aria-code 0.52.0__py3-none-any.whl

aria_code/apps/cli/intent_router.py

@@ -0,0 +1,276 @@
+"""Canonical lightweight intent routing for CLI services.
+
+This module keeps routing side-effect free.  It does not call tools, install
+packages, or fetch data; it only classifies a user message into stable service
+intents that other layers can reuse.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Iterable
+
+
+def _contains_any(text: str, keywords: Iterable[str]) -> bool:
+    return any(keyword in text for keyword in keywords)
+
+
+def _first_token(text: str) -> str:
+    stripped = text.strip()
+    if not stripped:
+        return ""
+    return stripped.split(maxsplit=1)[0].lower()
+
+
+def _add_unique(items: list[str], name: str) -> None:
+    if name not in items:
+        items.append(name)
+
+
+@dataclass(frozen=True)
+class IntentRoute:
+    message: str
+    primary: str
+    intents: tuple[str, ...]
+    services: tuple[str, ...]
+    explicit_code: bool = False
+    visual_artifact: bool = False
+    market_related: bool = False
+
+    @property
+    def wants_market_prefetch(self) -> bool:
+        return self.market_related and not self.visual_artifact and self.primary != "general"
+
+    @property
+    def allows_code_autorun(self) -> bool:
+        return self.primary in {"code", "strategy", "backtest"} and self.explicit_code
+
+
+COMMAND_INTENTS = {
+    "/chart": "chart",
+    "/dashboard": "dashboard",
+    "/report": "report",
+    "/team": "market_research",
+    "/analyze": "market_analysis",
+    "/ta": "market_analysis",
+    "/quote": "market_snapshot",
+    "/backtest": "backtest",
+    "/auto-strategy": "strategy",
+    "/strategy": "strategy",
+    "/ui": "ui_artifact",
+    "/vision": "vision",
+    "/upload-image": "vision",
+    "/screenshot": "screenshot",
+    "/browser": "browser",
+    "/file": "file_analysis",
+    "/github": "github",
+    "/mcp": "mcp",
+    "/cloud": "cloud",
+}
+
+
+def detect_intents(message: str) -> tuple[str, ...]:
+    low = message.lower().strip()
+    cmd = _first_token(low)
+    intents: list[str] = []
+
+    mapped = COMMAND_INTENTS.get(cmd)
+    if mapped:
+        _add_unique(intents, mapped)
+        # Arguments to /file are usually filenames.  Words such as “报告” or
+        # “分析” inside those paths describe the document, not a second market
+        # command, and must not pull in report/chart dependencies.
+        if cmd == "/file":
+            return tuple(intents)
+
+    if _contains_any(low, ("k线图", "k线", "k-line", "kline", "candlestick", "走势图", "图表", "chart", "plot")):
+        _add_unique(intents, "chart")
+    if _contains_any(low, ("看板", "晨报", "日报", "dashboard", "market board", "持仓看板")):
+        _add_unique(intents, "dashboard")
+    if _contains_any(low, ("研究报告", "财报", "报告", "report", "研报")):
+        _add_unique(intents, "report")
+    if _contains_any(low, ("回测", "backtest", "收益曲线", "最大回撤")):
+        _add_unique(intents, "backtest")
+    if _contains_any(low, ("策略代码", "写策略", "量化策略", "交易策略", "strategy code", "trading bot")):
+        _add_unique(intents, "strategy")
+    if _contains_any(low, ("技术分析", "技术指标", "rsi", "macd", "均线", "支撑", "阻力")):
+        _add_unique(intents, "market_analysis")
+    if _contains_any(low, ("股票", "行情", "持仓", "portfolio", "quote", "market data")):
+        _add_unique(intents, "market_snapshot")
+    if _contains_any(low, ("图片", "图像", "截图", "上传图片", "分析图片", "识别图片", "image", "screenshot")):
+        _add_unique(intents, "vision")
+    if _contains_any(low, ("浏览器", "网页截图", "打开网页", "browser", "playwright")):
+        _add_unique(intents, "browser")
+    if _contains_any(low, ("pdf", "docx", "word", "excel", "xlsx", "csv", "文件分析", "上传文件")):
+        _add_unique(intents, "file_analysis")
+    if _contains_any(low, ("github", "pull request", "pr ", "issue", "ci")):
+        _add_unique(intents, "github")
+    if _contains_any(low, ("mcp", "server", "tools", "skills")):
+        _add_unique(intents, "mcp")
+    if _contains_any(low, ("ollama", "本地模型", "local model")):
+        _add_unique(intents, "local_model")
+    if _contains_any(low, ("阿里云", "aliyun", "cloud service", "云端服务")):
+        _add_unique(intents, "cloud")
+    if _contains_any(low, ("搜索", "联网", "上网查", "web search", "google", "查一下", "搜一下", "最新消息")):
+        _add_unique(intents, "web_search")
+    if _contains_any(low, ("加密货币", "比特币", "以太坊", "bitcoin", "btc", "eth", "crypto", "币安", "binance", "okx", "资金费率", "funding rate")):
+        _add_unique(intents, "crypto")
+    if _contains_any(low, ("外汇", "汇率", "美元兑", "欧元兑", "eur/", "usd/", "jpy", "forex", "fx ")):
+        _add_unique(intents, "forex")
+    if _contains_any(low, ("期货", "商品", "黄金", "白银", "原油", "铜价", "天然气", "commodity", "futures")):
+        _add_unique(intents, "commodity")
+    if _contains_any(low, ("a股", "沪深", "上证", "深证", "创业板", "科创板")):
+        _add_unique(intents, "ashare")
+    if _contains_any(low, ("港股", "恒生", "h股", ".hk")):
+        _add_unique(intents, "hk_market")
+    if _contains_any(low, ("美股", "纳斯达克", "纽交所", "标普500", "道琼斯", "us stock")):
+        _add_unique(intents, "us_market")
+    if _contains_any(low, ("足球", "球赛", "比分预测", "世界杯", "欧洲杯", "英超", "西甲", "football", "soccer", "world cup", "premier league")):
+        _add_unique(intents, "sports")
+
+    if "file_analysis" in intents and "report" in intents and cmd != "/report":
+        intents = [intent for intent in intents if intent != "report"]
+
+    return tuple(intents)
+
+
+def _service_names(intents: tuple[str, ...]) -> tuple[str, ...]:
+    services: list[str] = []
+
+    def service(name: str) -> None:
+        _add_unique(services, name)
+
+    if any(i in intents for i in ("market_snapshot", "market_analysis", "chart", "dashboard", "report", "backtest", "strategy", "market_research", "ashare", "hk_market", "us_market", "forex", "commodity")):
+        service("market_data")
+    if "chart" in intents:
+        service("chart_renderer")
+    if "dashboard" in intents:
+        service("dashboard_generator")
+    if "report" in intents:
+        service("report_generator")
+    if "backtest" in intents or "strategy" in intents:
+        service("backtest_engine")
+    if "vision" in intents:
+        service("vision_input")
+    if "screenshot" in intents:
+        service("screenshot")
+    if "browser" in intents:
+        service("browser")
+    if "file_analysis" in intents:
+        service("file_parser")
+    if "github" in intents:
+        service("github_cli")
+    if "mcp" in intents:
+        service("mcp")
+    if "local_model" in intents:
+        service("local_llm")
+    if "cloud" in intents:
+        service("cloud_runtime")
+    if "web_search" in intents:
+        service("web_search")
+    if "crypto" in intents:
+        service("crypto_data")
+    if "forex" in intents:
+        service("forex_data")
+    if "commodity" in intents:
+        service("commodity_data")
+    if "ashare" in intents:
+        service("ashare_data")
+    if "hk_market" in intents:
+        service("hk_market_data")
+    if "us_market" in intents:
+        service("us_market_data")
+    if "sports" in intents:
+        service("sports_data")
+    return tuple(services)
+
+
+def build_intent_route(message: str) -> IntentRoute:
+    low = message.lower().strip()
+    intents = detect_intents(message)
+    conceptual_market_question = (
+        any(topic in low for topic in ("新闻面", "技术面", "基本面"))
+        and any(marker in low for marker in (
+            "有用吗", "是否有用", "区别", "关系", "怎么结合", "如何结合",
+            "你觉得", "为什么",
+        ))
+        and not any(marker in low for marker in (
+            "最新", "今天", "当前", "实时", "查一下", "搜一下", "搜索",
+            "/news", "/quote", "/analyze",
+        ))
+    )
+    explicit_code = _contains_any(low, (
+        "代码", "脚本", "python", "程序", "实现", "开发", "修改文件",
+        "写代码", "编写代码", "策略代码", "保存为.py", ".py",
+        "script", "code", "program", "implement", "edit file", "write file",
+    ))
+    try:
+        try:
+            from aria_code.intent_classifier import (
+                INTENT_ANALYSIS,
+                INTENT_CODING,
+                INTENT_FINANCE,
+                INTENT_GENERAL,
+                INTENT_REALTIME,
+                classify_intent_sync,
+                is_visual_market_artifact_request,
+            )
+        except ImportError:
+            from intent_classifier import (  # type: ignore[no-redef]
+                INTENT_ANALYSIS,
+                INTENT_CODING,
+                INTENT_FINANCE,
+                INTENT_GENERAL,
+                INTENT_REALTIME,
+                classify_intent_sync,
+                is_visual_market_artifact_request,
+            )
+        classifier_intent = classify_intent_sync(message)
+        visual_artifact = bool(is_visual_market_artifact_request(message))
+    except Exception:
+        INTENT_ANALYSIS = "analysis"
+        INTENT_CODING = "coding"
+        INTENT_FINANCE = "finance"
+        INTENT_GENERAL = "general"
+        INTENT_REALTIME = "realtime"
+        classifier_intent = INTENT_FINANCE
+        visual_artifact = any(i in intents for i in ("chart", "dashboard", "report", "ui_artifact"))
+
+    if conceptual_market_question:
+        # Methodology/capability questions need an explanation, not live market
+        # calls.  Treating the word "新闻面" as a news lookup caused the whole
+        # sentence to be forwarded to search providers as the query.
+        primary = "general"
+    elif intents:
+        primary = intents[0]
+    elif classifier_intent == INTENT_REALTIME:
+        primary = "market_snapshot"
+    elif classifier_intent == INTENT_ANALYSIS:
+        primary = "market_analysis"
+    elif classifier_intent == INTENT_CODING:
+        primary = "code"
+    elif classifier_intent == INTENT_GENERAL:
+        primary = "general"
+    elif classifier_intent == INTENT_FINANCE:
+        primary = "finance"
+    else:
+        primary = "finance"
+
+    market_related = not conceptual_market_question and (
+        any(i in intents for i in (
+            "market_snapshot", "market_analysis", "chart", "dashboard", "report",
+            "backtest", "strategy", "market_research", "ashare", "hk_market",
+            "us_market", "crypto", "forex", "commodity",
+        ))
+        or classifier_intent in {INTENT_ANALYSIS, INTENT_REALTIME, INTENT_FINANCE}
+    )
+
+    return IntentRoute(
+        message=message,
+        primary=primary,
+        intents=intents,
+        services=_service_names(intents),
+        explicit_code=explicit_code,
+        visual_artifact=visual_artifact,
+        market_related=market_related,
+    )

aria_code/apps/cli/intent_signals.py

@@ -0,0 +1,344 @@
+"""Signal extraction and intent precedence for message classification.
+
+Replaces a keyword-ordering scheme that had accumulated six layers of
+counter-patches (labelled "Bug ①"–"Bug ⑥" in intent_classifier.py) with no
+regression tests behind any of them.  That design failed in four recurring
+ways, all of which this module addresses directly:
+
+  - **Substring collisions.**  The coding keyword "repo" matched inside
+    "report_2024.docx", so asking about a document was classified as a code
+    task.  ASCII terms are now matched on word boundaries.
+  - **Spacing variants.**  The artifact keyword "k线" missed "K 线图", so a
+    chart request fell through to general.  CJK terms are now matched against
+    a whitespace-stripped form of the message.
+  - **Rules cancelling each other.**  A question word suppressed the coding
+    rule, so "这段代码为什么报错" — a debugging request — became a general
+    question.  Signals are now extracted independently and combined by one
+    explicit precedence table, rather than by the order of early returns.
+  - **Generic words as intent.**  Bare "分析"/"研究"/"评估" are ordinary
+    Chinese and were treated as market-analysis intent, held back only by an
+    entity allowlist.  A market intent now *requires* a concrete entity.
+
+The precedence table is the whole decision: read it top to bottom to know how
+any message is classified.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+
+INTENT_CODING = "coding"
+INTENT_ANALYSIS = "analysis"
+INTENT_REALTIME = "realtime"
+INTENT_GENERAL = "general"
+INTENT_FINANCE = "finance"
+
+_CJK = r"一-鿿"
+_ASCII_TERM_RE = re.compile(r"^[a-z0-9][a-z0-9 .+#/_-]*$")
+
+
+def _normalize(text: str) -> tuple[str, str]:
+    """Return (spaced, compact) lowercase forms of *text*.
+
+    ``spaced`` collapses whitespace runs and is used for boundary-anchored
+    ASCII matching.  ``compact`` removes whitespace entirely and is used for
+    CJK terms, so "K 线图" and "K线图" are the same message.
+    """
+    low = (text or "").strip().lower()
+    spaced = re.sub(r"\s+", " ", low)
+    compact = re.sub(r"\s+", "", low)
+    return spaced, compact
+
+
+def _matches(term: str, spaced: str, compact: str) -> bool:
+    """True if *term* occurs in the message under the right matching rule."""
+    term = term.strip().lower()
+    if not term:
+        return False
+    if _ASCII_TERM_RE.match(term):
+        # Word boundaries stop "repo" from matching inside "report".
+        return re.search(rf"(?<![a-z0-9]){re.escape(term)}(?![a-z0-9])", spaced) is not None
+    return re.sub(r"\s+", "", term) in compact
+
+
+def _any(terms, spaced: str, compact: str) -> bool:
+    return any(_matches(t, spaced, compact) for t in terms)
+
+
+# ── Term inventories ────────────────────────────────────────────────────────
+# Each list answers one question.  Terms live in exactly one list, so a change
+# has one obvious place to go.
+
+_CODE_TERMS = (
+    # Actions on code
+    "写代码", "编写代码", "重构", "调试", "报错", "修复", "改一下", "改成",
+    "实现", "开发", "部署", "上线", "打包", "编译", "跑一下", "运行一下",
+    "单元测试", "集成测试", "代码审查", "代码审核", "审核代码", "审查代码",
+    "refactor", "debug", "implement", "deploy", "compile", "lint",
+    "unit test", "integration test", "code review", "stack trace", "traceback",
+    "bug", "fix", "patch", "commit", "merge", "rebase", "pull request",
+    # Code nouns
+    "代码", "脚本", "函数", "方法", "类", "模块", "接口", "仓库", "代码库",
+    "前端", "后端", "数据库", "编译器", "依赖", "报错信息", "异常",
+    "script", "function", "module", "repo", "repository", "codebase",
+    "frontend", "backend", "database", "api", "sdk", "cli",
+    "python", "typescript", "javascript", "rust", "golang", "java",
+    "react", "vue", "next.js", "nextjs", "svelte", "swiftui", "flutter",
+    "fastapi", "django", "flask", "express", "pytest", "docker", "kubernetes",
+    # Build requests.  "写一个 AAPL 动量策略" is a request to produce code, even
+    # though the subject is a ticker — without these it fell through to the
+    # market branches and was answered as prose.
+    "写一个", "写个", "做一个", "搭一个", "建一个", "编写", "创建", "生成",
+    "策略", "交易策略", "量化策略", "回测", "选股",
+    "write a", "create a", "build a", "generate", "backtest", "strategy",
+    # How people actually report a defect.  The list above named tools and
+    # actions ("重构", "pytest", "traceback") but not the plain sentence a
+    # person writes when something is broken, so "这个项目的测试挂了，找出原因
+    # 并修好" classified as GENERAL — and the general prompt does not tell the
+    # model to act, so a 7B model with every tool available simply asked for
+    # more details and changed nothing.  An eval suite run caught this at 0/5.
+    #
+    # The bias here is deliberate: an unsure classification should fail toward
+    # capability. Sending "测试一下这个想法" down the coding path costs a longer
+    # system prompt; sending "测试挂了，修好" to general costs the entire task,
+    # silently.
+    "测试挂", "测试失败", "测试不过", "测试没过", "跑不通", "跑不过",
+    "修好", "修一下", "修掉", "改好", "没通过", "挂了",
+    "断言", "崩溃", "空指针", "越界", "校验", "入参", "边界条件", "打日志",
+    "failing test", "test fails", "tests fail", "test failure", "broken",
+    "crash", "assertion", "exception", "stacktrace", "regression",
+    "edge case", "off-by-one", "validation", "refactoring",
+)
+
+# Words that name a project/product artefact rather than code.  When these
+# appear without a code term, the task is documentation or product work.
+_PRODUCT_TERMS = (
+    "产品", "需求", "文档", "架构设计", "商业模式", "市场定位", "用户体验",
+    "运营", "品牌", "文案", "白皮书", "商业计划",
+)
+
+_ARTIFACT_TERMS = (
+    "图表", "走势图", "k线图", "k线", "k-line", "kline", "candlestick",
+    "蜡烛图", "热力图", "heatmap", "chart", "plot", "dashboard",
+    "看板", "晨报", "日报", "周报", "月报", "报表",
+)
+
+_MARKET_CONTEXT_TERMS = (
+    "股票", "股价", "行情", "市场", "美股", "港股", "a股", "指数", "大盘",
+    "持仓", "portfolio", "回测", "财报", "earnings", "基金", "etf",
+    "资产", "组合", "市值", "backtest", "market data", "stock",
+)
+
+_REALTIME_TERMS = (
+    "今天", "今日", "现在", "当前", "此刻", "最新", "最近", "实时",
+    "多少钱", "多少点", "是多少", "什么价", "现价", "报价", "涨跌", "涨幅",
+    "新闻", "消息", "快讯", "今天的", "盘中",
+    "today", "now", "current", "latest", "price", "quote", "news",
+    "market cap", "how much",
+)
+
+_ANALYSIS_TERMS = (
+    "分析", "研究", "评估", "研判", "看法", "怎么样", "如何",
+    "技术面", "基本面", "走势", "趋势", "估值",
+    "analysis", "analyze", "outlook", "valuation", "technical analysis",
+    "fundamental",
+)
+
+_CONCEPTUAL_TERMS = (
+    "什么是", "什么叫", "是什么", "为什么", "怎么理解", "如何理解",
+    "解释", "定义", "概念", "原理", "介绍一下", "区别", "举例",
+    "what is", "what are", "how does", "explain", "define",
+    "difference between", "tell me about",
+)
+
+_MACRO_TERMS = (
+    "宏观", "宏观经济", "宏观政策", "宏观角度", "货币政策", "财政政策",
+    "gdp", "通胀", "通货膨胀", "cpi", "ppi", "利率", "加息", "降息",
+    "美联储", "央行", "收益率曲线", "国债", "产业政策", "行业监管",
+)
+
+_REALTY_TERMS = (
+    "房价", "楼市", "房产", "房地产", "租金", "二手房", "商铺", "户型",
+)
+
+_RECOMMENDATION_TERMS = (
+    "应该买", "应该卖", "该不该", "要不要", "值不值", "是否值得", "值得投资",
+    "建议买", "建议卖", "能买吗", "可以买吗",
+)
+
+_FINANCE_CONCEPT_TERMS = (
+    "dcf", "wacc", "capm", "sharpe", "sortino", "ebitda", "beta", "alpha",
+    "var", "cvar", "drawdown", "black-scholes", "期权", "期货", "衍生品",
+    "套利", "量化", "quant", "因子", "夏普比率", "最大回撤",
+    "市盈率", "市净率", "估值模型", "pe", "pb", "roe", "roa",
+)
+
+_METRIC_TERMS = (
+    "pe", "pb", "ps", "市盈率", "市净率", "市销率", "eps", "净利润",
+    "营收", "市值", "股息", "分红", "毛利率", "roe", "roa",
+)
+
+_FILE_EXT_RE = re.compile(
+    r"[^\s]+\.(?:docx?|pdf|xlsx?|pptx?|txt|csv|json|ya?ml|md|log|ipynb)\b"
+)
+_CODE_EXT_RE = re.compile(r"[^\s]+\.(?:py|js|ts|tsx|jsx|go|rs|java|rb|c|cpp|h|sh)\b")
+
+# Concrete, named financial entities.  A market intent requires one of these
+# (or a resolvable ticker) — generic category words like "美股" deliberately do
+# not qualify, because they appear throughout macro discussion.
+_ENTITY_TERMS = (
+    "苹果", "谷歌", "英伟达", "微软", "特斯拉", "亚马逊", "腾讯", "阿里",
+    "百度", "比亚迪", "茅台", "招商银行", "中国平安", "华为", "小米",
+    "美团", "京东", "字节", "宁德时代", "中芯国际",
+    "apple", "google", "nvidia", "microsoft", "tesla", "amazon", "meta",
+    "netflix", "palantir", "snowflake", "mongodb",
+    "比特币", "以太坊", "bitcoin", "ethereum",
+    "纳斯达克", "标普500", "道琼斯", "沪深300", "中证500", "恒生指数",
+)
+
+_TICKER_RE = re.compile(r"(?<![A-Za-z0-9])[A-Z]{2,5}(?![A-Za-z0-9])")
+_A_SHARE_RE = re.compile(r"(?<!\d)(?:60|68|00|30)\d{4}(?!\d)")
+
+
+@dataclass(frozen=True)
+class IntentSignals:
+    """Independent observations about a message, with no precedence applied."""
+
+    entity: bool
+    ticker: bool
+    market_context: bool
+    code: bool
+    product: bool
+    file_ref: bool
+    code_file_ref: bool
+    artifact: bool
+    realtime: bool
+    analysis: bool
+    conceptual: bool
+    macro: bool
+    realty: bool
+    recommendation: bool
+    finance_concept: bool
+    metric: bool
+
+    @property
+    def any_entity(self) -> bool:
+        return self.entity or self.ticker
+
+
+def extract_signals(message: str) -> IntentSignals:
+    """Observe *message* without deciding anything about it."""
+    spaced, compact = _normalize(message)
+    raw = message or ""
+    return IntentSignals(
+        entity=_any(_ENTITY_TERMS, spaced, compact),
+        ticker=bool(_TICKER_RE.search(raw) or _A_SHARE_RE.search(raw)),
+        market_context=_any(_MARKET_CONTEXT_TERMS, spaced, compact),
+        code=_any(_CODE_TERMS, spaced, compact),
+        product=_any(_PRODUCT_TERMS, spaced, compact),
+        file_ref=bool(_FILE_EXT_RE.search(raw)),
+        code_file_ref=bool(_CODE_EXT_RE.search(raw)),
+        artifact=_any(_ARTIFACT_TERMS, spaced, compact),
+        realtime=_any(_REALTIME_TERMS, spaced, compact),
+        analysis=_any(_ANALYSIS_TERMS, spaced, compact),
+        conceptual=_any(_CONCEPTUAL_TERMS, spaced, compact),
+        macro=_any(_MACRO_TERMS, spaced, compact),
+        realty=_any(_REALTY_TERMS, spaced, compact),
+        recommendation=_any(_RECOMMENDATION_TERMS, spaced, compact),
+        finance_concept=_any(_FINANCE_CONCEPT_TERMS, spaced, compact),
+        metric=_any(_METRIC_TERMS, spaced, compact),
+    )
+
+
+def is_visual_artifact_request(message: str) -> bool:
+    """True for chart/dashboard/report requests that should be built, not looked up."""
+    spaced, _ = _normalize(message)
+    if spaced.startswith(("/chart", "/dashboard", "/report")):
+        return True
+    signals = extract_signals(message)
+    if not signals.artifact:
+        return False
+    return signals.market_context or signals.any_entity or signals.finance_concept
+
+
+def classify(message: str) -> str:
+    """Classify *message* into one INTENT_* label.
+
+    The precedence below is the entire decision. Each rule states why it
+    outranks the ones after it.
+    """
+    if not (message or "").strip():
+        return INTENT_GENERAL
+
+    s = extract_signals(message)
+
+    # 1. A chart/dashboard/report over market data is something to *build*.
+    #    It needs the code path, not a quote lookup.
+    if is_visual_artifact_request(message):
+        return INTENT_CODING
+
+    # 2. A source file named outright is a code task regardless of phrasing.
+    if s.code_file_ref:
+        return INTENT_CODING
+
+    # 3. A document named outright is a reading task, not a market query —
+    #    even though "report_2024.docx" contains market-ish letters.
+    if s.file_ref:
+        return INTENT_CODING if s.code else INTENT_GENERAL
+
+    # 4. Code signals win over question phrasing.  "这段代码为什么报错" is a
+    #    debugging request; the old order let "为什么" demote it to general.
+    if s.code:
+        return INTENT_CODING
+
+    # 5. Real estate never routes to the stock templates, which expect market
+    #    data that does not exist for a housing question.
+    if s.realty:
+        return INTENT_GENERAL
+
+    # 6. Macro topics are discussion unless a specific company is named.
+    if s.macro and not s.any_entity:
+        return INTENT_GENERAL
+
+    # 7. "What is X" wants an explanation, not a lookup — but only when it is
+    #    not asking about a specific company's live numbers.
+    if s.conceptual and not (s.any_entity and (s.realtime or s.metric)):
+        return INTENT_GENERAL
+
+    # 8. "Should I buy X" is advice, not a chart.
+    if s.recommendation:
+        return INTENT_FINANCE if (s.any_entity or s.market_context) else INTENT_GENERAL
+
+    # 9. Live data wins over analysis: "分析苹果今天的市场" needs a fetch, or
+    #    the model answers a price question from memory.
+    if s.realtime and (s.any_entity or s.market_context or s.metric):
+        return INTENT_REALTIME
+
+    # 10. Market analysis requires a concrete entity.  Bare "分析"/"研究" is
+    #     ordinary Chinese and must not reach the stock-analysis template.
+    if s.analysis and s.any_entity:
+        return INTENT_ANALYSIS
+
+    # 11. Finance vocabulary without an entity is finance chat.
+    if s.finance_concept or (s.market_context and s.analysis):
+        return INTENT_FINANCE
+
+    # 12. A bare live-data question with no market subject ("最近有什么新闻").
+    if s.realtime:
+        return INTENT_REALTIME
+
+    return INTENT_GENERAL
+
+
+__all__ = [
+    "INTENT_ANALYSIS",
+    "INTENT_CODING",
+    "INTENT_FINANCE",
+    "INTENT_GENERAL",
+    "INTENT_REALTIME",
+    "IntentSignals",
+    "classify",
+    "extract_signals",
+    "is_visual_artifact_request",
+]

aria_code/apps/cli/lifecycle_hooks.py

@@ -0,0 +1,48 @@
+"""Simple lifecycle shell hooks for the CLI adapter."""
+
+from __future__ import annotations
+
+import os
+import subprocess
+from pathlib import Path
+from typing import Mapping
+
+
+def run_event_hook(
+    event: str,
+    *,
+    config_dir: Path,
+    cwd: Path | None = None,
+    env_extra: Mapping[str, str] | None = None,
+    timeout: int = 10,
+) -> None:
+    """Run executable hook scripts for a lifecycle event.
+
+    Script locations are intentionally simple for compatibility:
+    `<config_dir>/hooks/<event>.sh` and `<cwd>/.aria/hooks/<event>.sh`.
+    Hook failures are non-fatal.
+    """
+    root = cwd or Path.cwd()
+    dirs = [
+        config_dir / "hooks",
+        root / ".aria" / "hooks",
+    ]
+    env = dict(os.environ)
+    env["ARIA_EVENT"] = event
+    if env_extra:
+        env.update({str(k): str(v) for k, v in env_extra.items()})
+    for hook_dir in dirs:
+        script = hook_dir / f"{event}.sh"
+        if not script.exists() or script.stat().st_size <= 0:
+            continue
+        try:
+            subprocess.run(
+                [str(script)],
+                env=env,
+                timeout=timeout,
+                capture_output=True,
+                text=True,
+                check=False,
+            )
+        except Exception:
+            continue