feat(sse): route GitHub Copilot Claude models through native /v1/messages (#7223)

* feat(sse): route GitHub Copilot Claude models through native /v1/messages

GitHub Copilot's /chat/completions and /responses endpoints never surface
prompt-cache token counts (cached_tokens) for Claude models, and round-tripping
Claude tool_use/tool_result/thinking content blocks through the OpenAI shape
is lossy. Copilot also exposes an Anthropic-native /v1/messages shim that
reports cached_tokens correctly and accepts native content blocks as-is.

Tag each github registry claude-* model with targetFormat: "claude" so
chatCore.ts translates the request to Anthropic-native shape before the
executor ever sees it (the same mechanism opencode/zen's Qwen entries and
opencode/go already use), and teach the github executor's buildUrl() /
buildHeaders() to dispatch those models at the new messagesUrl
(api.githubcopilot.com/v1/messages) with the required anthropic-version
header. transformRequest() now skips its /chat/completions-only quirks
(content-part flattening, trailing-assistant-prefill drop, the
response_format-as-system-prompt workaround) for the native path — the first
would destroy native tool_use/tool_result blocks, the prefill drop is
unnecessary because the real Anthropic API supports assistant prefill, and
the response_format workaround is superseded by the generic openai-to-claude
translator's own JSON-mode handling.

Co-authored-by: luoyide <ydhome.code@gmail.com>
Inspired-by: https://github.com/decolua/9router/pull/2608

* chore(changelog): fragment for #7223

* fix(sse): green PR #7223 CI — complexity ratchet + stale test expectations

- Extract applyChatCompletionsOnlyQuirks() and resolveInitiatorHeader()
  out of GithubExecutor.transformRequest()/buildHeaders() so the two
  methods drop back under the complexity/cognitive-complexity ratchets
  (2058/891 -> 2056/890, matching the frozen baseline). No behavior
  change — same guards, just relocated.
- Update 4 pre-existing unit tests that hard-coded now-native claude-*
  Copilot ids (claude-sonnet-4.5/4.6) to exercise the /chat/completions
  legacy path via an unregistered id (claude-sonnet-4), matching the
  sibling test already using that pattern. These ids now intentionally
  route to the native /v1/messages shim added by this PR, which
  correctly skips the /chat/completions-only workarounds these tests
  were built to verify — the native path's own coverage lives in
  github-copilot-claude-native-messages.test.ts.
- Split the routing invariant test (copilot-gemini-claude-route-no-responses.test.ts)
  into a Claude case (expects /v1/messages) and a Gemini case (still
  expects /chat/completions), reflecting the intentional routing change.

---------

Co-authored-by: luoyide <ydhome.code@gmail.com>

changelog.d/features/7223-github-copilot-claude-native-messages.md

@@ -0,0 +1 @@
+- **feat(sse):** GitHub Copilot Claude models now route through Copilot's native `/v1/messages` endpoint (prompt-cache token counts, no more lossy tool-call round-trip). (thanks @yidecode)

open-sse/config/providerRegistry.ts

@@ -39,6 +39,9 @@ export function generateLegacyProviders(): Record<string, LegacyProvider> {
     if (entry.responsesBaseUrl) {
       p.responsesBaseUrl = entry.responsesBaseUrl;
     }
+    if (entry.messagesUrl) {
+      p.messagesUrl = entry.messagesUrl;
+    }
     if (entry.requestDefaults) {
       p.requestDefaults = entry.requestDefaults;
     }

open-sse/config/providers/registry/github/index.ts

@@ -12,6 +12,12 @@ export const githubProvider: RegistryEntry = {
   executor: "github",
   baseUrl: "https://api.githubcopilot.com/chat/completions",
   responsesBaseUrl: "https://api.githubcopilot.com/responses",
+  // Anthropic-native shim: the only Copilot endpoint that surfaces prompt-cache
+  // token counts (cached_tokens) for Claude models, and avoids round-tripping
+  // tool_use/tool_result/thinking content blocks through the OpenAI shape.
+  // Routed via each claude-* model's targetFormat: "claude" below (see
+  // executors/github.ts buildUrl/buildHeaders). Port of decolua/9router#2608.
+  messagesUrl: "https://api.githubcopilot.com/v1/messages",
   authType: "oauth",
   authHeader: "bearer",
   // GitHub Copilot is a public device-flow OAuth client: it has a public client_id but
@@ -24,16 +30,23 @@ export const githubProvider: RegistryEntry = {
   },
   defaultContextLength: 128000,
   headers: getGitHubCopilotChatHeaders(),
+  // All claude-* entries below carry targetFormat: "claude" so chatCore.ts
+  // translates the request to Anthropic-native shape before the executor ever
+  // sees it, and the github executor's buildUrl()/buildHeaders() route them at
+  // messagesUrl (/v1/messages) instead of /chat/completions. Port of
+  // decolua/9router#2608 (author: yidecode) — see executors/github.ts.
   models: [
     {
       id: "claude-fable-5",
       name: "Claude Fable 5",
+      targetFormat: "claude",
       contextLength: 1000000,
       maxOutputTokens: 64000,
     },
     {
       id: "claude-opus-4.8-fast",
       name: "Claude Opus 4.8 (fast mode)",
+      targetFormat: "claude",
       contextLength: 1000000,
       maxOutputTokens: 64000,
       unsupportedParams: ["temperature", "top_p", "top_k"],
@@ -41,6 +54,7 @@ export const githubProvider: RegistryEntry = {
     {
       id: "claude-opus-4.8",
       name: "Claude Opus 4.8",
+      targetFormat: "claude",
       contextLength: 1000000,
       maxOutputTokens: 64000,
       unsupportedParams: ["temperature", "top_p", "top_k"],
@@ -48,36 +62,42 @@ export const githubProvider: RegistryEntry = {
     {
       id: "claude-opus-4.7",
       name: "Claude Opus 4.7",
+      targetFormat: "claude",
       contextLength: 1000000,
       maxOutputTokens: 64000,
     },
     {
       id: "claude-sonnet-4.6",
       name: "Claude Sonnet 4.6",
+      targetFormat: "claude",
       contextLength: 1000000,
       maxOutputTokens: 64000,
     },
     {
       id: "claude-opus-4.5",
       name: "Claude Opus 4.5",
+      targetFormat: "claude",
       contextLength: 200000,
       maxOutputTokens: 32000,
     },
     {
       id: "claude-sonnet-5",
       name: "Claude Sonnet 5",
+      targetFormat: "claude",
       contextLength: 1000000,
       maxOutputTokens: 64000,
     },
     {
       id: "claude-sonnet-4.5",
       name: "Claude Sonnet 4.5",
+      targetFormat: "claude",
       contextLength: 200000,
       maxOutputTokens: 32000,
     },
     {
       id: "claude-haiku-4.5",
       name: "Claude Haiku 4.5",
+      targetFormat: "claude",
       contextLength: 200000,
       maxOutputTokens: 32000,
     },

open-sse/config/providers/shared.ts

@@ -107,6 +107,10 @@ export interface RegistryEntry {
   /** Override base URL used only for API key validation (e.g., opencode-go validates on zen/v1) */
   testKeyBaseUrl?: string;
   responsesBaseUrl?: string;
+  /** Anthropic-native /v1/messages endpoint (e.g. GitHub Copilot's shim) used
+   *  for models tagged `targetFormat: "claude"` on an otherwise openai-format
+   *  provider — see registry/github/index.ts. */
+  messagesUrl?: string;
   urlSuffix?: string;
   urlBuilder?: (base: string, model: string, stream: boolean) => string;
   authType: string;
@@ -174,6 +178,7 @@ export interface LegacyProvider {
   baseUrl?: string;
   baseUrls?: string[];
   responsesBaseUrl?: string;
+  messagesUrl?: string;
   headers?: Record<string, string>;
   requestDefaults?: ProviderRequestDefaults;
   clientId?: string;

open-sse/executors/base.ts

@@ -117,6 +117,7 @@ export type ProviderConfig = {
   baseUrl?: string;
   baseUrls?: string[];
   responsesBaseUrl?: string;
+  messagesUrl?: string;
   chatPath?: string;
   clientVersion?: string;
   clientId?: string;

open-sse/executors/github.ts

@@ -41,6 +41,16 @@ export class GithubExecutor extends BaseExecutor {
 
   buildUrl(model: string, _stream: boolean, _urlIndex = 0) {
     const targetFormat = getModelTargetFormat("gh", model);
+    // Claude models: route to Copilot's Anthropic-native /v1/messages shim — the
+    // only Copilot endpoint that surfaces prompt-cache token counts for Claude and
+    // avoids a lossy round-trip of tool_use/tool_result/thinking content blocks
+    // through the OpenAI shape. Driven by the registry's per-model targetFormat
+    // (see registry/github/index.ts), which chatCore.ts also uses to translate the
+    // request to Claude shape before the executor ever sees it.
+    // Port of decolua/9router#2608 (author: yidecode).
+    if (targetFormat === "claude" && this.config.messagesUrl) {
+      return this.config.messagesUrl;
+    }
     // 9router#102: Copilot Codex models advertise supported_endpoints: ["/responses"]
     // and 400 on /chat/completions. Route any *-codex id to /responses even when it
     // isn't in the curated registry, so newly-shipped Codex models work out of the box.
@@ -93,6 +103,15 @@ export class GithubExecutor extends BaseExecutor {
     const sourceBody = body && typeof body === "object" ? body : {};
     const modifiedBody = { ...sourceBody };
 
+    // Claude models arrive here already translated to Anthropic-native shape by
+    // chatCore.ts (registry targetFormat: "claude" — see registry/github/index.ts)
+    // and are dispatched at /v1/messages (buildUrl above), which behaves like the
+    // real Anthropic API. None of the /chat/completions-only quirks below apply —
+    // content-part flattening would destroy native tool_use/tool_result/thinking
+    // blocks, and the native endpoint (unlike Copilot's /chat/completions) honors
+    // assistant-message prefill. Port of decolua/9router#2608 (author: yidecode).
+    const isClaudeNative = getModelTargetFormat("gh", model) === "claude";
+
     if (Array.isArray(sourceBody.input)) {
       modifiedBody.input = sanitizeResponsesInputItems(sourceBody.input, false);
     }
@@ -110,14 +129,6 @@ export class GithubExecutor extends BaseExecutor {
       });
     }
 
-    if (modifiedBody.response_format && model.toLowerCase().includes("claude")) {
-      modifiedBody.messages = this.injectResponseFormat(
-        Array.isArray(modifiedBody.messages) ? modifiedBody.messages : [],
-        modifiedBody.response_format
-      );
-      delete modifiedBody.response_format;
-    }
-
     if (Array.isArray(modifiedBody.tools) && modifiedBody.tools.length > 128) {
       modifiedBody.tools = modifiedBody.tools.slice(0, 128);
     }
@@ -136,29 +147,13 @@ export class GithubExecutor extends BaseExecutor {
       delete modifiedBody.temperature;
     }
 
-    // GitHub Copilot /chat/completions only accepts {type:'text'} or {type:'image_url'}
-    // content parts. Clients like Cursor IDE pass through Anthropic-shape parts
-    // (tool_use, tool_result, thinking) untouched when using Claude models, which makes
-    // the endpoint return: "type has to be either 'image_url' or 'text'" (HTTP 400).
-    // Serialize unknown part types as text, drop empty parts, and collapse to null when
-    // every part is stripped (assistant messages whose only content was tool_calls).
-    // Port from 9router#220 (fixes 9router#219).
-    if (Array.isArray(modifiedBody.messages)) {
-      modifiedBody.messages = modifiedBody.messages.map((msg: any) =>
-        this.sanitizeChatCompletionsMessage(msg)
-      );
-    }
-
-    // GitHub Copilot's /chat/completions endpoint rejects a conversation that ends
-    // with an assistant message: "This model does not support assistant message
-    // prefill. The conversation must end with a user message." (HTTP 400). Anthropic
-    // clients such as newest Claude Desktop send a trailing assistant turn as a
-    // prefill seed — the Anthropic API honors it, but Copilot does not. Drop it here,
-    // scoped to the GitHub executor only (the shared translator/contextManager and
-    // other providers that DO honor prefill are untouched).
-    // Port of 9router#2143 (author: Manuel <baslr@users.noreply.github.com>).
-    if (Array.isArray(modifiedBody.messages)) {
-      modifiedBody.messages = this.dropTrailingAssistantPrefill(modifiedBody.messages);
+    // The quirks below (response_format-as-system-prompt, content-part flattening,
+    // trailing-assistant-prefill drop) are all workarounds for /chat/completions-only
+    // limitations. They either don't apply to Claude-shape bodies or actively corrupt
+    // them, so they are skipped entirely for the native /v1/messages path. Port of
+    // decolua/9router#2608 (author: yidecode) — see class doc comment above.
+    if (!isClaudeNative) {
+      this.applyChatCompletionsOnlyQuirks(model, modifiedBody);
     }
 
     // Config-driven strip of params unsupported by the target provider/model.
@@ -171,6 +166,46 @@ export class GithubExecutor extends BaseExecutor {
     return modifiedBody;
   }
 
+  // GitHub Copilot's /chat/completions endpoint has several quirks that the native
+  // /v1/messages shim doesn't share — extracted from transformRequest so the native
+  // path (the common case for Claude models going forward) doesn't pay their branch
+  // cost. Mutates modifiedBody in place.
+  private applyChatCompletionsOnlyQuirks(model: string, modifiedBody): void {
+    // Claude models on /chat/completions don't support response_format — inject the
+    // instruction as a system message instead. Port from 9router (see
+    // injectResponseFormat above).
+    if (modifiedBody.response_format && model.toLowerCase().includes("claude")) {
+      modifiedBody.messages = this.injectResponseFormat(
+        Array.isArray(modifiedBody.messages) ? modifiedBody.messages : [],
+        modifiedBody.response_format
+      );
+      delete modifiedBody.response_format;
+    }
+
+    if (!Array.isArray(modifiedBody.messages)) return;
+
+    // GitHub Copilot /chat/completions only accepts {type:'text'} or {type:'image_url'}
+    // content parts. Clients like Cursor IDE pass through Anthropic-shape parts
+    // (tool_use, tool_result, thinking) untouched when using Claude models, which makes
+    // the endpoint return: "type has to be either 'image_url' or 'text'" (HTTP 400).
+    // Serialize unknown part types as text, drop empty parts, and collapse to null when
+    // every part is stripped (assistant messages whose only content was tool_calls).
+    // Port from 9router#220 (fixes 9router#219).
+    modifiedBody.messages = modifiedBody.messages.map((msg: any) =>
+      this.sanitizeChatCompletionsMessage(msg)
+    );
+
+    // GitHub Copilot's /chat/completions endpoint rejects a conversation that ends
+    // with an assistant message: "This model does not support assistant message
+    // prefill. The conversation must end with a user message." (HTTP 400). Anthropic
+    // clients such as newest Claude Desktop send a trailing assistant turn as a
+    // prefill seed — the Anthropic API honors it, but Copilot does not. Drop it here,
+    // scoped to the GitHub executor only (the shared translator/contextManager and
+    // other providers that DO honor prefill are untouched).
+    // Port of 9router#2143 (author: Manuel <baslr@users.noreply.github.com>).
+    modifiedBody.messages = this.dropTrailingAssistantPrefill(modifiedBody.messages);
+  }
+
   private sanitizeChatCompletionsMessage(msg: any): any {
     if (!msg || typeof msg !== "object") return msg;
     // String content and missing content (e.g. assistant w/ only tool_calls) pass through.
@@ -235,17 +270,38 @@ export class GithubExecutor extends BaseExecutor {
   buildHeaders(
     credentials: ProviderCredentials,
     stream = true,
-    clientHeaders?: Record<string, string> | null
+    clientHeaders?: Record<string, string> | null,
+    model?: string
   ): Record<string, string> {
     const token = this.getCopilotToken(credentials) || credentials.accessToken;
+    const initiator = this.resolveInitiatorHeader(clientHeaders);
 
-    // Forward the client's x-initiator header when present. OpenCode and other
-    // Copilot-aware clients use this to distinguish user-initiated turns
-    // (x-initiator: user) from autonomous tool-call continuations
-    // (x-initiator: agent). GitHub Copilot's billing treats "agent" turns as
-    // free, so forwarding the value avoids burning a premium request on every
-    // tool-call round-trip.  Fall back to "user" when the header is absent to
-    // preserve the existing default behaviour.
+    const headers: Record<string, string> = {
+      ...getGitHubCopilotChatHeaders(stream ? "text/event-stream" : "application/json", initiator),
+      Authorization: `Bearer ${token}`,
+      "x-request-id":
+        crypto.randomUUID?.() || `${Date.now()}-${Math.random().toString(36).slice(2)}`,
+    };
+
+    // Claude models routed to the Anthropic-native /v1/messages shim require the
+    // anthropic-version header (harmless no-op on /chat/completions and /responses,
+    // but /v1/messages rejects the request without it). Port of decolua/9router#2608.
+    if (model && getModelTargetFormat("gh", model) === "claude") {
+      headers["anthropic-version"] = "2023-06-01";
+    }
+
+    return headers;
+  }
+
+  // Forward the client's x-initiator header when present. OpenCode and other
+  // Copilot-aware clients use this to distinguish user-initiated turns
+  // (x-initiator: user) from autonomous tool-call continuations
+  // (x-initiator: agent). GitHub Copilot's billing treats "agent" turns as
+  // free, so forwarding the value avoids burning a premium request on every
+  // tool-call round-trip. Falls back to "user" when the header is absent to
+  // preserve the existing default behaviour. Extracted from buildHeaders so
+  // header assembly stays the one place that reads it.
+  private resolveInitiatorHeader(clientHeaders?: Record<string, string> | null): string {
     let clientInitiator = clientHeaders?.["x-initiator"] || clientHeaders?.["X-Initiator"];
     if (!clientInitiator && clientHeaders) {
       for (const key in clientHeaders) {
@@ -255,15 +311,7 @@ export class GithubExecutor extends BaseExecutor {
         }
       }
     }
-    const initiator =
-      clientInitiator === "agent" || clientInitiator === "user" ? clientInitiator : "user";
-
-    return {
-      ...getGitHubCopilotChatHeaders(stream ? "text/event-stream" : "application/json", initiator),
-      Authorization: `Bearer ${token}`,
-      "x-request-id":
-        crypto.randomUUID?.() || `${Date.now()}-${Math.random().toString(36).slice(2)}`,
-    };
+    return clientInitiator === "agent" || clientInitiator === "user" ? clientInitiator : "user";
   }
 
   async refreshCopilotToken(githubAccessToken, log) {

tests/unit/copilot-gemini-claude-route-no-responses.test.ts

@@ -27,6 +27,7 @@ import type { RegistryModel } from "../../open-sse/config/providerRegistry.ts";
 
 const CHAT_URL = "https://api.githubcopilot.com/chat/completions";
 const RESPONSES_URL = "https://api.githubcopilot.com/responses";
+const MESSAGES_URL = "https://api.githubcopilot.com/v1/messages";
 
 function getGithubModel(modelId: string): RegistryModel {
   const model = PROVIDER_MODELS["gh"]?.find((entry) => entry.id === modelId);
@@ -35,7 +36,7 @@ function getGithubModel(modelId: string): RegistryModel {
 }
 
 describe("GithubExecutor — Gemini/Claude must never hit /responses (port 9router#1536)", () => {
-  it("routes registered Claude/Gemini Copilot models to chat/completions", () => {
+  it("routes registered Claude Copilot models to the native /v1/messages shim (port decolua/9router#2608)", () => {
     const exec = new GithubExecutor();
     for (const id of [
       "claude-haiku-4.5",
@@ -43,14 +44,18 @@ describe("GithubExecutor — Gemini/Claude must never hit /responses (port 9rout
       "claude-sonnet-4.6",
       "claude-sonnet-5",
       "claude-fable-5",
-      "claude-opus-4.6",
       "claude-opus-4.7",
       "claude-opus-4.8",
       "claude-opus-4.8-fast",
       "claude-opus-4.5",
-      "gemini-3.1-pro-preview",
-      "gemini-3.5-flash",
     ]) {
+      assert.equal(exec.buildUrl(id, false), MESSAGES_URL, `${id} must route to /v1/messages`);
+    }
+  });
+
+  it("routes registered Gemini Copilot models to chat/completions", () => {
+    const exec = new GithubExecutor();
+    for (const id of ["gemini-3.1-pro-preview", "gemini-3.5-flash"]) {
       assert.equal(exec.buildUrl(id, false), CHAT_URL, `${id} must route to chat/completions`);
     }
   });

tests/unit/executor-github-prefill-sanitize.test.ts

@@ -87,15 +87,21 @@ test("dropTrailingAssistantPrefill is null/empty safe", () => {
 
 test("GithubExecutor.transformRequest drops the trailing assistant prefill end-to-end", () => {
   const executor = new GithubExecutor();
+  // Use an unregistered claude-* id so getModelTargetFormat("gh", ...) resolves
+  // to null and this stays on the /chat/completions path this test targets.
+  // Registered claude-* ids (e.g. "claude-sonnet-4.6") now carry
+  // targetFormat:"claude" (native /v1/messages, which supports prefill — port
+  // of decolua/9router#2608, see github-copilot-claude-native-messages.test.ts)
+  // and intentionally skip this drop.
   const body = {
-    model: "claude-sonnet-4.6",
+    model: "claude-sonnet-4",
     messages: [
       { role: "user", content: "Hi" },
       { role: "assistant", content: "Here is the answer:" },
     ],
   };
 
-  const out = executor.transformRequest("claude-sonnet-4.6", body, false, {});
+  const out = executor.transformRequest("claude-sonnet-4", body, false, {});
 
   assert.equal(out.messages.length, 1);
   assert.equal(out.messages[0].role, "user");

tests/unit/executor-github.test.ts

@@ -161,7 +161,13 @@ test("GithubExecutor.transformRequest sanitizes Anthropic-shape content parts (t
     ],
   };
 
-  const result = executor.transformRequest("claude-sonnet-4.6", body, true, {});
+  // Use an unregistered claude-* id (not "claude-sonnet-4.6"/etc.) so
+  // getModelTargetFormat("gh", ...) resolves to null and this stays on the
+  // /chat/completions path this test targets. Registered claude-* ids now
+  // carry targetFormat:"claude" (native /v1/messages — port of
+  // decolua/9router#2608, see github-copilot-claude-native-messages.test.ts)
+  // and intentionally skip this sanitization.
+  const result = executor.transformRequest("claude-sonnet-4", body, true, {});
 
   // user message keeps text + image_url parts untouched
   assert.equal(result.messages[0].content[0].type, "text");

tests/unit/github-copilot-claude-native-messages.test.ts

@@ -0,0 +1,131 @@
+// GitHub Copilot exposes an Anthropic-native `/v1/messages` shim alongside its
+// OpenAI-shape `/chat/completions` and `/responses` endpoints. Only the native
+// shim surfaces prompt-cache token counts (`cached_tokens`) for Claude models —
+// /chat/completions silently drops them, and round-tripping Claude tool_use /
+// tool_result / thinking content blocks through the OpenAI shape is lossy.
+//
+// Port of upstream decolua/9router#2608 (author: yidecode), adapted to
+// OmniRoute's architecture: instead of the executor doing its own
+// translateRequest/translateResponse + manual SSE TransformStream (9router has
+// no generic per-model targetFormat mechanism), OmniRoute already has a
+// registry-driven `targetFormat` field (see opencode/zen's Qwen entries,
+// opencode/go) that makes chatCore.ts translate the request to Claude shape
+// *before* the executor ever sees it, and translate the response back
+// generically afterwards. So the actual port is: (1) tag the github registry's
+// claude-* models with targetFormat:"claude", (2) teach the github executor's
+// buildUrl()/buildHeaders() to route those models at the new messagesUrl with
+// an anthropic-version header, and (3) gate the executor's /chat/completions-only
+// request transforms (content-part flattening, trailing-assistant-prefill drop,
+// response_format-as-system-prompt workaround) off for the native path, since
+// they either don't apply to Claude-shape bodies or actively corrupt them.
+
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const { GithubExecutor } = await import("../../open-sse/executors/github.ts");
+const { getModelTargetFormat } = await import("../../open-sse/config/providerModels.ts");
+
+test("registry: claude-* github models resolve targetFormat 'claude'", () => {
+  for (const model of ["claude-opus-4.8", "claude-sonnet-4.6", "claude-haiku-4.5"]) {
+    assert.equal(
+      getModelTargetFormat("gh", model),
+      "claude",
+      `${model} must resolve to the claude target format so chatCore translates natively`
+    );
+  }
+});
+
+test("registry: non-claude github models keep their existing targetFormat", () => {
+  assert.equal(getModelTargetFormat("gh", "gpt-5.4"), "openai-responses");
+  assert.equal(getModelTargetFormat("gh", "gpt-4o-mini"), null);
+});
+
+test("buildUrl: claude models route to the native /v1/messages endpoint", () => {
+  const executor = new GithubExecutor();
+  const url = executor.buildUrl("claude-opus-4.8", true);
+  assert.equal(url, "https://api.githubcopilot.com/v1/messages");
+});
+
+test("buildUrl: gpt codex/responses models still route to /responses", () => {
+  const executor = new GithubExecutor();
+  const url = executor.buildUrl("gpt-5.4", true);
+  assert.match(url, /\/responses$/);
+});
+
+test("buildUrl: plain gpt models still route to /chat/completions", () => {
+  const executor = new GithubExecutor();
+  const url = executor.buildUrl("gpt-4o-mini", true);
+  assert.equal(url, executor.config.baseUrl);
+  assert.match(url, /\/chat\/completions$/);
+});
+
+test("buildHeaders: claude-native requests carry anthropic-version", () => {
+  const executor = new GithubExecutor();
+  const headers = executor.buildHeaders({ accessToken: "tok" }, true, null, "claude-opus-4.8");
+  assert.equal(headers["anthropic-version"], "2023-06-01");
+});
+
+test("buildHeaders: non-claude requests do not carry anthropic-version", () => {
+  const executor = new GithubExecutor();
+  const headers = executor.buildHeaders({ accessToken: "tok" }, true, null, "gpt-4o-mini");
+  assert.equal(headers["anthropic-version"], undefined);
+});
+
+test("transformRequest: claude-native path preserves native tool_use/tool_result content blocks", () => {
+  const executor = new GithubExecutor();
+  const body = {
+    model: "claude-opus-4.8",
+    system: "you are a helpful assistant",
+    messages: [
+      {
+        role: "assistant",
+        content: [{ type: "tool_use", id: "t1", name: "search", input: { q: "hi" } }],
+      },
+      {
+        role: "user",
+        content: [{ type: "tool_result", tool_use_id: "t1", content: "result text" }],
+      },
+    ],
+  };
+  const result = executor.transformRequest("claude-opus-4.8", body, true, {});
+  // Pre-port: sanitizeChatCompletionsMessage flattened every non-text/image_url
+  // part to {type:"text", text: ...}, destroying the tool_use/tool_result blocks
+  // Anthropic's native /v1/messages endpoint actually needs.
+  assert.equal((result.messages[0].content[0] as { type: string }).type, "tool_use");
+  assert.equal((result.messages[1].content[0] as { type: string }).type, "tool_result");
+});
+
+test("transformRequest: claude-native path keeps a trailing assistant message (prefill)", () => {
+  const executor = new GithubExecutor();
+  const body = {
+    model: "claude-opus-4.8",
+    messages: [
+      { role: "user", content: "hi" },
+      { role: "assistant", content: "Sure, here is" },
+    ],
+  };
+  const result = executor.transformRequest("claude-opus-4.8", body, true, {});
+  // Pre-port: dropTrailingAssistantPrefill removed the trailing assistant turn
+  // because Copilot's /chat/completions rejects prefill — but the native
+  // /v1/messages endpoint is real Anthropic-compatible and supports it.
+  assert.equal(result.messages.length, 2);
+  assert.equal(result.messages[1].role, "assistant");
+});
+
+test("transformRequest: non-claude (chat/completions) path still flattens tool_use content and drops prefill", () => {
+  const executor = new GithubExecutor();
+  const body = {
+    model: "gpt-4o-mini",
+    messages: [
+      {
+        role: "assistant",
+        content: [{ type: "tool_use", id: "t1", name: "search", input: {} }],
+      },
+      { role: "user", content: "hi" },
+      { role: "assistant", content: "trailing prefill" },
+    ],
+  };
+  const result = executor.transformRequest("gpt-4o-mini", body, true, {});
+  assert.equal((result.messages[0].content[0] as { type: string }).type, "text");
+  assert.equal(result.messages.length, 2, "trailing assistant message must still be dropped");
+});

tests/unit/provider-models-config.test.ts

@@ -89,8 +89,15 @@ test("GitHub Copilot registry reflects the current supported model lineup", () =
 
   assert.deepEqual(ids, [...GITHUB_COPILOT_MODEL_ALLOWLIST]);
   assert.equal(getModelTargetFormat("gh", "gpt-5.3-codex"), "openai-responses");
+  // "claude-opus-4.6" is not a real Copilot model id (unlike claude-sonnet-4.6);
+  // it never appears in the registry, so its target format stays null.
   assert.equal(getModelTargetFormat("gh", "claude-opus-4.6"), null);
-  assert.equal(getModelTargetFormat("gh", "claude-opus-4.8-fast"), null);
+  // Claude models route through Copilot's Anthropic-native /v1/messages shim
+  // (executors/github.ts) — the only endpoint that surfaces prompt-cache token
+  // counts for Claude and avoids a lossy tool_use/tool_result round-trip through
+  // the OpenAI shape. Port of decolua/9router#2608.
+  assert.equal(getModelTargetFormat("gh", "claude-opus-4.8-fast"), "claude");
+  assert.equal(getModelTargetFormat("gh", "claude-sonnet-4.6"), "claude");
   assert.equal(getModelTargetFormat("gh", "gemini-3.5-flash"), null);
   assert.equal(getModelTargetFormat("gh", "kimi-k2.7-code"), null);
   assert.equal(ids.includes("gpt-4"), false);

tests/unit/t27-github-copilot-response-format.test.ts

@@ -32,7 +32,13 @@ test("T27: Claude + response_format=json_object injects system instruction and s
     response_format: { type: "json_object" },
   };
 
-  const transformed = executor.transformRequest("claude-sonnet-4.5", request, false, {});
+  // Use an unregistered claude-* id so getModelTargetFormat("gh", ...) resolves
+  // to null and this stays on the /chat/completions path this test targets.
+  // Registered claude-* ids (e.g. "claude-sonnet-4.5") now carry
+  // targetFormat:"claude" (native /v1/messages, which doesn't need this
+  // response_format-as-system-prompt workaround — port of decolua/9router#2608,
+  // see github-copilot-claude-native-messages.test.ts) and intentionally skip it.
+  const transformed = executor.transformRequest("claude-sonnet-4", request, false, {});
 
   assert.equal(transformed.response_format, undefined);
   assert.equal(transformed.messages[0].role, "system");