@pi-unipi/unipi 2.3.0 → 2.4.0

package/packages/image/src/register-providers.ts

@@ -0,0 +1,220 @@
+/**
+ * @pi-unipi/image — Bridge pi's chat providers into pi-ai's images collection
+ *
+ * pi-ai ships exactly one image provider (openrouter), so out of the box image
+ * generation demands an OpenRouter account even when the user has half a dozen
+ * other providers configured. pi's own registry knows those providers and their
+ * credentials, so we re-register each one as an *images* provider backed by the
+ * single generic OpenAI-compatible adapter.
+ *
+ * The result: any OpenAI-compatible provider the user configures in pi can
+ * generate and edit images with no image-specific setup, and no per-provider
+ * code here.
+ *
+ * ## Why capability detection stays heuristic
+ * pi's model registry cannot tell us which models emit images.
+ * `provider-composer.ts` builds each registered model from an explicit field
+ * list — `{id, name, api, provider, baseUrl, reasoning, input, cost,
+ * contextWindow, maxTokens, headers, compat}` — so an extension that attaches
+ * `output: ["image"]` has it silently dropped. `ProviderModelConfig` has no
+ * `output` field at all. Hence `looksLikeImageGenerator()` name-matching, plus
+ * explicit "provider/model-id" entry as the always-available escape hatch.
+ */
+
+import * as imagesApi from "./openai-images-api.js";
+import {
+  getImagesModels,
+  listRegistryImageGenModels,
+  type ChatModelRegistry,
+  type ImageGenModel,
+} from "./models.js";
+
+/** pi-ai's `createImagesProvider`, kept structural to avoid type coupling. */
+interface CreateImagesProviderFn {
+  (input: {
+    id: string;
+    name?: string;
+    auth: unknown;
+    models: readonly ImageGenModel[];
+    api: unknown;
+  }): unknown;
+}
+
+/** The subset of a registry provider we need. */
+interface ProviderLike {
+  id: string;
+  name?: string;
+  baseUrl?: string;
+}
+
+let registered = false;
+
+/** Reset registration state. Test-only. */
+export function __resetRegistrationForTests(): void {
+  registered = false;
+}
+
+/**
+ * Group discovered generator models by provider, attaching the provider's
+ * baseUrl so the adapter knows where to POST.
+ */
+export function groupModelsByProvider(
+  models: ImageGenModel[],
+  baseUrlFor: (provider: string) => string | undefined,
+): Map<string, { baseUrl: string; models: ImageGenModel[] }> {
+  const grouped = new Map<string, { baseUrl: string; models: ImageGenModel[] }>();
+
+  for (const model of models) {
+    const baseUrl = model.baseUrl ?? baseUrlFor(model.provider);
+    // Without an endpoint the adapter cannot issue a request; skip rather than
+    // register a provider that is guaranteed to fail.
+    if (!baseUrl) continue;
+
+    let entry = grouped.get(model.provider);
+    if (!entry) {
+      entry = { baseUrl, models: [] };
+      grouped.set(model.provider, entry);
+    }
+    entry.models.push({ ...model, baseUrl });
+  }
+
+  return grouped;
+}
+
+/** Read a provider's baseUrl out of pi's registry. */
+function providerBaseUrlLookup(
+  registry: ChatModelRegistry,
+): (provider: string) => string | undefined {
+  const cache = new Map<string, string | undefined>();
+
+  return (provider: string) => {
+    if (cache.has(provider)) return cache.get(provider);
+
+    let baseUrl: string | undefined;
+    try {
+      const models = (registry.getAvailable?.() ?? registry.getAll()) as Array<{
+        provider?: string;
+        baseUrl?: string;
+      }>;
+      baseUrl = models.find((m) => m?.provider === provider && m.baseUrl)?.baseUrl;
+    } catch {
+      baseUrl = undefined;
+    }
+
+    cache.set(provider, baseUrl);
+    return baseUrl;
+  };
+}
+
+/**
+ * Register every pi provider that looks capable of image generation into
+ * pi-ai's images collection.
+ *
+ * Idempotent and best-effort: a failure here must never break the extension,
+ * since generation still works for pi-ai's built-in providers.
+ *
+ * @returns the provider ids registered.
+ */
+export async function registerRegistryImageProviders(
+  registry: ChatModelRegistry | undefined,
+  options?: { force?: boolean },
+): Promise<string[]> {
+  if (!registry) return [];
+  if (registered && !options?.force) return [];
+
+  const images = await getImagesModels();
+  if (!images) return [];
+
+  // `setProvider` is on MutableImagesModels; the built-in collection provides
+  // it, but guard in case a future pi-ai hands back an immutable one.
+  const mutable = images as unknown as {
+    setProvider?: (provider: unknown) => void;
+    getProvider?: (id: string) => unknown;
+  };
+  if (typeof mutable.setProvider !== "function") return [];
+
+  let createImagesProvider: CreateImagesProviderFn;
+  try {
+    const mod = (await import("@earendil-works/pi-ai")) as unknown as {
+      createImagesProvider?: CreateImagesProviderFn;
+    };
+    if (typeof mod.createImagesProvider !== "function") return [];
+    createImagesProvider = mod.createImagesProvider;
+  } catch {
+    return [];
+  }
+
+  const discovered = listRegistryImageGenModels(registry);
+  if (discovered.length === 0) {
+    registered = true;
+    return [];
+  }
+
+  const grouped = groupModelsByProvider(discovered, providerBaseUrlLookup(registry));
+  const added: string[] = [];
+
+  for (const [providerId, { models }] of grouped) {
+    // Never shadow a provider pi-ai serves natively — its own implementation
+    // is better informed than our generic adapter.
+    try {
+      if (mutable.getProvider?.(providerId)) continue;
+    } catch {
+      // Treat a lookup failure as "not present" and attempt registration.
+    }
+
+    try {
+      const provider = createImagesProvider({
+        id: providerId,
+        name: providerId,
+        models,
+        api: imagesApi,
+        auth: {
+          apiKey: {
+            name: `${providerId} API key`,
+            // Resolve through pi's own auth storage so the user never logs in
+            // twice. `resolve` MUST return an AuthResult (`{ auth: {...} }`);
+            // returning a bare key fails silently at request time.
+            resolve: async () => {
+              const key = await resolveProviderKey(registry, providerId);
+              return key ? { auth: { apiKey: key }, source: `pi:${providerId}` } : undefined;
+            },
+          },
+        },
+      });
+
+      mutable.setProvider(provider);
+      added.push(providerId);
+    } catch {
+      // One bad provider must not stop the rest.
+    }
+  }
+
+  registered = true;
+  return added;
+}
+
+/** Resolve a provider key from pi's auth storage, falling back to the env. */
+async function resolveProviderKey(
+  registry: ChatModelRegistry,
+  provider: string,
+): Promise<string | undefined> {
+  try {
+    const key = await registry.getApiKeyForProvider?.(provider);
+    if (key) return key;
+  } catch {
+    // Fall through to the environment.
+  }
+  const envName = `${provider.toUpperCase().replace(/[^A-Z0-9]/g, "_")}_API_KEY`;
+  return process.env[envName] || undefined;
+}
+
+/** Provider ids pi-ai can currently generate with, after registration. */
+export function registeredProviderIds(images: {
+  getProviders?: () => ReadonlyArray<{ id: string }>;
+}): string[] {
+  try {
+    return (images.getProviders?.() ?? []).map((p) => p.id);
+  } catch {
+    return [];
+  }
+}

package/packages/image/src/tools.ts

@@ -12,7 +12,9 @@ import { IMAGE_TOOLS } from "@pi-unipi/core";
 
 import { generateImage } from "./generate.js";
 import { loadImage } from "./image-source.js";
+import { registerRegistryImageProviders } from "./register-providers.js";
 import {
+  findProviderBaseUrl,
   formatModelRef,
   listAllImageGenModels,
   resolveImageGenModel,
@@ -84,12 +86,15 @@ function registerGenerateTool(pi: ExtensionAPI): void {
     name: IMAGE_TOOLS.GENERATE,
     label: "Generate Image",
     description:
-      "Generate an image from a text prompt using an image model. " +
-      "The image is returned inline and, when enabled, saved to disk.",
+      "Generate an image from a text prompt, or edit an existing image by " +
+      "passing `image`. The result is returned inline and, when enabled, saved to disk.",
     promptSnippet: "Generate an image from a text prompt.",
     promptGuidelines: [
       "Use image_generate to create images from a text description.",
       "Write a detailed prompt — subject, style, composition and lighting all help.",
+      "Pass `image` to edit an existing image instead of generating a new one.",
+      "Editing regenerates the whole image, so unmentioned details may change.",
+      "Describe what you DO want; negation is unreliable in image models.",
       "Omit model to use the one configured in /unipi:image-settings.",
       "Generated images cost money per call; do not regenerate without being asked.",
     ],
@@ -97,6 +102,13 @@ function registerGenerateTool(pi: ExtensionAPI): void {
       prompt: Type.String({
         description: "Description of the image to generate. Be specific.",
       }),
+      image: Type.Optional(
+        Type.String({
+          description:
+            "Source image to edit: a local file path, data: URL, or base64 data. " +
+            "When set, the model edits this image instead of generating from scratch.",
+        }),
+      ),
       model: Type.Optional(
         Type.String({
           description:
@@ -109,22 +121,41 @@ function registerGenerateTool(pi: ExtensionAPI): void {
       try {
         const config = loadConfig();
         const registry = getRegistry(ctx);
+        // Bridge pi's own providers into pi-ai's images collection so the user
+        // is not forced onto OpenRouter. Idempotent and best-effort.
+        await registerRegistryImageProviders(registry);
         // Include image models contributed by registered providers, so the
         // tool can resolve anything the settings picker offers.
         const models = await listAllImageGenModels(registry);
 
         const requested = params.model?.trim() || config.generate.model;
-        const resolved = resolveImageGenModel(requested, models);
-        if (typeof resolved === "string") return errorResult(resolved);
+        const maybeResolved = resolveImageGenModel(requested, models);
+        if (typeof maybeResolved === "string") return errorResult(maybeResolved);
+
+        // A model may arrive without an endpoint — notably a user-typed
+        // "provider/model-id", accepted at face value. Fill it in from the
+        // registry so the adapter knows where to POST.
+        const registryBaseUrl = maybeResolved.baseUrl
+          ? undefined
+          : findProviderBaseUrl(registry, maybeResolved.provider);
+        const resolved = registryBaseUrl
+          ? { ...maybeResolved, baseUrl: registryBaseUrl }
+          : maybeResolved;
 
         // pi-ai resolves image auth from its own credential store; only fall
         // back to pi's chat-provider key when that comes up empty.
         const fallbackKey = await resolveApiKey(registry, resolved.provider);
 
+        // An input image switches the request into edit mode.
+        const sourceImage = params.image?.trim()
+          ? loadImage(params.image, ctx.cwd ?? process.cwd())
+          : undefined;
+
         const result = await generateImage({
           prompt: params.prompt,
           model: resolved,
           ...(fallbackKey ? { apiKey: fallbackKey } : {}),
+          ...(sourceImage ? { inputImage: sourceImage } : {}),
           signal,
           outputDir: config.generate.saveToDisk ? getOutputDir(config) : undefined,
         });
@@ -134,8 +165,8 @@ function registerGenerateTool(pi: ExtensionAPI): void {
           .filter((path): path is string => Boolean(path));
 
         const summary = [
-          `Generated ${result.images.length} image${result.images.length === 1 ? "" : "s"} ` +
-            `with ${formatModelRef(resolved)}.`,
+          `${sourceImage ? "Edited" : "Generated"} ${result.images.length} ` +
+            `image${result.images.length === 1 ? "" : "s"} with ${formatModelRef(resolved)}.`,
           saved.length > 0 ? `Saved to:\n${saved.map((p) => `  ${p}`).join("\n")}` : "",
           config.generate.saveToDisk && saved.length === 0
             ? "Could not write to the output directory — returning the image inline only."

package/packages/image/src/tui/settings-dialog.ts

@@ -23,6 +23,7 @@ import {
   type ChatModelRegistry,
 } from "../models.js";
 import { ImageModelSelectorOverlay, type SelectableModel } from "./model-selector.js";
+import { registerRegistryImageProviders } from "../register-providers.js";
 
 const EXIT = "__exit__";
 
@@ -239,6 +240,10 @@ async function collectModels(
     .modelRegistry;
 
   if (kind === "generate") {
+    // Bridge pi's providers in first, so a model the user can actually run is
+    // not flagged "no image route" purely because we had not registered it yet.
+    await registerRegistryImageProviders(registry);
+
     // Include models from providers registered by other extensions, not just
     // pi-ai's built-in OpenRouter catalog.
     const models = await listAllImageGenModels(registry);
@@ -253,7 +258,7 @@ async function collectModels(
       // so flag them rather than letting the user pick a dead option.
       unavailable:
         generating.length > 0 && !generating.includes(m.provider)
-          ? "cannot generate"
+          ? "no image route"
           : undefined,
     }));
   }

package/packages/info-screen/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/info-screen",
-  "version": "2.3.0",
+  "version": "2.4.0",
   "description": "Dashboard and module registry for Unipi — configurable info overlay with tabbed groups",
   "type": "module",
   "main": "index.ts",
@@ -33,7 +33,7 @@
     "access": "public"
   },
   "dependencies": {
-    "@pi-unipi/core": "2.3.0"
+    "@pi-unipi/core": "2.4.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": "^0.80.0",

package/packages/input-shortcuts/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/input-shortcuts",
-  "version": "2.3.0",
+  "version": "2.4.0",
   "description": "Keyboard shortcuts for stash/restore, undo/redo, clipboard, and thinking toggle — chord-based overlay system",
   "type": "module",
   "main": "src/index.ts",
@@ -33,7 +33,7 @@
     "access": "public"
   },
   "dependencies": {
-    "@pi-unipi/core": "2.3.0"
+    "@pi-unipi/core": "2.4.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": "^0.80.0",

package/packages/kanboard/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/kanboard",
-  "version": "2.3.0",
+  "version": "2.4.0",
   "description": "Visualization layer for unipi workflow — HTTP server with htmx/Alpine.js UI, modular parsers, TUI overlay, and kanban board",
   "type": "module",
   "main": "index.ts",
@@ -39,7 +39,7 @@
     "access": "public"
   },
   "dependencies": {
-    "@pi-unipi/core": "2.3.0"
+    "@pi-unipi/core": "2.4.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": "^0.80.0",

package/packages/mcp/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/mcp",
-  "version": "2.3.0",
+  "version": "2.4.0",
   "description": "MCP server management extension for Pi coding agent — browse, add, configure, and use MCP servers",
   "type": "module",
   "main": "src/index.ts",
@@ -27,7 +27,7 @@
     "README.md"
   ],
   "dependencies": {
-    "@pi-unipi/core": "2.3.0"
+    "@pi-unipi/core": "2.4.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": "^0.80.0",

package/packages/memory/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/memory",
-  "version": "2.3.0",
+  "version": "2.4.0",
   "description": "Persistent cross-session memory with MemPalace backend (auto-installed) and SQLite fallback for Pi coding agent",
   "type": "module",
   "main": "index.ts",
@@ -43,8 +43,8 @@
     "better-sqlite3": "^12.9.0",
     "sqlite-vec": "^0.1.9",
     "js-yaml": "^4.1.0",
-    "@pi-unipi/core": "2.3.0",
-    "@pi-unipi/info-screen": "2.3.0"
+    "@pi-unipi/core": "2.4.0",
+    "@pi-unipi/info-screen": "2.4.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": "^0.80.0",

package/packages/milestone/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/milestone",
-  "version": "2.3.0",
+  "version": "2.4.0",
   "description": "Lifecycle layer for project-level goals — MILESTONES.md tracking, session hooks, auto-sync",
   "type": "module",
   "main": "index.ts",
@@ -29,7 +29,7 @@
     "access": "public"
   },
   "dependencies": {
-    "@pi-unipi/core": "2.3.0"
+    "@pi-unipi/core": "2.4.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": "^0.80.0",

package/packages/notify/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/notify",
-  "version": "2.3.0",
+  "version": "2.4.0",
   "description": "Cross-platform notification extension for Pi — native OS, Gotify, and Telegram notifications for agent lifecycle events",
   "type": "module",
   "main": "index.ts",
@@ -34,7 +34,7 @@
     "access": "public"
   },
   "dependencies": {
-    "@pi-unipi/core": "2.3.0",
+    "@pi-unipi/core": "2.4.0",
     "node-notifier": "^10.0.1"
   },
   "peerDependencies": {

package/packages/ralph/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/ralph",
-  "version": "2.3.0",
+  "version": "2.4.0",
   "description": "Long-running iterative development loops for Pi coding agent",
   "type": "module",
   "main": "index.ts",
@@ -27,8 +27,8 @@
     "access": "public"
   },
   "dependencies": {
-    "@pi-unipi/core": "2.3.0",
-    "@pi-unipi/info-screen": "2.3.0"
+    "@pi-unipi/core": "2.4.0",
+    "@pi-unipi/info-screen": "2.4.0"
   },
   "peerDependencies": {
     "@earendil-works/pi-ai": "^0.80.0",

package/packages/subagents/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-unipi/subagents",
-  "version": "2.3.0",
+  "version": "2.4.0",
   "description": "Subagents for UniPi — parallel execution, file locking, workflow integration",
   "type": "module",
   "main": "dist/index.js",
@@ -17,9 +17,9 @@
     "test": "npx tsx --test src/__tests__/*.test.ts"
   },
   "dependencies": {
-    "@pi-unipi/core": "2.3.0",
-    "@pi-unipi/workflow": "2.3.0",
-    "@pi-unipi/info-screen": "2.3.0",
+    "@pi-unipi/core": "2.4.0",
+    "@pi-unipi/workflow": "2.4.0",
+    "@pi-unipi/info-screen": "2.4.0",
     "@earendil-works/pi-agent-core": "^0.80.0"
   },
   "devDependencies": {