[{"_1":2,"_180":-5,"_181":-5},"loaderData",{"_3":4,"_96":97},"root",{"_5":6,"_37":38,"_69":70},"ctx",{"_7":8,"_23":-5,"_24":25,"_26":27,"_28":20,"_29":20,"_30":20,"_31":-5,"_32":33,"_34":20,"_35":20,"_36":20},"auth",{"_9":-5,"_10":-5,"_11":-5,"_12":-5,"_13":-5,"_14":-5,"_15":-5,"_16":-5,"_17":-5,"_18":-5,"_19":20,"_21":22},"sessionClaims","sessionId","sessionStatus","userId","actor","orgId","orgRole","orgSlug","orgPermissions","factorVerificationAge","isAuthenticated",false,"tokenType","session_token","user","preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesChatPanelState","notificationCount",0,"flagDenols","isBusinessOrgMember","isHigherTownieTier","ENV",{"_39":40,"_41":42,"_43":44,"_45":46,"_47":48,"_49":50,"_51":52,"_53":54,"_55":56,"_57":58,"_59":60,"_61":62,"_63":64,"_65":66,"_67":68},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","VAL_RUNTIME_NAME","prod","RENDER_GIT_COMMIT","37154e1a6452dbe4d1c1c1b5325456d6c3f27b9b","RENDER_GIT_BRANCH","main","CLERK_PUBLISHABLE_KEY","pk_live_Y2xlcmsudmFsLnRvd24k","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","API_URL","https://api.val.town","clerkState",{"_71":72},"__internal_clerk_state",{"_73":74,"_75":48,"_76":77,"_78":77,"_79":20,"_80":77,"_81":77,"_82":77,"_83":77,"_84":85,"_94":20,"_95":20},"__clerk_ssr_state",{"_21":22,"_9":-5,"_10":-5,"_11":-5,"_12":-5,"_13":-5,"_14":-5,"_15":-5,"_16":-5,"_17":-5,"_18":-5,"_19":20},"__publishableKey","__proxyUrl","","__domain","__isSatellite","__signInUrl","__signUpUrl","__afterSignInUrl","__afterSignUpUrl","__clerk_debug",{"_86":20,"_19":20,"_87":77,"_88":89,"_90":77,"_91":48,"_92":20,"_93":77},"isSignedIn","proxyUrl","reason","session-token-and-uat-missing","message","publishableKey","isSatellite","domain","__telemetryDisabled","__telemetryDebug","routes/embed.x.$handle.$name.$",{"_98":99,"_100":101,"_102":103,"_104":105,"_158":159,"_177":178,"_179":-5},"handle","isaacgreenmann","name","beltbackend-3631","path","main.ts","project",{"_106":107,"_100":101,"_108":109,"_110":-5,"_111":20,"_112":113,"_114":-5,"_115":-5,"_12":116,"_117":-5,"_118":20,"_119":-5,"_120":-5,"_121":33,"_14":122,"_123":-5,"_124":-5,"_125":-5,"_126":-5,"_127":128,"_129":128,"_130":131,"_132":131,"_133":131,"_134":128,"_135":128,"_23":136,"_139":140,"_144":145,"_157":20},"id","b4ffc9c0-33a4-11f1-8e9d-42b51c65c3df","createdAt",["D",1775691792950],"deletedAt","legacyEnvVars","privacy","public","description","imageUrl","cf543f84-0db2-4b4f-a39b-58281a21b5ff","pinnedFileId","pinPreview","pinMigrated","screenshotVersion","likeCount","019d6f58-ab5b-76c5-9a0f-ec878657e051","tursoName","tursoHostname","tursoMeasuredAt","pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm",{"_98":99,"_137":138},"profileImageUrl","https://avatars.githubusercontent.com/u/103331631?v=4","org",{"_98":99,"_141":138,"_12":116,"_142":143},"avatar","tier","pro","branch",{"_146":147,"_100":46,"_106":148,"_149":107,"_150":-5,"_151":-5,"_152":-5,"_153":-5,"_154":-5,"_155":101,"_156":99},"version",15,"b50094e0-33a4-11f1-8e9d-42b51c65c3df","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","isOwned","file",{"_106":160,"_100":103,"_161":162,"_163":-5,"_164":148,"_165":147,"_166":167,"_108":168,"_146":169,"_170":171,"_172":173,"_174":20,"_175":-5,"_102":176},"67e761f8-345c-11f1-98e2-42b51c65c3df","fileId","b50311de-33a4-11f1-8e9d-42b51c65c3df","parentId","branchId","branchVersion","latest",true,["D",1775770690986],14,"fileType","http","content","/**\n * Val Town HTTP handler — personal screenshot → triple-LLM → TTS backend.\n *\n * Required environment variables (Val Town secrets / Deno.env):\n * ANTHROPIC_API_KEY — Anthropic API key\n * OPENAI_API_KEY — OpenAI API key (chat + TTS)\n * GOOGLE_API_KEY — Google AI (Gemini) API key\n * SHARED_SECRET — must match X-Key header on all requests\n *\n * Images are stored in Val Town blob (keys like belt-img/...); SQLite JSON only\n * holds short blob_key + mime_type so statements stay under the ~1MB limit.\n * Captures accept common raster formats (sniffed from bytes and/or Content-Type).\n * Non–vision-safe formats (BMP, TIFF, HEIC, …) are transcoded to PNG via\n * npm:imagescript before calling Claude/OpenAI/Gemini.\n */\n\n\n\nimport { blob } from \"https://esm.town/v/std/blob\";\nimport { sqlite } from \"https://esm.town/v/std/sqlite\";\n\nconst USER_PROMPT = \"First describe the question / problem you see. Then please solve fully and to the best of your ability.\";\n\ntype ActiveLlm = \"claude\" | \"openai\" | \"gemini\";\n\ntype ClaudeContentBlock =\n | { type: \"text\"; text: string }\n | {\n type: \"image\";\n source: { type: \"base64\"; media_type: string; data: string };\n };\n\ntype OpenAIContentPart =\n | { type: \"text\"; text: string }\n | { type: \"image_url\"; image_url: { url: string } };\n\ntype GeminiPart =\n | { text: string }\n | { inline_data: { mime_type: string; data: string } };\n\ntype GeminiContent = {\n role: \"user\" | \"model\";\n parts: GeminiPart[];\n};\n\n/** Stored in SQLite instead of inlined base64 / data URLs */\ntype ImageRefBlock = {\n type: \"image_ref\";\n blob_key: string;\n /** e.g. image/png — used when hydrating APIs; omitted in old rows (sniff from blob) */\n mime_type?: string;\n};\n\ntype GeminiImageRefPart = {\n image_ref: { blob_key: string; mime_type?: string };\n};\n\ntype ClaudeStoredMessage = {\n role: \"user\" | \"assistant\";\n content: string | (ClaudeContentBlock | ImageRefBlock)[];\n};\n\ntype OpenAIStoredMessage = {\n role: \"user\" | \"assistant\";\n content: string | (OpenAIContentPart | ImageRefBlock)[];\n};\n\ntype GeminiStoredContent = {\n role: \"user\" | \"model\";\n parts: (GeminiPart | GeminiImageRefPart)[];\n};\n\ntype ClaudeMessage = {\n role: \"user\" | \"assistant\";\n content: string | ClaudeContentBlock[];\n};\n\ntype OpenAIMessage = {\n role: \"user\" | \"assistant\";\n content: string | OpenAIContentPart[];\n};\n\nfunction arrayBufferToBase64(buffer: ArrayBuffer): string {\n const bytes = new Uint8Array(buffer);\n let binary = \"\";\n for (let i = 0; i < bytes.byteLength; i++) {\n binary += String.fromCharCode(bytes[i]!);\n }\n return btoa(binary);\n}\n\nfunction jsonParse(s: string, fallback: T): T {\n try {\n return JSON.parse(s) as T;\n } catch {\n return fallback;\n }\n}\n\n/** Val Town / libsql may return each row as a value array parallel to `columns`, not an object. */\ntype SqlResult = {\n columns?: string[];\n rows: unknown[];\n};\n\nfunction rowToRecord>(\n columns: string[],\n row: unknown,\n): T {\n if (row == null) {\n return {} as T;\n }\n if (Array.isArray(row) && columns.length > 0) {\n const o: Record = {};\n for (let i = 0; i < columns.length; i++) {\n const key = columns[i]!;\n if (i < row.length) o[key] = (row as unknown[])[i];\n }\n return o as T;\n }\n if (typeof row === \"object\" && !Array.isArray(row)) {\n return row as T;\n }\n return {} as T;\n}\n\nfunction sqlRows>(result: SqlResult): T[] {\n const cols = result.columns ?? [];\n return result.rows.map((row) => rowToRecord(cols, row));\n}\n\n// Audio columns store Val Town blob keys (short strings), not base64 MP3.\nfunction ensureTable(): Promise {\n return sqlite.execute(`\n CREATE TABLE IF NOT EXISTS sessions (\n id TEXT PRIMARY KEY,\n created_at INTEGER NOT NULL,\n updated_at INTEGER NOT NULL,\n status TEXT NOT NULL,\n active_llm TEXT NOT NULL,\n claude_messages TEXT NOT NULL,\n openai_messages TEXT NOT NULL,\n gemini_contents TEXT NOT NULL,\n claude_audio TEXT NOT NULL,\n openai_audio TEXT NOT NULL,\n gemini_audio TEXT NOT NULL\n )\n `);\n}\n\nfunction sessionAudioBlobKey(\n sessionId: string,\n which: \"claude\" | \"openai\" | \"gemini\",\n): string {\n return `belt-audio/${sessionId}/${which}.mp3`;\n}\n\nasync function deleteSessionAudioBlobs(sessionId: string): Promise {\n for (const which of [\"claude\", \"openai\", \"gemini\"] as const) {\n const key = sessionAudioBlobKey(sessionId, which);\n try {\n await blob.delete(key);\n } catch {\n // ignore missing / race\n }\n }\n}\n\nasync function readBlobAsBytes(key: string): Promise {\n if (!key) return null;\n try {\n const data = await blob.get(key);\n if (data == null) return null;\n if (data instanceof ArrayBuffer) return new Uint8Array(data);\n if (data instanceof Uint8Array) return data;\n const ab = await new Response(data as BodyInit).arrayBuffer();\n return new Uint8Array(ab);\n } catch (e) {\n const name = e instanceof Error ? e.name : \"\";\n if (name === \"ValTownBlobNotFoundError\") return null;\n throw e;\n }\n}\n\nfunction fileExtForMime(mime: string): string {\n const m = normalizeImageMime(mime);\n const map: Record = {\n \"image/jpeg\": \"jpg\",\n \"image/jpg\": \"jpg\",\n \"image/png\": \"png\",\n \"image/gif\": \"gif\",\n \"image/webp\": \"webp\",\n \"image/bmp\": \"bmp\",\n \"image/x-ms-bmp\": \"bmp\",\n \"image/tiff\": \"tiff\",\n \"image/tif\": \"tiff\",\n \"image/heic\": \"heic\",\n \"image/heif\": \"heif\",\n \"image/avif\": \"avif\",\n \"image/x-icon\": \"ico\",\n \"image/vnd.microsoft.icon\": \"ico\",\n };\n return map[m] ?? \"img\";\n}\n\nfunction sessionImageBlobKey(\n sessionId: string,\n imageId: string,\n ext: string,\n): string {\n const safe = ext.replace(/[^a-z0-9]/gi, \"\") || \"img\";\n return `belt-img/${sessionId}/${imageId}.${safe}`;\n}\n\n/** Canonical image/* subtype for APIs */\nfunction normalizeImageMime(mime: string): string {\n const m = mime.split(\";\")[0]!.trim().toLowerCase();\n if (m === \"image/jpg\" || m === \"image/pjpeg\") return \"image/jpeg\";\n return m;\n}\n\n/**\n * Detect image format from magic bytes. Returns null if unknown.\n */\nfunction sniffImageMimeType(u8: Uint8Array): string | null {\n if (u8.length < 12) return null;\n // JPEG\n if (u8[0] === 0xff && u8[1] === 0xd8 && u8[2] === 0xff) return \"image/jpeg\";\n // PNG\n if (\n u8[0] === 0x89 && u8[1] === 0x50 && u8[2] === 0x4e && u8[3] === 0x47\n ) {\n return \"image/png\";\n }\n // GIF\n if (\n u8[0] === 0x47 && u8[1] === 0x49 && u8[2] === 0x46 && u8[3] === 0x38\n ) {\n return \"image/gif\";\n }\n // BMP\n if (u8[0] === 0x42 && u8[1] === 0x4d) return \"image/bmp\";\n // WebP: RIFF....WEBP\n if (\n u8[0] === 0x52 && u8[1] === 0x49 && u8[2] === 0x46 && u8[3] === 0x46 &&\n u8[8] === 0x57 && u8[9] === 0x45 && u8[10] === 0x42 && u8[11] === 0x50\n ) {\n return \"image/webp\";\n }\n // TIFF\n if (\n (u8[0] === 0x49 && u8[1] === 0x49 && u8[2] === 0x2a && u8[3] === 0x00) ||\n (u8[0] === 0x4d && u8[1] === 0x4d && u8[2] === 0x00 && u8[3] === 0x2a)\n ) {\n return \"image/tiff\";\n }\n // ICO\n if (u8[0] === 0x00 && u8[1] === 0x00 && u8[2] === 0x01 && u8[3] === 0x00) {\n return \"image/x-icon\";\n }\n // ISO BMFF (HEIF/AVIF): ....ftypxxxx\n if (u8.length >= 16 && u8[4] === 0x66 && u8[5] === 0x74 && u8[6] === 0x79 &&\n u8[7] === 0x70\n ) {\n const brand = String.fromCharCode(u8[8]!, u8[9]!, u8[10]!, u8[11]!);\n if (\n brand === \"heic\" || brand === \"heix\" || brand === \"hevc\" ||\n brand === \"mif1\" || brand === \"msf1\"\n ) {\n return \"image/heic\";\n }\n if (brand === \"avif\" || brand === \"avis\") return \"image/avif\";\n }\n return null;\n}\n\n/**\n * Prefer magic bytes; fall back to Content-Type; last resort image/png (vision APIs tolerate PNG).\n */\nfunction resolveCaptureImageMime(\n u8: Uint8Array,\n contentTypeHeader: string | null,\n): string {\n const sniffed = sniffImageMimeType(u8);\n if (sniffed) return sniffed;\n const raw = contentTypeHeader?.split(\";\")[0]?.trim() ?? \"\";\n const h = normalizeImageMime(raw);\n if (h.startsWith(\"image/\") && h.length > 6) {\n return h;\n }\n return \"image/png\";\n}\n\nfunction isImageRefBlock(x: unknown): x is ImageRefBlock {\n return (\n typeof x === \"object\" &&\n x !== null &&\n (x as ImageRefBlock).type === \"image_ref\" &&\n typeof (x as ImageRefBlock).blob_key === \"string\"\n );\n}\n\nfunction isGeminiImageRefPart(x: unknown): x is GeminiImageRefPart {\n return (\n typeof x === \"object\" &&\n x !== null &&\n \"image_ref\" in x &&\n typeof (x as GeminiImageRefPart).image_ref === \"object\" &&\n (x as GeminiImageRefPart).image_ref !== null &&\n typeof (x as GeminiImageRefPart).image_ref.blob_key === \"string\"\n );\n}\n\nfunction bytesToBase64(bytes: Uint8Array): string {\n const u8 = new Uint8Array(bytes.byteLength);\n u8.set(bytes);\n return arrayBufferToBase64(u8.buffer);\n}\n\n/** OpenAI vision only allows these; Claude/Gemini handle the same set reliably when we normalize. */\nconst VISION_API_MIMES = new Set([\n \"image/png\",\n \"image/jpeg\",\n \"image/gif\",\n \"image/webp\",\n]);\n\n// deno-lint-ignore no-explicit-any\nlet imagescriptModule: any = null;\nasync function getImageScript() {\n if (!imagescriptModule) {\n imagescriptModule = await import(\"npm:imagescript@1.3.0\");\n }\n return imagescriptModule;\n}\n\n/**\n * Decodes arbitrary raster bytes and re-encodes as PNG when the MIME is not\n * one of png/jpeg/gif/webp (providers reject BMP, TIFF, ICO, AVIF, HEIC, etc.).\n */\nasync function ensureVisionApiImage(\n raw: Uint8Array,\n mime: string,\n): Promise<{ bytes: Uint8Array; mime: string }> {\n const m = normalizeImageMime(mime);\n if (VISION_API_MIMES.has(m)) {\n return { bytes: raw, mime: m };\n }\n try {\n const mod = await getImageScript();\n type Img = { encode: (c?: number) => Promise };\n let surface: Img;\n try {\n surface = (await mod.Image.decode(raw)) as Img;\n } catch {\n const d = await mod.decode(raw, true);\n if (mod.GIF && d instanceof mod.GIF && d[0]) {\n surface = d[0] as Img;\n } else {\n surface = d as Img;\n }\n }\n const encoded = await surface.encode();\n const out = encoded instanceof Uint8Array\n ? encoded\n : new Uint8Array(encoded);\n return { bytes: out, mime: \"image/png\" };\n } catch (e) {\n console.log(\"ensureVisionApiImage: transcode failed, mime was\", m, e);\n throw new Error(\n `Could not convert image to PNG/JPEG/GIF/WebP (${m}). Try a PNG or JPEG screenshot.`,\n );\n }\n}\n\ntype VisionPrep = { bytes: Uint8Array; mime: string };\n\ntype HydrateCaches = {\n /** Dedup parallel blob reads + vision transcoding across the 3 hydrators */\n raw: Map>;\n vision: Map>;\n};\n\nasync function loadImageRefForApi(\n block: ImageRefBlock,\n caches: HydrateCaches,\n): Promise<{ b64: string; mime: string }> {\n const key = block.blob_key;\n let rawP = caches.raw.get(key);\n if (!rawP) {\n rawP = readBlobAsBytes(key);\n caches.raw.set(key, rawP);\n }\n const raw = await rawP;\n if (!raw || raw.length === 0) {\n return { b64: \"\", mime: \"image/png\" };\n }\n let mime = block.mime_type ? normalizeImageMime(block.mime_type) : \"\";\n if (!mime.startsWith(\"image/\")) {\n mime = sniffImageMimeType(raw) ?? \"\";\n }\n if (!mime.startsWith(\"image/\")) mime = \"image/jpeg\";\n\n let prepP = caches.vision.get(key);\n if (!prepP) {\n prepP = ensureVisionApiImage(raw, mime);\n caches.vision.set(key, prepP);\n }\n const prep = await prepP;\n return { b64: bytesToBase64(prep.bytes), mime: prep.mime };\n}\n\nasync function hydrateClaudeStored(\n messages: ClaudeStoredMessage[],\n caches: HydrateCaches,\n): Promise {\n const out: ClaudeMessage[] = [];\n for (const m of messages) {\n if (m.role === \"assistant\") {\n out.push({\n role: \"assistant\",\n content: typeof m.content === \"string\"\n ? m.content\n : (m.content as ClaudeContentBlock[]),\n });\n continue;\n }\n const c = m.content;\n if (typeof c === \"string\") {\n out.push({ role: \"user\", content: c });\n continue;\n }\n const blocks: ClaudeContentBlock[] = [];\n for (const block of c) {\n if (isImageRefBlock(block)) {\n const { b64: data, mime } = await loadImageRefForApi(block, caches);\n blocks.push({\n type: \"image\",\n source: {\n type: \"base64\",\n media_type: mime,\n data,\n },\n });\n } else if (\n typeof block === \"object\" &&\n block !== null &&\n (block as { type?: string }).type === \"image\"\n ) {\n blocks.push(block as ClaudeContentBlock);\n } else if (\n typeof block === \"object\" &&\n block !== null &&\n (block as { type?: string }).type === \"text\"\n ) {\n blocks.push(block as ClaudeContentBlock);\n }\n }\n out.push({ role: \"user\", content: blocks });\n }\n return out;\n}\n\nasync function hydrateOpenAIStored(\n messages: OpenAIStoredMessage[],\n caches: HydrateCaches,\n): Promise {\n const out: OpenAIMessage[] = [];\n for (const m of messages) {\n if (m.role === \"assistant\") {\n out.push({\n role: \"assistant\",\n content: typeof m.content === \"string\"\n ? m.content\n : (m.content as OpenAIContentPart[]),\n });\n continue;\n }\n const c = m.content;\n if (typeof c === \"string\") {\n out.push({ role: \"user\", content: c });\n continue;\n }\n const parts: OpenAIContentPart[] = [];\n for (const part of c) {\n if (isImageRefBlock(part)) {\n const { b64: data, mime } = await loadImageRefForApi(part, caches);\n parts.push({\n type: \"image_url\",\n image_url: { url: `data:${mime};base64,${data}` },\n });\n } else if (\n typeof part === \"object\" &&\n part !== null &&\n (part as { type?: string }).type === \"image_url\"\n ) {\n parts.push(part as OpenAIContentPart);\n } else if (\n typeof part === \"object\" &&\n part !== null &&\n (part as { type?: string }).type === \"text\"\n ) {\n parts.push(part as OpenAIContentPart);\n }\n }\n out.push({ role: \"user\", content: parts });\n }\n return out;\n}\n\nasync function hydrateGeminiStored(\n contents: GeminiStoredContent[],\n caches: HydrateCaches,\n): Promise {\n const out: GeminiContent[] = [];\n for (const m of contents) {\n if (m.role === \"model\") {\n out.push({ role: \"model\", parts: m.parts as GeminiPart[] });\n continue;\n }\n const parts: GeminiPart[] = [];\n for (const p of m.parts) {\n if (isGeminiImageRefPart(p)) {\n const { b64: data, mime } = await loadImageRefForApi(\n {\n type: \"image_ref\",\n blob_key: p.image_ref.blob_key,\n mime_type: p.image_ref.mime_type,\n },\n caches,\n );\n parts.push({ inline_data: { mime_type: mime, data } });\n } else {\n parts.push(p as GeminiPart);\n }\n }\n out.push({ role: \"user\", parts });\n }\n return out;\n}\n\n/** Keeps each SQLite execute payload small (one large JSON column per statement). */\nasync function persistSessionMessagesSplit(\n id: string,\n updatedAt: number,\n extra: { status?: string },\n claude: ClaudeStoredMessage[],\n openai: OpenAIStoredMessage[],\n gemini: GeminiStoredContent[],\n): Promise {\n if (extra.status !== undefined) {\n await sqlite.execute({\n sql: `UPDATE sessions SET updated_at = ?, status = ?, claude_messages = ? WHERE id = ?`,\n args: [updatedAt, extra.status, JSON.stringify(claude), id],\n });\n } else {\n await sqlite.execute({\n sql: `UPDATE sessions SET updated_at = ?, claude_messages = ? WHERE id = ?`,\n args: [updatedAt, JSON.stringify(claude), id],\n });\n }\n await sqlite.execute({\n sql: `UPDATE sessions SET openai_messages = ? WHERE id = ?`,\n args: [JSON.stringify(openai), id],\n });\n await sqlite.execute({\n sql: `UPDATE sessions SET gemini_contents = ? WHERE id = ?`,\n args: [JSON.stringify(gemini), id],\n });\n}\n\nfunction unauthorized(): Response {\n return new Response(JSON.stringify({ error: \"Unauthorized\" }), {\n status: 401,\n headers: { \"Content-Type\": \"application/json\" },\n });\n}\n\nfunction badRequest(message: string): Response {\n return new Response(JSON.stringify({ error: message }), {\n status: 400,\n headers: { \"Content-Type\": \"application/json\" },\n });\n}\n\nfunction notFound(): Response {\n return new Response(JSON.stringify({ error: \"Not found\" }), {\n status: 404,\n headers: { \"Content-Type\": \"application/json\" },\n });\n}\n\nfunction json(data: unknown, status = 200): Response {\n return new Response(JSON.stringify(data), {\n status,\n headers: { \"Content-Type\": \"application/json\" },\n });\n}\n\nfunction checkAuth(req: Request): boolean {\n const secret = Deno.env.get(\"SHARED_SECRET\");\n if (!secret) return false;\n return req.headers.get(\"X-Key\") === secret;\n}\n\nfunction latestClaudeAssistant(messages: ClaudeStoredMessage[]): string {\n for (let i = messages.length - 1; i >= 0; i--) {\n const m = messages[i]!;\n if (m.role !== \"assistant\") continue;\n const c = m.content;\n if (typeof c === \"string\") return c;\n if (Array.isArray(c)) {\n for (const b of c) {\n if (b.type === \"text\" && \"text\" in b) return b.text;\n }\n }\n }\n return \"\";\n}\n\nfunction latestOpenAIAssistant(messages: OpenAIStoredMessage[]): string {\n for (let i = messages.length - 1; i >= 0; i--) {\n const m = messages[i]!;\n if (m.role !== \"assistant\") continue;\n const c = m.content;\n if (typeof c === \"string\") return c;\n if (Array.isArray(c)) {\n for (const b of c) {\n if (b.type === \"text\" && \"text\" in b) return b.text;\n }\n }\n }\n return \"\";\n}\n\nfunction latestGeminiAssistant(contents: GeminiStoredContent[]): string {\n for (let i = contents.length - 1; i >= 0; i--) {\n const m = contents[i]!;\n if (m.role !== \"model\") continue;\n for (const p of m.parts) {\n if (\"text\" in p && p.text) return p.text;\n }\n }\n return \"\";\n}\n\nfunction previewFromActive(\n active: ActiveLlm,\n claude: ClaudeStoredMessage[],\n openai: OpenAIStoredMessage[],\n gemini: GeminiStoredContent[],\n): string {\n let text = \"\";\n if (active === \"claude\") text = latestClaudeAssistant(claude);\n else if (active === \"openai\") text = latestOpenAIAssistant(openai);\n else text = latestGeminiAssistant(gemini);\n return text.length <= 60 ? text : text.slice(0, 60);\n}\n\nasync function callClaude(\n messages: ClaudeMessage[],\n): Promise {\n try {\n const key = Deno.env.get(\"ANTHROPIC_API_KEY\");\n if (!key) return \"Error: ANTHROPIC_API_KEY is not set\";\n const res = await fetch(\"https://api.anthropic.com/v1/messages\", {\n method: \"POST\",\n headers: {\n \"x-api-key\": key,\n \"anthropic-version\": \"2023-06-01\",\n \"content-type\": \"application/json\",\n },\n body: JSON.stringify({\n model: \"claude-sonnet-4-5\",\n max_tokens: 4096,\n messages,\n }),\n });\n const raw = await res.text();\n if (!res.ok) return `Error: ${res.status} ${raw}`;\n const data = JSON.parse(raw) as {\n content?: { type?: string; text?: string }[];\n };\n const text = data.content?.[0]?.text;\n if (typeof text !== \"string\" || !text.length) {\n return \"Error: empty or unexpected Claude response\";\n }\n return text;\n } catch (e) {\n return `Error: ${e instanceof Error ? e.message : String(e)}`;\n }\n}\n\nasync function callOpenAIChat(\n messages: OpenAIMessage[],\n): Promise {\n try {\n const key = Deno.env.get(\"OPENAI_API_KEY\");\n if (!key) return \"Error: OPENAI_API_KEY is not set\";\n const res = await fetch(\"https://api.openai.com/v1/chat/completions\", {\n method: \"POST\",\n headers: {\n Authorization: `Bearer ${key}`,\n \"content-type\": \"application/json\",\n },\n body: JSON.stringify({\n model: \"gpt-4o\",\n max_tokens: 4096,\n messages,\n }),\n });\n const raw = await res.text();\n if (!res.ok) return `Error: ${res.status} ${raw}`;\n const data = JSON.parse(raw) as {\n choices?: { message?: { content?: string } }[];\n };\n const text = data.choices?.[0]?.message?.content;\n if (typeof text !== \"string\" || !text.length) {\n return \"Error: empty or unexpected OpenAI response\";\n }\n return text;\n } catch (e) {\n return `Error: ${e instanceof Error ? e.message : String(e)}`;\n }\n}\n\nasync function callGemini(contents: GeminiContent[]): Promise {\n try {\n const key = Deno.env.get(\"GOOGLE_API_KEY\");\n if (!key) return \"Error: GOOGLE_API_KEY is not set\";\n const url =\n `https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent?key=${\n encodeURIComponent(key)\n }`;\n const res = await fetch(url, {\n method: \"POST\",\n headers: { \"content-type\": \"application/json\" },\n body: JSON.stringify({ contents }),\n });\n const raw = await res.text();\n if (!res.ok) return `Error: ${res.status} ${raw}`;\n const data = JSON.parse(raw) as {\n candidates?: { content?: { parts?: { text?: string }[] } }[];\n };\n const text = data.candidates?.[0]?.content?.parts?.[0]?.text;\n if (typeof text !== \"string\" || !text.length) {\n return \"Error: empty or unexpected Gemini response\";\n }\n return text;\n } catch (e) {\n return `Error: ${e instanceof Error ? e.message : String(e)}`;\n }\n}\n\nasync function callOpenAITts(input: string): Promise {\n try {\n const key = Deno.env.get(\"OPENAI_API_KEY\");\n if (!key) return null;\n const res = await fetch(\"https://api.openai.com/v1/audio/speech\", {\n method: \"POST\",\n headers: {\n Authorization: `Bearer ${key}`,\n \"content-type\": \"application/json\",\n },\n body: JSON.stringify({\n model: \"tts-1\",\n voice: \"alloy\",\n input,\n }),\n });\n if (!res.ok) {\n console.log(\"TTS HTTP error\", res.status, await res.text());\n return null;\n }\n return await res.arrayBuffer();\n } catch (e) {\n console.log(\"TTS exception\", e);\n return null;\n }\n}\n\n/** Legacy sessions stored raw base64 MP3 in SQLite; keep for GET /audio only. */\nfunction legacyBase64Mp3ToBytes(b64: string): Uint8Array | null {\n try {\n const binary = atob(b64);\n const out = new Uint8Array(binary.length);\n for (let i = 0; i < binary.length; i++) {\n out[i] = binary.charCodeAt(i);\n }\n return out;\n } catch {\n return null;\n }\n}\n\nasync function uploadMp3ToBlob(\n sessionId: string,\n which: \"claude\" | \"openai\" | \"gemini\",\n buf: ArrayBuffer | null,\n): Promise {\n if (!buf || buf.byteLength === 0) return \"\";\n const key = sessionAudioBlobKey(sessionId, which);\n try {\n await blob.set(key, buf);\n return key;\n } catch (e) {\n console.log(\"blob.set failed\", which, e);\n return \"\";\n }\n}\n\ntype SessionRow = {\n id: string;\n created_at: number;\n updated_at: number;\n status: string;\n active_llm: string;\n claude_messages: string;\n openai_messages: string;\n gemini_contents: string;\n /** Blob key (belt-audio/...) or legacy base64 MP3 */\n claude_audio: string;\n openai_audio: string;\n gemini_audio: string;\n};\n\nasync function loadSession(id: string): Promise {\n const r = (await sqlite.execute({\n sql: `SELECT * FROM sessions WHERE id = ?`,\n args: [id],\n })) as SqlResult;\n if (!r.rows.length) return null;\n const row = rowToRecord(r.columns ?? [], r.rows[0]);\n if (row.id == null || row.id === \"\") return null;\n return row;\n}\n\nfunction normalizeActiveLlm(s: string): ActiveLlm {\n if (s === \"openai\" || s === \"gemini\") return s;\n return \"claude\";\n}\n\nexport default async function (req: Request): Promise {\n await ensureTable();\n\n if (!checkAuth(req)) {\n return unauthorized();\n }\n\n const url = new URL(req.url);\n let pathname = url.pathname;\n if (pathname.length > 1 && pathname.endsWith(\"/\")) {\n pathname = pathname.slice(0, -1);\n }\n\n // GET /sessions\n if (req.method === \"GET\" && pathname === \"/sessions\") {\n const r = (await sqlite.execute({\n sql: `SELECT id, created_at, updated_at, status, active_llm, claude_messages, openai_messages, gemini_contents FROM sessions ORDER BY updated_at DESC LIMIT 20`,\n args: [],\n })) as SqlResult;\n const rows = sqlRows<{\n id: string;\n created_at: number;\n updated_at: number;\n status: string;\n active_llm: string;\n claude_messages: string;\n openai_messages: string;\n gemini_contents: string;\n }>(r);\n const out = rows.map((row) => {\n const active = normalizeActiveLlm(row.active_llm);\n const claude = jsonParse(row.claude_messages, []);\n const openai = jsonParse(row.openai_messages, []);\n const gemini = jsonParse(row.gemini_contents, []);\n return {\n id: row.id,\n created_at: row.created_at,\n status: row.status,\n active_llm: active,\n preview: previewFromActive(active, claude, openai, gemini),\n };\n });\n return json(out);\n }\n\n // GET /session/:id\n const sessionMatch = pathname.match(/^\\/session\\/([^/]+)$/);\n if (req.method === \"GET\" && sessionMatch) {\n const id = sessionMatch[1]!;\n const row = await loadSession(id);\n if (!row) return notFound();\n const claude = jsonParse(row.claude_messages, []);\n const openai = jsonParse(row.openai_messages, []);\n const gemini = jsonParse(row.gemini_contents, []);\n const active = normalizeActiveLlm(row.active_llm);\n return json({\n id: row.id,\n created_at: row.created_at,\n updated_at: row.updated_at,\n status: row.status,\n active_llm: active,\n claude_text: latestClaudeAssistant(claude),\n openai_text: latestOpenAIAssistant(openai),\n gemini_text: latestGeminiAssistant(gemini),\n audio_url: `/audio/${id}`,\n });\n }\n\n // GET /audio/:id\n const audioMatch = pathname.match(/^\\/audio\\/([^/]+)$/);\n if (req.method === \"GET\" && audioMatch) {\n const id = audioMatch[1]!;\n const row = await loadSession(id);\n if (!row) return notFound();\n const active = normalizeActiveLlm(row.active_llm);\n let b64 = \"\";\n if (active === \"claude\") b64 = row.claude_audio;\n else if (active === \"openai\") b64 = row.openai_audio;\n else b64 = row.gemini_audio;\n let bytes: Uint8Array | null = null;\n if (b64.startsWith(\"belt-audio/\")) {\n bytes = await readBlobAsBytes(b64);\n } else if (b64) {\n bytes = legacyBase64Mp3ToBytes(b64);\n }\n if (!bytes || bytes.length === 0) {\n return new Response(null, {\n status: 204,\n headers: { \"Content-Type\": \"audio/mpeg\" },\n });\n }\n return new Response(bytes.slice(), {\n headers: { \"Content-Type\": \"audio/mpeg\" },\n });\n }\n\n // POST /switch-llm/:id\n const switchMatch = pathname.match(/^\\/switch-llm\\/([^/]+)$/);\n if (req.method === \"POST\" && switchMatch) {\n const id = switchMatch[1]!;\n const row = await loadSession(id);\n if (!row) return notFound();\n const order: ActiveLlm[] = [\"claude\", \"openai\", \"gemini\"];\n const cur = normalizeActiveLlm(row.active_llm);\n const idx = order.indexOf(cur);\n const next = order[(idx + 1) % 3]!;\n const now = Date.now();\n await sqlite.execute({\n sql: `UPDATE sessions SET active_llm = ?, updated_at = ? WHERE id = ?`,\n args: [next, now, id],\n });\n return json({ active_llm: next, audio_url: `/audio/${id}` });\n }\n\n // POST /capture\n if (req.method === \"POST\" && pathname === \"/capture\") {\n const bodyBuf = await req.arrayBuffer();\n if (!bodyBuf.byteLength) {\n return badRequest(\"Expected non-empty image body\");\n }\n const u8 = new Uint8Array(bodyBuf);\n const imageMime = resolveCaptureImageMime(\n u8,\n req.headers.get(\"Content-Type\"),\n );\n const imageExt = fileExtForMime(imageMime);\n\n const existingId = req.headers.get(\"X-Session-Id\");\n let id: string;\n let created_at: number;\n let claude: ClaudeStoredMessage[];\n let openai: OpenAIStoredMessage[];\n let gemini: GeminiStoredContent[];\n\n if (existingId) {\n const row = await loadSession(existingId);\n if (!row) return notFound();\n id = row.id;\n created_at = row.created_at;\n claude = jsonParse(row.claude_messages, []);\n openai = jsonParse(row.openai_messages, []);\n gemini = jsonParse(row.gemini_contents, []);\n } else {\n id = crypto.randomUUID();\n created_at = Date.now();\n claude = [];\n openai = [];\n gemini = [];\n }\n\n const imageId = crypto.randomUUID();\n const imageKey = sessionImageBlobKey(id, imageId, imageExt);\n try {\n await blob.set(imageKey, bodyBuf);\n } catch (e) {\n console.log(\"blob.set image failed\", e);\n return badRequest(\"Failed to store image\");\n }\n\n const claudeUser: ClaudeStoredMessage = {\n role: \"user\",\n content: [\n { type: \"image_ref\", blob_key: imageKey, mime_type: imageMime },\n { type: \"text\", text: USER_PROMPT },\n ],\n };\n const openaiUser: OpenAIStoredMessage = {\n role: \"user\",\n content: [\n { type: \"image_ref\", blob_key: imageKey, mime_type: imageMime },\n { type: \"text\", text: USER_PROMPT },\n ],\n };\n const geminiUser: GeminiStoredContent = {\n role: \"user\",\n parts: [\n { image_ref: { blob_key: imageKey, mime_type: imageMime } },\n { text: USER_PROMPT },\n ],\n };\n\n claude.push(claudeUser);\n openai.push(openaiUser);\n gemini.push(geminiUser);\n\n const nowProc = Date.now();\n\n if (existingId) {\n await persistSessionMessagesSplit(\n id,\n nowProc,\n { status: \"processing\" },\n claude,\n openai,\n gemini,\n );\n } else {\n await sqlite.execute({\n sql: `INSERT INTO sessions (id, created_at, updated_at, status, active_llm, claude_messages, openai_messages, gemini_contents, claude_audio, openai_audio, gemini_audio) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,\n args: [\n id,\n created_at,\n nowProc,\n \"processing\",\n \"claude\",\n JSON.stringify(claude),\n JSON.stringify(openai),\n JSON.stringify(gemini),\n \"\",\n \"\",\n \"\",\n ],\n });\n }\n\n console.log(\"received capture\", {\n session_id: id,\n follow_up: !!existingId,\n image_mime: imageMime,\n });\n console.log(\"calling LLMs\");\n\n const hydrateCaches: HydrateCaches = {\n raw: new Map(),\n vision: new Map(),\n };\n let claudeApi: ClaudeMessage[];\n let openaiApi: OpenAIMessage[];\n let geminiApi: GeminiContent[];\n try {\n const triple = await Promise.all([\n hydrateClaudeStored(claude, hydrateCaches),\n hydrateOpenAIStored(openai, hydrateCaches),\n hydrateGeminiStored(gemini, hydrateCaches),\n ]);\n claudeApi = triple[0]!;\n openaiApi = triple[1]!;\n geminiApi = triple[2]!;\n } catch (e) {\n const msg = e instanceof Error ? e.message : String(e);\n console.log(\"hydrate failed\", msg);\n return badRequest(msg);\n }\n\n const llmSettled = await Promise.allSettled([\n callClaude(claudeApi),\n callOpenAIChat(openaiApi),\n callGemini(geminiApi),\n ]);\n\n const claudeText =\n llmSettled[0]!.status === \"fulfilled\"\n ? llmSettled[0].value\n : `Error: ${String(llmSettled[0]!.reason)}`;\n const openaiText =\n llmSettled[1]!.status === \"fulfilled\"\n ? llmSettled[1].value\n : `Error: ${String(llmSettled[1]!.reason)}`;\n const geminiText =\n llmSettled[2]!.status === \"fulfilled\"\n ? llmSettled[2].value\n : `Error: ${String(llmSettled[2]!.reason)}`;\n\n claude.push({ role: \"assistant\", content: claudeText });\n openai.push({ role: \"assistant\", content: openaiText });\n gemini.push({\n role: \"model\",\n parts: [{ text: geminiText }],\n });\n\n console.log(\"LLMs done\");\n\n await persistSessionMessagesSplit(\n id,\n Date.now(),\n {},\n claude,\n openai,\n gemini,\n );\n\n console.log(\"calling TTS\");\n\n const ttsSettled = await Promise.allSettled([\n callOpenAITts(claudeText),\n callOpenAITts(openaiText),\n callOpenAITts(geminiText),\n ]);\n\n const claudeBuf =\n ttsSettled[0]!.status === \"fulfilled\" ? ttsSettled[0].value : null;\n const openaiBuf =\n ttsSettled[1]!.status === \"fulfilled\" ? ttsSettled[1].value : null;\n const geminiBuf =\n ttsSettled[2]!.status === \"fulfilled\" ? ttsSettled[2].value : null;\n\n console.log(\"TTS done\");\n\n await deleteSessionAudioBlobs(id);\n\n const uploadSettled = await Promise.allSettled([\n uploadMp3ToBlob(id, \"claude\", claudeBuf),\n uploadMp3ToBlob(id, \"openai\", openaiBuf),\n uploadMp3ToBlob(id, \"gemini\", geminiBuf),\n ]);\n\n const claudeAudioKey =\n uploadSettled[0]!.status === \"fulfilled\" ? uploadSettled[0].value : \"\";\n const openaiAudioKey =\n uploadSettled[1]!.status === \"fulfilled\" ? uploadSettled[1].value : \"\";\n const geminiAudioKey =\n uploadSettled[2]!.status === \"fulfilled\" ? uploadSettled[2].value : \"\";\n\n console.log(\"audio blobs written\");\n\n const readyTime = Date.now();\n await sqlite.execute({\n sql: `UPDATE sessions SET updated_at = ?, status = ?, claude_audio = ?, openai_audio = ?, gemini_audio = ? WHERE id = ?`,\n args: [\n readyTime,\n \"ready\",\n claudeAudioKey,\n openaiAudioKey,\n geminiAudioKey,\n id,\n ],\n });\n\n console.log(\"ready\", { session_id: id });\n\n return json({ session_id: id, status: \"ready\" });\n }\n\n return new Response(\"Not found\", { status: 404 });\n}","deleted","label",[103],"permalink","https://www.val.town/x/isaacgreenmann/beltbackend-3631/code/main.ts","extension","actionData","errors"]