codsh-bundle 0.6.1 → 0.8.0

package/lib/types/vision.d.ts

@@ -4,14 +4,16 @@
  * DeepSeek Vision routes receive first-class image blocks before this module
  * is involved. For text-only routes such as Flash and Pro, codsh gives an
  * image two honest lives: it is always saved to a stable file the agent's
- * tools can touch — inspect, commit, embed. And when a vision sidecar is
- * configured (`CODSH_VISION_*`: any OpenAI-compatible multimodal endpoint),
- * the image is also described into text the model can actually read: everything in it
- * transcribed, structure narrated. Both ride the same message the person
+ * tools can touch — inspect, commit, embed. An explicit `CODSH_VISION_*`
+ * sidecar, or DeepSeek's built-in Vision Exp bridge when no sidecar is set,
+ * can also describe it into text the model can actually read: everything in
+ * it transcribed, structure narrated. Both ride the same message the person
  * sent, so they persist in durable history and survive `--resume`.
  * @module codsh-bundle/src/vision
  */
-import type { EncodedImageAttachment, ImageAttachmentLimits } from '@deepseek-ai/dsh-attachment/types';
+import type { EncodedImageAttachment, ImageAttachmentLimits, ImageAttachmentRef } from '@deepseek-ai/dsh-attachment/types';
+import type { GenerateOptions, LlmCallConfig, ModelModality, StreamChunk } from '@deepseek-ai/dsh-llm';
+import type { SessionId } from '@deepseek-ai/dsh-session';
 /** A vision sidecar: an OpenAI-compatible endpoint that can see. */
 export interface VisionConfig {
     /** The API base, e.g. `https://open.bigmodel.cn/api/paas/v4`. */
@@ -21,6 +23,16 @@ export interface VisionConfig {
     /** The multimodal model to ask. */
     model: string;
 }
+/** The sibling route that lends sight to DeepSeek's text-only models. */
+export declare const DEEPSEEK_VISION_MODEL = "deepseek-v4-flash-vision-exp";
+/** The provider-owned preparation seam needed for one auxiliary vision call. */
+export interface VisionLlm {
+    prepareCall(config: LlmCallConfig, signal?: AbortSignal): Promise<{
+        readonly config: LlmCallConfig;
+        readonly inputModalities?: readonly ModelModality[];
+        stream(options: GenerateOptions): AsyncIterable<StreamChunk>;
+    }>;
+}
 /**
  * The sidecar from the environment, or undefined when none is configured.
  * @param env - the process environment.
@@ -36,6 +48,23 @@ export declare function visionConfigFromEnv(env: Record<string, string | undefin
  * @throws on timeout, a non-2xx answer, or an answer with no text.
  */
 export declare function describeImage(image: EncodedImageAttachment, config: VisionConfig, signal?: AbortSignal): Promise<string>;
+/**
+ * Ask DeepSeek's native vision sibling to describe one durable image.
+ *
+ * This is deliberately a one-shot auxiliary call: it receives no conversation
+ * history, system prompt, or tools, and its answer is returned as plain text
+ * for the selected text model's ordinary user turn.
+ * @param image - the image reference already admitted to the durable store.
+ * @param llm - the provider-neutral runtime serving the DeepSeek route.
+ * @param options - optional caller cancellation and request attribution.
+ * @returns the visible text emitted by the vision model.
+ * @throws when the route cannot see, the provider fails, the call is aborted,
+ *   or the response contains no visible text.
+ */
+export declare function describeImageWithLlm(image: ImageAttachmentRef, llm: VisionLlm, options?: {
+    signal?: AbortSignal;
+    sessionId?: SessionId;
+}): Promise<string>;
 /**
  * The upstream store's default admission limits, for use when no store is
  * mounted: the sidecar payload is bounded by the same line either way.

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "codsh-bundle",
   "description": "The codsh runtime: the interactive TTY surface and code-cli agent preset, installed into a dsh profile. Users install codsh-cli (the launcher) — this package is what it registers.",
-  "version": "0.6.1",
+  "version": "0.8.0",
   "type": "module",
   "license": "MIT",
   "main": "lib/index.js",
@@ -50,6 +50,7 @@
     "@deepseek-ai/dsh-lsp-stdio": "^0.1.1-rc.2",
     "@deepseek-ai/dsh-persona": "^0.1.1-rc.2",
     "@deepseek-ai/dsh-plan-mode": "^0.1.1-rc.2",
+    "@deepseek-ai/dsh-skill": "^0.1.1-rc.2",
     "@deepseek-ai/dsh-skill-filesystem": "^0.1.1-rc.2",
     "@deepseek-ai/dsh-terminal": "^0.1.1-rc.2",
     "@deepseek-ai/dsh-terminal-bash": "^0.1.1-rc.2",