每包 POST body 追加到 `samples/raw/{matchid}/gsi.jsonl`，
+行格式 `{"t": <unix>, "payload": <原文>}`。内容上限仍是 cfg 订阅字段 + 普通玩家视角；
+不能靠落盘补出双方 pick。CLI：`--dump-gsi` / `--no-dump-gsi`。
+
 ---

 ## 3. 关键实测（2026-07-25）
@@ -174,6 +180,7 @@ climperor/
 | `gsi.revise_gain` | 选人阶段改判所需分数增益 | 0.15 |
 | `gsi.strategy_tail_polls` | 决策阶段继续视觉轮询 | 8 |
 | `gsi.strategy_gsi_wait` | 视觉结束后等待本人 GSI 英雄 | 3.0 |
+| `gsi.dump_payloads` | 全量 GSI JSONL 落盘 | true |

 ---

@@ -182,6 +189,7 @@ climperor/
 ```
 Dota 2 (-gamestateintegration)
  → POST → gsi_watch.py :3223
+  → samples/raw/{matchid}/gsi.jsonl（可选）
  → DraftSession 轮询 recognize + roles/grid/modes
  → results/draft_*.json
 ```
@@ -34,7 +34,7 @@ python gsi_watch.py
 ```

 进入英雄选择后自动跟踪；凑齐 10 人或超时后写出 `results/draft_<时间戳>.json`（该目录已 gitignore）。
-会话截图按对局落在 `samples/raw/<matchid>/`。
+会话截图按对局落在 `samples/raw/<matchid>/`；默认同时把每包 GSI JSON 追加到同目录 `gsi.jsonl`（`--no-dump-gsi` 可关）。

 > 主菜单收不到 GSI 是正常的：客户端**第一次载入对局后**才开始推送。看到 `[gsi] connected` 才算链路通。

@@ -10,6 +10,7 @@ Notes:
    - Frames are saved as PNG at native resolution, named cap_HHMMSS.png.
 """

+import json
 import sys
 import time
 from pathlib import Path
@@ -19,6 +20,7 @@ import mss
 import numpy as np

 RAW_DIR = Path(__file__).parent / "samples" / "raw"
+GSI_JSONL = "gsi.jsonl"


 def raw_dir_for_match(match_id: str | None = None) -> Path:
@@ -32,6 +34,20 @@ def raw_dir_for_match(match_id: str | None = None) -> Path:
    return RAW_DIR / safe


+def append_gsi_payload(match_id: str | None, payload: dict) -> Path:
+    """Append one full GSI POST body to samples/raw/{match_id}/gsi.jsonl.
+
+    Each line is {"t": <unix seconds>, "payload": <original body>}.
+    """
+    out = raw_dir_for_match(match_id)
+    out.mkdir(parents=True, exist_ok=True)
+    path = out / GSI_JSONL
+    record = {"t": time.time(), "payload": payload}
+    with path.open("a", encoding="utf-8") as f:
+        f.write(json.dumps(record, ensure_ascii=False) + "\n")
+    return path
+
+
 def grab_frame(sct=None) -> np.ndarray:
    """Grab the primary monitor as a BGR image."""
    if sct is None:
@@ -119,7 +119,8 @@
    "strategy_tail_polls": 8,
    "strategy_gsi_wait": 3.0,
    "capture_interval": 1.0,
-    "target_slots": 10
+    "target_slots": 10,
+    "dump_payloads": true
  },
  "calibrated_from": "draft_141704.png"
 }
@@ -6,6 +6,8 @@ Usage:
    python gsi_watch.py --once            # capture+recognize right now, no GSI
    python gsi_watch.py --port 3223
    python gsi_watch.py --states HERO_SELECTION,STRATEGY_TIME
+    python gsi_watch.py --dump-gsi        # force full GSI JSONL logging
+    python gsi_watch.py --no-dump-gsi     # disable it

 When the game enters hero selection the watcher follows the whole draft,
 polling the screen and logging each pick as it is revealed (All Pick reveals
@@ -14,6 +16,9 @@ which lane role you queued for. The result lands in results/draft_<ts>.json.

 If config.json has no calibrated slots yet the watcher runs in capture-only
 mode: it still saves frames to samples/raw/<matchid>/ so you can calibrate from them.
+
+With gsi.dump_payloads (default on), every POST body is appended to
+samples/raw/<matchid>/gsi.jsonl for later analysis.
 """

 import json
@@ -25,7 +30,7 @@ from pathlib import Path

 import mss

-from capture import grab_frame, raw_dir_for_match, save_frame
+from capture import append_gsi_payload, grab_frame, raw_dir_for_match, save_frame
 from common import ROOT, load_config, load_template_library
 from draft_session import HERO_SELECTION, DraftSession, describe, gsi_slot, loc
 from recognize import recognize_image
@@ -52,6 +57,7 @@ class Watcher:
        gsi = cfg.get("gsi", {})
        self.trigger_states = set(gsi.get("trigger_states", ["DOTA_GAMERULES_STATE_STRATEGY_TIME"]))
        self.interval = gsi.get("capture_interval", 1.0)
+        self.dump_payloads = bool(gsi.get("dump_payloads", True))

        self.library = load_template_library() if calibrated else []
        self.last_state: str | None = None
@@ -59,7 +65,9 @@ class Watcher:
        self.handled_matches: set[str] = set()
        self.connected = False
        self.busy = threading.Lock()
+        self.dump_lock = threading.Lock()
        self.self_info: dict = {}
+        self._dump_announced: set[str] = set()

    def on_payload(self, payload: dict) -> None:
        if not self.connected:
@@ -79,6 +87,8 @@ class Watcher:
        state = (payload.get("map") or {}).get("game_state")
        match_id = str((payload.get("map") or {}).get("matchid") or "no-match")
        self.last_match_id = match_id
+        if self.dump_payloads:
+            self._dump_gsi(match_id, payload)
        if state is None:
            # main menu / no active match
            if self.last_state is not None:
@@ -108,6 +118,17 @@ class Watcher:
            return
        threading.Thread(target=self.track, args=(match_id, state, key), daemon=True).start()

+    def _dump_gsi(self, match_id: str, payload: dict) -> None:
+        """Persist the full POST body; announce the path once per match folder."""
+        try:
+            with self.dump_lock:
+                path = append_gsi_payload(match_id, payload)
+            if match_id not in self._dump_announced:
+                self._dump_announced.add(match_id)
+                print(f"[gsi] dumping payloads -> {path}", flush=True)
+        except OSError as e:
+            print(f"[gsi] dump failed: {e}", flush=True)
+
    def track(self, match_id: str, state: str, key: str) -> dict | None:
        """Follow the draft from here to the end, recording every reveal."""
        if not self.busy.acquire(blocking=False):
@@ -225,6 +246,10 @@ def main() -> None:
        print()

    watcher = Watcher(cfg, calibrated, verbose)
+    if "--dump-gsi" in args:
+        watcher.dump_payloads = True
+    if "--no-dump-gsi" in args:
+        watcher.dump_payloads = False

    if "--states" in args:
        names = args[args.index("--states") + 1].split(",")
@@ -251,6 +276,7 @@ def main() -> None:
        )
    print(f"listening on http://127.0.0.1:{port}/  (Ctrl+C to stop)")
    print(f"trigger states: {', '.join(sorted(watcher.trigger_states))}")
+    print(f"gsi dump   : {'on -> samples/raw/<matchid>/gsi.jsonl' if watcher.dump_payloads else 'off'}")
    if calibrated:
        print(f"template library: {len(watcher.library)} entries")
    print("waiting for Dota 2 ... (needs -gamestateintegration launch option)")