cartethyia 2.7.3 → 2.7.5

package/src/providers/integrations/github/github-discovery.ts

@@ -29,6 +29,7 @@ import {
   GITHUB_PROVIDER_ID,
   GITHUB_ENDPOINT_PATH,
   GITHUB_RESPONSES_ENDPOINT_PATH,
+  GITHUB_MESSAGES_ENDPOINT_PATH,
   GITHUB_AUTO_MODELS,
   recordGithubModelRoutes,
 } from "./github";
@@ -47,12 +48,11 @@ function positiveLimit(value: unknown): number | undefined {
  * The surfaces a row answers on, normalized to the canonical paths this
  * gateway registers.
  *
- * A row that supports neither chat nor responses is skipped rather than
- * defaulted: Copilot also serves a native Anthropic `/v1/messages` surface for
- * some SKUs, and this adapter speaks only the OpenAI wire. Registering such a
- * row as chat would publish a route that fails on first dispatch.
+ * Copilot serves three OpenAI/Anthropic surfaces: `/chat/completions`,
+ * `/responses`, and `/v1/messages`. A row that supports none of them is
+ * skipped rather than defaulted.
  */
-function wireFamilyFor(entry: Record<string, unknown>): "chat" | "responses" | undefined {
+function wireFamilyFor(entry: Record<string, unknown>): "chat" | "responses" | "messages" | undefined {
   const declared = boundedUpstreamArray(entry.supported_endpoints);
   if (!declared) return "chat";
   const endpoints = new Set<string>();
@@ -60,11 +60,12 @@ function wireFamilyFor(entry: Record<string, unknown>): "chat" | "responses" | u
     const label = sanitizeUpstreamLabel(value);
     if (label !== undefined) endpoints.add(label.replace(/^\/v1(?=\/)/, ""));
   }
-  // Chat is preferred when a row supports both: it is the surface every
-  // OpenAI-compatible client already speaks, and the upstream answers it for
-  // the same SKU. Responses is used only when chat is absent.
+  // Chat is preferred when a row supports it: it is the surface every
+  // OpenAI-compatible client already speaks. Responses and Messages are used
+  // only when chat is absent, in that order.
   if (endpoints.has("/chat/completions")) return "chat";
   if (endpoints.has("/responses")) return "responses";
+  if (endpoints.has("/messages")) return "messages";
   return undefined;
 }
 
@@ -75,6 +76,34 @@ function supports(entry: Record<string, unknown>, key: string): boolean {
   return declared?.[key] === true;
 }
 
+/**
+ * Price deltas from the account's `/models` row when it publishes `token_prices`
+ * (dollars per 1M tokens). This is the real per-account price Copilot serves,
+ * which the shared models.dev catalog cannot know; when absent the base-catalog
+ * estimate is used instead.
+ */
+function copilotCost(entry: Record<string, unknown>): ModelDefinition["cost"] {
+  const billing = isRecord(entry.billing) ? entry.billing : undefined;
+  const prices = isRecord(billing?.token_prices) ? billing.token_prices : undefined;
+  const defaultTier = isRecord(prices?.default) ? prices.default : undefined;
+  const input = positiveLimit(defaultTier?.input_price);
+  const output = positiveLimit(defaultTier?.output_price);
+  if (input === undefined && output === undefined) {
+    return modelsDevCatalog.costFor(GITHUB_PROVIDER_ID, String(entry.id ?? ""));
+  }
+  return {
+    input: input ?? null,
+    output: output ?? null,
+    ...(typeof (defaultTier as Record<string, unknown> | undefined)?.cache_read_price === "number"
+      ? { cache_read: (defaultTier as { cache_read_price?: number }).cache_read_price }
+      : {}),
+    ...(typeof (defaultTier as Record<string, unknown> | undefined)?.cache_write_price === "number"
+      ? { cache_write: (defaultTier as { cache_write_price?: number }).cache_write_price }
+      : {}),
+    pricing_model: "pay-per-use",
+  };
+}
+
 function toModelDefinition(entry: Record<string, unknown>): ModelDefinition | undefined {
   const modelId = sanitizeUpstreamLabel(entry.id);
   if (modelId === undefined) return undefined;
@@ -89,16 +118,32 @@ function toModelDefinition(entry: Record<string, unknown>): ModelDefinition | un
   // clamped the same way every other catalog row is.
   const outputLimit = Math.min(declaredOutput, contextLimit);
   const vision = supports(entry, "vision");
+  // `reasoning_effort` is a list (e.g. ["low","medium","high","xhigh","max"])
+  // when the model supports effort levels; absence means the model does not
+  // expose an effort knob. Carrying it lets routing clamp to the model's real
+  // ladder instead of the provider-wide `max→high` cap.
+  const supportsObj = isRecord(capabilities?.supports) ? capabilities.supports : undefined;
+  const reasoningEfforts = Array.isArray(supportsObj?.reasoning_effort)
+    ? (supportsObj.reasoning_effort as unknown[])
+        .filter((level): level is string => typeof level === "string" && level.length > 0)
+    : undefined;
+  const cost = copilotCost(entry);
   return {
     modelId,
     wireFamily,
-    endpointPath: wireFamily === "responses" ? GITHUB_RESPONSES_ENDPOINT_PATH : GITHUB_ENDPOINT_PATH,
+    endpointPath:
+      wireFamily === "responses"
+        ? GITHUB_RESPONSES_ENDPOINT_PATH
+        : wireFamily === "messages"
+          ? GITHUB_MESSAGES_ENDPOINT_PATH
+          : GITHUB_ENDPOINT_PATH,
     contextLimit,
     outputLimit,
     modalities: { input: vision ? ["text", "image"] : ["text"], output: ["text"] },
     reasoning: true,
+    ...(reasoningEfforts === undefined ? {} : { reasoningEfforts }),
     toolCall: true,
-    cost: modelsDevCatalog.costFor(GITHUB_PROVIDER_ID, modelId),
+    cost,
   };
 }
 

package/src/providers/integrations/github/github-oauth.ts

@@ -43,13 +43,14 @@ export const GITHUB_DEFAULT_API_HOST = "api.individual.githubcopilot.com" as con
  * The API rejects a request that does not identify a Copilot client, so these
  * are a protocol requirement rather than cosmetic fingerprinting — unlike the
  * Codex/Claude CLI identities elsewhere in this repository, which exist to look
- * like a first-party client. The version pair is the Copilot Chat extension's,
- * which is what the upstream expects for this client id.
+ * like a first-party client. The version pair is the Copilot Chat extension's
+ * (0.72.0 on VS Code 1.144), which is what the upstream expects for this
+ * client id.
  */
 export const GITHUB_REQUEST_HEADERS: Readonly<Record<string, string>> = Object.freeze({
-  "user-agent": "GitHubCopilotChat/0.70.0",
-  "editor-version": "vscode/1.142.0",
-  "editor-plugin-version": "copilot-chat/0.70.0",
+  "user-agent": "GitHubCopilotChat/0.72.0",
+  "editor-version": "vscode/1.144.0",
+  "editor-plugin-version": "copilot-chat/0.72.0",
   "copilot-integration-id": "vscode-chat",
   "x-github-api-version": "2026-08-01",
 });

package/src/providers/integrations/github/github.ts

@@ -44,17 +44,34 @@ export const GITHUB_ENDPOINT_PATH = "/chat/completions" as const;
  */
 export const GITHUB_RESPONSES_ENDPOINT_PATH = "/responses" as const;
 
+/**
+ * Anthropic Messages surface. Copilot serves the premium Claude family
+ * (`claude-sonnet-*`, `claude-opus-*`, `claude-haiku-*`, `claude-fable-*`)
+ * on the native `/v1/messages` wire, and some SKUs are messages-only. The
+ * adapter's shared pipeline already speaks the messages codec; this constant
+ * lets discovery route those rows here instead of dropping them.
+ */
+export const GITHUB_MESSAGES_ENDPOINT_PATH = "/v1/messages" as const;
+
 
 
 /**
- * The highest reasoning effort Copilot serves. A caller asking for `max` or
- * `xhigh` is answered at `high` rather than refused: the model has no level
- * above it, and the request is otherwise valid.
+ * The highest reasoning effort Copilot serves as a fallback when a model does
+ * not declare its own ladder. A caller asking for `max` or `xhigh` is answered
+ * at `high` rather than refused: the model has no level above it, and the
+ * request is otherwise valid.
  */
 const GITHUB_REASONING_EFFORT_CAP = "high";
 const GITHUB_CAPPED_EFFORTS = new Set(["max", "xhigh"]);
 
-/** Rewrites an effort above Copilot's ceiling to `high`, on either wire. */
+/**
+ * Rewrites an effort above Copilot's ceiling to the highest level we know the
+ * model accepts. The shared encoder already pinned the effort to the model's
+ * declared ladder (`resolveThinkingForModel` reads `reasoningEfforts` from the
+ * catalog); this hook only guards a bare fallback so a caller asking for
+ * `max`/`xhigh` on a model without a declared ladder is answered at `high`
+ * rather than refused.
+ */
 function clampGithubReasoningEffort(payload: Record<string, unknown>): void {
   const effort = payload["reasoning_effort"];
   if (typeof effort === "string" && GITHUB_CAPPED_EFFORTS.has(effort)) {
@@ -88,12 +105,6 @@ export const GITHUB_EXCLUDED_MODEL_IDS: ReadonlySet<string> = new Set([
   "gpt-6-luna-medium",
   "gpt-6-luna-high",
   "kimi-k3",
-  "kimi-k3-base",
-  "kimi-k3-copilot",
-  "mai-code-1-flash-4th",
-  "mai-code-1-flash-secondary",
-  "mai-code-1-flash-tertiary",
-  "mai-code-1.1-flash",
 ]);
 
 const modelRoutesByCredential = new Map<string, ReadonlyMap<string, GithubModelRoute>>();
@@ -122,7 +133,10 @@ function parseRouteMap(raw: string): Map<string, GithubModelRoute> | undefined {
       if (typeof modelId !== "string" || route === null || typeof route !== "object") continue;
       const wireFamily = (route as { wireFamily?: unknown }).wireFamily;
       const endpointPath = (route as { endpointPath?: unknown }).endpointPath;
-      if ((wireFamily === "chat" || wireFamily === "responses") && typeof endpointPath === "string") {
+      if (
+        (wireFamily === "chat" || wireFamily === "responses" || wireFamily === "messages") &&
+        typeof endpointPath === "string"
+      ) {
         map.set(modelId, { wireFamily, endpointPath });
       }
     }
@@ -162,7 +176,7 @@ async function writeSharedModelRoutes(
 }
 
 export interface GithubModelRoute {
-  readonly wireFamily: "chat" | "responses";
+  readonly wireFamily: "chat" | "responses" | "messages";
   readonly endpointPath: string;
 }
 const GITHUB_AUTO_SESSION_CACHE_MAX_ENTRIES = 50;
@@ -187,8 +201,8 @@ export function recordGithubModelRoutes(
   if (key === undefined) return;
   const routes = models
     .filter(
-      (model): model is ModelDefinition & { readonly wireFamily: "chat" | "responses" } =>
-        model.wireFamily === "chat" || model.wireFamily === "responses",
+      (model): model is ModelDefinition & { readonly wireFamily: "chat" | "responses" | "messages" } =>
+        model.wireFamily === "chat" || model.wireFamily === "responses" || model.wireFamily === "messages",
     )
     .map((model) => [
       model.modelId.toLowerCase(),
@@ -304,6 +318,9 @@ function routeFromAutoModel(value: unknown): GithubModelRoute | undefined {
   if (endpoints.has("/responses")) {
     return { wireFamily: "responses", endpointPath: GITHUB_RESPONSES_ENDPOINT_PATH };
   }
+  if (endpoints.has("/messages")) {
+    return { wireFamily: "messages", endpointPath: GITHUB_MESSAGES_ENDPOINT_PATH };
+  }
   return undefined;
 }
 
@@ -411,7 +428,7 @@ class GithubAdapter extends OpenAICompatibleAdapter {
 
     const secret = context.credential.secret;
     if (secret === undefined || secret.length === 0) {
-      throw new GatewayError("authentication_failed", 401, "GitHub Copilot credential is missing");
+      throw new GatewayError("authentication_failed", 401, "GitHub Copilot credential is missing (request {requestId}). Wrong key, right door — check the credential 🔑");
     }
     const credential = new TextDecoder().decode(secret);
     const autoInput = autoPrompt(request);
@@ -424,7 +441,7 @@ class GithubAdapter extends OpenAICompatibleAdapter {
       : this.#getCachedAutoDecision(cacheKey, autoInput.hasImage);
     if (autoDecision === undefined) {
       if (autoInput.prompt.length === 0) {
-        throw new GatewayError("invalid_request", 400, "Copilot Auto requires a text prompt");
+        throw new GatewayError("invalid_request", 400, "Copilot Auto requires a text prompt (request {requestId}). Check the shape of what you sent — then resend 🪞");
       }
       autoDecision = await this.#selectAutoModel(
         autoInput.prompt,
@@ -462,7 +479,7 @@ class GithubAdapter extends OpenAICompatibleAdapter {
       throw new GatewayError(
         "upstream_unprocessable",
         502,
-        "Copilot Auto selected a model without a discovered inference endpoint",
+        "Copilot Auto selected a model without a discovered inference endpoint (request {requestId}). The upstream came back empty-handed — retry 🌙",
         { selected_model: selectedModelId },
         "upstream",
       );
@@ -498,7 +515,7 @@ class GithubAdapter extends OpenAICompatibleAdapter {
       ? undefined
       : parseGithubCredential(new TextDecoder().decode(context.credential.secret));
     if (credential === undefined) {
-      throw new GatewayError("authentication_failed", 401, "GitHub Copilot credential is missing");
+      throw new GatewayError("authentication_failed", 401, "GitHub Copilot credential is missing (request {requestId}). Wrong key, right door — check the credential 🔑");
     }
     const payload: Record<string, unknown> = { prompt, tier };
     if (hasImage) payload.has_image = true;
@@ -522,7 +539,7 @@ class GithubAdapter extends OpenAICompatibleAdapter {
       throw new GatewayError(
         "upstream_unprocessable",
         502,
-        "Copilot Auto returned an invalid routing response",
+        "Copilot Auto returned an invalid routing response (request {requestId}). The upstream came back empty-handed — retry 🌙",
         {},
         "upstream",
       );
@@ -545,7 +562,7 @@ class GithubAdapter extends OpenAICompatibleAdapter {
       throw new GatewayError(
         "upstream_unprocessable",
         502,
-        "Copilot Auto did not return a selected model",
+        "Copilot Auto did not return a selected model (request {requestId}). The upstream came back empty-handed — retry 🌙",
         {},
         "upstream",
       );
@@ -640,10 +657,20 @@ class GithubAdapter extends OpenAICompatibleAdapter {
       "editor-device-id": await getOrCreateInstallId("github"),
       "openai-intent": "conversation-agent",
       "x-agent-task-id": requestId,
+      "x-interaction-id": requestId,
       "x-initiator": isAgentInitiated(request) ? "agent" : "user",
       "x-interaction-type": "conversation-agent",
       "x-request-id": requestId,
+      "x-vscode-user-agent-library-version": "vscode-chat/0.72.0",
     });
+    // The copy Claude family is served on the native Anthropic Messages wire,
+    // which gates feature betas behind `anthropic-beta` (mirroring the
+    // Copilot extension's `getAnthropicBetaHeader`). Adaptive thinking is the
+    // baseline the model roster declares, so it is requested for every messages
+    // route; the header is scoped to that wire and never sent elsewhere.
+    if (candidate?.wire_family === "messages") {
+      headers["anthropic-beta"] = "interleaved-thinking-2025-05-14";
+    }
     if (allParts(request).some(containsImage)) {
       headers["copilot-vision-request"] = "true";
     } else {