poker-cli 0.11.2__py3-none-any.whl

poker_cli/coach.py

@@ -0,0 +1,233 @@
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+import shutil
+import subprocess
+import tempfile
+
+from .models import HandResult
+from .usage import ModelUsageTracker, parse_claude_result, parse_codex_usage
+
+
+class CoachError(RuntimeError):
+    pass
+
+
+class LocalCLICoach:
+    """Run post-hand coaching through an already authenticated local agent CLI."""
+
+    def __init__(
+        self,
+        provider: str,
+        *,
+        timeout: int = 180,
+        codex_model: str | None = None,
+        claude_model: str | None = None,
+        usage_tracker: ModelUsageTracker | None = None,
+    ) -> None:
+        if provider not in {"auto", "codex", "claude"}:
+            raise ValueError(f"unknown coach provider: {provider}")
+        self.provider = self._resolve_provider(provider)
+        self.timeout = timeout
+        self.codex_model = codex_model or os.getenv(
+            "POKER_CODEX_MODEL", "gpt-5.4-mini"
+        )
+        self.claude_model = claude_model or os.getenv(
+            "POKER_CLAUDE_MODEL", "haiku"
+        )
+        self.usage_tracker = usage_tracker
+
+    @staticmethod
+    def _resolve_provider(provider: str) -> str:
+        if provider != "auto":
+            if shutil.which(provider) is None:
+                raise CoachError(f"找不到本地 {provider} 命令")
+            return provider
+        if shutil.which("codex"):
+            return "codex"
+        if shutil.which("claude"):
+            return "claude"
+        raise CoachError("找不到本地 codex 或 claude 命令")
+
+    def review(self, result: HandResult) -> str:
+        prompt = build_coach_prompt(result)
+        if self.provider == "codex":
+            return self._review_with_codex(prompt)
+        return self._review_with_claude(prompt)
+
+    def review_tournament(self, results: list[HandResult], outcome: str) -> str:
+        prompt = build_tournament_coach_prompt(results, outcome)
+        if self.provider == "codex":
+            return self._review_with_codex(prompt)
+        return self._review_with_claude(prompt)
+
+    def _review_with_codex(self, prompt: str) -> str:
+        with tempfile.TemporaryDirectory(prefix="poker-coach-") as temp_dir:
+            output_path = Path(temp_dir) / "review.txt"
+            command = [
+                "codex",
+                "exec",
+                "--ephemeral",
+                "--ignore-user-config",
+                "--sandbox",
+                "read-only",
+                "--skip-git-repo-check",
+                "--color",
+                "never",
+                "--json",
+                "--output-last-message",
+                str(output_path),
+            ]
+            if self.codex_model:
+                command.extend(("-m", self.codex_model))
+            command.extend(("-c", 'model_reasoning_effort="low"', "-"))
+            completed = self._run(command, prompt, cwd=temp_dir)
+            usage = parse_codex_usage(completed.stdout)
+            if usage is not None and self.usage_tracker is not None:
+                self.usage_tracker.record("codex", usage)
+            if not output_path.exists():
+                raise CoachError("Codex 没有生成复盘结果")
+            answer = output_path.read_text(encoding="utf-8").strip()
+            if not answer:
+                raise CoachError("Codex 返回了空复盘")
+            return answer
+
+    def _review_with_claude(self, prompt: str) -> str:
+        command = [
+            "claude",
+            "-p",
+            "--safe-mode",
+            "--tools",
+            "",
+            "--permission-mode",
+            "dontAsk",
+            "--no-session-persistence",
+            "--model",
+            self.claude_model,
+            "--effort",
+            "low",
+            "--system-prompt",
+            "Give concise, practical poker coaching. Do not use tools.",
+            "--output-format",
+            "json",
+        ]
+        completed = self._run(command, prompt)
+        answer, usage = parse_claude_result(completed.stdout)
+        if usage is not None and self.usage_tracker is not None:
+            self.usage_tracker.record("claude", usage)
+        if not answer:
+            raise CoachError("Claude Code 返回了空复盘")
+        return answer
+
+    def _run(
+        self,
+        command: list[str],
+        prompt: str,
+        *,
+        cwd: str | None = None,
+    ) -> subprocess.CompletedProcess[str]:
+        try:
+            completed = subprocess.run(
+                command,
+                input=prompt,
+                text=True,
+                capture_output=True,
+                timeout=self.timeout,
+                cwd=cwd,
+                check=False,
+            )
+        except subprocess.TimeoutExpired as error:
+            raise CoachError(f"{self.provider} 复盘超时") from error
+        except OSError as error:
+            raise CoachError(f"无法启动 {self.provider}: {error}") from error
+        if completed.returncode != 0:
+            detail = completed.stderr.strip().splitlines()
+            suffix = f"：{detail[-1]}" if detail else ""
+            raise CoachError(f"{self.provider} 复盘失败{suffix}")
+        return completed
+
+
+def build_coach_prompt(result: HandResult) -> str:
+    decisions = [
+        {
+            "street": decision.street.value,
+            "hero_hole_cards": list(decision.hole_cards),
+            "community_cards": list(decision.community_cards),
+            "pot_before_action": decision.pot_before,
+            "amount_to_call": decision.to_call,
+            "hero_action": decision.action.kind.value,
+            "hero_raise_to": decision.action.amount or None,
+            "estimated_equity_vs_random_range": round(decision.equity, 4),
+            "pot_odds": round(decision.pot_odds, 4),
+            "local_baseline": decision.recommendation,
+        }
+        for decision in result.hero_decisions
+    ]
+    packet = {
+        "hand_id": result.hand_id,
+        "board": [card.code() for card in result.board],
+        "result": result.summary,
+        "showdown": result.showdown,
+        "public_action_history": result.history,
+        "hero_decisions": decisions,
+    }
+    return (
+        "你是一名严谨的德州扑克教练。请根据下面这一手牌为玩家做中文复盘。\n"
+        "要求：\n"
+        "1. 先给一句结论，再按关键决策点分析；总长度控制在 500 字以内。\n"
+        "2. 区分底池赔率、范围推断、位置、下注尺度与结果导向，不能因为最后输赢倒推行动好坏。\n"
+        "3. 数据里的胜率是假设对手为随机范围的本地蒙特卡洛估计，只能作为参考。\n"
+        "4. 没有提供完整范围和求解器结果时，不得声称某个动作是精确 GTO；请说明建议的置信度。\n"
+        "5. 最后给一条下次可以直接执行的训练建议。\n"
+        "6. 不要调用工具、不要读取本地文件，只分析输入内容。\n\n"
+        f"牌局数据：\n{json.dumps(packet, ensure_ascii=False, indent=2)}"
+    )
+
+
+def build_tournament_coach_prompt(results: list[HandResult], outcome: str) -> str:
+    decisions = [
+        {
+            "hand_id": result.hand_id,
+            "street": decision.street.value,
+            "hero_hole_cards": list(decision.hole_cards),
+            "community_cards": list(decision.community_cards),
+            "pot_before_action": decision.pot_before,
+            "amount_to_call": decision.to_call,
+            "hero_action": decision.action.kind.value,
+            "hero_raise_to": decision.action.amount or None,
+            "estimated_equity_vs_random_ranges": round(decision.equity, 4),
+            "pot_odds": round(decision.pot_odds, 4),
+            "local_baseline": decision.recommendation,
+        }
+        for result in results
+        for decision in result.hero_decisions
+    ]
+    deviations = [
+        decision
+        for decision in decisions
+        if decision["hero_action"] != decision["local_baseline"]
+    ]
+    key_decisions = sorted(
+        deviations or decisions,
+        key=lambda item: item["pot_before_action"],
+        reverse=True,
+    )[:20]
+    packet = {
+        "outcome": outcome,
+        "hands_played": len(results),
+        "decision_count": len(decisions),
+        "key_decisions": key_decisions,
+    }
+    return (
+        "你是一名严谨的德州扑克锦标赛教练。请对玩家的整届单桌锦标赛做中文复盘。\n"
+        "要求：\n"
+        "1. 先总结三个贯穿整届比赛的倾向，再分析最多五个关键决策。\n"
+        "2. 讨论位置、有效筹码、底池赔率、范围与锦标赛阶段；避免结果导向。\n"
+        "3. 本地胜率基于随机多人范围，不是精确范围或 GTO/ICM 求解结果。\n"
+        "4. 没有充分证据时明确降低置信度，不要虚构对手手牌。\n"
+        "5. 最后给出三项下一届比赛可直接执行的训练目标；总长度 1000 字以内。\n"
+        "6. 不调用工具、不读取文件，只分析输入数据。\n\n"
+        f"锦标赛数据：\n{json.dumps(packet, ensure_ascii=False, indent=2)}"
+    )