dsh-any-deepseek 0.0.10 → 0.0.12

package/lib/adapter.d.ts

@@ -27,6 +27,7 @@ import { LlmAdapter } from '@deepseek-ai/dsh-llm';
 import type { GenerateOptions, LlmModelInfo, LlmProviderInfo, LlmResolvedModelInfo, PreparedAdapterCall, StreamChunk, TokenUsage } from '@deepseek-ai/dsh-llm';
 import { credentialRef } from '@deepseek-ai/dsh-credentials';
 import { type Api, type AssistantMessage, type AssistantMessageEvent, type Context as PiContext, type Model } from '@earendil-works/pi-ai';
+import { openAICompletionsApi } from '@earendil-works/pi-ai/api/openai-completions.lazy';
 import { type ChannelModality } from './presets.js';
 import type { ResolvedChannel } from './channels.js';
 import type { LedgerEntry } from './usage-ledger.js';
@@ -163,6 +164,15 @@ export declare function requestHeadersFor(channel: ResolvedChannel, sessionId: u
  * @returns a pi-ai chat model describing that relay's view of the model.
  */
 export declare function buildPiModel(channel: ResolvedChannel): Model<Api>;
+/** The shape `createProvider` accepts per protocol (an object with `stream`/`streamSimple`). */
+type ApiStreams = ReturnType<typeof openAICompletionsApi>;
+/**
+ * Build the per-protocol implementation map for the pi-ai provider.
+ *
+ * Every supported protocol is constructed eagerly because the factories are lazy internally
+ * (they load the SDK on first use), so this costs one closure each, not one SDK download.
+ */
+export declare function apiImplementations(): Record<string, ApiStreams>;
 /** One resolved image, keyed by attachment id. */
 export type ImageBytes = Map<string, {
     data: string;

package/lib/client.js

@@ -115,6 +115,32 @@ window.__ModuleLoader__.load({
 			if (value < 1e6) return `${(value / 1e3).toFixed(value < 1e4 ? 1 : 0)}K`;
 			return `${(value / 1e6).toFixed(value < 1e7 ? 2 : 1)}M`;
 		}
+		/**
+		* Format a throughput figure, or an em dash when it is unmeasurable.
+		*
+		* 🔴 `null` is NOT zero: a group with no first-token measurement has no speed, and printing `0`
+		* would claim the relay produced nothing — a different and much worse statement. The em dash
+		* says "not measured", which is the truth.
+		* @param value - tokens per second, or null/undefined.
+		* @returns e.g. `42 tok/s`, or `—`.
+		*/
+		function formatSpeed(value) {
+			if (value === null || value === void 0 || !Number.isFinite(value) || value <= 0) return "—";
+			return `${value >= 100 ? value.toFixed(0) : value.toFixed(1)} tok/s`;
+		}
+		/**
+		* Format a duration in ms compactly.
+		* @param value - milliseconds, or null/undefined.
+		* @returns e.g. `820ms` / `1.4s` / `2m 05s`, or `—`.
+		*/
+		function formatDuration(value) {
+			if (value === null || value === void 0 || !Number.isFinite(value) || value < 0) return "—";
+			if (value < 1e3) return `${Math.round(value)}ms`;
+			const seconds = value / 1e3;
+			if (seconds < 60) return `${seconds.toFixed(1)}s`;
+			const minutes = Math.floor(seconds / 60);
+			return `${minutes}m ${String(Math.round(seconds - minutes * 60)).padStart(2, "0")}s`;
+		}
 		//#endregion
 		//#region \0ad-css:src/client/SettingsPage
 		const css = "/* Panel styles.\n *\n * 🔴 There are NO `--dsw-alias-*` design tokens on this host (measured 2026-10-21: the document\n * declares 29 custom properties and not one is a colour). This file used to reference a dozen of\n * them, every one falling back to a hardcoded LIGHT value (`rgba(0,0,0,0.12)` and friends) — so\n * on the dark theme the borders and dividers were effectively invisible.\n *\n * ⇒ Derive everything from `currentColor`, which the host sets per theme. `color-mix` gives the\n * low-contrast borders and fills that a token would have provided, and it follows the theme\n * automatically instead of assuming one palette.\n */\n\n.ad-page {\n  display: flex;\n  flex-direction: column;\n  gap: 20px;\n  padding: 4px 2px 12px;\n}\n\n/* One place to change the two theme-derived values the whole panel uses. */\n.ad-page {\n  --ad-line: color-mix(in srgb, currentColor 18%, transparent);\n  --ad-line-strong: color-mix(in srgb, currentColor 32%, transparent);\n  --ad-fill-subtle: color-mix(in srgb, currentColor 7%, transparent);\n  /* The current row's fill: visible against both themes because it is derived from the text\n     colour the host already picked, not from an assumed palette. */\n  --ad-fill-current: color-mix(in srgb, currentColor 11%, transparent);\n  --ad-danger: #e5534b;\n}\n\n.ad-head {\n  display: flex;\n  flex-direction: column;\n  gap: 6px;\n}\n\n.ad-title {\n  margin: 0;\n  font-size: 15px;\n  font-weight: 600;\n}\n\n.ad-intro {\n  margin: 0;\n  font-size: 12px;\n  line-height: 1.6;\n  opacity: 0.72;\n}\n\n.ad-error {\n  border: 1px solid var(--ad-danger);\n  border-radius: 6px;\n  padding: 8px 10px;\n  font-size: 12px;\n}\n\n.ad-section {\n  display: flex;\n  flex-direction: column;\n  gap: 10px;\n}\n\n.ad-sectionTitle {\n  margin: 0;\n  font-size: 12px;\n  font-weight: 600;\n  letter-spacing: 0.04em;\n  text-transform: uppercase;\n  opacity: 0.6;\n}\n\n.ad-rows {\n  list-style: none;\n  margin: 0;\n  padding: 0;\n  display: flex;\n  flex-direction: column;\n  gap: 6px;\n}\n\n.ad-row {\n  display: flex;\n  align-items: flex-start;\n  justify-content: space-between;\n  gap: 12px;\n  border: 1px solid var(--ad-line);\n  border-radius: 8px;\n  padding: 10px 12px;\n  /*\n   * The left edge is reserved on EVERY row so the current one does not shift its content when\n   * selected — a row that reindents on click makes the list feel like it moved.\n   */\n  border-left: 3px solid transparent;\n}\n\n/*\n * 🔴 The current channel is marked three ways, not one.\n *\n * It used to be a small badge plus a barely-different border, and the owner could not tell at a\n * glance which relay was live — glanceability is the whole point of a single-choice list. So:\n * a colour bar down the left edge, a stronger border, and a tinted fill. All derived from\n * `currentColor`, because this host ships no colour tokens (measured: 29 custom properties, none\n * of them a colour), so any hardcoded value would assume the wrong theme.\n */\n.ad-row-current {\n  border-color: var(--ad-line-strong);\n  border-left-color: currentColor;\n  background: var(--ad-fill-current);\n}\n\n/* The 当前 badge keeps its place as the textual confirmation, for when colour is not enough. */\n.ad-row-current .ad-badge {\n  border-color: currentColor;\n  opacity: 1;\n  font-weight: 600;\n}\n\n/* A rename button must not look like the row's own selection affordance. */\n.ad-renameBtn {\n  margin-left: auto;\n  flex: none;\n}\n\n.ad-rowMain {\n  display: flex;\n  flex-direction: column;\n  gap: 4px;\n  min-width: 0;\n}\n\n.ad-rowTitleLine {\n  display: flex;\n  align-items: center;\n  gap: 6px;\n  font-size: 13px;\n  font-weight: 500;\n}\n\n.ad-rowTitle {\n  white-space: nowrap;\n}\n\n.ad-badge {\n  font-size: 10px;\n  padding: 1px 6px;\n  border-radius: 999px;\n  border: 1px solid currentColor;\n  opacity: 0.8;\n}\n\n.ad-warn {\n  color: #b7791f;\n  cursor: help;\n}\n\n.ad-rowMeta {\n  display: flex;\n  flex-wrap: wrap;\n  align-items: center;\n  gap: 6px;\n  font-size: 11px;\n  opacity: 0.66;\n  min-width: 0;\n}\n\n.ad-mono {\n  font-family: ui-monospace, SFMono-Regular, Menlo, monospace;\n  word-break: break-all;\n}\n\n.ad-sep {\n  opacity: 0.5;\n}\n\n.ad-rowActions {\n  display: flex;\n  align-items: center;\n  gap: 8px;\n  flex-shrink: 0;\n}\n\n.ad-btn {\n  font-size: 11px;\n  padding: 4px 10px;\n  border-radius: 6px;\n  border: 1px solid var(--ad-line);\n  background: transparent;\n  color: inherit;\n  cursor: pointer;\n}\n\n.ad-btn:disabled {\n  opacity: 0.45;\n  cursor: default;\n}\n\n.ad-btn-quiet {\n  opacity: 0.8;\n}\n\n.ad-probe {\n  font-size: 11px;\n  padding-top: 2px;\n}\n\n.ad-probe-ok {\n  color: #2f855a;\n}\n\n.ad-probe-bad {\n  color: #c53030;\n  word-break: break-word;\n}\n\n.ad-hint,\n.ad-empty,\n.ad-note {\n  margin: 0;\n  font-size: 11px;\n  line-height: 1.6;\n  opacity: 0.66;\n}\n\n.ad-dot {\n  width: 8px;\n  height: 8px;\n  border-radius: 50%;\n  flex-shrink: 0;\n  display: inline-block;\n}\n\n.ad-dot-on {\n  background: #2f855a;\n}\n\n.ad-dot-off {\n  border: 1px solid currentColor;\n  opacity: 0.5;\n}\n\n.ad-dot-unknown {\n  background: linear-gradient(90deg, currentColor 50%, transparent 50%);\n  opacity: 0.5;\n}\n\n/* The channel row is a radio \"pick\" zone: the whole left side is clickable, and the radio sits\n   at the start of the line so \"which one is live\" is readable down a single column. */\n.ad-rowPick {\n  display: flex;\n  align-items: flex-start;\n  gap: 10px;\n  flex: 1;\n  min-width: 0;\n  cursor: pointer;\n}\n\n.ad-rowPick input[type='radio'] {\n  flex: none;\n  /* Optically centred against the first text line rather than the row box. */\n  margin-top: 3px;\n  accent-color: var(--ad-line-strong);\n  cursor: pointer;\n}\n\n/* The switch is gone: selection is the only axis (see `channels.ts`). Keeping the rules would\n   invite someone to bring the second axis back. */\n\n.ad-tabs {\n  display: flex;\n  gap: 6px;\n}\n\n.ad-tab {\n  font-size: 11px;\n  padding: 3px 10px;\n  border-radius: 999px;\n  border: 1px solid var(--ad-line);\n  background: transparent;\n  color: inherit;\n  cursor: pointer;\n}\n\n.ad-tab-on {\n  border-color: var(--ad-line-strong);\n  font-weight: 600;\n}\n\n.ad-tab-refresh {\n  margin-left: auto;\n  opacity: 0.75;\n}\n\n.ad-table {\n  width: 100%;\n  border-collapse: collapse;\n  font-size: 11px;\n}\n\n.ad-table th,\n.ad-table td {\n  padding: 5px 8px;\n  border-bottom: 1px solid var(--ad-line);\n  text-align: left;\n}\n\n.ad-table th {\n  font-weight: 500;\n  opacity: 0.6;\n}\n\n.ad-num {\n  text-align: right;\n  font-family: ui-monospace, SFMono-Regular, Menlo, monospace;\n}\n\n.ad-total td {\n  font-weight: 600;\n  border-bottom: none;\n}\n\n/* ── add-channel form ─────────────────────────────────────────────────────── */\n\n.ad-addForm {\n  display: flex;\n  flex-direction: column;\n  gap: 8px;\n  border: 1px dashed var(--ad-line);\n  border-radius: 8px;\n  padding: 10px 12px;\n}\n\n.ad-field {\n  display: flex;\n  flex-direction: column;\n  gap: 3px;\n  font-size: 11px;\n}\n\n.ad-field > span {\n  opacity: 0.7;\n}\n\n.ad-input {\n  font-size: 11px;\n  padding: 4px 8px;\n  border-radius: 6px;\n  border: 1px solid var(--ad-line);\n  background: transparent;\n  color: inherit;\n  font-family: inherit;\n}\n\n/*\n * 🔴 A native `<select>` cannot be themed by leaving the background transparent.\n *\n * The popup list is drawn by the browser, not by the closed control's box: the options inherit\n * nothing from it, so `background: transparent` made them render in the platform default (light)\n * palette against this dark theme — reported by the owner and measured on the real host.\n *\n * 🔴 SECOND correction (measured 2026-10-21): this file used `var(--dsw-alias-color-*, #fallback)`\n * throughout, and **none of those variables exist**. The host declares only 29 custom properties\n * in total and not one is a colour token, so every `var()` here was falling back — and a fallback\n * that happens to be set inside a shorthand can be silently dropped (the select's background\n * computed to `rgba(0,0,0,0)`, i.e. still transparent, which is exactly the bug).\n *\n * ⇒ Use only values this file controls, derived from `currentColor` (the host sets a light text\n * colour on both themes) plus explicit translucent layers. No host token is assumed.\n */\nselect.ad-input {\n  /* Tells the engine which palette to use for the popup. This part DID work. */\n  color-scheme: dark light;\n  /*\n   * Paint the closed box with a layer that follows the surrounding text colour, so it matches\n   * whichever theme is active instead of hardcoding a dark value into a light theme.\n   */\n  background-color: color-mix(in srgb, currentColor 8%, transparent);\n  color: inherit;\n  cursor: pointer;\n  padding-right: 22px;\n  /* Room for the chevron drawn below, so text never runs under it. */\n  appearance: none;\n  background-image: linear-gradient(45deg, transparent 50%, currentColor 50%),\n    linear-gradient(135deg, currentColor 50%, transparent 50%);\n  background-position: calc(100% - 12px) calc(50% + 1px), calc(100% - 8px) calc(50% + 1px);\n  background-size: 4px 4px, 4px 4px;\n  background-repeat: no-repeat;\n}\n\nselect.ad-input:hover:not(:disabled) {\n  background-color: color-mix(in srgb, currentColor 14%, transparent);\n}\n\nselect.ad-input:focus-visible {\n  outline: 2px solid currentColor;\n  outline-offset: 1px;\n}\n\n/*\n * The popup itself follows `color-scheme`; these keep the option rows explicit for engines that\n * honour per-option colours. `color-scheme: dark` above is what actually stops the light-on-dark\n * list, because the popup is native chrome rather than a styled element.\n */\nselect.ad-input option {\n  background-color: Canvas;\n  color: CanvasText;\n}\n\n.ad-fieldError {\n  margin: 0;\n  font-size: 11px;\n  color: #c53030;\n}\n\n/* ── key row ──────────────────────────────────────────────────────────────── */\n\n/*\n * The key line states a FACT and offers the fix beside it: \"configured / not configured\",\n * the reference name the secret is filed under, and one text button that opens the input.\n *\n * The value is never shown — the harness credential seam answers \"configured?\" and refuses to\n * answer anything else, so there is nothing to render even if we wanted to.\n */\n.ad-keyLine {\n  display: flex;\n  flex-wrap: wrap;\n  align-items: center;\n  gap: 6px;\n  font-size: 11px;\n  padding-top: 2px;\n}\n\n.ad-keyOn {\n  color: #2f855a;\n}\n\n.ad-keyOff {\n  color: #b7791f;\n}\n\n/* The reference name, not a secret: monospace so it lines up and misreads as a name, not a value. */\n.ad-keyName {\n  font-size: 10px;\n  padding: 1px 5px;\n  border-radius: 4px;\n  background: var(--ad-fill-subtle);\n  opacity: 0.8;\n}\n\n/* A text-shaped action. Styled as a link because it opens an inline editor, not a dialog. */\n.ad-link {\n  font-size: 11px;\n  padding: 0;\n  border: none;\n  background: none;\n  color: inherit;\n  text-decoration: underline;\n  text-underline-offset: 2px;\n  opacity: 0.75;\n  cursor: pointer;\n}\n\n.ad-link:hover {\n  opacity: 1;\n}\n\n/* The inline key editor: one masked field and one button, on the same line as the fact above. */\n.ad-keyForm {\n  display: flex;\n  align-items: center;\n  gap: 6px;\n  padding-top: 4px;\n}\n\n.ad-keyForm .ad-input {\n  flex: 1;\n  min-width: 0;\n}\n\n/* A destructive action is quiet until hovered: it should be findable, not inviting. */\n.ad-btn-danger {\n  opacity: 0.6;\n}\n\n.ad-btn-danger:hover:not(:disabled) {\n  opacity: 1;\n  border-color: var(--ad-danger);\n  color: var(--ad-danger);\n}\n";
@@ -767,6 +793,16 @@ window.__ModuleLoader__.load({
 									/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", {
 										className: "ad-num",
 										children: "合计"
+									}),
+									/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", {
+										className: "ad-num",
+										title: "生成阶段速度（不含首字等待）",
+										children: "速度"
+									}),
+									/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", {
+										className: "ad-num",
+										title: "首字平均等待时间",
+										children: "首字"
 									})
 								] }) }), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("tbody", { children: [stats.rows.map((row) => /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("tr", { children: [
 									/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: row.label ?? row.key }),
@@ -797,6 +833,14 @@ window.__ModuleLoader__.load({
 									/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", {
 										className: "ad-num",
 										children: formatTokens(row.aggregate.total)
+									}),
+									/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", {
+										className: "ad-num",
+										children: formatSpeed(row.tokensPerSecond)
+									}),
+									/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", {
+										className: "ad-num",
+										children: formatDuration(row.firstTokenMs)
 									})
 								] }, row.key)), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("tr", {
 									className: "ad-total",
@@ -829,6 +873,14 @@ window.__ModuleLoader__.load({
 										/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", {
 											className: "ad-num",
 											children: formatTokens(stats.total.total)
+										}),
+										/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", {
+											className: "ad-num",
+											children: formatSpeed(stats.totalSpeed)
+										}),
+										/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", {
+											className: "ad-num",
+											children: formatDuration(stats.totalFirstTokenMs)
 										})
 									]
 								})] })]

package/lib/http.d.ts

@@ -59,6 +59,21 @@ export interface WebServerLike {
         handler: (req: IncomingMessage, res: ServerResponse) => void | Promise<void>;
     }) => unknown;
 }
+/**
+ * The URL path and body a probe uses for one protocol.
+ *
+ * 🔴 **The probe used to hardcode `/chat/completions`.** With the Responses API now selectable
+ * per channel (owner directive 2026-10-22: 「优先支持并使用 responses api」), that would have
+ * probed the WRONG endpoint for every Responses channel and reported a verdict about a protocol
+ * the channel does not use. "A probe must send what a real request sends" already cost one round
+ * over headers; the rule applies unchanged to the path.
+ * @param channel - the channel being probed.
+ * @returns the path suffix and request body for its protocol.
+ */
+export declare function probeRequest(channel: ResolvedChannel): {
+    path: string;
+    body: unknown;
+};
 /**
  * Probe one channel with a minimal real request.
  *
@@ -83,6 +98,25 @@ export declare function probeChannel(channel: ResolvedChannel, key: string, sess
     message?: string;
     responded?: string;
 }>;
+/**
+ * The body's `error` value, when it is actually an error.
+ *
+ * Returns `undefined` for a missing field AND for an explicit `null` — both mean "no error".
+ * A non-null value (object or string) is returned as JSON text for the panel to show.
+ * @param text - the raw response body.
+ * @returns the error as text, or undefined when the body reports no error.
+ */
+export declare function errorFieldValue(text: string): string | undefined;
+/**
+ * What a successful probe response says the relay called its model.
+ *
+ * Three protocols, three response shapes — the old reader understood only completions
+ * (`choices[0].message.content`), so a Responses or Anthropic 200 came back with an empty verdict
+ * and the panel showed a blank success. A verdict nobody can read is not a verdict.
+ * @param text - the raw response body.
+ * @returns the reported model name, or the first output text; empty when neither is present.
+ */
+export declare function reportedName(text: string): string;
 /**
  * Register the panel's routes.
  * @param ctx - plugin context (only used to bind the disposer when the host offers one).

package/lib/index.js

@@ -5,6 +5,8 @@ import "@deepseek-ai/dsh-credentials";
 import { createHash } from "node:crypto";
 import { createModels, createProvider } from "@earendil-works/pi-ai";
 import { openAICompletionsApi } from "@earendil-works/pi-ai/api/openai-completions.lazy";
+import { openAIResponsesApi } from "@earendil-works/pi-ai/api/openai-responses.lazy";
+import { anthropicMessagesApi } from "@earendil-works/pi-ai/api/anthropic-messages.lazy";
 import { appendFile, mkdir, readFile, readdir } from "node:fs/promises";
 //#region src/config.ts
 /**
@@ -250,6 +252,22 @@ const OPENCODE_SESSION_FALLBACK = "dsh-any-deepseek";
 *
 * The id is what a channel entry stores (`preset: opencode-go`), so it is part of the user's
 * configuration surface: renaming one breaks saved channels. Add, do not rename.
+*
+* ## Why every preset prefers the Responses API
+*
+* Owner directive (2026-10-22): 「对于各家查阅资料，优先支持并使用 responses api(这个api更先进)」.
+* Responses is the newer surface: it carries reasoning items as first-class output, reports
+* `output_tokens_details.reasoning_tokens`, and supports server-side state/tool primitives the
+* completions shape cannot express.
+*
+* 🔴 **But "prefer" is not "assume".** A relay that answers `/responses` with 404 would break
+* every request if this were a blanket switch, so the rule is: a preset declares Responses ONLY
+* where the endpoint is documented to serve it; everything else keeps completions, which is the
+* lowest common denominator every relay speaks. `api` stays a per-channel field, so a user can
+* override any of them in the panel without waiting for a release.
+*
+* DeepSeek's own docs confirm `/responses` with `deepseek-flash`, including `input_image` — so
+* the official preset is the first to move and is the reference case.
 */
 const CHANNEL_PRESETS = {
 	"opencode-go": {
@@ -286,7 +304,7 @@ const CHANNEL_PRESETS = {
 	},
 	"deepseek-official": {
 		displayName: "DeepSeek Official",
-		api: "openai-completions",
+		api: "openai-responses",
 		baseURL: "https://api.deepseek.com",
 		model: "deepseek-flash",
 		...V41_FLASH_CAPS
@@ -784,6 +802,38 @@ function buildPiModel(channel) {
 	};
 }
 /**
+* The transport implementation for one wire protocol.
+*
+* 🔴 **This table is the fix for a silent hole.** The provider was built with
+* `api: openAICompletionsApi()` — a SINGLE implementation. `createProvider` treats a single
+* implementation as "serves every model" (`models.js:529-532`: `single = input.api &&
+* typeof input.api.stream === 'function'`), so `model.api` was never consulted and every channel
+* went out over the completions wire **no matter what its `api` field said**. A channel
+* configured for `openai-responses` therefore produced completions requests: the setting was
+* accepted, shown in the panel, and silently ignored. Same class as the other defects this
+* plugin has paid for — accepted, rendered, and not in effect.
+*
+* Passing a MAP instead makes `createProvider` dispatch on `model.api` (`models.js:555`
+* `apiFor`), and a model naming an unimplemented protocol fails loudly with
+* `has no API implementation for "<api>"` rather than silently using the wrong one.
+*/
+const API_FACTORIES = {
+	"openai-completions": openAICompletionsApi,
+	"openai-responses": openAIResponsesApi,
+	"anthropic-messages": anthropicMessagesApi
+};
+/**
+* Build the per-protocol implementation map for the pi-ai provider.
+*
+* Every supported protocol is constructed eagerly because the factories are lazy internally
+* (they load the SDK on first use), so this costs one closure each, not one SDK download.
+*/
+function apiImplementations() {
+	const table = {};
+	for (const [protocol, factory] of Object.entries(API_FACTORIES)) table[protocol] = factory();
+	return table;
+}
+/**
 * Collect the bytes for every image in a request.
 *
 * 🔴 A separate pass, because the conversion itself is synchronous and the store is not. This is
@@ -1272,6 +1322,20 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 			usage: null
 		};
 		let recorded = false;
+		/**
+		* Time to the first streamed block, in ms from request start.
+		*
+		* 🔴 Owner request (2026-10-22): 「能加上 token 速度统计吗」. Throughput cannot be recovered
+		* from the totals alone — `output / totalMs` folds in queueing and time-to-first-token, so a
+		* relay with slow first-token but fast generation looks identical to a uniformly slow one.
+		* Recording the split makes the two visible separately, and both numbers are already
+		* measurable at this boundary: the first `block-start` is the moment generation began.
+		*
+		* Guards that matter: it is measured ONCE (first block only), and only for a call that got
+		* that far — an error before any block leaves it undefined rather than reporting 0, which
+		* would read as "instant first token".
+		*/
+		let firstTokenMs;
 		const finish = async () => {
 			if (recorded || channel === void 0) return;
 			recorded = true;
@@ -1281,6 +1345,7 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 				model: channel.model,
 				...options.sessionId === void 0 ? {} : { sessionId: String(options.sessionId) },
 				...options.purpose === void 0 ? {} : { purpose: options.purpose },
+				...firstTokenMs === void 0 ? {} : { firstTokenMs },
 				ok: outcome.ok,
 				ms: Math.max(0, Math.round(this.deps.elapsed() - startedAt)),
 				usage: outcome.usage === null ? null : {
@@ -1331,7 +1396,7 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 				resolve: async () => ({ auth: { apiKey: key } })
 			} },
 			models: [model],
-			api: openAICompletionsApi()
+			api: apiImplementations()
 		}));
 		/**
 		* Per-request headers.
@@ -1363,7 +1428,10 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 			const events = models.streamSimple(model, toPiContext(options, images), streamOptions);
 			for await (const chunk of toStreamChunks(events, (observed) => {
 				outcome = observed;
-			})) yield chunk;
+			})) {
+				if (firstTokenMs === void 0 && chunk.type === "block-start") firstTokenMs = Math.max(0, Math.round(this.deps.elapsed() - startedAt));
+				yield chunk;
+			}
 		} catch (error) {
 			const message = error instanceof Error ? error.message : String(error);
 			outcome = {
@@ -1433,6 +1501,33 @@ function failureChunk(failure) {
 * exposes the sum explicitly, because a table that showed only `input` would read as if input
 * were being under-reported.
 */
+/**
+* Output tokens per second generated, or null when unmeasurable.
+*
+* 🔴 **Why not `output / total` (owner request 2026-10-22: 「能加上 token 速度统计吗」).**
+* One ratio collapses queueing and generation together, so a relay that thinks for 3s and then
+* streams fast looks identical to a uniformly slow one — and those two call for different fixes
+* (switch relay vs. shrink the prompt). The split is measured per request and this divides only
+* by the generating half.
+*
+* Averaged as a RATIO OF SUMS rather than a mean of per-request speeds: a mean lets one
+* five-token reply dominate, while the sum ratio answers "across this group, how fast did tokens
+* actually come out".
+* @param aggregate - one group's figures.
+* @returns tokens per second, or null when nothing was timed or nothing was generated.
+*/
+function outputTokensPerSecond(aggregate) {
+	if (aggregate.generationMs <= 0 || aggregate.timedOutput <= 0) return null;
+	return aggregate.timedOutput * 1e3 / aggregate.generationMs;
+}
+/**
+* Mean time to first token, in ms, or null when nothing was timed.
+* @param aggregate - one group's figures.
+*/
+function meanFirstTokenMs(aggregate) {
+	if (aggregate.timed <= 0) return null;
+	return aggregate.ttftMs / aggregate.timed;
+}
 /** Empty aggregate, used both as an accumulator seed and as a zero row. */
 function emptyAggregate() {
 	return {
@@ -1445,7 +1540,11 @@ function emptyAggregate() {
 		cacheWrite: 0,
 		reasoning: 0,
 		billableInput: 0,
-		total: 0
+		total: 0,
+		timed: 0,
+		ttftMs: 0,
+		generationMs: 0,
+		timedOutput: 0
 	};
 }
 /** Fold one entry into an aggregate accumulator (mutates and returns it, for cheap loops). */
@@ -1464,8 +1563,29 @@ function addToAggregate(into, entry) {
 	into.reasoning += usage.reasoning;
 	into.billableInput += usage.input + usage.cacheRead + usage.cacheWrite;
 	into.total += usage.input + usage.cacheRead + usage.cacheWrite + usage.output;
+	recordSpeed(into, entry);
 	return into;
 }
+/**
+* Fold one entry's speed measurement into the accumulators.
+*
+* Extraction is a bare subtraction, which can go negative when a relay reports a `ms` shorter
+* than its first token (clock granularity, or a resume counted differently). A negative
+* generation time would poison the ratio, so such an entry contributes to `timed`/`ttftMs` but
+* not to `generationMs`/`timedOutput` — the honest reading is "we measured a start, not a rate".
+* @param into - the group accumulator.
+* @param entry - one finished request.
+*/
+function recordSpeed(into, entry) {
+	const firstTokenMs = entry.firstTokenMs;
+	if (firstTokenMs === void 0) return;
+	into.timed += 1;
+	into.ttftMs += firstTokenMs;
+	const generationMs = entry.ms - firstTokenMs;
+	if (generationMs <= 0 || entry.usage === null) return;
+	into.generationMs += generationMs;
+	into.timedOutput += entry.usage.output;
+}
 /** `YYYY-MM` for a timestamp, in the timestamp's own offset. */
 function monthKey(at) {
 	return `${at.getFullYear()}-${`${at.getMonth() + 1}`.padStart(2, "0")}`;
@@ -1641,6 +1761,10 @@ function totalOf(rows) {
 		total.reasoning += row.aggregate.reasoning;
 		total.billableInput += row.aggregate.billableInput;
 		total.total += row.aggregate.total;
+		total.timed += row.aggregate.timed;
+		total.ttftMs += row.aggregate.ttftMs;
+		total.generationMs += row.aggregate.generationMs;
+		total.timedOutput += row.aggregate.timedOutput;
 	}
 	return total;
 }
@@ -1678,6 +1802,51 @@ function queryDate(value) {
 	return Number.isNaN(parsed.getTime()) ? void 0 : parsed;
 }
 /**
+* The URL path and body a probe uses for one protocol.
+*
+* 🔴 **The probe used to hardcode `/chat/completions`.** With the Responses API now selectable
+* per channel (owner directive 2026-10-22: 「优先支持并使用 responses api」), that would have
+* probed the WRONG endpoint for every Responses channel and reported a verdict about a protocol
+* the channel does not use. "A probe must send what a real request sends" already cost one round
+* over headers; the rule applies unchanged to the path.
+* @param channel - the channel being probed.
+* @returns the path suffix and request body for its protocol.
+*/
+function probeRequest(channel) {
+	if (channel.api === "openai-responses") return {
+		path: "/responses",
+		body: {
+			model: channel.model,
+			input: "ping",
+			max_output_tokens: 256,
+			stream: false
+		}
+	};
+	if (channel.api === "anthropic-messages") return {
+		path: "/messages",
+		body: {
+			model: channel.model,
+			max_tokens: 16,
+			messages: [{
+				role: "user",
+				content: "ping"
+			}]
+		}
+	};
+	return {
+		path: "/chat/completions",
+		body: {
+			model: channel.model,
+			messages: [{
+				role: "user",
+				content: "ping"
+			}],
+			max_tokens: 1,
+			stream: false
+		}
+	};
+}
+/**
 * Probe one channel with a minimal real request.
 *
 * This spends tokens, which is why it is an explicit route rather than something the panel does
@@ -1696,25 +1865,21 @@ function queryDate(value) {
 */
 async function probeChannel(channel, key, sessionId, signal) {
 	const started = Date.now();
-	const url = `${channel.baseURL.replace(/\/+$/, "")}/chat/completions`;
+	const request = probeRequest(channel);
+	const url = `${channel.baseURL.replace(/\/+$/, "")}${request.path}`;
 	try {
 		const response = await fetch(url, {
 			method: "POST",
 			headers: {
 				"content-type": "application/json",
-				authorization: `Bearer ${key}`,
+				...channel.api === "anthropic-messages" ? {
+					"x-api-key": key,
+					"anthropic-version": "2023-06-01"
+				} : { authorization: `Bearer ${key}` },
 				...channel.headers,
 				...requestHeadersFor(channel, sessionId)
 			},
-			body: JSON.stringify({
-				model: channel.model,
-				messages: [{
-					role: "user",
-					content: "ping"
-				}],
-				max_tokens: 1,
-				stream: false
-			}),
+			body: JSON.stringify(request.body),
 			...signal === void 0 ? {} : { signal }
 		});
 		const ms = Date.now() - started;
@@ -1725,23 +1890,17 @@ async function probeChannel(channel, key, sessionId, signal) {
 			code: probeCode(response.status, text),
 			message: text.slice(0, 400)
 		};
-		if (text.includes("\"error\"")) return {
+		const bodyError = errorFieldValue(text);
+		if (bodyError !== void 0) return {
 			ok: false,
 			ms,
 			code: "SERVER",
-			message: text.slice(0, 400)
+			message: bodyError.slice(0, 400)
 		};
-		let responded = "";
-		try {
-			const parsed = JSON.parse(text);
-			responded = typeof parsed.model === "string" ? parsed.model : typeof parsed.choices?.[0]?.message?.content === "string" ? parsed.choices[0].message.content : "";
-		} catch {
-			responded = "";
-		}
 		return {
 			ok: true,
 			ms,
-			responded
+			responded: reportedName(text)
 		};
 	} catch (error) {
 		return {
@@ -1752,6 +1911,61 @@ async function probeChannel(channel, key, sessionId, signal) {
 		};
 	}
 }
+/**
+* The body's `error` value, when it is actually an error.
+*
+* Returns `undefined` for a missing field AND for an explicit `null` — both mean "no error".
+* A non-null value (object or string) is returned as JSON text for the panel to show.
+* @param text - the raw response body.
+* @returns the error as text, or undefined when the body reports no error.
+*/
+function errorFieldValue(text) {
+	let parsed;
+	try {
+		parsed = JSON.parse(text);
+	} catch {
+		return text.includes("\"error\"") ? text : void 0;
+	}
+	const candidate = parsed?.error;
+	if (candidate === void 0 || candidate === null) return void 0;
+	return typeof candidate === "string" ? candidate : JSON.stringify(candidate);
+}
+/**
+* What a successful probe response says the relay called its model.
+*
+* Three protocols, three response shapes — the old reader understood only completions
+* (`choices[0].message.content`), so a Responses or Anthropic 200 came back with an empty verdict
+* and the panel showed a blank success. A verdict nobody can read is not a verdict.
+* @param text - the raw response body.
+* @returns the reported model name, or the first output text; empty when neither is present.
+*/
+function reportedName(text) {
+	let parsed;
+	try {
+		parsed = JSON.parse(text);
+	} catch {
+		return "";
+	}
+	const asString = (value) => typeof value === "string" ? value : "";
+	const model = asString(parsed.model);
+	if (model !== "") return model;
+	const output = parsed.output;
+	if (Array.isArray(output)) for (const item of output) {
+		const content = item.content;
+		if (!Array.isArray(content)) continue;
+		for (const part of content) {
+			const value = asString(part.text);
+			if (value !== "") return value;
+		}
+	}
+	const choices = parsed.choices;
+	if (Array.isArray(choices)) {
+		const first = choices[0];
+		const value = asString(first?.message?.content);
+		if (value !== "") return value;
+	}
+	return "";
+}
 /** Map an HTTP status from a probe onto the same codes the runtime uses. */
 function probeCode(status, body) {
 	if (status === 401 || status === 403) return "AUTH";
@@ -1837,8 +2051,11 @@ function registerStatsRoutes(ctx, webServer, deps) {
 				const labels = new Map(deps.channels().map((channel) => [channel.id, channel.label]));
 				const named = rows.map((row) => ({
 					...row,
-					label: group === "day" ? row.key : labels.get(row.key) ?? row.key
+					label: group === "day" ? row.key : labels.get(row.key) ?? row.key,
+					tokensPerSecond: outputTokensPerSecond(row.aggregate),
+					firstTokenMs: meanFirstTokenMs(row.aggregate)
 				}));
+				const totalAggregate = totalOf(rows);
 				json(res, 200, {
 					range: {
 						from: from?.toISOString() ?? null,
@@ -1846,8 +2063,15 @@ function registerStatsRoutes(ctx, webServer, deps) {
 						group
 					},
 					rows: named,
-					total: totalOf(rows),
-					note: "input is uncached input only; billable input = input + cacheRead + cacheWrite"
+					total: totalAggregate,
+					/**
+					* The footer's speed figures, computed from the TOTAL aggregate rather than by
+					* averaging the row values: a ratio of sums is the only one where a quiet channel
+					* cannot outweigh a chatty one. Same functions, one definition.
+					*/
+					totalSpeed: outputTokensPerSecond(totalAggregate),
+					totalFirstTokenMs: meanFirstTokenMs(totalAggregate),
+					note: "input is uncached input only; billable input = input + cacheRead + cacheWrite; tokensPerSecond counts generation time only (total minus time-to-first-token)"
 				});
 			}
 		},

package/lib/presets.d.ts

@@ -145,6 +145,22 @@ export declare const OPENCODE_SESSION_FALLBACK = "dsh-any-deepseek";
  *
  * The id is what a channel entry stores (`preset: opencode-go`), so it is part of the user's
  * configuration surface: renaming one breaks saved channels. Add, do not rename.
+ *
+ * ## Why every preset prefers the Responses API
+ *
+ * Owner directive (2026-10-22): 「对于各家查阅资料，优先支持并使用 responses api(这个api更先进)」.
+ * Responses is the newer surface: it carries reasoning items as first-class output, reports
+ * `output_tokens_details.reasoning_tokens`, and supports server-side state/tool primitives the
+ * completions shape cannot express.
+ *
+ * 🔴 **But "prefer" is not "assume".** A relay that answers `/responses` with 404 would break
+ * every request if this were a blanket switch, so the rule is: a preset declares Responses ONLY
+ * where the endpoint is documented to serve it; everything else keeps completions, which is the
+ * lowest common denominator every relay speaks. `api` stays a per-channel field, so a user can
+ * override any of them in the panel without waiting for a release.
+ *
+ * DeepSeek's own docs confirm `/responses` with `deepseek-flash`, including `input_image` — so
+ * the official preset is the first to move and is the reference case.
  */
 export declare const CHANNEL_PRESETS: Readonly<Record<string, ChannelPreset>>;
 /** Every preset id, in table order — the panel's picker source. */

package/lib/usage-ledger.d.ts

@@ -71,7 +71,42 @@ export interface LedgerAggregate {
     billableInput: number;
     /** `billableInput + output`. */
     total: number;
+    /**
+     * Requests carrying a first-token measurement — the denominator for both speed figures.
+     *
+     * Kept separate from `requests` on purpose: not every request has one (an error before the
+     * first block has no first token, and ledger lines written before this field existed do not
+     * either). Dividing by `requests` would silently understate every average.
+     */
+    timed: number;
+    /** Sum of observed time-to-first-token, over `timed` requests only. */
+    ttftMs: number;
+    /** Sum of generation time (`ms - firstTokenMs`), over `timed` requests only. */
+    generationMs: number;
+    /** Sum of `output` over `timed` requests — the numerator for generation throughput. */
+    timedOutput: number;
 }
+/**
+ * Output tokens per second generated, or null when unmeasurable.
+ *
+ * 🔴 **Why not `output / total` (owner request 2026-10-22: 「能加上 token 速度统计吗」).**
+ * One ratio collapses queueing and generation together, so a relay that thinks for 3s and then
+ * streams fast looks identical to a uniformly slow one — and those two call for different fixes
+ * (switch relay vs. shrink the prompt). The split is measured per request and this divides only
+ * by the generating half.
+ *
+ * Averaged as a RATIO OF SUMS rather than a mean of per-request speeds: a mean lets one
+ * five-token reply dominate, while the sum ratio answers "across this group, how fast did tokens
+ * actually come out".
+ * @param aggregate - one group's figures.
+ * @returns tokens per second, or null when nothing was timed or nothing was generated.
+ */
+export declare function outputTokensPerSecond(aggregate: LedgerAggregate): number | null;
+/**
+ * Mean time to first token, in ms, or null when nothing was timed.
+ * @param aggregate - one group's figures.
+ */
+export declare function meanFirstTokenMs(aggregate: LedgerAggregate): number | null;
 /** Aggregation result for one grouping. */
 export interface LedgerSummary {
     /** Group key (channel id, or `YYYY-MM-DD`). */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-any-deepseek",
-  "version": "0.0.10",
+  "version": "0.0.12",
   "description": "One DeepSeek V4.1 Flash provider for DeepSeek Harness: paste a key per channel, switch channels manually, token usage booked per real upstream.",
   "license": "MIT",
   "repository": {