@pi-unipi/image 2.12.0 → 2.14.0

package/README.md

@@ -7,7 +7,7 @@ Image generation and image recognition tools for the agent.
 | Tool | Description |
 |------|-------------|
 | `image_generate` | Generate an image from a text prompt. Returned inline and saved to disk. |
-| `image_recognize` | Analyze an image with a vision model. Accepts a file path, `data:` URL, or base64. |
+| `image_recognize` | Analyze an image with a vision model. Accepts a file path, `data:` URL, or base64. Automatically hidden while the session model itself has vision. |
 
 ## Commands
 
@@ -63,6 +63,13 @@ works. Supported: PNG, JPEG, GIF, WebP. Remote URLs are not fetched.
 Prefer file paths — inlining base64 into the conversation is far more
 expensive in tokens.
 
+**Vision models don't get this tool.** When the session's current model already
+accepts image input, `image_recognize` is dropped from the active tool set —
+the model reads images natively through pi's own tools, so a separate
+recognition round-trip through another model would only duplicate that ability
+and burn context. Switch to a text-only model (via `/model`) and the tool comes
+back automatically.
+
 ## Configuration
 
 `~/.unipi/config/image/config.json`:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/image",
-  "version": "2.12.0",
+  "version": "2.14.0",
   "description": "Image generation and image recognition tools for the Pi coding agent",
   "type": "module",
   "main": "src/index.ts",
@@ -34,7 +34,7 @@
     "access": "public"
   },
   "dependencies": {
-    "@pi-unipi/core": "2.12.0"
+    "@pi-unipi/core": "2.14.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-ai": "^0.84.0",

package/src/index.ts

@@ -20,7 +20,13 @@ import {
 
 import { registerImageCommands } from "./commands.js";
 import { registerImageTools } from "./tools.js";
-import { listImageGenModels, listVisionModels, type ChatModelRegistry } from "./models.js";
+import {
+  applyRecognizeGating,
+  isVisionModel,
+  listImageGenModels,
+  listVisionModels,
+  type ChatModelRegistry,
+} from "./models.js";
 import { registerRegistryImageProviders } from "./register-providers.js";
 import { loadConfig } from "./settings.js";
 
@@ -37,10 +43,27 @@ function getInfoRegistry() {
   ).__unipi_info_registry;
 }
 
+/**
+ * Hide image_recognize when the session model can natively see images, and
+ * restore it when a text-only model takes over. Returns whether the tool is
+ * provided after gating.
+ */
+function applyVisionGating(pi: ExtensionAPI, model: unknown): boolean {
+  const active = pi.getActiveTools();
+  const next = applyRecognizeGating(active, model, IMAGE_TOOLS.RECOGNIZE);
+  if (next !== active) pi.setActiveTools(next);
+  return next.includes(IMAGE_TOOLS.RECOGNIZE);
+}
+
 export default function (pi: ExtensionAPI) {
   registerImageTools(pi);
   registerImageCommands(pi);
 
+  pi.on("model_select", (event) => {
+    if (!loadConfig().recognize.enabled) return;
+    applyVisionGating(pi, event.model);
+  });
+
   pi.on("session_start", async (_event, ctx) => {
     const config = loadConfig();
 
@@ -51,9 +74,16 @@ export default function (pi: ExtensionAPI) {
       (ctx as unknown as { modelRegistry?: ChatModelRegistry }).modelRegistry,
     ).catch(() => undefined);
 
+    // A vision-capable session model reads images itself, so image_recognize
+    // would only duplicate that ability. Drop it from the active tool set;
+    // the model_select handler above restores it when a text-only model is
+    // chosen later in the same session.
+    const recognizeProvided =
+      config.recognize.enabled && applyVisionGating(pi, ctx.model);
+
     const tools: string[] = [];
     if (config.generate.enabled) tools.push(IMAGE_TOOLS.GENERATE);
-    if (config.recognize.enabled) tools.push(IMAGE_TOOLS.RECOGNIZE);
+    if (recognizeProvided) tools.push(IMAGE_TOOLS.RECOGNIZE);
 
     emitEvent(pi, UNIPI_EVENTS.MODULE_READY, {
       name: MODULES.IMAGE,
@@ -92,9 +122,11 @@ export default function (pi: ExtensionAPI) {
           .modelRegistry;
         const vision = chatRegistry ? listVisionModels(chatRegistry) : [];
 
-        const recognize = current.recognize.enabled
-          ? current.recognize.model || "Session model"
-          : "Disabled";
+        const recognize = !current.recognize.enabled
+          ? "Disabled"
+          : isVisionModel(ctx.model)
+            ? "Hidden (model has vision)"
+            : current.recognize.model || "Session model";
 
         return {
           generate: { value: generate },

package/src/models.ts

@@ -383,7 +383,7 @@ export function listVisionModels(registry: ChatModelRegistry): VisionModel[] {
   return models.filter(isVisionModel);
 }
 
-function isVisionModel(model: unknown): model is VisionModel {
+export function isVisionModel(model: unknown): model is VisionModel {
   if (model === null || typeof model !== "object") return false;
   const candidate = model as Partial<VisionModel>;
   if (typeof candidate.id !== "string" || typeof candidate.provider !== "string") {
@@ -392,6 +392,29 @@ function isVisionModel(model: unknown): model is VisionModel {
   return Array.isArray(candidate.input) && candidate.input.includes("image");
 }
 
+/**
+ * Active tool names after hiding `recognizeTool` for a vision-capable model.
+ *
+ * A model that accepts image input can read images natively (pi's own read
+ * tool hands it the pixels), so a separate image_recognize round-trip through
+ * another model only duplicates the ability and burns system-prompt context.
+ * Text-only models get the tool back. Models that do not declare their input
+ * modalities are treated as non-vision, matching `isVisionModel`.
+ */
+export function applyRecognizeGating(
+  active: string[],
+  model: unknown,
+  recognizeTool: string,
+): string[] {
+  const vision = isVisionModel(model);
+  const present = active.includes(recognizeTool);
+  // Already correct: hidden for a vision model, or provided for a text-only one.
+  if (vision !== present) return active;
+  return vision
+    ? active.filter((name) => name !== recognizeTool)
+    : [...active, recognizeTool];
+}
+
 /**
  * Resolve a vision-model reference, restricted to image-capable models.
  *