From f52c88f6a3abdb6f3ad9b013782923e2a6b8fb11 Mon Sep 17 00:00:00 2001
From: Felipe Teodoro <felipe.teodoro0@gmail.com>
Date: Fri, 7 Aug 2026 22:50:01 -0300
Subject: [PATCH] feat(ai): add Moonshot (Kimi) as an AI provider

Adds 'moonshot' to the AiModelProvider union with four models: kimi-k3,
kimi-k2.7-code, kimi-k2.7-code-highspeed, and kimi-k2.6. Routes through
the workshop backend's openai-completions path against api.moonshot.ai/v1
because Moonshot is not a Cloudflare first-party provider (no AI Gateway
route exists for it). Adds a local MOONSHOT_MODEL_COSTS lookup so that
the missing pi-ai catalog entry doesn't fall back to ZERO_COST and skew
BYOK accounting. Also wires up MOONSHOT_API_KEY/MOONSHOT_API_URL through
run-dev-server for local dev defaults.
---
 packages/workshop-backend/src/ai-models.ts    | 54 +++++++++++++++++++
 .../src/chat-attachment-validation.ts         |  1 +
 packages/workshop-backend/src/env.d.ts        |  6 +++
 packages/workshop-shared/src/api.ts           | 13 ++++-
 run-dev-server.js                             |  2 +
 5 files changed, 75 insertions(+), 1 deletion(-)

diff --git a/packages/workshop-backend/src/ai-models.ts b/packages/workshop-backend/src/ai-models.ts
index 25d1b759..40c8683f 100644
--- a/packages/workshop-backend/src/ai-models.ts
+++ b/packages/workshop-backend/src/ai-models.ts
@@ -115,6 +115,21 @@ const API_STREAMS: Record<string, StreamFunction<Api, SimpleStreamOptions>> = {
 
 const ZERO_COST: ModelCost = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
 
+// Per-model Moonshot billing (USD per 1k tokens), reasoning capability, and input modalities.
+// Source: https://platform.moonshot.ai/docs/pricing/chat — verify against live page before shipping.
+// Used because Moonshot isn't in pi-ai's built-in catalog, so catalogModel() returns undefined
+// and we look up locally to avoid falling back to ZERO_COST (which would skew BYOK accounting).
+const MOONSHOT_MODEL_COSTS: Record<string, {
+  cost: ModelCost,
+  reasoning: boolean,
+  input: ("text" | "image")[],
+}> = {
+  "kimi-k3":                  { cost: { input: 0.002,  output: 0.008,  cacheRead: 0, cacheWrite: 0 }, reasoning: true,  input: ["text"] },
+  "kimi-k2.7-code":           { cost: { input: 0.0015, output: 0.006,  cacheRead: 0, cacheWrite: 0 }, reasoning: true,  input: ["text"] },
+  "kimi-k2.7-code-highspeed": { cost: { input: 0.0015, output: 0.006,  cacheRead: 0, cacheWrite: 0 }, reasoning: true,  input: ["text"] },
+  "kimi-k2.6":                { cost: { input: 0.0006, output: 0.002,  cacheRead: 0, cacheWrite: 0 }, reasoning: false, input: ["text"] },
+};
+
 // Consult pi's builtin catalog for cost/compat metadata of a known model id. Unknown models are
 // fine (synthesized with zero cost). Import per-provider, not providers/all.
 function catalogModel(provider: AiModelConfig["provider"], modelId: string): Model<Api> | undefined {
@@ -124,6 +139,7 @@ function catalogModel(provider: AiModelConfig["provider"], modelId: string): Mod
     case "google": return (GOOGLE_MODELS as Record<string, Model<Api>>)[modelId];
     case "cloudflare": return (CLOUDFLARE_WORKERS_AI_MODELS as Record<string, Model<Api>>)[modelId];
     case "ollama": return undefined;
+    case "moonshot": return undefined;  // Moonshot is OpenAI-compatible; pi doesn't ship a catalog.
     default: return undefined;
   }
 }
@@ -231,6 +247,22 @@ function gatewayNativeModel(config: AiModelConfig, gatewayUrl: string): Model<Ap
         ...window,
         compat: workersAiCompat(catalog),
       };
+    case "moonshot":
+      // Moonshot is OpenAI-compatible; route through the gateway's /compat layer when the
+      // gateway is configured. Currently inactive for this workshop (no moonshot provider in
+      // the AI Gateway catalog), but kept symmetrical with `getModelDirect` so the direct
+      // path mirrors the gateway-routed one.
+      return {
+        id: config.model,
+        name: catalog?.name ?? config.model,
+        api: "openai-completions",
+        provider: "moonshot",
+        baseUrl: `${gatewayUrl}/compat`,
+        reasoning: true,
+        input: ["text"],
+        cost: catalog?.cost ?? ZERO_COST,
+        ...window,
+      };
     default:
       return undefined;
   }
@@ -587,6 +619,28 @@ function getModelDirect(config: AiModelConfig, sessionAffinity?: string): ModelH
         apiKey: config.apiToken,
         sessionAffinity,
       });
+    case "moonshot": {
+      // Direct Moonshot hit (no AI Gateway). Moonshot exposes an OpenAI-compatible
+      // /v1/chat/completions endpoint, so we route through pi's openai-completions API impl.
+      // Cost / reasoning / input are per-model: lookup locally so pi-ai's missing catalog
+      // (catalogModel returns undefined) doesn't fall back to ZERO_COST.
+      const spec = MOONSHOT_MODEL_COSTS[config.model];
+      return makeHandle({
+        model: {
+          id: config.model,
+          name: catalog?.name ?? config.model,
+          api: "openai-completions",
+          provider: "moonshot",
+          baseUrl: config.apiUrl ?? "https://api.moonshot.ai/v1",
+          reasoning: spec?.reasoning ?? true,
+          input: spec?.input ?? ["text"],
+          cost: spec?.cost ?? ZERO_COST,
+          ...window,
+        },
+        apiKey: config.apiToken,
+        sessionAffinity,
+      });
+    }
     default:
       config.provider satisfies never;
       throw new Error(`Unknown provider "${config.provider}".`);
diff --git a/packages/workshop-backend/src/chat-attachment-validation.ts b/packages/workshop-backend/src/chat-attachment-validation.ts
index 4f236a67..58037f95 100644
--- a/packages/workshop-backend/src/chat-attachment-validation.ts
+++ b/packages/workshop-backend/src/chat-attachment-validation.ts
@@ -38,6 +38,7 @@ const ATTACHMENT_SUPPORT_BY_PROVIDER = {
   google: isTextImageOrPdfMime,
   cloudflare: isTextOrImageMime,
   ollama: isTextOrImageMime,
+  moonshot: isTextOrImageMime,
 } satisfies Record<AiModelProvider, (mimeType: string) => boolean>;
 
 function sanitizeChatAttachmentMimeType(mimeType: string | undefined): string {
diff --git a/packages/workshop-backend/src/env.d.ts b/packages/workshop-backend/src/env.d.ts
index 3077a96d..895e16c1 100644
--- a/packages/workshop-backend/src/env.d.ts
+++ b/packages/workshop-backend/src/env.d.ts
@@ -83,6 +83,12 @@ declare global {
       // Minimum connected-account balance (USD) to proceed via BYOK. Defaults to
       // MINIMUM_CLOUDFLARE_BALANCE.
       MINIMUM_CLOUDFLARE_BALANCE?: string;
+
+      // Moonshot AI provider (Kimi). The workshop hits api.moonshot.ai directly because Moonshot
+      // is not a Cloudflare first-party provider (no AI Gateway route for it). These env vars
+      // configure direct-mode defaults; per-model configurations in the user DO override them.
+      MOONSHOT_API_KEY?: string;
+      MOONSHOT_API_URL?: string;
     }
   }
 }
diff --git a/packages/workshop-shared/src/api.ts b/packages/workshop-shared/src/api.ts
index 7cee2e78..aac2c46b 100644
--- a/packages/workshop-shared/src/api.ts
+++ b/packages/workshop-shared/src/api.ts
@@ -917,7 +917,7 @@ export type CloudflareAccountOption = {
 };
 
 // Supported AI providers.
-export type AiModelProvider = "openai" | "anthropic" | "google" | "cloudflare" | "ollama";
+export type AiModelProvider = "openai" | "anthropic" | "google" | "cloudflare" | "ollama" | "moonshot";
 
 // Information about the AI gateway configuration. Returned by `AuthenticatedApi.getAiConfig()`.
 export type AiGatewayInfo = {
@@ -985,6 +985,17 @@ export const SUGGESTED_MODELS: Record<
   },
   "ollama": {
   },
+  "moonshot": {
+    // Moonshot AI — Kimi models. Hosted at api.moonshot.ai via OpenAI-compatible
+    // /v1/chat/completions; the workshop hits it directly (bypassing AI Gateway since
+    // Moonshot is not a Cloudflare first-party provider). contextWindow values per
+    // https://platform.moonshot.ai/docs/pricing/chat (K3: 262144, K2.7-code: 262144,
+    // K2.6: 131072).
+    "kimi-k3":               {name: "Kimi K3 (Moonshot)",              contextWindow: 262144, outputLimit: 32768},
+    "kimi-k2.7-code":        {name: "Kimi K2.7 Code (Moonshot)",       contextWindow: 262144, outputLimit: 32768},
+    "kimi-k2.7-code-highspeed": {name: "Kimi K2.7 Code Highspeed (Moonshot)", contextWindow: 262144, outputLimit: 32768},
+    "kimi-k2.6":             {name: "Kimi K2.6 (Moonshot)",            contextWindow: 131072, outputLimit: 8192},
+  },
 };
 
 // Metadata about a workspace (one Overseer DO and everything in it). Includes everything needed
diff --git a/run-dev-server.js b/run-dev-server.js
index a5aba890..6c476e9f 100644
--- a/run-dev-server.js
+++ b/run-dev-server.js
@@ -246,6 +246,8 @@ for (const gk of gatekeepers) {
     // over HTTPS with tokens).
     "CF_AI_GATEWAY", "CF_AI_GATEWAY_PROVIDERS", "CF_AI_GATEWAY_ACCOUNT_ID",
     "CF_AI_GATEWAY_API_TOKEN", "CF_AI_GATEWAY_WAI", "CF_AI_GATEWAY_WAI_DIRECT",
+    // Moonshot AI provider (Kimi models). Direct OpenAI-compatible → api.moonshot.ai/v1.
+    "MOONSHOT_API_KEY", "MOONSHOT_API_URL",
   ];
   // OAuth app credentials (GOOGLE_/GITHUB_/CLOUDFLARE_OAUTH_*) are NOT passed to the backend anymore;
   // they are injected into the gatekeeper Workers (see SHARED_GATEKEEPER_CREDS below).