dsh-any-deepseek 0.0.8 → 0.0.9

package/lib/adapter.d.ts

@@ -27,12 +27,51 @@ 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 { type ChannelModality } from './presets.js';
 import type { ResolvedChannel } from './channels.js';
 import type { LedgerEntry } from './usage-ledger.js';
 /** The single provider route this plugin owns. */
 export declare const PROVIDER_ID = "any-deepseek";
 /** Display name for the route in provider selectors. */
 export declare const PROVIDER_NAME = "Any DeepSeek";
+/**
+ * Input modalities this provider's one model accepts.
+ *
+ * 🔴 **Declared at registration, because the harness asks the model — not the channel.**
+ *
+ * `LlmModelInfo.inputModalities` is documented as "absent means unknown" (`dsh-llm/lib/types/
+ * types.d.ts:330-331`), and the host reads it at dispatch to decide whether an image survives
+ * projection at all: `if (modelInfo.inputModalities !== undefined &&
+ * !modelInfo.inputModalities.includes('image') && …) projectImagesForTextModel(…)`
+ * (`dsh-llm/lib/index.js:2337`). Omitting it therefore did not mean "we never send images"; it
+ * meant a user's screenshot was replaced by a placeholder *before* this plugin ever saw it, which
+ * is why the model answered as though the message had been text.
+ *
+ * The value is the MODEL's capability, so it does not depend on which channel is selected, on
+ * whether a key exists, or on whether `current` is set — the same rule as the catalog itself.
+ * A relay that genuinely cannot carry images declares `input: ['text']` on its own channel, and
+ * the transport (`pi-ai`) downgrades the image per channel; the harness-level answer stays the
+ * model's.
+ */
+export declare const MODEL_INPUT_MODALITIES: readonly ChannelModality[];
+/**
+ * The reasoning levels in force for one channel.
+ *
+ * 🔴 **A channel that says nothing inherits the model's table, and that is the whole fix for the
+ * missing thinking control.**
+ *
+ * The old code read `channel.reasoningEfforts` alone. That value only existed once a channel
+ * declared it — so during first-run setup (no channel, or a channel written before the panel
+ * offered levels) the answer was "no reasoning at all", and the composer's strength selector was
+ * absent. But the model always had levels; the CHANNEL was merely silent about them.
+ *
+ * An empty map counts as silence, not as "explicitly unsupported": `z.dict()` produces `{}` when
+ * the panel writes the field, and treating that as a negative capability is how the picker broke
+ * with `INVALID_MODEL_REASONING` in the first place.
+ * @param channel - the selected channel, when there is one.
+ * @returns the levels to advertise and to send, keyed by level name.
+ */
+export declare function effectiveReasoningEfforts(channel: ResolvedChannel | undefined): Record<string, string | null>;
 /** What the adapter needs from the plugin, so it can be tested without a harness context. */
 export interface AdapterDeps {
     /** Enabled, resolved channels in configuration order. */
@@ -56,11 +95,10 @@ export interface AdapterDeps {
      * seam an image cannot be sent at all. When it is missing, image content is replaced by a text
      * placeholder rather than silently dropped: the model should know an image existed even when it
      * cannot see it.
-     * @param attachmentId - durable id from the message block.
-     * @param mediaType - the block's declared media type.
+     * @param ref - the block's attachment reference, passed through WHOLE.
      * @returns base64 bytes and the media type to put on the wire.
      */
-    readImage?: (attachmentId: string, mediaType: string) => Promise<{
+    readImage?: (ref: HarnessImageRef) => Promise<{
         data: string;
         mediaType: string;
     } | undefined>;
@@ -69,6 +107,15 @@ export interface AdapterDeps {
     /** Monotonic-ish millisecond clock for durations. */
     elapsed: () => number;
 }
+/** A durable image reference as a harness message block carries it. */
+export interface HarnessImageRef {
+    attachmentId: string;
+    mediaType: string;
+    bytes?: number;
+    width?: number;
+    height?: number;
+    name?: string;
+}
 /** One finished call, as the adapter observed it. */
 interface CallOutcome {
     ok: boolean;
@@ -94,6 +141,19 @@ interface CallOutcome {
  * @returns a stable, non-empty session id.
  */
 export declare function opencodeSessionId(sessionId: unknown): string;
+/**
+ * Build the per-request headers one channel needs.
+ *
+ * 🔴 **Shared by BOTH paths that talk to a relay** — the runtime `stream()` and the panel's
+ * `/probe`. They used to differ: `stream()` injected the OpenCode session header and `/probe`
+ * did not, so "测试" reported 400 `MissingSessionID` for a channel that worked in real use. A
+ * probe that checks something other than what a real request does is not a probe; it is a second
+ * implementation of the truth, and the two will disagree.
+ * @param channel - the resolved channel.
+ * @param sessionId - the harness session id, when the caller has one.
+ * @returns headers to send, on top of whatever the channel already declares.
+ */
+export declare function requestHeadersFor(channel: ResolvedChannel, sessionId: unknown): Record<string, string>;
 /**
  * Build the pi-ai model entry for one channel.
  *
@@ -178,6 +238,9 @@ export declare class AnyDeepSeekAdapter extends LlmAdapter {
      * ⇒ The catalog answers "what does this provider serve", which is a property of the plugin.
      * Key presence is answered per request (`MISSING_CREDENTIAL`) and per row in the panel
      * (`keyConfigured`), where the user can actually act on it.
+     *
+     * The same reasoning extends to `inputModalities`: it describes the model this route serves,
+     * so it is reported here unconditionally rather than derived from a channel.
      */
     listModels(_provider: string): Promise<LlmModelInfo[]>;
     resolveModel(provider: string, model: string, _signal?: AbortSignal): Promise<LlmResolvedModelInfo>;

package/lib/http.d.ts

@@ -64,12 +64,19 @@ export interface WebServerLike {
  *
  * This spends tokens, which is why it is an explicit route rather than something the panel does
  * on render. The request asks for one output token, so the cost is a prompt plus a token.
+ *
+ * 🔴 **The headers come from `requestHeadersFor`, the same function the runtime uses.** They used
+ * to be built here separately, which meant this probe sent no OpenCode session header while real
+ * requests did — so "测试" reported 400 `MissingSessionID` on a channel that worked in use. A
+ * probe that does not send what a real request sends is not testing the channel.
  * @param channel - the channel to probe.
  * @param key - its resolved credential.
+ * @param sessionId - session identity for per-request headers (the probe has no conversation of
+ *   its own, so callers pass the selected channel's own stable value).
  * @param signal - cancellation.
  * @returns a verdict and, on success, what the relay reported.
  */
-export declare function probeChannel(channel: ResolvedChannel, key: string, signal?: AbortSignal): Promise<{
+export declare function probeChannel(channel: ResolvedChannel, key: string, sessionId?: string, signal?: AbortSignal): Promise<{
     ok: boolean;
     ms: number;
     code?: string;

package/lib/index.js

@@ -170,6 +170,13 @@ const DEFAULT_MAX_TOKENS = 32768;
 * images" — it meant the user's image never reached the model at all, silently. The relay side is
 * now real: `adapter.ts` collects image bytes through the durable attachment store and maps them
 * to pi-ai `image` content, and falls back to a visible text placeholder when it cannot.
+*
+* ## Why this is exported
+*
+* The adapter must report the SAME capabilities through `listModels()` / `resolveModel()`, which
+* describe the advertised model rather than any one channel. Reaching for this constant (instead
+* of re-listing the numbers there) keeps the catalog and the exact-model answer from drifting —
+* and they are read by different host paths, so a drift is silent.
 */
 const V41_FLASH_CAPS = {
 	contextWindow: 1e6,
@@ -654,6 +661,48 @@ const PROVIDER_ID = "any-deepseek";
 /** Display name for the route in provider selectors. */
 const PROVIDER_NAME = "Any DeepSeek";
 /**
+* Input modalities this provider's one model accepts.
+*
+* 🔴 **Declared at registration, because the harness asks the model — not the channel.**
+*
+* `LlmModelInfo.inputModalities` is documented as "absent means unknown" (`dsh-llm/lib/types/
+* types.d.ts:330-331`), and the host reads it at dispatch to decide whether an image survives
+* projection at all: `if (modelInfo.inputModalities !== undefined &&
+* !modelInfo.inputModalities.includes('image') && …) projectImagesForTextModel(…)`
+* (`dsh-llm/lib/index.js:2337`). Omitting it therefore did not mean "we never send images"; it
+* meant a user's screenshot was replaced by a placeholder *before* this plugin ever saw it, which
+* is why the model answered as though the message had been text.
+*
+* The value is the MODEL's capability, so it does not depend on which channel is selected, on
+* whether a key exists, or on whether `current` is set — the same rule as the catalog itself.
+* A relay that genuinely cannot carry images declares `input: ['text']` on its own channel, and
+* the transport (`pi-ai`) downgrades the image per channel; the harness-level answer stays the
+* model's.
+*/
+const MODEL_INPUT_MODALITIES = V41_FLASH_CAPS.input;
+/**
+* The reasoning levels in force for one channel.
+*
+* 🔴 **A channel that says nothing inherits the model's table, and that is the whole fix for the
+* missing thinking control.**
+*
+* The old code read `channel.reasoningEfforts` alone. That value only existed once a channel
+* declared it — so during first-run setup (no channel, or a channel written before the panel
+* offered levels) the answer was "no reasoning at all", and the composer's strength selector was
+* absent. But the model always had levels; the CHANNEL was merely silent about them.
+*
+* An empty map counts as silence, not as "explicitly unsupported": `z.dict()` produces `{}` when
+* the panel writes the field, and treating that as a negative capability is how the picker broke
+* with `INVALID_MODEL_REASONING` in the first place.
+* @param channel - the selected channel, when there is one.
+* @returns the levels to advertise and to send, keyed by level name.
+*/
+function effectiveReasoningEfforts(channel) {
+	const declared = channel?.reasoningEfforts;
+	if (declared !== void 0 && Object.keys(declared).length > 0) return declared;
+	return V41_FLASH_CAPS.reasoningEfforts;
+}
+/**
 * Derive the value for the OpenCode session header.
 *
 * 🔴 This is THIS PLUGIN's job, not the harness's. `/zen/go` is one relay's private requirement;
@@ -677,6 +726,28 @@ function opencodeSessionId(sessionId) {
 	return createHash("sha256").update(raw).digest("hex").slice(0, 32);
 }
 /**
+* Build the per-request headers one channel needs.
+*
+* 🔴 **Shared by BOTH paths that talk to a relay** — the runtime `stream()` and the panel's
+* `/probe`. They used to differ: `stream()` injected the OpenCode session header and `/probe`
+* did not, so "测试" reported 400 `MissingSessionID` for a channel that worked in real use. A
+* probe that checks something other than what a real request does is not a probe; it is a second
+* implementation of the truth, and the two will disagree.
+* @param channel - the resolved channel.
+* @param sessionId - the harness session id, when the caller has one.
+* @returns headers to send, on top of whatever the channel already declares.
+*/
+function requestHeadersFor(channel, sessionId) {
+	const headers = {};
+	if (channel.headers !== void 0 && Object.hasOwn(channel.headers, "x-opencode-client")) {
+		const session = opencodeSessionId(sessionId);
+		headers[OPENCODE_SESSION_HEADER] = session;
+		headers["X-Session-ID"] = session;
+		headers["User-Agent"] = CLIENT_USER_AGENT;
+	}
+	return headers;
+}
+/**
 * Build the pi-ai model entry for one channel.
 *
 * The advertised id (`deepseek-v4.1-flash`) never reaches the wire here: each channel's own
@@ -685,8 +756,12 @@ function opencodeSessionId(sessionId) {
 * @returns a pi-ai chat model describing that relay's view of the model.
 */
 function buildPiModel(channel) {
-	const thinkingLevelMap = channel.reasoningEfforts === void 0 ? void 0 : Object.fromEntries(Object.entries(channel.reasoningEfforts).filter(([, value]) => value !== null));
-	const hasThinking = thinkingLevelMap !== void 0 && Object.keys(thinkingLevelMap).length > 0;
+	const thinkingLevelMap = (() => {
+		const declared = {};
+		for (const [level, wire] of Object.entries(effectiveReasoningEfforts(channel))) if (typeof wire === "string") declared[level] = wire;
+		return Object.keys(declared).length === 0 ? void 0 : declared;
+	})();
+	const hasThinking = thinkingLevelMap !== void 0;
 	return {
 		id: channel.model,
 		name: channel.label,
@@ -732,21 +807,18 @@ async function collectImages(options, readImage) {
 		for (const block of message.content) {
 			const candidate = block;
 			if (candidate.type !== "image") continue;
-			const id = candidate.attachment?.attachmentId;
-			const mediaType = candidate.attachment?.mediaType;
-			if (typeof id !== "string" || typeof mediaType !== "string") continue;
-			if (images.has(id)) continue;
-			pending.push({
-				id,
-				mediaType
-			});
+			const ref = candidate.attachment;
+			if (ref === void 0 || typeof ref.attachmentId !== "string") continue;
+			if (images.has(ref.attachmentId)) continue;
+			if (pending.some((item) => item.attachmentId === ref.attachmentId)) continue;
+			pending.push(ref);
 		}
 	}
-	const resolved = await Promise.all(pending.map(async (item) => {
+	const resolved = await Promise.all(pending.map(async (ref) => {
 		try {
-			return [item.id, await readImage(item.id, item.mediaType)];
+			return [ref.attachmentId, await readImage(ref)];
 		} catch {
-			return [item.id, void 0];
+			return [ref.attachmentId, void 0];
 		}
 	}));
 	for (const [id, value] of resolved) if (value !== void 0) images.set(id, value);
@@ -1145,6 +1217,9 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 	* ⇒ The catalog answers "what does this provider serve", which is a property of the plugin.
 	* Key presence is answered per request (`MISSING_CREDENTIAL`) and per row in the panel
 	* (`keyConfigured`), where the user can actually act on it.
+	*
+	* The same reasoning extends to `inputModalities`: it describes the model this route serves,
+	* so it is reported here unconditionally rather than derived from a channel.
 	*/
 	async listModels(_provider) {
 		const selected = this.selected();
@@ -1154,12 +1229,13 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 			provider: PROVIDER_ID,
 			id: ADVERTISED_MODEL_ID,
 			name: ADVERTISED_MODEL_NAME,
-			description: label === void 0 ? "no channel configured yet" : `via ${label}`
+			description: label === void 0 ? "no channel configured yet" : `via ${label}`,
+			inputModalities: [...MODEL_INPUT_MODALITIES]
 		}];
 	}
 	async resolveModel(provider, model, _signal) {
 		const channel = this.selected();
-		const efforts = Object.entries(channel?.reasoningEfforts ?? {}).filter(([, wire]) => wire !== null).map(([id]) => id);
+		const efforts = Object.entries(effectiveReasoningEfforts(channel)).filter(([, wire]) => wire !== null).map(([id]) => id);
 		const reasoning = efforts.length === 0 ? {} : { reasoning: { efforts: efforts.map((id) => ({
 			id: ReasoningEffortId(id),
 			name: id
@@ -1171,6 +1247,7 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 			description: channel === void 0 ? void 0 : `via ${channel.label}`,
 			context: { contextWindow: channel?.contextWindow ?? 1e6 },
 			defaultMaxTokens: channel?.maxTokens,
+			inputModalities: [...MODEL_INPUT_MODALITIES],
 			...reasoning
 		};
 	}
@@ -1269,14 +1346,10 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 		*    for the same conversation, which is all the endpoint's affinity key needs.
 		*  · `User-Agent` must identify THIS client rather than the generic SDK. pi-ai sets its own,
 		*    so the preset's static headers alone can never satisfy that rule.
+		*
+		* The construction lives in `requestHeadersFor` so the panel's probe sends the same thing.
 		*/
-		const requestHeaders = {};
-		if (channel.headers !== void 0 && Object.hasOwn(channel.headers, "x-opencode-client")) {
-			const session = opencodeSessionId(options.sessionId);
-			requestHeaders[OPENCODE_SESSION_HEADER] = session;
-			requestHeaders["X-Session-ID"] = session;
-			requestHeaders["User-Agent"] = CLIENT_USER_AGENT;
-		}
+		const requestHeaders = requestHeadersFor(channel, options.sessionId);
 		const streamOptions = {
 			apiKey: key,
 			signal: options.signal,
@@ -1609,12 +1682,19 @@ function queryDate(value) {
 *
 * This spends tokens, which is why it is an explicit route rather than something the panel does
 * on render. The request asks for one output token, so the cost is a prompt plus a token.
+*
+* 🔴 **The headers come from `requestHeadersFor`, the same function the runtime uses.** They used
+* to be built here separately, which meant this probe sent no OpenCode session header while real
+* requests did — so "测试" reported 400 `MissingSessionID` on a channel that worked in use. A
+* probe that does not send what a real request sends is not testing the channel.
 * @param channel - the channel to probe.
 * @param key - its resolved credential.
+* @param sessionId - session identity for per-request headers (the probe has no conversation of
+*   its own, so callers pass the selected channel's own stable value).
 * @param signal - cancellation.
 * @returns a verdict and, on success, what the relay reported.
 */
-async function probeChannel(channel, key, signal) {
+async function probeChannel(channel, key, sessionId, signal) {
 	const started = Date.now();
 	const url = `${channel.baseURL.replace(/\/+$/, "")}/chat/completions`;
 	try {
@@ -1623,7 +1703,8 @@ async function probeChannel(channel, key, signal) {
 			headers: {
 				"content-type": "application/json",
 				authorization: `Bearer ${key}`,
-				...channel.headers
+				...channel.headers,
+				...requestHeadersFor(channel, sessionId)
 			},
 			body: JSON.stringify({
 				model: channel.model,
@@ -1784,7 +1865,7 @@ function registerStatsRoutes(ctx, webServer, deps) {
 					});
 					return;
 				}
-				const verdict = await probeChannel(channel, key);
+				const verdict = await probeChannel(channel, key, `probe:${channel.id}`);
 				json(res, verdict.ok ? 200 : 502, verdict);
 			}
 		}
@@ -1906,25 +1987,28 @@ function apply(ctx, config) {
 	* 2048px keeps the harness's normal form, and the byte ceiling is the one the host's pi-ai
 	* adapter uses for the same purpose.
 	*/
-	const readImage = async (attachmentId, mediaType) => {
+	const readImage = async (ref) => {
 		const attachments = service(ctx, "attachments");
-		if (attachments?.readImageRequest === void 0) return void 0;
+		if (attachments?.readImageRequest === void 0) {
+			console.warn("[any-deepseek] 图片无法读取：本组合没有装配 attachments 服务（readImageRequest 不可用）");
+			return;
+		}
 		try {
-			const stored = await attachments.readImageRequest({
-				attachmentId,
-				mediaType
-			}, {
+			const stored = await attachments.readImageRequest(ref, {
 				width: 2048,
 				height: 2048,
 				maxBytes: 4194304
 			});
-			if (stored?.data === void 0) return void 0;
+			if (stored?.data === void 0) {
+				console.warn(`[any-deepseek] 图片 ${ref.attachmentId} 读取结果为空`);
+				return;
+			}
 			return {
 				data: Buffer.from(stored.data).toString("base64"),
-				mediaType: stored.mediaType ?? mediaType
+				mediaType: stored.mediaType ?? ref.mediaType
 			};
 		} catch (error) {
-			console.warn(`[any-deepseek] 读不到图片 ${attachmentId}: ${error instanceof Error ? error.message : String(error)}`);
+			console.warn(`[any-deepseek] 读不到图片 ${ref.attachmentId}: ${error instanceof Error ? error.message : String(error)}`);
 			return;
 		}
 	};

package/lib/presets.d.ts

@@ -81,6 +81,41 @@ export declare const ADVERTISED_MODEL_NAME = "DeepSeek V4.1 Flash";
  */
 export declare const DEFAULT_CONTEXT_WINDOW = 262144;
 export declare const DEFAULT_MAX_TOKENS = 32768;
+/**
+ * Capability numbers shared by every DeepSeek V4.1 Flash deployment we know of.
+ *
+ * Kept as one constant because these are properties of the *model*, not of a relay; a relay
+ * that disagrees corrects them in its own preset.
+ *
+ * ## Why `input` lists image
+ *
+ * The model is multimodal, and the harness only projects an image into a request when the model
+ * entry declares the capability. Declaring text-only therefore did not mean "we never send
+ * images" — it meant the user's image never reached the model at all, silently. The relay side is
+ * now real: `adapter.ts` collects image bytes through the durable attachment store and maps them
+ * to pi-ai `image` content, and falls back to a visible text placeholder when it cannot.
+ *
+ * ## Why this is exported
+ *
+ * The adapter must report the SAME capabilities through `listModels()` / `resolveModel()`, which
+ * describe the advertised model rather than any one channel. Reaching for this constant (instead
+ * of re-listing the numbers there) keeps the catalog and the exact-model answer from drifting —
+ * and they are read by different host paths, so a drift is silent.
+ */
+export declare const V41_FLASH_CAPS: {
+    contextWindow: number;
+    maxTokens: number;
+    input: ChannelModality[];
+    /**
+     * Selectable thinking levels, in pi-ai's vocabulary.
+     *
+     * The VALUES are the wire spellings this model family accepts (`reasoning_effort`); `null`
+     * would mark a level as unsupported. pi-ai consumes this map directly as `thinkingLevelMap`,
+     * which is why this plugin does not implement a translation layer of its own — a second
+     * mapping would inevitably drift from the one the transport actually applies.
+     */
+    reasoningEfforts: Record<string, string | null>;
+};
 /**
  * The user agent a request to OpenCode Go identifies itself with.
  *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-any-deepseek",
-  "version": "0.0.8",
+  "version": "0.0.9",
   "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": {