feat(ai): dynamic Routstr model lineup, vision icons, and image attachments

app/.monicon/icons.js

@@ -1029,5 +1029,20 @@ module.exports = {
     "svg": "<svg viewBox=\"0 0 58 44\" width=\"1em\" height=\"1em\" ><path fill=\"currentColor\" fill-rule=\"evenodd\" clip-rule=\"evenodd\" d=\"M0 44V0H14.7304V13.4775L21.2348 0H35.9652V13.4775L42.4696 0H57.2V44H42.4696V30.5225L35.9652 44H21.2348V30.5225L14.7304 44H0ZM12.4348 2.29565H2.29565V39.2432L12.4348 18.2342V2.29565ZM44.7652 41.7043H54.9044V4.75676L44.7652 25.7658V41.7043ZM34.5241 41.7043L53.5431 2.29565H43.9107L24.8917 41.7043H34.5241ZM32.3083 2.29565H22.6759L3.65691 41.7043H13.2893L32.3083 2.29565ZM33.6696 4.75676L23.5304 25.7658V39.2432L33.6696 18.2342V4.75676Z\"/></svg>",
     "width": 16,
     "height": 16
+  },
+  "ri:google-fill": {
+    "svg": "<svg xmlns=\"http://www.w3.org/2000/svg\" width=\"1em\" height=\"1em\" viewBox=\"0 0 24 24\"><path fill=\"currentColor\" d=\"M3.064 7.51A10 10 0 0 1 12 2c2.695 0 4.959.991 6.69 2.605l-2.867 2.868C14.786 6.482 13.468 5.977 12 5.977c-2.605 0-4.81 1.76-5.595 4.123c-.2.6-.314 1.24-.314 1.9s.114 1.3.314 1.9c.786 2.364 2.99 4.123 5.595 4.123c1.345 0 2.49-.355 3.386-.955a4.6 4.6 0 0 0 1.996-3.018H12v-3.868h9.418c.118.654.182 1.336.182 2.045c0 3.046-1.09 5.61-2.982 7.35C16.964 21.105 14.7 22 12 22A9.996 9.996 0 0 1 2 12c0-1.614.386-3.14 1.064-4.49\"/></svg>",
+    "width": "1em",
+    "height": "1em"
+  },
+  "mdi:image-outline": {
+    "svg": "<svg xmlns=\"http://www.w3.org/2000/svg\" width=\"1em\" height=\"1em\" viewBox=\"0 0 24 24\"><path fill=\"currentColor\" d=\"M19 19H5V5h14m0-2H5a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2V5a2 2 0 0 0-2-2m-5.04 9.29l-2.75 3.54l-1.96-2.36L6.5 17h11z\"/></svg>",
+    "width": "1em",
+    "height": "1em"
+  },
+  "mdi:cloud-off-outline": {
+    "svg": "<svg xmlns=\"http://www.w3.org/2000/svg\" width=\"1em\" height=\"1em\" viewBox=\"0 0 24 24\"><path fill=\"currentColor\" d=\"M19.8 22.6L17.15 20H6.5q-2.3 0-3.9-1.6T1 14.5q0-1.92 1.19-3.42q1.19-1.51 3.06-1.93q.08-.2.15-.39q.1-.19.15-.41L1.4 4.2l1.4-1.4l18.4 18.4M6.5 18h8.65L7.1 9.95q-.05.28-.07.55q-.03.23-.03.5h-.5q-1.45 0-2.47 1.03Q3 13.05 3 14.5T4.03 17q1.02 1 2.47 1m15.1.75l-1.45-1.4q.43-.35.64-.81T21 15.5q0-1.05-.73-1.77q-.72-.73-1.77-.73H17v-2q0-2.07-1.46-3.54Q14.08 6 12 6q-.67 0-1.3.16q-.63.17-1.2.52L8.05 5.23q.88-.6 1.86-.92Q10.9 4 12 4q2.93 0 4.96 2.04Q19 8.07 19 11q1.73.2 2.86 1.5q1.14 1.28 1.14 3q0 1-.37 1.81q-.38.84-1.03 1.44m-6.77-6.72\"/></svg>",
+    "width": "1em",
+    "height": "1em"
   }
 };

app/__tests__/aiSendAssembly.test.ts

@@ -0,0 +1,207 @@
+/**
+ * Pins the multimodal request-assembly contract shared by `useAiSend`'s
+ * send and retry flows (`assembleApiMessages` is the ONE assembly seam, so
+ * testing it pins both), the routstrStore persistence tolerance for the
+ * new attachment/lineup fields, and the vision-aware candidate filter's
+ * building blocks in `format.ts`.
+ */
+import {
+  measureMessageContent,
+  type RoutstrChatMessage,
+  type RoutstrModel,
+} from '@/shared/lib/routstr/api';
+import type { ChatAttachment, RoutstrMessage } from '@/shared/stores/profile/routstrStore';
+import {
+  MAX_INLINE_IMAGES,
+  MAX_INLINE_IMAGE_TURNS,
+  assembleApiMessages,
+  stripImageParts,
+} from '@/features/ai/lib/assembleApiMessages';
+import {
+  entryForSlot,
+  estimateTurnCostSatsFromPricing,
+  resolveCandidateEntries,
+  resolveSelectedEntry,
+} from '@/features/ai/lib/format';
+import { deriveLineup } from '@/shared/lib/routstr/lineup';
+import fixture from './fixtures/routstr-models.fixture.json';
+
+const att = (n: number): ChatAttachment => ({
+  localUri: `file:///photos/img-${n}.jpg`,
+  mimeType: 'image/jpeg',
+  width: 100,
+  height: 100,
+});
+
+const msg = (
+  id: string,
+  role: 'user' | 'assistant',
+  content: string,
+  attachments?: ChatAttachment[]
+): RoutstrMessage => ({ id, role, content, timestamp: 1, attachments });
+
+/** Encoder stub: data-URL derived from the URI; records call counts. */
+function makeEncoder(missing: Set<string> = new Set()) {
+  const calls = new Map<string, number>();
+  const encode = async (attachment: ChatAttachment) => {
+    calls.set(attachment.localUri, (calls.get(attachment.localUri) ?? 0) + 1);
+    if (missing.has(attachment.localUri)) return null;
+    return `data:image/jpeg;base64,${attachment.localUri}`;
+  };
+  return { encode, calls };
+}
+
+const imageParts = (m: RoutstrChatMessage): number =>
+  typeof m.content === 'string' ? 0 : m.content.filter((p) => p.type === 'image_url').length;
+
+describe('assembleApiMessages', () => {
+  it('keeps text-only turns as plain string content (pre-image wire shape unchanged)', async () => {
+    const { encode } = makeEncoder();
+    const { messages, imageCount } = await assembleApiMessages(
+      [msg('u1', 'user', 'hi'), msg('a1', 'assistant', 'hello'), msg('u2', 'user', 'ok')],
+      encode
+    );
+    expect(imageCount).toBe(0);
+    expect(messages).toEqual([
+      { role: 'user', content: 'hi' },
+      { role: 'assistant', content: 'hello' },
+      { role: 'user', content: 'ok' },
+    ]);
+  });
+
+  it('expands attachment turns into text + image_url content parts', async () => {
+    const { encode } = makeEncoder();
+    const { messages, imageCount } = await assembleApiMessages(
+      [msg('u1', 'user', 'what is this?', [att(1), att(2)])],
+      encode
+    );
+    expect(imageCount).toBe(2);
+    expect(messages[0].content).toEqual([
+      { type: 'text', text: 'what is this?' },
+      { type: 'image_url', image_url: { url: `data:image/jpeg;base64,${att(1).localUri}` } },
+      { type: 'image_url', image_url: { url: `data:image/jpeg;base64,${att(2).localUri}` } },
+    ]);
+  });
+
+  it(`inlines images only from the newest ${MAX_INLINE_IMAGE_TURNS} attachment turns; older image turns assemble text-only`, async () => {
+    const { encode, calls } = makeEncoder();
+    const path = [
+      msg('u1', 'user', 'first', [att(1)]),
+      msg('a1', 'assistant', 'r1'),
+      msg('u2', 'user', 'second', [att(2)]),
+      msg('a2', 'assistant', 'r2'),
+      msg('u3', 'user', 'third', [att(3)]),
+      msg('a3', 'assistant', 'r3'),
+    ];
+    const { messages, imageCount } = await assembleApiMessages(path, encode);
+    expect(imageCount).toBe(2);
+    expect(messages[0]).toEqual({ role: 'user', content: 'first' }); // outside the window
+    expect(imageParts(messages[2])).toBe(1);
+    expect(imageParts(messages[4])).toBe(1);
+    // The windowed-out attachment is never even read from disk.
+    expect(calls.has(att(1).localUri)).toBe(false);
+  });
+
+  it(`caps inline images at ${MAX_INLINE_IMAGES} per request, newest turns first`, async () => {
+    const { encode } = makeEncoder();
+    const path = [
+      msg('u1', 'user', 'many', [att(1), att(2), att(3)]),
+      msg('a1', 'assistant', 'r'),
+      msg('u2', 'user', 'more', [att(4), att(5), att(6)]),
+    ];
+    const { messages, imageCount } = await assembleApiMessages(path, encode);
+    expect(imageCount).toBe(MAX_INLINE_IMAGES);
+    expect(imageParts(messages[2])).toBe(3); // newest turn gets its full set
+    expect(imageParts(messages[0])).toBe(1); // older turn gets the remainder
+  });
+
+  it('degrades a missing/expired local URI to text-only with the other images intact', async () => {
+    const { encode } = makeEncoder(new Set([att(1).localUri]));
+    const { messages, imageCount } = await assembleApiMessages(
+      [msg('u1', 'user', 'broken + ok', [att(1), att(2)])],
+      encode
+    );
+    expect(imageCount).toBe(1);
+    expect(imageParts(messages[0])).toBe(1);
+    // The text part survives regardless.
+    expect((messages[0].content as { type: string }[])[0]).toEqual({
+      type: 'text',
+      text: 'broken + ok',
+    });
+  });
+
+  it('send/retry parity is structural: identical path in → identical wire messages out', async () => {
+    const path = [msg('u1', 'user', 'q', [att(1)]), msg('a1', 'assistant', 'r')];
+    const a = await assembleApiMessages(path, makeEncoder().encode);
+    const b = await assembleApiMessages(path, makeEncoder().encode);
+    expect(a).toEqual(b);
+  });
+
+  it('stripImageParts degrades a multimodal request to plain text (no-vision-candidate fallback)', async () => {
+    const { encode } = makeEncoder();
+    const { messages } = await assembleApiMessages(
+      [msg('u1', 'user', 'look', [att(1), att(2)]), msg('a1', 'assistant', 'reply')],
+      encode
+    );
+    const stripped = stripImageParts(messages);
+    expect(stripped).toEqual([
+      { role: 'user', content: 'look' },
+      { role: 'assistant', content: 'reply' },
+    ]);
+    expect(measureMessageContent(stripped).imageParts).toBe(0);
+  });
+
+  it('measureMessageContent counts text chars + image parts without serialising payloads', async () => {
+    const { encode } = makeEncoder();
+    const { messages } = await assembleApiMessages(
+      [msg('u1', 'user', '12345', [att(1)]), msg('a1', 'assistant', '678')],
+      encode
+    );
+    expect(measureMessageContent(messages)).toEqual({ textChars: 8, imageParts: 1 });
+  });
+});
+
+describe('vision-aware candidate resolution (format.ts over the derived lineup)', () => {
+  const { lineup } = deriveLineup(fixture.data as unknown as RoutstrModel[]);
+
+  it('resolves a (provider, tier) slot and returns null only when no lineup exists', () => {
+    expect(resolveSelectedEntry('openai', 'auto', 1_000_000, lineup)?.modelId).toBe('gpt-5.5');
+    expect(resolveSelectedEntry('openai', 'auto', 1_000_000, null)).toBeNull();
+  });
+
+  it('falls back across providers in the same tier before other tiers', () => {
+    const chain = resolveCandidateEntries('claude', 'max', lineup);
+    expect(chain[0]?.modelId).toBe(entryForSlot(lineup, 'claude', 'max')?.modelId);
+    // Next candidates are the other providers' max cells, in provider order.
+    expect(chain[1]?.modelId).toBe(entryForSlot(lineup, 'openai', 'max')?.modelId);
+    expect(chain.length).toBe(12); // every filled cell exactly once
+    expect(new Set(chain.map((e) => e.modelId)).size).toBe(12);
+  });
+
+  it('prefers the first affordable candidate, else the primary at any cost', () => {
+    // Balance below every max_cost → primary comes back unaffordable-first.
+    const broke = resolveSelectedEntry('claude', 'max', 0, lineup);
+    expect(broke?.modelId).toBe(entryForSlot(lineup, 'claude', 'max')?.modelId);
+    // A balance that can only afford the cheapest cells skips to them.
+    const cheap = resolveSelectedEntry('openai', 'max', 700, lineup);
+    expect(cheap).not.toBeNull();
+    expect(cheap!.satsPricing.max_cost!).toBeLessThanOrEqual(700);
+  });
+
+  it('every selected entry carries visionInput so the send path can filter a chain with images', () => {
+    const chain = resolveCandidateEntries('google', 'max', lineup);
+    expect(chain.every((e) => typeof e.visionInput === 'boolean')).toBe(true);
+    expect(chain.filter((e) => e.visionInput).length).toBeGreaterThan(0);
+  });
+
+  it('adds per-image fees into the turn estimate (Gemini undercount fix)', () => {
+    const gemini = entryForSlot(lineup, 'google', 'max')!;
+    const base = estimateTurnCostSatsFromPricing(gemini.satsPricing);
+    const withImages = estimateTurnCostSatsFromPricing(gemini.satsPricing, 2);
+    expect(base).not.toBeNull();
+    expect(withImages!).toBeCloseTo(base! + gemini.satsPricing.image! * 2, 6);
+    // Models without an image fee are unchanged.
+    const noFee = estimateTurnCostSatsFromPricing({ ...gemini.satsPricing, image: null }, 2);
+    expect(noFee).toBeCloseTo(base!, 6);
+  });
+});