@deepwhale/llm 1.0.0

package/dist/pricing-config.js

@@ -0,0 +1,186 @@
+/**
+ * @deepwhale/llm — Pricing 配置 (Sprint 1b.5 Step 1)
+ *
+ * 把原本 hardcode 在 `types.ts:computeCostBreakdown` 的 V4-Flash pricing 抽到
+ * ship-in + 用户可覆盖的 TOML 配置。Per-model currency (DeepSeek CNY, Anthropic USD),
+ * 不在 UI 层做汇率换算 — UI 从 `cost_currency` 字段读 symbol (¥ / $)。
+ *
+ * ## 设计原则 (R7 / Step 0 拍板, 2026-06-03)
+ *
+ * 1. `loadPricingConfig(userPath)`:
+ *    - 显式传 `userPath` 但文件不存在或非法 → **抛 PricingConfigParseError** (不静默)
+ *    - 不传 → 加载 ship-in `pricing.default.toml`
+ *
+ * 2. `computeCost(pricing, model, ...)` 是**纯函数**: 无 console / 无 logger / 无 IO。
+ *    3 种返回路径 (`CostBreakdownResult` 联合类型):
+ *    - 整体 `undefined`: 没 `cached_tokens` → 4 字段全 absent
+ *    - `{ base, no cost }`: 有 cache 但 pricing 找不到 model → 有 cache_hit_rate / tokens_uncached, 缺 cost
+ *    - 完整 4 字段: 找到
+ *
+ * 3. Warning 责任在 caller (client/mode 边界), 纯函数零副作用。
+ *
+ * 4. UI/RPC: `cost_turn` absent 时字段不显示 (absent 不是 null), 读 `cost_currency` 决 symbol。
+ */
+import { readFile } from 'node:fs/promises';
+import { fileURLToPath } from 'node:url';
+import { dirname, resolve } from 'node:path';
+import { parse as parseToml } from 'smol-toml';
+/**
+ * pricing TOML 解析/加载错误. `userPath` 显式传但文件不存在/非法时抛.
+ */
+export class PricingConfigParseError extends Error {
+    userPath;
+    name = 'PricingConfigParseError';
+    constructor(message, userPath, options) {
+        super(message);
+        this.userPath = userPath;
+        if (options?.cause !== undefined) {
+            this.cause = options.cause;
+        }
+    }
+}
+// ---- 纯函数 (无 console, 无 IO) ----
+/**
+ * 解析 TOML string → PricingConfig. 纯函数, 无 IO.
+ *
+ * 校验:
+ * - 必须有 `[models.xxx]` 块 (至少 1 个)
+ * - 每个 model 4 字段全必填 (cache_miss_per_m / cache_hit_per_m / completion_per_m / currency)
+ * - currency 必须在 'CNY' | 'USD' enum
+ * - 价格数字必须非负有限
+ *
+ * @throws PricingConfigParseError — malformed / 缺字段 / currency 错 / 数字非有限
+ */
+export function parsePricingConfig(toml) {
+    let raw;
+    try {
+        raw = parseToml(toml);
+    }
+    catch (cause) {
+        throw new PricingConfigParseError(`TOML parse failed: ${cause.message}`, '<inline>', { cause });
+    }
+    if (raw === null || typeof raw !== 'object' || Array.isArray(raw)) {
+        throw new PricingConfigParseError('TOML root must be a table', '<inline>');
+    }
+    const root = raw;
+    const modelsRaw = root['models'];
+    if (modelsRaw === undefined || modelsRaw === null) {
+        throw new PricingConfigParseError('Missing required section: [models.*]', '<inline>');
+    }
+    if (typeof modelsRaw !== 'object' || Array.isArray(modelsRaw)) {
+        throw new PricingConfigParseError('Section [models.*] must be a table', '<inline>');
+    }
+    const models = {};
+    for (const [modelId, value] of Object.entries(modelsRaw)) {
+        if (value === null || typeof value !== 'object' || Array.isArray(value)) {
+            throw new PricingConfigParseError(`[models.${modelId}] must be a table`, '<inline>');
+        }
+        const block = value;
+        const cacheMiss = block['cache_miss_per_m'];
+        const cacheHit = block['cache_hit_per_m'];
+        const completion = block['completion_per_m'];
+        const currency = block['currency'];
+        for (const [field, val] of [
+            ['cache_miss_per_m', cacheMiss],
+            ['cache_hit_per_m', cacheHit],
+            ['completion_per_m', completion],
+        ]) {
+            if (typeof val !== 'number' || !Number.isFinite(val) || val < 0) {
+                throw new PricingConfigParseError(`[models.${modelId}].${field} must be a non-negative finite number, got: ${JSON.stringify(val)}`, '<inline>');
+            }
+        }
+        if (currency !== 'CNY' && currency !== 'USD') {
+            throw new PricingConfigParseError(`[models.${modelId}].currency must be 'CNY' or 'USD', got: ${JSON.stringify(currency)}`, '<inline>');
+        }
+        models[modelId] = {
+            cache_miss_per_m: cacheMiss,
+            cache_hit_per_m: cacheHit,
+            completion_per_m: completion,
+            currency,
+        };
+    }
+    if (Object.keys(models).length === 0) {
+        throw new PricingConfigParseError('At least one model required in [models.*]', '<inline>');
+    }
+    return { models };
+}
+/**
+ * 纯函数, 无 console / 无 logger / 无 IO. 3 种返回路径见 `CostBreakdownResult`.
+ *
+ * 公式: tokens_uncached = max(0, prompt - cached)
+ *       cache_hit_rate  = cached / prompt  (prompt=0 时 0, 避免除零)
+ *       cost_turn       = tokens_uncached * cache_miss_per_m/1e6
+ *                        + cached * cache_hit_per_m/1e6
+ *                        + completion * completion_per_m/1e6
+ *
+ * 把 /M 转 /token 必须除 1e6, 不要直接当 /token 用 (会大 1000×).
+ */
+export function computeCost(pricing, model, promptTokens, completionTokens, cachedTokens) {
+    if (cachedTokens === undefined) {
+        return undefined;
+    }
+    const tokensUncached = Math.max(0, promptTokens - cachedTokens);
+    const hitRate = promptTokens > 0 ? cachedTokens / promptTokens : 0;
+    // model undefined 或 pricing 找不到 model → 走 R7 中间路径
+    // (有 cached 但 cost 字段 absent). 不静默 fallback, 不抛错, 不写 0.
+    const modelPricing = model !== undefined ? pricing?.models[model] : undefined;
+    if (modelPricing === undefined) {
+        return {
+            cache_hit_rate: hitRate,
+            tokens_uncached: tokensUncached,
+        };
+    }
+    const costTurn = tokensUncached * (modelPricing.cache_miss_per_m / 1_000_000) +
+        cachedTokens * (modelPricing.cache_hit_per_m / 1_000_000) +
+        completionTokens * (modelPricing.completion_per_m / 1_000_000);
+    return {
+        cache_hit_rate: hitRate,
+        tokens_uncached: tokensUncached,
+        cost_turn: costTurn,
+        cost_currency: modelPricing.currency,
+    };
+}
+// ---- IO 函数 (一次启动加载, 注入到 client) ----
+/**
+ * 加载 pricing config. 三种来源优先级 (高→低):
+ * 1. `userPath` 显式指定且存在 → 解析用户文件
+ * 2. `userPath` 显式指定但不存在或非法 → **抛 PricingConfigParseError** (不静默 fallback)
+ * 3. `userPath` 未传 → 加载 ship-in `pricing.default.toml`
+ *
+ * hot path 不应调 (走 client.pricing 字段). Step 2 的 AnthropicClient 启动时调一次.
+ */
+export async function loadPricingConfig(userPath) {
+    if (userPath === undefined) {
+        // 走 ship-in default: 跟 pricing-config.ts 同目录, build 后复制到 dist/.
+        const here = dirname(fileURLToPath(import.meta.url));
+        // dist/pricing-config.js → dist/pricing.default.toml
+        const defaultPath = resolve(here, 'pricing.default.toml');
+        return await loadFromPath(defaultPath);
+    }
+    return await loadFromPath(userPath);
+}
+async function loadFromPath(path) {
+    let tomlText;
+    try {
+        tomlText = await readFile(path, 'utf-8');
+    }
+    catch (cause) {
+        const err = cause;
+        if (err.code === 'ENOENT') {
+            throw new PricingConfigParseError(`Pricing config not found at: ${path}`, path, { cause });
+        }
+        throw new PricingConfigParseError(`Failed to read pricing config at ${path}: ${err.message}`, path, { cause });
+    }
+    try {
+        return parsePricingConfig(tomlText);
+    }
+    catch (cause) {
+        if (cause instanceof PricingConfigParseError) {
+            // 重新抛, 用真实 path 替换 inner 错误的 '<inline>' 占位符.
+            // 保留原始 userPath (readonly 字段, 不用 spread 因为 class 自带字段).
+            throw new PricingConfigParseError(cause.message.replace('<inline>', path), path, { cause });
+        }
+        throw cause;
+    }
+}
+//# sourceMappingURL=pricing-config.js.map

package/dist/pricing-config.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"pricing-config.js","sourceRoot":"","sources":["../src/pricing-config.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,OAAO,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AAC5C,OAAO,EAAE,aAAa,EAAE,MAAM,UAAU,CAAC;AACzC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AAC7C,OAAO,EAAE,KAAK,IAAI,SAAS,EAAE,MAAM,WAAW,CAAC;AA8C/C;;GAEG;AACH,MAAM,OAAO,uBAAwB,SAAQ,KAAK;IAIrC;IAHO,IAAI,GAAG,yBAAkC,CAAC;IAC5D,YACE,OAAe,EACN,QAAgB,EACzB,OAA6B;QAE7B,KAAK,CAAC,OAAO,CAAC,CAAC;QAHN,aAAQ,GAAR,QAAQ,CAAQ;QAIzB,IAAI,OAAO,EAAE,KAAK,KAAK,SAAS,EAAE,CAAC;YAChC,IAA4B,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QACtD,CAAC;IACH,CAAC;CACF;AAED,kCAAkC;AAElC;;;;;;;;;;GAUG;AACH,MAAM,UAAU,kBAAkB,CAAC,IAAY;IAC7C,IAAI,GAAY,CAAC;IACjB,IAAI,CAAC;QACH,GAAG,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC;IACxB,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,MAAM,IAAI,uBAAuB,CAAC,sBAAuB,KAAe,CAAC,OAAO,EAAE,EAAE,UAAU,EAAE,EAAE,KAAK,EAAE,CAAC,CAAC;IAC7G,CAAC;IAED,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,GAAG,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE,CAAC;QAClE,MAAM,IAAI,uBAAuB,CAAC,2BAA2B,EAAE,UAAU,CAAC,CAAC;IAC7E,CAAC;IAED,MAAM,IAAI,GAAG,GAA8B,CAAC;IAC5C,MAAM,SAAS,GAAG,IAAI,CAAC,QAAQ,CAAC,CAAC;IAEjC,IAAI,SAAS,KAAK,SAAS,IAAI,SAAS,KAAK,IAAI,EAAE,CAAC;QAClD,MAAM,IAAI,uBAAuB,CAAC,sCAAsC,EAAE,UAAU,CAAC,CAAC;IACxF,CAAC;IACD,IAAI,OAAO,SAAS,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,EAAE,CAAC;QAC9D,MAAM,IAAI,uBAAuB,CAAC,oCAAoC,EAAE,UAAU,CAAC,CAAC;IACtF,CAAC;IAED,MAAM,MAAM,GAAiC,EAAE,CAAC;IAChD,KAAK,MAAM,CAAC,OAAO,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,SAAoC,CAAC,EAAE,CAAC;QACpF,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;YACxE,MAAM,IAAI,uBAAuB,CAC/B,WAAW,OAAO,mBAAmB,EACrC,UAAU,CACX,CAAC;QACJ,CAAC;QACD,MAAM,KAAK,GAAG,KAAgC,CAAC;QAC/C,MAAM,SAAS,GAAG,KAAK,CAAC,kBAAkB,CAAC,CAAC;QAC5C,MAAM,QAAQ,GAAG,KAAK,CAAC,iBAAiB,CAAC,CAAC;QAC1C,MAAM,UAAU,GAAG,KAAK,CAAC,kBAAkB,CAAC,CAAC;QAC7C,MAAM,QAAQ,GAAG,KAAK,CAAC,UAAU,CAAC,CAAC;QAEnC,KAAK,MAAM,CAAC,KAAK,EAAE,GAAG,CAAC,IAAI;YACzB,CAAC,kBAAkB,EAAE,SAAS,CAAC;YAC/B,CAAC,iBAAiB,EAAE,QAAQ,CAAC;YAC7B,CAAC,kBAAkB,EAAE,UAAU,CAAC;SACxB,EAAE,CAAC;YACX,IAAI,OAAO,GAAG,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,GAAG,CAAC,IAAI,GAAG,GAAG,CAAC,EAAE,CAAC;gBAChE,MAAM,IAAI,uBAAuB,CAC/B,WAAW,OAAO,KAAK,KAAK,+CAA+C,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,EAAE,EAChG,UAAU,CACX,CAAC;YACJ,CAAC;QACH,CAAC;QACD,IAAI,QAAQ,KAAK,KAAK,IAAI,QAAQ,KAAK,KAAK,EAAE,CAAC;YAC7C,MAAM,IAAI,uBAAuB,CAC/B,WAAW,OAAO,2CAA2C,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,EAAE,EACvF,UAAU,CACX,CAAC;QACJ,CAAC;QACD,MAAM,CAAC,OAAO,CAAC,GAAG;YAChB,gBAAgB,EAAE,SAAmB;YACrC,eAAe,EAAE,QAAkB;YACnC,gBAAgB,EAAE,UAAoB;YACtC,QAAQ;SACT,CAAC;IACJ,CAAC;IAED,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACrC,MAAM,IAAI,uBAAuB,CAAC,2CAA2C,EAAE,UAAU,CAAC,CAAC;IAC7F,CAAC;IAED,OAAO,EAAE,MAAM,EAAE,CAAC;AACpB,CAAC;AAED;;;;;;;;;;GAUG;AACH,MAAM,UAAU,WAAW,CACzB,OAAkC,EAClC,KAA0B,EAC1B,YAAoB,EACpB,gBAAwB,EACxB,YAAgC;IAEhC,IAAI,YAAY,KAAK,SAAS,EAAE,CAAC;QAC/B,OAAO,SAAS,CAAC;IACnB,CAAC;IAED,MAAM,cAAc,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,YAAY,GAAG,YAAY,CAAC,CAAC;IAChE,MAAM,OAAO,GAAG,YAAY,GAAG,CAAC,CAAC,CAAC,CAAC,YAAY,GAAG,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC;IAEnE,kDAAkD;IAClD,wDAAwD;IACxD,MAAM,YAAY,GAAG,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,OAAO,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;IAC9E,IAAI,YAAY,KAAK,SAAS,EAAE,CAAC;QAC/B,OAAO;YACL,cAAc,EAAE,OAAO;YACvB,eAAe,EAAE,cAAc;SAChC,CAAC;IACJ,CAAC;IAED,MAAM,QAAQ,GACZ,cAAc,GAAG,CAAC,YAAY,CAAC,gBAAgB,GAAG,SAAS,CAAC;QAC5D,YAAY,GAAG,CAAC,YAAY,CAAC,eAAe,GAAG,SAAS,CAAC;QACzD,gBAAgB,GAAG,CAAC,YAAY,CAAC,gBAAgB,GAAG,SAAS,CAAC,CAAC;IAEjE,OAAO;QACL,cAAc,EAAE,OAAO;QACvB,eAAe,EAAE,cAAc;QAC/B,SAAS,EAAE,QAAQ;QACnB,aAAa,EAAE,YAAY,CAAC,QAAQ;KACrC,CAAC;AACJ,CAAC;AAED,uCAAuC;AAEvC;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,iBAAiB,CAAC,QAAiB;IACvD,IAAI,QAAQ,KAAK,SAAS,EAAE,CAAC;QAC3B,gEAAgE;QAChE,MAAM,IAAI,GAAG,OAAO,CAAC,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC;QACrD,qDAAqD;QACrD,MAAM,WAAW,GAAG,OAAO,CAAC,IAAI,EAAE,sBAAsB,CAAC,CAAC;QAC1D,OAAO,MAAM,YAAY,CAAC,WAAW,CAAC,CAAC;IACzC,CAAC;IACD,OAAO,MAAM,YAAY,CAAC,QAAQ,CAAC,CAAC;AACtC,CAAC;AAED,KAAK,UAAU,YAAY,CAAC,IAAY;IACtC,IAAI,QAAgB,CAAC;IACrB,IAAI,CAAC;QACH,QAAQ,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE,OAAO,CAAC,CAAC;IAC3C,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,MAAM,GAAG,GAAG,KAA8B,CAAC;QAC3C,IAAI,GAAG,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;YAC1B,MAAM,IAAI,uBAAuB,CAC/B,gCAAgC,IAAI,EAAE,EACtC,IAAI,EACJ,EAAE,KAAK,EAAE,CACV,CAAC;QACJ,CAAC;QACD,MAAM,IAAI,uBAAuB,CAC/B,oCAAoC,IAAI,KAAK,GAAG,CAAC,OAAO,EAAE,EAC1D,IAAI,EACJ,EAAE,KAAK,EAAE,CACV,CAAC;IACJ,CAAC;IACD,IAAI,CAAC;QACH,OAAO,kBAAkB,CAAC,QAAQ,CAAC,CAAC;IACtC,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,IAAI,KAAK,YAAY,uBAAuB,EAAE,CAAC;YAC7C,6CAA6C;YAC7C,wDAAwD;YACxD,MAAM,IAAI,uBAAuB,CAC/B,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,UAAU,EAAE,IAAI,CAAC,EACvC,IAAI,EACJ,EAAE,KAAK,EAAE,CACV,CAAC;QACJ,CAAC;QACD,MAAM,KAAK,CAAC;IACd,CAAC;AACH,CAAC"}

package/dist/pricing.default.toml

@@ -0,0 +1,46 @@
+# deepwhale 默认 pricing (2026-06)
+# 用户可在 ~/.deepwhale/pricing.toml 覆盖 (--pricing-config <path> 后续可加, 1b.5 走 env)。
+# **per-model currency**: DeepSeek CNY, Anthropic USD。不在 UI 层做汇率换算。
+#
+# **当前生产 model 2 个** (Q 拍板 2026-06-03): flash + pro.
+# Only production DeepSeek models are supported. (Q 拍板 2026-06-04, Task 2 Gap 4 留痕)
+# 不再加 deepseek-chat / deepseek-reasoner 老 alias — 那些是 1a/1b era 历史,
+# 2026-07-24 后服务端 hard-fail, 老 session replay 走 R7 中间路径 (base 2 字段, cost 字段 absent).
+#
+# **信源唯一性**: DeepSeek pricing 永远以 https://api-docs.deepseek.com/quick_start/pricing
+# 为唯一信源。Q 在 2026-06-03 review 时贴出官方 "模型细节" 页截屏纠正我之前 commit 用的
+# 0.5/0.1/1.0 (来自 1b 时代印象 + 二手源, 未核对官方). 实际官方:
+# - cache_miss_per_m = input 缓存未命中 (1M tokens)
+# - cache_hit_per_m  = input 缓存命中 (1M tokens)
+# - completion_per_m = output (1M tokens)
+# 注意: 1b follow-up P1 改的 0.01/0.1 fix 是 **公式 1000× off** — pricing 数字本身的 0.1
+# 看上去像 0.1 ¥/M 但其实只是 0.5→0.1 cache hit 巧合小, 真正的官方是 0.02¥/M (5× off 真实价).
+
+[models.deepseek-v4-flash]
+cache_miss_per_m = 1.0     # ¥1 /M tokens (input 缓存未命中, 官方 api-docs.deepseek.com 2026-06)
+cache_hit_per_m  = 0.02    # ¥0.02 /M tokens (input 缓存命中, 官方)
+completion_per_m = 2.0     # ¥2 /M tokens (output, 官方)
+currency         = "CNY"
+
+[models.deepseek-v4-pro]
+cache_miss_per_m = 3.0     # ¥3 /M tokens (input 缓存未命中, 官方)
+cache_hit_per_m  = 0.025   # ¥0.025 /M tokens (input 缓存命中, 官方)
+completion_per_m = 6.0     # ¥6 /M tokens (output, 官方)
+currency         = "CNY"
+
+# Anthropic Claude Sonnet 4.5 (2026-06 official first-party API pricing)
+# https://platform.claude.com/docs/en/about-claude/pricing
+[models.claude-sonnet-4-5]
+cache_miss_per_m = 3.0    # $3 /M tokens (input)
+cache_hit_per_m  = 0.30   # $0.30 /M tokens (cache read, 1h TTL — 不是 5min ephemeral $0.06)
+completion_per_m = 15.0   # $15 /M tokens
+currency         = "USD"
+
+# Anthropic Claude Opus 4.5 (2026-06 official first-party API pricing)
+# P1 plan correction (2026-06-03): user 核对官方 pricing 页后纠正
+# 之前值 $24/$0.48/$120 是错的 (类似某种旧/特殊模式价格, 不适合 first-party default)
+[models.claude-opus-4-5]
+cache_miss_per_m = 5.0    # $5 /M
+cache_hit_per_m  = 0.50   # $0.50 /M (cache read, 1h TTL — 不是 5min ephemeral $0.48)
+completion_per_m = 25.0   # $25 /M (不是 $120)
+currency         = "USD"

package/dist/types.d.ts

@@ -0,0 +1,255 @@
+/**
+ * @deepwhale/llm — 公共类型
+ *
+ * Sprint 0.3 范围：非流式 chat + 5 个 LLMError 子类。
+ * Sprint 1a 扩展：system/tool role + tool_calls + stream() + usage + retry metadata。
+ * Sprint 1b 再加：prompt cache、cost accounting、canonical schema。
+ */
+import type { Brand } from '@deepwhale/core';
+/** 模型 ID 的品牌类型，避免把 'deepseek-chat' 当成任意 string 传来传去。 */
+export type ModelId = Brand<string, 'ModelId'>;
+/** Role 联合 — Sprint 1a 加 'system' / 'tool'(tool loop 需要)。 */
+export type Role = 'system' | 'user' | 'assistant' | 'tool';
+/**
+ * LLM 调用的 tool 调用描述。Sprint 1a 最小：
+ * - id: 由 LLM 返回,tool 响应里 echo 回去
+ * - name: tool 名(去 registry 查)
+ * - args: tool 参数对象(已 JSON 解析,不是 string)
+ *
+ * Sprint 1b+ 会在 tool_loop 里用这个 type guard args 是否合法(schema 校验)。
+ */
+export interface ToolCall {
+    readonly id: string;
+    readonly name: string;
+    readonly args: Record<string, unknown>;
+}
+/**
+ * 单条 chat message。Sprint 1a:
+ * - role: 加 'system' / 'tool'
+ * - content: assistant 在有 tool_calls 时可为空字符串(OAI spec)
+ * - tool_calls: assistant 消息携带
+ * - tool_call_id: tool 消息携带,echo 上面那个 id
+ * - name: tool 消息携带,工具名(OAI 协议要求,便于审计)
+ */
+export interface ChatMessage {
+    role: Role;
+    content: string;
+    tool_calls?: ReadonlyArray<ToolCall>;
+    tool_call_id?: string;
+    name?: string;
+}
+/**
+ * Token usage 来自 LLM 响应。Sprint 1a: OAI 标准 + cached_tokens。
+ * Sprint 1b: 加 cache_hit_rate / cost_turn / tokens_uncached (可观测性)。
+ * Sprint 1b.5: 加 cost_currency (¥ / $) 让 UI 决 symbol, 不在 UI 层做汇率换算.
+ *
+ * OAI 标准字段 + DeepSeek 扩展:
+ * - prompt_tokens / completion_tokens / total_tokens
+ * - cached_tokens (DeepSeek V4 起, 命中 cache 的 token 数)
+ *
+ * Sprint 1b 扩展 (Prefix-cache 可观测性):
+ * - cache_hit_rate: 0..1, cached_tokens / prompt_tokens。LLM 不返 cached_tokens 时 undefined。
+ * - cost_turn: 本次 turn 估算费用, 按 pricing config 算 (Sprint 1b.5 抽 config.toml)。
+ * - tokens_uncached: prompt_tokens - cached_tokens, 方便人眼扫读"实际新付的 token"。
+ *
+ * Sprint 1b.5 扩展 (per-model currency, 不引入汇率换算):
+ * - cost_currency: 'CNY' | 'USD' — 跟 cost_turn 同步存在或同步 absent。
+ *   UI 读这个字段决定显示 ¥ 还是 $, 不在数字层做换算。
+ *   cost_turn absent 时 cost_currency 也必须 absent (不变量)。
+ */
+export interface Usage {
+    prompt_tokens: number;
+    completion_tokens: number;
+    total_tokens: number;
+    cached_tokens?: number;
+    /** 0..1, cached_tokens / prompt_tokens。LLM 不返 cached_tokens 时 undefined。 */
+    cache_hit_rate?: number;
+    /**
+     * 本 turn 估算费用, pricing config 算 (Sprint 1b.5 抽 config.toml)。
+     * 数字单位: currency (CNY ¥ 或 USD $). 不在 UI 层做汇率换算。
+     */
+    cost_turn?: number;
+    /** Currency 跟 cost_turn 同步存在, UI 读这个决 symbol (¥ / $). */
+    cost_currency?: 'CNY' | 'USD';
+    /** prompt_tokens - cached_tokens, "实际新付"的 token 数, 方便人眼扫读。 */
+    tokens_uncached?: number;
+}
+/**
+ * 旧的 `computeCostBreakdown` (Sprint 1b, hardcode V4-Flash pricing) 已整体迁移到
+ * `pricing-config.ts:computeCost`, 同时:
+ * - 接收 `PricingConfig` + `ModelId` 参数 (per-model currency)
+ * - 接收 `pricing: undefined` (显式表达"没 pricing 也走纯函数 2 字段路径")
+ * - 返 `CostBreakdownResult` 联合类型 (3 种路径, 显式表达无 pricing 也能算 cache 字段)
+ * - 行为兼容: 公式 (V4-Flash pricing 1.0/0.5/0.1 等数字没变) + 边界 (cached=undefined → undefined)
+ *
+ * R7 设计原则: pricing 是观测, 不让成功 LLM 响应因"缺价格表"失败; silent fallback 制造假成本更危险。
+ * 缺 pricing 时不静默 fallback 到硬编码, 也不抛错, 走"base 2 字段, cost 字段 absent"路径。
+ * 找不到 model 的 warning 责任在 caller (client/mode 边界), 纯函数零副作用。
+ *
+ * Sprint 1a: cache_hit_rate 仅做总 cost 估算, 没暴露。Sprint 1b 起可观测。
+ */
+/** chat() 完整调用的返回值。Sprint 1a 加 tool_calls + usage。 */
+export interface ChatResult {
+    model: ModelId;
+    content: string;
+    tool_calls?: ReadonlyArray<ToolCall>;
+    usage?: Usage;
+    /**
+     * finish_reason:
+     * - 'stop': 自然结束
+     * - 'tool_calls': LLM 决定调工具
+     * - 'length': 触达 max_tokens
+     * - 'content_filter': 触发安全过滤
+     */
+    finish_reason?: 'stop' | 'tool_calls' | 'length' | 'content_filter';
+}
+/**
+ * 流式 chunk。Sprint 1a 极简:
+ * - delta.content: 增量文本(可能为空,如纯 tool_calls 增量)
+ * - delta.tool_calls: 增量 tool_call(Sprint 1a 一次性返回完整,Sprint 1b+ 再支持 incremental)
+ * - usage: 只在最后一个 chunk 出现(OAI 协议)
+ * - finish_reason: 同 ChatResult
+ */
+export interface ChatChunk {
+    delta: {
+        content?: string;
+        tool_calls?: ReadonlyArray<ToolCall>;
+    };
+    usage?: Usage;
+    finish_reason?: ChatResult['finish_reason'];
+}
+/**
+ * LLM 客户端的抽象接口。Sprint 1a 加 stream() + tools/tool_choice 字段。
+ *
+ * - chat(): 非流式,一次性返回完整结果
+ * - stream(): 流式,逐 chunk 回调(onChunk 同步;onComplete 异步 return final)
+ *
+ * @throws APIKeyMissingError — API key 未设置
+ * @throws LLMRateLimitError — 429(retry 透传后仍未恢复)
+ * @throws LLMAuthError      — 401/403
+ * @throws LLMNetworkError   — 网络/DNS 失败
+ * @throws LLMUnknownError   — 其它 5xx 或 JSON 解析失败
+ * @throws LLMStreamError    — SSE 解析中途断流
+ */
+export interface LLMClient {
+    readonly model: ModelId;
+    /**
+     * 发送一组 messages,拿到助手完整回复(非流式)。
+     *
+     * Sprint 1a 新增:
+     * - tools: 工具 schema 列表(LLM 看到后可能决定调)
+     * - tool_choice: 'auto' / 'none' / 'required'(OAI 标准)
+     *
+     * retry: 内部对 429 / 5xx / network error 自动指数退避(默认 3 次),
+     *         全失败才抛 LLMError。Sprint 1a 简化为固定 3 次。
+     */
+    chat(messages: ChatMessage[], options?: {
+        signal?: AbortSignal;
+        tools?: ReadonlyArray<LLMToolSchema>;
+        tool_choice?: 'auto' | 'none' | 'required';
+    }): Promise<ChatResult>;
+    /**
+     * 流式 chat。Sprint 1a 新增。
+     *
+     * 用法:
+     *   for await (const chunk of client.stream(messages, { onChunk })) {
+     *     // chunk = ChatChunk
+     *   }
+     *
+     * 也支持 callback 模式(更简单,适合 REPL):
+     *   await client.stream(messages, { onChunk: (c) => out.write(c.delta.content ?? '') });
+     *
+     * retry 行为: 内部 stream 重连在 Sprint 1a **不做**(流式断流语义复杂),
+     * 整个 stream 失败抛 LLMError 让 caller 决定(Sprint 1b+ 补断点续传)。
+     */
+    stream(messages: ChatMessage[], options: {
+        signal?: AbortSignal;
+        tools?: ReadonlyArray<LLMToolSchema>;
+        tool_choice?: 'auto' | 'none' | 'required';
+        onChunk: (chunk: ChatChunk) => void;
+    }): Promise<ChatResult>;
+}
+/**
+ * Sprint 1a 极简:复用 OpenAI function-calling 协议 + ToolInputSchema 兼容形态。
+ *
+ * 不用 import @deepwhale/coding-agent 避免循环依赖 — Tool 那边把 schema 转成
+ * 这个形态即可(structural typing,运行时不影响)。
+ */
+export interface LLMToolSchema {
+    name: string;
+    description: string;
+    parameters: LLMToolParametersSchema;
+}
+/** OAI function-calling 参数 schema(简化形态,跟 coding-agent 的 ToolInputSchema 兼容)。 */
+export interface LLMToolParametersSchema {
+    readonly type: 'object';
+    readonly properties: Record<string, LLMToolParamSchema>;
+    readonly required?: ReadonlyArray<string>;
+}
+export type LLMToolParamSchema = {
+    type: 'string';
+    description: string;
+    enum?: ReadonlyArray<string>;
+} | {
+    type: 'number';
+    description: string;
+    minimum?: number;
+    maximum?: number;
+} | {
+    type: 'boolean';
+    description: string;
+} | {
+    type: 'array';
+    description: string;
+    items: LLMToolParamSchema;
+};
+/** Sprint 1a：SSE 流中途断流专用。Sprint 1b 再加 retry/续传。 */
+export declare class LLMStreamError extends Error implements LLMError {
+    readonly name: "LLMStreamError";
+    readonly isLLMError: true;
+    constructor(message: string, options?: {
+        cause?: unknown;
+    });
+}
+export interface LLMError {
+    readonly isLLMError: true;
+    readonly name: string;
+    readonly message: string;
+    readonly cause?: unknown;
+}
+/** 类型守卫：通过 isLLMError 标签识别 LLM 派生的 Error。 */
+export declare function isLLMError(err: unknown): err is LLMError;
+export declare class APIKeyMissingError extends Error implements LLMError {
+    readonly name: "APIKeyMissingError";
+    readonly isLLMError: true;
+    constructor(message: string);
+}
+export declare class LLMRateLimitError extends Error implements LLMError {
+    readonly name: "LLMRateLimitError";
+    readonly isLLMError: true;
+    readonly status: 429;
+    constructor(message: string);
+}
+export declare class LLMAuthError extends Error implements LLMError {
+    readonly name: "LLMAuthError";
+    readonly isLLMError: true;
+    readonly status: number;
+    constructor(status: number, message: string);
+}
+export declare class LLMNetworkError extends Error implements LLMError {
+    readonly name: "LLMNetworkError";
+    readonly isLLMError: true;
+    constructor(message: string, options?: {
+        cause?: unknown;
+    });
+}
+export declare class LLMUnknownError extends Error implements LLMError {
+    readonly name: "LLMUnknownError";
+    readonly isLLMError: true;
+    readonly status?: number;
+    constructor(message: string, options?: {
+        cause?: unknown;
+        status?: number;
+    });
+}
+//# sourceMappingURL=types.d.ts.map

package/dist/types.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,iBAAiB,CAAC;AAE7C,wDAAwD;AACxD,MAAM,MAAM,OAAO,GAAG,KAAK,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;AAE/C,6DAA6D;AAC7D,MAAM,MAAM,IAAI,GAAG,QAAQ,GAAG,MAAM,GAAG,WAAW,GAAG,MAAM,CAAC;AAE5D;;;;;;;GAOG;AACH,MAAM,WAAW,QAAQ;IACvB,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACxC;AAED;;;;;;;GAOG;AACH,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,IAAI,CAAC;IACX,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,CAAC,EAAE,aAAa,CAAC,QAAQ,CAAC,CAAC;IACrC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,MAAM,WAAW,KAAK;IACpB,aAAa,EAAE,MAAM,CAAC;IACtB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,YAAY,EAAE,MAAM,CAAC;IACrB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,4EAA4E;IAC5E,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB;;;OAGG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,yDAAyD;IACzD,aAAa,CAAC,EAAE,KAAK,GAAG,KAAK,CAAC;IAC9B,8DAA8D;IAC9D,eAAe,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED;;;;;;;;;;;;;GAaG;AAEH,sDAAsD;AACtD,MAAM,WAAW,UAAU;IACzB,KAAK,EAAE,OAAO,CAAC;IACf,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,CAAC,EAAE,aAAa,CAAC,QAAQ,CAAC,CAAC;IACrC,KAAK,CAAC,EAAE,KAAK,CAAC;IACd;;;;;;OAMG;IACH,aAAa,CAAC,EAAE,MAAM,GAAG,YAAY,GAAG,QAAQ,GAAG,gBAAgB,CAAC;CACrE;AAED;;;;;;GAMG;AACH,MAAM,WAAW,SAAS;IACxB,KAAK,EAAE;QACL,OAAO,CAAC,EAAE,MAAM,CAAC;QACjB,UAAU,CAAC,EAAE,aAAa,CAAC,QAAQ,CAAC,CAAC;KACtC,CAAC;IACF,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,aAAa,CAAC,EAAE,UAAU,CAAC,eAAe,CAAC,CAAC;CAC7C;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,WAAW,SAAS;IACxB,QAAQ,CAAC,KAAK,EAAE,OAAO,CAAC;IAExB;;;;;;;;;OASG;IACH,IAAI,CACF,QAAQ,EAAE,WAAW,EAAE,EACvB,OAAO,CAAC,EAAE;QACR,MAAM,CAAC,EAAE,WAAW,CAAC;QACrB,KAAK,CAAC,EAAE,aAAa,CAAC,aAAa,CAAC,CAAC;QACrC,WAAW,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,CAAC;KAC5C,GACA,OAAO,CAAC,UAAU,CAAC,CAAC;IAEvB;;;;;;;;;;;;;OAaG;IACH,MAAM,CACJ,QAAQ,EAAE,WAAW,EAAE,EACvB,OAAO,EAAE;QACP,MAAM,CAAC,EAAE,WAAW,CAAC;QACrB,KAAK,CAAC,EAAE,aAAa,CAAC,aAAa,CAAC,CAAC;QACrC,WAAW,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,CAAC;QAC3C,OAAO,EAAE,CAAC,KAAK,EAAE,SAAS,KAAK,IAAI,CAAC;KACrC,GACA,OAAO,CAAC,UAAU,CAAC,CAAC;CACxB;AAED;;;;;GAKG;AACH,MAAM,WAAW,aAAa;IAC5B,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,uBAAuB,CAAC;CACrC;AAED,gFAAgF;AAChF,MAAM,WAAW,uBAAuB;IACtC,QAAQ,CAAC,IAAI,EAAE,QAAQ,CAAC;IACxB,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,kBAAkB,CAAC,CAAC;IACxD,QAAQ,CAAC,QAAQ,CAAC,EAAE,aAAa,CAAC,MAAM,CAAC,CAAC;CAC3C;AAED,MAAM,MAAM,kBAAkB,GAC1B;IAAE,IAAI,EAAE,QAAQ,CAAC;IAAC,WAAW,EAAE,MAAM,CAAC;IAAC,IAAI,CAAC,EAAE,aAAa,CAAC,MAAM,CAAC,CAAA;CAAE,GACrE;IAAE,IAAI,EAAE,QAAQ,CAAC;IAAC,WAAW,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,MAAM,CAAA;CAAE,GAC3E;IAAE,IAAI,EAAE,SAAS,CAAC;IAAC,WAAW,EAAE,MAAM,CAAA;CAAE,GACxC;IAAE,IAAI,EAAE,OAAO,CAAC;IAAC,WAAW,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,kBAAkB,CAAA;CAAE,CAAC;AAEtE,mDAAmD;AACnD,qBAAa,cAAe,SAAQ,KAAM,YAAW,QAAQ;IAC3D,SAAkB,IAAI,EAAG,gBAAgB,CAAU;IACnD,QAAQ,CAAC,UAAU,EAAG,IAAI,CAAU;gBACxB,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,OAAO,CAAA;KAAE;CAM3D;AAQD,MAAM,WAAW,QAAQ;IACvB,QAAQ,CAAC,UAAU,EAAE,IAAI,CAAC;IAC1B,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,KAAK,CAAC,EAAE,OAAO,CAAC;CAC1B;AAED,6CAA6C;AAC7C,wBAAgB,UAAU,CAAC,GAAG,EAAE,OAAO,GAAG,GAAG,IAAI,QAAQ,CAExD;AAED,qBAAa,kBAAmB,SAAQ,KAAM,YAAW,QAAQ;IAC/D,SAAkB,IAAI,EAAG,oBAAoB,CAAU;IACvD,QAAQ,CAAC,UAAU,EAAG,IAAI,CAAU;gBACxB,OAAO,EAAE,MAAM;CAG5B;AAED,qBAAa,iBAAkB,SAAQ,KAAM,YAAW,QAAQ;IAC9D,SAAkB,IAAI,EAAG,mBAAmB,CAAU;IACtD,QAAQ,CAAC,UAAU,EAAG,IAAI,CAAU;IACpC,QAAQ,CAAC,MAAM,EAAG,GAAG,CAAU;gBACnB,OAAO,EAAE,MAAM;CAG5B;AAED,qBAAa,YAAa,SAAQ,KAAM,YAAW,QAAQ;IACzD,SAAkB,IAAI,EAAG,cAAc,CAAU;IACjD,QAAQ,CAAC,UAAU,EAAG,IAAI,CAAU;IACpC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;gBACZ,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM;CAI5C;AAED,qBAAa,eAAgB,SAAQ,KAAM,YAAW,QAAQ;IAC5D,SAAkB,IAAI,EAAG,iBAAiB,CAAU;IACpD,QAAQ,CAAC,UAAU,EAAG,IAAI,CAAU;gBACxB,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,OAAO,CAAA;KAAE;CAM3D;AAED,qBAAa,eAAgB,SAAQ,KAAM,YAAW,QAAQ;IAC5D,SAAkB,IAAI,EAAG,iBAAiB,CAAU;IACpD,QAAQ,CAAC,UAAU,EAAG,IAAI,CAAU;IACpC,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;gBACb,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE;QAAE,KAAK,CAAC,EAAE,OAAO,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAA;KAAE;CAO5E"}

package/dist/types.js

@@ -0,0 +1,70 @@
+/**
+ * @deepwhale/llm — 公共类型
+ *
+ * Sprint 0.3 范围：非流式 chat + 5 个 LLMError 子类。
+ * Sprint 1a 扩展：system/tool role + tool_calls + stream() + usage + retry metadata。
+ * Sprint 1b 再加：prompt cache、cost accounting、canonical schema。
+ */
+/** Sprint 1a：SSE 流中途断流专用。Sprint 1b 再加 retry/续传。 */
+export class LLMStreamError extends Error {
+    name = 'LLMStreamError';
+    isLLMError = true;
+    constructor(message, options) {
+        super(message);
+        if (options?.cause !== undefined) {
+            this.cause = options.cause;
+        }
+    }
+}
+/** 类型守卫：通过 isLLMError 标签识别 LLM 派生的 Error。 */
+export function isLLMError(err) {
+    return err instanceof Error && err.isLLMError === true;
+}
+export class APIKeyMissingError extends Error {
+    name = 'APIKeyMissingError';
+    isLLMError = true;
+    constructor(message) {
+        super(message);
+    }
+}
+export class LLMRateLimitError extends Error {
+    name = 'LLMRateLimitError';
+    isLLMError = true;
+    status = 429;
+    constructor(message) {
+        super(message);
+    }
+}
+export class LLMAuthError extends Error {
+    name = 'LLMAuthError';
+    isLLMError = true;
+    status;
+    constructor(status, message) {
+        super(message);
+        this.status = status;
+    }
+}
+export class LLMNetworkError extends Error {
+    name = 'LLMNetworkError';
+    isLLMError = true;
+    constructor(message, options) {
+        super(message);
+        if (options?.cause !== undefined) {
+            this.cause = options.cause;
+        }
+    }
+}
+export class LLMUnknownError extends Error {
+    name = 'LLMUnknownError';
+    isLLMError = true;
+    status;
+    constructor(message, options) {
+        super(message);
+        if (options?.cause !== undefined) {
+            this.cause = options.cause;
+        }
+        if (options?.status !== undefined)
+            this.status = options.status;
+    }
+}
+//# sourceMappingURL=types.js.map

package/dist/types.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAiNH,mDAAmD;AACnD,MAAM,OAAO,cAAe,SAAQ,KAAK;IACrB,IAAI,GAAG,gBAAyB,CAAC;IAC1C,UAAU,GAAG,IAAa,CAAC;IACpC,YAAY,OAAe,EAAE,OAA6B;QACxD,KAAK,CAAC,OAAO,CAAC,CAAC;QACf,IAAI,OAAO,EAAE,KAAK,KAAK,SAAS,EAAE,CAAC;YAChC,IAA4B,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QACtD,CAAC;IACH,CAAC;CACF;AAeD,6CAA6C;AAC7C,MAAM,UAAU,UAAU,CAAC,GAAY;IACrC,OAAO,GAAG,YAAY,KAAK,IAAK,GAAgC,CAAC,UAAU,KAAK,IAAI,CAAC;AACvF,CAAC;AAED,MAAM,OAAO,kBAAmB,SAAQ,KAAK;IACzB,IAAI,GAAG,oBAA6B,CAAC;IAC9C,UAAU,GAAG,IAAa,CAAC;IACpC,YAAY,OAAe;QACzB,KAAK,CAAC,OAAO,CAAC,CAAC;IACjB,CAAC;CACF;AAED,MAAM,OAAO,iBAAkB,SAAQ,KAAK;IACxB,IAAI,GAAG,mBAA4B,CAAC;IAC7C,UAAU,GAAG,IAAa,CAAC;IAC3B,MAAM,GAAG,GAAY,CAAC;IAC/B,YAAY,OAAe;QACzB,KAAK,CAAC,OAAO,CAAC,CAAC;IACjB,CAAC;CACF;AAED,MAAM,OAAO,YAAa,SAAQ,KAAK;IACnB,IAAI,GAAG,cAAuB,CAAC;IACxC,UAAU,GAAG,IAAa,CAAC;IAC3B,MAAM,CAAS;IACxB,YAAY,MAAc,EAAE,OAAe;QACzC,KAAK,CAAC,OAAO,CAAC,CAAC;QACf,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACvB,CAAC;CACF;AAED,MAAM,OAAO,eAAgB,SAAQ,KAAK;IACtB,IAAI,GAAG,iBAA0B,CAAC;IAC3C,UAAU,GAAG,IAAa,CAAC;IACpC,YAAY,OAAe,EAAE,OAA6B;QACxD,KAAK,CAAC,OAAO,CAAC,CAAC;QACf,IAAI,OAAO,EAAE,KAAK,KAAK,SAAS,EAAE,CAAC;YAChC,IAA4B,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QACtD,CAAC;IACH,CAAC;CACF;AAED,MAAM,OAAO,eAAgB,SAAQ,KAAK;IACtB,IAAI,GAAG,iBAA0B,CAAC;IAC3C,UAAU,GAAG,IAAa,CAAC;IAC3B,MAAM,CAAU;IACzB,YAAY,OAAe,EAAE,OAA8C;QACzE,KAAK,CAAC,OAAO,CAAC,CAAC;QACf,IAAI,OAAO,EAAE,KAAK,KAAK,SAAS,EAAE,CAAC;YAChC,IAA4B,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QACtD,CAAC;QACD,IAAI,OAAO,EAAE,MAAM,KAAK,SAAS;YAAE,IAAI,CAAC,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAClE,CAAC;CACF"}

package/package.json

@@ -0,0 +1,28 @@
+{
+  "name": "@deepwhale/llm",
+  "version": "1.0.0",
+  "type": "module",
+  "main": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js"
+    }
+  },
+  "files": [
+    "dist",
+    "src/pricing.default.toml"
+  ],
+  "dependencies": {
+    "@anthropic-ai/sdk": "^0.40.1",
+    "@deepwhale/core": "^1.0.0",
+    "smol-toml": "^1.6.1"
+  },
+  "scripts": {
+    "build": "tsc -b && node scripts/copy-toml.mjs",
+    "copy-toml": "node scripts/copy-toml.mjs",
+    "dev": "tsc -b --watch",
+    "clean": "rm -rf dist .tsbuildinfo"
+  }
+}

package/src/pricing.default.toml

@@ -0,0 +1,46 @@
+# deepwhale 默认 pricing (2026-06)
+# 用户可在 ~/.deepwhale/pricing.toml 覆盖 (--pricing-config <path> 后续可加, 1b.5 走 env)。
+# **per-model currency**: DeepSeek CNY, Anthropic USD。不在 UI 层做汇率换算。
+#
+# **当前生产 model 2 个** (Q 拍板 2026-06-03): flash + pro.
+# Only production DeepSeek models are supported. (Q 拍板 2026-06-04, Task 2 Gap 4 留痕)
+# 不再加 deepseek-chat / deepseek-reasoner 老 alias — 那些是 1a/1b era 历史,
+# 2026-07-24 后服务端 hard-fail, 老 session replay 走 R7 中间路径 (base 2 字段, cost 字段 absent).
+#
+# **信源唯一性**: DeepSeek pricing 永远以 https://api-docs.deepseek.com/quick_start/pricing
+# 为唯一信源。Q 在 2026-06-03 review 时贴出官方 "模型细节" 页截屏纠正我之前 commit 用的
+# 0.5/0.1/1.0 (来自 1b 时代印象 + 二手源, 未核对官方). 实际官方:
+# - cache_miss_per_m = input 缓存未命中 (1M tokens)
+# - cache_hit_per_m  = input 缓存命中 (1M tokens)
+# - completion_per_m = output (1M tokens)
+# 注意: 1b follow-up P1 改的 0.01/0.1 fix 是 **公式 1000× off** — pricing 数字本身的 0.1
+# 看上去像 0.1 ¥/M 但其实只是 0.5→0.1 cache hit 巧合小, 真正的官方是 0.02¥/M (5× off 真实价).
+
+[models.deepseek-v4-flash]
+cache_miss_per_m = 1.0     # ¥1 /M tokens (input 缓存未命中, 官方 api-docs.deepseek.com 2026-06)
+cache_hit_per_m  = 0.02    # ¥0.02 /M tokens (input 缓存命中, 官方)
+completion_per_m = 2.0     # ¥2 /M tokens (output, 官方)
+currency         = "CNY"
+
+[models.deepseek-v4-pro]
+cache_miss_per_m = 3.0     # ¥3 /M tokens (input 缓存未命中, 官方)
+cache_hit_per_m  = 0.025   # ¥0.025 /M tokens (input 缓存命中, 官方)
+completion_per_m = 6.0     # ¥6 /M tokens (output, 官方)
+currency         = "CNY"
+
+# Anthropic Claude Sonnet 4.5 (2026-06 official first-party API pricing)
+# https://platform.claude.com/docs/en/about-claude/pricing
+[models.claude-sonnet-4-5]
+cache_miss_per_m = 3.0    # $3 /M tokens (input)
+cache_hit_per_m  = 0.30   # $0.30 /M tokens (cache read, 1h TTL — 不是 5min ephemeral $0.06)
+completion_per_m = 15.0   # $15 /M tokens
+currency         = "USD"
+
+# Anthropic Claude Opus 4.5 (2026-06 official first-party API pricing)
+# P1 plan correction (2026-06-03): user 核对官方 pricing 页后纠正
+# 之前值 $24/$0.48/$120 是错的 (类似某种旧/特殊模式价格, 不适合 first-party default)
+[models.claude-opus-4-5]
+cache_miss_per_m = 5.0    # $5 /M
+cache_hit_per_m  = 0.50   # $0.50 /M (cache read, 1h TTL — 不是 5min ephemeral $0.48)
+completion_per_m = 25.0   # $25 /M (不是 $120)
+currency         = "USD"