@thincoder/core 0.10.12 → 0.10.13

package/provider/google.mjs

@@ -5,12 +5,14 @@
  */
 
 import { proxyFetch } from "../proxy.mjs"
+import { withUserAgent } from "../client-identity.mjs"
 import { requestWithRetry } from "./retry.mjs"
 import { effectiveFetchTimeoutMs, effectiveBodyIdleMs } from "./core.mjs"
 import { abortError, timeoutError } from "../abort-provenance.mjs"
 import { terminateBody } from "../stream-destroy.mjs"
 import { requestBudget } from "../config.mjs"
 import { DEFAULT_BODY_IDLE_MS } from "../config-defaults.mjs"
+import { assertPromptFits } from "./errors.mjs"
 
 /** OpenAI 语义 tool_choice → Gemini FunctionCallingConfig（2026-08-31 能力层）。 */
 function mapFunctionCallingConfig(choice) {
@@ -77,7 +79,7 @@ export function convertMessages(messages) {
 
 /** Build and send a Gemini chat request. Returns the same shape as core.mjs chat.
  *  2026-08-31 会诊 #6：接入 rateGate/recordRate（原实现完全绕过 TPM/RPM 闸门）。 */
-export async function chat(provider, { messages, tools, onToken, onReasoning, onWait, signal, toolChoice }) {
+export async function chat(provider, { messages, tools, onToken, onReasoning, onWait, signal, toolChoice, logCtx }) {
   const systemMessages = messages.filter((m) => m.role === "system")
   const contents = convertMessages(messages)
 
@@ -118,7 +120,10 @@ export async function chat(provider, { messages, tools, onToken, onReasoning, on
 
   // 会诊 #6：TPM/RPM 闸门 + 记账
   const { rateGate, recordRate, estimateRequestTokens } = await import("./rate.mjs")
-  const estimated = estimateRequestTokens({ messages })
+  // §6.17 发送面净判：估准面补 `tools`（原只投影 messages）；body 用 `maxOutputTokens`、无 `max_tokens` 键 ⇒ 减项恒 0
+  const estimated = estimateRequestTokens({ messages, tools })
+  const estPrompt = estimated - (body.max_tokens ?? 0)
+  assertPromptFits(provider, estPrompt, logCtx)
   await rateGate(provider, estimated, onWait, signal)
 
   // 2026-08-31：5xx/网络与 OpenAI 格式统一退避重试链（原完全无重试——Gemini 高峰
@@ -126,7 +131,7 @@ export async function chat(provider, { messages, tools, onToken, onReasoning, on
   const response = await requestWithRetry(
     () => proxyFetch(url, {
       method: "POST",
-      headers: { ...(provider.headers ?? {}), "Content-Type": "application/json" }, // 定制头展开（PROVIDER.md §21）——定制头在前、内置头在后：内置头胜出
+      headers: { ...withUserAgent(provider.headers), "Content-Type": "application/json" }, // 定制头 + 产品标识头（UA 不可被渠道覆盖——PROVIDER.md §6.17）；内置头在后故同名胜出
       body: JSON.stringify(body),
       // 2026-09-01：同 core.mjs——绝对墙钟废除；响应头阶段 fetchTimeoutMs（600s 默认），body 阶段读侧 idle 管
       signal,
@@ -145,6 +150,7 @@ export async function chat(provider, { messages, tools, onToken, onReasoning, on
       content: result.content,
       reasoning: result.reasoning,
       finishReason: result.finishReason ?? null, // §6.31.7 F-DA5（原三 transport 恒 null——字段值收正）
+      estimated: estPrompt, // §6.17 自校准采样（den 与净判同口径）
       usage: {
         prompt_tokens: usage.prompt_tokens ?? 0,
         completion_tokens: usage.completion_tokens ?? 0,
@@ -154,7 +160,7 @@ export async function chat(provider, { messages, tools, onToken, onReasoning, on
     }
   }
 
-  return { content: result.content, reasoning: result.reasoning, finishReason: result.finishReason ?? null, toolCalls: result.toolCalls }
+  return { content: result.content, reasoning: result.reasoning, finishReason: result.finishReason ?? null, estimated: estPrompt, toolCalls: result.toolCalls }
 }
 
 /** Gemini `finishReason` ⇒ 核统一 `finishReason`（§6.31.7 F-DA5——归一只喂既有分类；

package/provider/list-models.mjs

@@ -14,6 +14,7 @@
  * 解析保持防御性（字段缺失即跳过该项）。候选不过滤非对话模型（embedding 等——§16.6 #14）。
  */
 import { proxyFetch } from "../proxy.mjs"
+import { withUserAgent } from "../client-identity.mjs"
 import { resolveProviderSecrets } from "../env-ref.mjs"
 import { positiveInt, recordListedContext } from "../model-context.mjs"
 
@@ -74,7 +75,7 @@ function listedWindowOf(row) {
 
 async function listOpenai(provider, signal) {
   const url = `${provider.baseURL}/models`
-  const data = await fetchJson(provider, url, { ...(provider.headers ?? {}), Authorization: `Bearer ${provider.apiKey ?? ""}` }, signal)
+  const data = await fetchJson(provider, url, { ...withUserAgent(provider.headers), Authorization: `Bearer ${provider.apiKey ?? ""}` }, signal)
   // L3 写面的收集段（PROVIDER.md §6.15）：解析 `data[]` 时顺带收 `id → 窗口`——三字段按序取
   // 第一个**合法**值（R25 ∥ R26），非法相当于该字段缺席；三字段都不可得 = 该模型列表未报（不进表）。
   const ids = []
@@ -92,7 +93,7 @@ async function listOpenai(provider, signal) {
 }
 
 async function listAnthropic(provider, signal) {
-  const headers = { ...(provider.headers ?? {}), "x-api-key": provider.apiKey ?? "", "anthropic-version": ANTHROPIC_VERSION }
+  const headers = { ...withUserAgent(provider.headers), "x-api-key": provider.apiKey ?? "", "anthropic-version": ANTHROPIC_VERSION }
   const base = `${provider.baseURL}/models?limit=1000`
   const out = []
   let afterId = null
@@ -108,7 +109,7 @@ async function listAnthropic(provider, signal) {
 }
 
 async function listGoogle(provider, signal) {
-  const headers = { ...(provider.headers ?? {}) }
+  const headers = { ...withUserAgent(provider.headers) }
   const base = `${provider.baseURL}/models?key=${encodeURIComponent(provider.apiKey ?? "")}&pageSize=1000`
   const out = []
   let pageToken = null

package/provider/responses.mjs

@@ -13,10 +13,11 @@
  */
 import { effectiveFetchTimeoutMs, effectiveBodyIdleMs } from "./core.mjs"
 import { proxyFetch } from "../proxy.mjs"
+import { withUserAgent } from "../client-identity.mjs"
 import { requestWithRetry } from "./retry.mjs"
 import { rateGate, recordRate, estimateRequestTokens } from "./rate.mjs"
 import { buildBody, normalizeUsage } from "./responses-request.mjs"
-import { classifyResponsesErrorCode } from "./errors.mjs"
+import { classifyResponsesErrorCode, assertPromptFits } from "./errors.mjs"
 // 同名 re-export（2026-09-29 structure-split-2 · 台账 #620——缝 = 同名再出口）：请求构造面出档 `provider/responses-request.mjs`；
 // `buildBody` / `isStoreRequiredHost` / `builtinToolsFor` 三名既有导出经本档转口——消费者 import 面零改。
 export { buildBody, isStoreRequiredHost, builtinToolsFor } from "./responses-request.mjs"
@@ -223,15 +224,18 @@ function handleEvent(ev, h) {
 }
 
 /** 主入口：请求 + 链状态推进（与 core.mjs 的 chat 同形返回）。 */
-export async function chat(provider, { messages, tools, onToken, onReasoning, onWait, signal, toolChoice, stateful }) {
+export async function chat(provider, { messages, tools, onToken, onReasoning, onWait, signal, toolChoice, stateful, logCtx }) {
   const { body: reqBody, previousResponseId, warnings, newChain } = buildBody(provider, messages, tools, { toolChoice, stateful })
   const body = { ...reqBody, ...(previousResponseId ? { previous_response_id: previousResponseId } : {}) }
 
   // rateGate/recordRate 对齐 core（responses body 无 messages 键——按本地全量 messages 估算）
-  const estimated = estimateRequestTokens({ messages })
+  // §6.17 发送面净判：估准面补 `tools`（原只投影 messages）；body 用 `max_output_tokens`、无 `max_tokens` 键 ⇒ 减项恒 0
+  const estimated = estimateRequestTokens({ messages, tools })
+  const estPrompt = estimated - (body.max_tokens ?? 0)
+  assertPromptFits(provider, estPrompt, logCtx)
   await rateGate(provider, estimated, onWait, signal)
-  // 定制头展开（PROVIDER.md §21）：provider.headers 在前、内置头在后——同名内置头胜出（三处 fetch 共用）
-  const headers = { ...(provider.headers ?? {}), "Content-Type": "application/json", Authorization: `Bearer ${provider.apiKey}` }
+  // 定制头 + 产品标识头（UA 不可被渠道覆盖——PROVIDER.md §6.17）：定制头在前、内置头在后——同名内置头胜出（三处 fetch 共用）
+  const headers = { ...withUserAgent(provider.headers), "Content-Type": "application/json", Authorization: `Bearer ${provider.apiKey}` }
 
   // round2 复验 #1（2026-08-31）：retry 层对 4xx 非可重试是 throw 而非返回——
   // requestWithRetry 从不返回 400/404 响应 → 下方 isChainInvalidError 分支原来不可达（D6 死代码）。
@@ -287,11 +291,17 @@ export async function chat(provider, { messages, tools, onToken, onReasoning, on
       { signal, onWait, buildMessage: (status, text) => `Responses API error ${status}: ${text}` },
     )
   }
-  return finish(provider, response, { onToken, onReasoning, signal, newChain, warnings, estimated })
+  return finish(provider, response, {
+    onToken, onReasoning, signal, newChain, warnings, estimated,
+    // §6.17 自校准口径：链式（`previous_response_id`）请求的 body 只带增量（`responses-request.mjs`
+    // `body.input = newOutputs`），而分母是本地全量估算 ⇒ 两者口径不同，本轮不采样。
+    estPrompt: previousResponseId ? undefined : estPrompt,
+  })
 }
 
-async function finish(provider, response, { onToken, onReasoning, signal, newChain, warnings, estimated }) {
+async function finish(provider, response, { onToken, onReasoning, signal, newChain, warnings, estimated, estPrompt }) {
   const result = await parseStream(response, { onToken, onReasoning, signal })
+  result.estimated = estPrompt // §6.17 自校准采样（den 与净判同口径）
   recordRate(provider, estimated, result.usage)
   // 链状态推进：completed 事件里 response.id 供同一 turn 后续增量；
   // 截断/失败/无 id（部分端点不回传）→ 链作废（后续全量，正确性优先）。

package/session-lifecycle.mjs

@@ -322,6 +322,9 @@ export function resetSessionState(agent) {
   agent._runStartHistoryLen = 0
   agent._lastPromptTokens = null
   agent._usageAtLen = null
+  // §6.17 估算自校准：新会话无样本（比 = 1）——跨会话不继承旧模型的估算偏差
+  agent._tok = null
+  agent._tokRatio = 1
   agent.planMode = false
   agent.goal = null
   agent._pendingReminders = []

package/token-window.mjs

@@ -138,11 +138,13 @@ export function contextUsage(agent, extras = {}) {
   const overhead = system + tools
   const historyTokens = estimateTokens(history)
   const measured = agent?._lastPromptTokens != null
+  // §6.3 / §6.17 估算自校准：估算值乘 `_tokRatio`；`_lastPromptTokens` 是服务端真值，不乘。
+  const ratio = agent?._tokRatio ?? 1
   return {
     basis: measured ? "measured" : "estimated",
     total: measured
-      ? agent._lastPromptTokens + estimateTokens(history.slice(agent._usageAtLen ?? history.length))
-      : historyTokens + overhead,
+      ? agent._lastPromptTokens + Math.round(estimateTokens(history.slice(agent._usageAtLen ?? history.length)) * ratio)
+      : Math.round((historyTokens + overhead) * ratio),
     history: historyTokens,
     system,
     tools,
@@ -153,6 +155,25 @@ export function contextUsage(agent, extras = {}) {
   }
 }
 
+/**
+ * 估算自校准采样（§6.17「估算自校准」）——池化比：`num += usage.prompt_tokens`、`den += estPrompt`
+ * （分子分母同取输入侧口径；`estPrompt` 由发送面随 result 带回，与净判同一个值）。
+ * `key = format:model` 变 ⇒ 归零重采（切模型自清，不接四个刷新点）；`_tok` / `_tokRatio` 只由本函数写，
+ * `/new` 与 run 装配复位。无样本 ⇒ 比 = 1（估算维持原口径）。
+ * 作用面 = `contextUsage` 两支承（§6.3）；实测基线 `_lastPromptTokens` 不乘。
+ */
+export function recordTokenSample(agent, provider, estPrompt, usage) {
+  const prompt = Number(usage?.prompt_tokens)
+  if (!agent || !(prompt > 0) || !(estPrompt > 0)) return
+  const key = `${provider?.format ?? "openai"}:${provider?.model ?? ""}`
+  if (agent._tok?.key !== key) agent._tok = { key, num: 0, den: 0 }
+  agent._tok.num += prompt
+  agent._tok.den += estPrompt
+  // 下界 = 1：已知偏差方向是估算偏低；比 < 1 会把全部估算支缩小、重开「该压没压」原始故障，
+  // 保守方向不取（仅一侧校准，零新魔数）。
+  agent._tokRatio = Math.max(1, agent._tok.num / agent._tok.den)
+}
+
 /**
  * 状态行百分比（§6.16.4「与状态行 ctx% 的单源关系」）——与 CLI 状态行（`render-frame.mjs` 式）与 VSC `ctxPercentForHistory` **同式**；
  * stats 报出的就是用户看到的那个百分比（不新增第三口径）；空历史 ⇒ 0（状态行的 `ctxPct > 0` 显示门在端侧）。