dsh-any-deepseek 0.0.7 → 0.0.8

package/lib/adapter.d.ts

@@ -48,6 +48,22 @@ export interface AdapterDeps {
     resolveKey: (ref: string) => Promise<string | undefined>;
     /** Record one finished request. Must never throw. */
     record: (entry: LedgerEntry) => Promise<void>;
+    /**
+     * Read one image's request bytes from the durable attachment store.
+     *
+     * 🔴 Optional, and its absence is a real state (a composition without the attachment service).
+     * An `ImageBlock` in the harness carries only an `attachmentId` — never bytes — so without this
+     * 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.
+     * @returns base64 bytes and the media type to put on the wire.
+     */
+    readImage?: (attachmentId: string, mediaType: string) => Promise<{
+        data: string;
+        mediaType: string;
+    } | undefined>;
     /** Clock, injectable so tests are deterministic. */
     now: () => Date;
     /** Monotonic-ish millisecond clock for durations. */
@@ -59,6 +75,25 @@ interface CallOutcome {
     code?: string;
     usage: TokenUsage | null;
 }
+/**
+ * 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;
+ * the host cannot know about it and must not be expected to stamp anything for it. Whatever the
+ * harness happens to put on a request is only an input to a decision made here.
+ *
+ * The id is a stable hash of the harness session id, so:
+ *  · the same conversation keeps one affinity key (what the endpoint wants), and
+ *  · the value is opaque rather than echoing an internal id to a third party.
+ *
+ * Two hard requirements, both learned from a 400:
+ *  · **never absent** — the header is always sent, so a request can never go out without it;
+ *  · **stable when there is no session** — falling back to a random value would give every
+ *    request a different affinity key, which is worse than a constant.
+ * @param sessionId - the harness session id, when it supplied one.
+ * @returns a stable, non-empty session id.
+ */
+export declare function opencodeSessionId(sessionId: unknown): string;
 /**
  * Build the pi-ai model entry for one channel.
  *
@@ -68,15 +103,35 @@ interface CallOutcome {
  * @returns a pi-ai chat model describing that relay's view of the model.
  */
 export declare function buildPiModel(channel: ResolvedChannel): Model<Api>;
+/** One resolved image, keyed by attachment id. */
+export type ImageBytes = Map<string, {
+    data: string;
+    mediaType: string;
+}>;
+/**
+ * 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
+ * also how the host's own pi-ai adapter does it, for the same reason.
+ *
+ * A block that cannot be read is left OUT of the map, and the converter then substitutes a text
+ * placeholder. That is deliberate: dropping the block would tell the model nothing, while a
+ * placeholder at least says an image was attached and names why it could not be shown. Skipping
+ * it silently is the failure mode this whole change exists to remove — the user sent an image and
+ * the model answered as though the message had been text only.
+ * @param options - the harness request.
+ * @param readImage - the attachment seam, when the composition has one.
+ * @returns resolved images keyed by attachment id.
+ */
+export declare function collectImages(options: GenerateOptions, readImage: AdapterDeps['readImage']): Promise<ImageBytes>;
 /**
  * Translate one harness request into a pi-ai context.
  *
- * The harness already projects images and files before an adapter sees them, so this is a
- * straightforward content mapping; anything the harness did not project is passed through.
  * @param options - the harness request.
+ * @param images - image bytes collected by {@link collectImages}; omitted for a text-only request.
  * @returns a pi-ai context plus the one-shot system prompt, when the caller set one.
  */
-export declare function toPiContext(options: GenerateOptions): PiContext;
+export declare function toPiContext(options: GenerateOptions, images?: ImageBytes): PiContext;
 /** Map pi-ai usage onto the harness's disjoint counters. */
 export declare function toTokenUsage(usage: AssistantMessage['usage']): TokenUsage;
 /**

package/lib/index.js

@@ -2,6 +2,7 @@ import { join } from "node:path";
 import z from "@deepseek-ai/schemastery";
 import { LlmAdapter, ReasoningEffortId, ToolCallId } from "@deepseek-ai/dsh-llm";
 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 { appendFile, mkdir, readFile, readdir } from "node:fs/promises";
@@ -162,17 +163,18 @@ const DEFAULT_MAX_TOKENS = 32768;
 * 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.
 *
-* ⚠️ `input` is `['text']` **on purpose, for now**. The model is multimodal, but declaring image
-* support is a promise this adapter cannot keep yet: a harness `ImageBlock` carries only an
-* `attachmentId`, and turning that into bytes requires the durable attachment service
-* (`ctx.attachments.readImageRequest`), which the adapter does not consult. Declaring images
-* without that path would be worse than declaring text — the harness would stop substituting its
-* placeholder text and the image would be silently dropped. See SESSION decision 38.
+* ## 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.
 */
 const V41_FLASH_CAPS = {
 	contextWindow: 1e6,
 	maxTokens: 384e3,
-	input: ["text"],
+	input: ["text", "image"],
 	/**
 	* Selectable thinking levels, in pi-ai's vocabulary.
 	*
@@ -201,21 +203,12 @@ const DEEPSEEK_THINKING_COMPAT = {
 	thinkingFormat: "deepseek"
 };
 /**
-* Headers the OpenCode Go endpoint requires, MINUS the session id.
-*
-* `/zen/go` answers 400 `Request is missing x-opencode-session` without it, and the official
-* documentation adds two requirements this preset has to respect:
-*
-*  1. the session id must be **stable per conversation** — a constant shared by every
-*     conversation violates the thing the header exists for (routing and cache affinity);
-*  2. the client must **identify itself with its own user agent** rather than a generic SDK name.
-*
-* So the session id is deliberately NOT here: `adapter.ts` injects it per request from the
-* harness's own session identity. What stays here is the static identity of this client.
+* Static headers the OpenCode Go endpoint wants on every request.
 *
-* ⚠️ The upstream documentation lists **DeepSeek Harness as a known-problematic client** for this
-* endpoint ("session information arrives on some model paths, but is missing on others"). Even
-* with everything here correct, expect this relay to be stricter than the others.
+* The session id is NOT here on purpose — it is injected per request, because the endpoint wants
+* it stable *per conversation* rather than per install. What lives here is the client identity,
+* which does not vary. See {@link OPENCODE_SESSION_FALLBACK} for what happens when the harness
+* gives us no session at all.
 */
 const OPENCODE_HEADERS = {
 	"x-opencode-client": "dsh",
@@ -224,12 +217,28 @@ const OPENCODE_HEADERS = {
 /**
 * The user agent a request to OpenCode Go identifies itself with.
 *
-* Requirement 2 above: a generic SDK name is exactly what they ask clients not to send.
+* The upstream asks clients to identify themselves rather than send a generic SDK name.
 */
 const CLIENT_USER_AGENT = "dsh-any-deepseek/1.0";
 /** Header carrying the per-conversation routing id on OpenCode Go. */
 const OPENCODE_SESSION_HEADER = "x-opencode-session";
 /**
+* Session id used when the harness gives us none.
+*
+* 🔴 **This fallback is load-bearing, not decoration**, and dropping it is a measured failure.
+*
+* The endpoint wants a per-conversation id, so the honest implementation is "use the harness's
+* `sessionId`". The problem is that the harness does not stamp one on every request — and when it
+* does not, an implementation that ONLY reads `options.sessionId` sends **no session header at
+* all**, which the endpoint rejects with 400 `MissingSessionID`: precisely the failure the header
+* exists to prevent. So the value is "harness session when present, this constant otherwise".
+*
+* A constant is a worse affinity key than a real per-session id and a far better one than none:
+* it still gives the upstream one consistent key per install. That is the same trade the sibling
+* `dsh-serenity-hooks` plugin made for the same endpoint, proved on a real host.
+*/
+const OPENCODE_SESSION_FALLBACK = "dsh-any-deepseek";
+/**
 * Presets keyed by a stable id.
 *
 * The id is what a channel entry stores (`preset: opencode-go`), so it is part of the user's
@@ -493,12 +502,42 @@ function resolveChannel(id, raw) {
 			input,
 			...(entry.reasoningEfforts ?? preset?.reasoningEfforts) === void 0 ? {} : { reasoningEfforts: entry.reasoningEfforts ?? preset?.reasoningEfforts },
 			...(entry.compat ?? preset?.compat) === void 0 ? {} : { compat: entry.compat ?? preset?.compat },
-			...(entry.headers ?? preset?.headers) === void 0 ? {} : { headers: entry.headers ?? preset?.headers },
+			/**
+			* 🔴 Headers MERGE per key; they do not replace.
+			*
+			* `entry.headers ?? preset?.headers` treated the two as alternatives, so a user who added
+			* one header of their own silently lost the preset's — and for OpenCode Go those are the
+			* headers the endpoint refuses requests without. Merging keeps the preset's required set
+			* and lets the entry win only on the keys it actually names.
+			*/
+			...mergeHeaders(preset?.headers, entry.headers) === void 0 ? {} : { headers: mergeHeaders(preset?.headers, entry.headers) },
 			warnings
 		}
 	};
 }
 /**
+* Merge preset headers with entry headers, the entry winning per key.
+*
+* HTTP field names are case-insensitive, so a user writing `X-Session-ID` must override a
+* preset's `x-session-id` rather than appearing alongside it as a second header — sending both
+* would leave the upstream to pick one, which is exactly the ambiguity the header exists to
+* resolve.
+* @param preset - headers the preset declares.
+* @param entry - headers the entry declares.
+* @returns the merged set, or undefined when neither side declares any.
+*/
+function mergeHeaders(preset, entry) {
+	if (entry === void 0) return preset === void 0 ? void 0 : { ...preset };
+	if (preset === void 0) return { ...entry };
+	const merged = {};
+	for (const [name, value] of Object.entries(preset)) merged[name] = value;
+	for (const [name, value] of Object.entries(entry)) {
+		for (const existing of Object.keys(merged)) if (existing !== name && existing.toLowerCase() === name.toLowerCase()) delete merged[existing];
+		merged[name] = value;
+	}
+	return merged;
+}
+/**
 * Resolve every channel in one configuration document.
 *
 * Problems are collected, never thrown: one broken channel must not make the rest of the
@@ -615,6 +654,29 @@ const PROVIDER_ID = "any-deepseek";
 /** Display name for the route in provider selectors. */
 const PROVIDER_NAME = "Any DeepSeek";
 /**
+* 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;
+* the host cannot know about it and must not be expected to stamp anything for it. Whatever the
+* harness happens to put on a request is only an input to a decision made here.
+*
+* The id is a stable hash of the harness session id, so:
+*  · the same conversation keeps one affinity key (what the endpoint wants), and
+*  · the value is opaque rather than echoing an internal id to a third party.
+*
+* Two hard requirements, both learned from a 400:
+*  · **never absent** — the header is always sent, so a request can never go out without it;
+*  · **stable when there is no session** — falling back to a random value would give every
+*    request a different affinity key, which is worse than a constant.
+* @param sessionId - the harness session id, when it supplied one.
+* @returns a stable, non-empty session id.
+*/
+function opencodeSessionId(sessionId) {
+	const raw = sessionId === void 0 || sessionId === null ? "" : String(sessionId).trim();
+	if (raw === "") return OPENCODE_SESSION_FALLBACK;
+	return createHash("sha256").update(raw).digest("hex").slice(0, 32);
+}
+/**
 * 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
@@ -647,14 +709,57 @@ function buildPiModel(channel) {
 	};
 }
 /**
+* 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
+* also how the host's own pi-ai adapter does it, for the same reason.
+*
+* A block that cannot be read is left OUT of the map, and the converter then substitutes a text
+* placeholder. That is deliberate: dropping the block would tell the model nothing, while a
+* placeholder at least says an image was attached and names why it could not be shown. Skipping
+* it silently is the failure mode this whole change exists to remove — the user sent an image and
+* the model answered as though the message had been text only.
+* @param options - the harness request.
+* @param readImage - the attachment seam, when the composition has one.
+* @returns resolved images keyed by attachment id.
+*/
+async function collectImages(options, readImage) {
+	const images = /* @__PURE__ */ new Map();
+	if (readImage === void 0) return images;
+	const pending = [];
+	for (const message of options.messages) {
+		if (!Array.isArray(message.content)) continue;
+		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 resolved = await Promise.all(pending.map(async (item) => {
+		try {
+			return [item.id, await readImage(item.id, item.mediaType)];
+		} catch {
+			return [item.id, void 0];
+		}
+	}));
+	for (const [id, value] of resolved) if (value !== void 0) images.set(id, value);
+	return images;
+}
+/**
 * Translate one harness request into a pi-ai context.
 *
-* The harness already projects images and files before an adapter sees them, so this is a
-* straightforward content mapping; anything the harness did not project is passed through.
 * @param options - the harness request.
+* @param images - image bytes collected by {@link collectImages}; omitted for a text-only request.
 * @returns a pi-ai context plus the one-shot system prompt, when the caller set one.
 */
-function toPiContext(options) {
+function toPiContext(options, images) {
 	const messages = options.messages.map((message) => {
 		if (message.role === "assistant") return {
 			role: "assistant",
@@ -691,7 +796,7 @@ function toPiContext(options) {
 		};
 		return {
 			role: "user",
-			content: textAndImages(message.content),
+			content: userContent(message.content, images),
 			timestamp: 0
 		};
 	});
@@ -731,14 +836,45 @@ function safeJson(raw) {
 		return {};
 	}
 }
-/** Map harness content blocks to pi-ai text/image blocks. */
-function textAndImages(content) {
+/**
+* Map one user message's blocks to pi-ai content.
+*
+* 🔴 An image block carries an `attachmentId`, never bytes — so it can only become an image when
+* {@link collectImages} resolved it. When it did not, the block becomes a TEXT PLACEHOLDER rather
+* than disappearing: the model should be told an image was attached and that it could not be
+* read. The previous version dropped such blocks entirely, which is how a user's screenshot
+* reached the plugin and the model answered as though the message had been text only.
+* @param content - harness content blocks.
+* @param images - resolved image bytes, when any were collected.
+* @returns pi-ai content blocks.
+*/
+function userContent(content, images) {
 	const out = [];
 	for (const block of content) {
 		const typed = block;
-		if (typed.type === "text" && typeof typed.text === "string") out.push({
+		if (typed.type === "text" && typeof typed.text === "string") {
+			out.push({
+				type: "text",
+				text: typed.text
+			});
+			continue;
+		}
+		if (typed.type !== "image") continue;
+		const id = typed.attachment?.attachmentId;
+		const resolved = typeof id === "string" ? images?.get(id) : void 0;
+		if (resolved !== void 0) {
+			out.push({
+				type: "image",
+				data: resolved.data,
+				mimeType: resolved.mediaType
+			});
+			continue;
+		}
+		if (typed.offloaded === true || typed.attachment?.offloaded === true) continue;
+		const name = typed.attachment?.name === void 0 ? "" : ` (${typed.attachment.name})`;
+		out.push({
 			type: "text",
-			text: typed.text
+			text: `[an image was attached${name}, but this route could not read its bytes]`
 		});
 	}
 	return out;
@@ -1125,20 +1261,20 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 		*
 		* Two things OpenCode Go requires cannot live in a preset, because neither is a constant:
 		*
-		*  · `x-opencode-session` must be STABLE PER CONVERSATION. A single shared literal satisfied
-		*    the letter of "send the header" while defeating its purpose (routing and prompt-cache
-		*    affinity), which is why the previous preset's fixed value produced 400s. The harness
-		*    already stamps every request with `sessionId`; that is the correct source.
+		*  · `x-opencode-session` must be STABLE PER CONVERSATION, and it must ALWAYS be present.
+		*    Both halves matter, and getting only the first produced a 400: an earlier version read
+		*    `options.sessionId` and sent nothing when it was absent — and the harness does not stamp
+		*    one on every request. So the value is a stable hash of the harness session id when we
+		*    have one, and a constant when we do not. A hash keeps the id opaque while staying stable
+		*    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.
 		*/
 		const requestHeaders = {};
 		if (channel.headers !== void 0 && Object.hasOwn(channel.headers, "x-opencode-client")) {
-			const session = options.sessionId === void 0 ? void 0 : String(options.sessionId);
-			if (session !== void 0 && session !== "") {
-				requestHeaders[OPENCODE_SESSION_HEADER] = session;
-				requestHeaders["X-Session-ID"] = session;
-			}
+			const session = opencodeSessionId(options.sessionId);
+			requestHeaders[OPENCODE_SESSION_HEADER] = session;
+			requestHeaders["X-Session-ID"] = session;
 			requestHeaders["User-Agent"] = CLIENT_USER_AGENT;
 		}
 		const streamOptions = {
@@ -1150,7 +1286,8 @@ var AnyDeepSeekAdapter = class extends LlmAdapter {
 			...Object.keys(requestHeaders).length === 0 ? {} : { headers: requestHeaders }
 		};
 		try {
-			const events = models.streamSimple(model, toPiContext(options), streamOptions);
+			const images = await collectImages(options, this.deps.readImage);
+			const events = models.streamSimple(model, toPiContext(options, images), streamOptions);
 			for await (const chunk of toStreamChunks(events, (observed) => {
 				outcome = observed;
 			})) yield chunk;
@@ -1757,10 +1894,45 @@ function apply(ctx, config) {
 		const value = unwrap((ctx.fiber?.config ?? config).current);
 		return typeof value === "string" ? value : void 0;
 	};
+	/**
+	* Read one image's request bytes from the harness attachment store.
+	*
+	* 🔴 This is the host seam, not something this plugin invents: a harness `ImageBlock` carries
+	* only an `attachmentId`, and `readImageRequest` is the documented way to turn that into bytes.
+	* Without this path the model cannot receive an image no matter what `input` declares — which
+	* was exactly this plugin's state until now, and why a screenshot the user sent never arrived.
+	*
+	* The target bounds are the store's own normalization request, not a guess about the relay:
+	* 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 attachments = service(ctx, "attachments");
+		if (attachments?.readImageRequest === void 0) return void 0;
+		try {
+			const stored = await attachments.readImageRequest({
+				attachmentId,
+				mediaType
+			}, {
+				width: 2048,
+				height: 2048,
+				maxBytes: 4194304
+			});
+			if (stored?.data === void 0) return void 0;
+			return {
+				data: Buffer.from(stored.data).toString("base64"),
+				mediaType: stored.mediaType ?? mediaType
+			};
+		} catch (error) {
+			console.warn(`[any-deepseek] 读不到图片 ${attachmentId}: ${error instanceof Error ? error.message : String(error)}`);
+			return;
+		}
+	};
 	const adapter = new AnyDeepSeekAdapter({
 		channels,
 		current,
 		resolveKey,
+		readImage,
 		record: makeRecorder(ctx, config),
 		now: () => /* @__PURE__ */ new Date(),
 		elapsed: () => Number(process.hrtime.bigint() / 1000000n)

package/lib/presets.d.ts

@@ -84,11 +84,27 @@ export declare const DEFAULT_MAX_TOKENS = 32768;
 /**
  * The user agent a request to OpenCode Go identifies itself with.
  *
- * Requirement 2 above: a generic SDK name is exactly what they ask clients not to send.
+ * The upstream asks clients to identify themselves rather than send a generic SDK name.
  */
 export declare const CLIENT_USER_AGENT = "dsh-any-deepseek/1.0";
 /** Header carrying the per-conversation routing id on OpenCode Go. */
 export declare const OPENCODE_SESSION_HEADER = "x-opencode-session";
+/**
+ * Session id used when the harness gives us none.
+ *
+ * 🔴 **This fallback is load-bearing, not decoration**, and dropping it is a measured failure.
+ *
+ * The endpoint wants a per-conversation id, so the honest implementation is "use the harness's
+ * `sessionId`". The problem is that the harness does not stamp one on every request — and when it
+ * does not, an implementation that ONLY reads `options.sessionId` sends **no session header at
+ * all**, which the endpoint rejects with 400 `MissingSessionID`: precisely the failure the header
+ * exists to prevent. So the value is "harness session when present, this constant otherwise".
+ *
+ * A constant is a worse affinity key than a real per-session id and a far better one than none:
+ * it still gives the upstream one consistent key per install. That is the same trade the sibling
+ * `dsh-serenity-hooks` plugin made for the same endpoint, proved on a real host.
+ */
+export declare const OPENCODE_SESSION_FALLBACK = "dsh-any-deepseek";
 /**
  * Presets keyed by a stable id.
  *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-any-deepseek",
-  "version": "0.0.7",
+  "version": "0.0.8",
   "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": {