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.

open-sse/executors/github.ts

@@ -129,14 +129,6 @@ export class GithubExecutor extends BaseExecutor {
       });
     }
 
-    if (!isClaudeNative && 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);
     }
@@ -155,31 +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). Skipped for the native /v1/messages
-    // path — those content parts ARE the native Claude shape and must survive intact.
-    if (!isClaudeNative && 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). Skipped for the native
-    // /v1/messages path, which — like the real Anthropic API — supports prefill.
-    // Port of 9router#2143 (author: Manuel <baslr@users.noreply.github.com>).
-    if (!isClaudeNative && 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.
@@ -192,6 +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.
@@ -260,25 +274,7 @@ export class GithubExecutor extends BaseExecutor {
     model?: string
   ): Record<string, string> {
     const token = this.getCopilotToken(credentials) || credentials.accessToken;
-
-    // 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.
-    let clientInitiator = clientHeaders?.["x-initiator"] || clientHeaders?.["X-Initiator"];
-    if (!clientInitiator && clientHeaders) {
-      for (const key in clientHeaders) {
-        if (key.toLowerCase() === "x-initiator") {
-          clientInitiator = clientHeaders[key];
-          break;
-        }
-      }
-    }
-    const initiator =
-      clientInitiator === "agent" || clientInitiator === "user" ? clientInitiator : "user";
+    const initiator = this.resolveInitiatorHeader(clientHeaders);
 
     const headers: Record<string, string> = {
       ...getGitHubCopilotChatHeaders(stream ? "text/event-stream" : "application/json", initiator),
@@ -297,6 +293,27 @@ export class GithubExecutor extends BaseExecutor {
     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) {
+        if (key.toLowerCase() === "x-initiator") {
+          clientInitiator = clientHeaders[key];
+          break;
+        }
+      }
+    }
+    return clientInitiator === "agent" || clientInitiator === "user" ? clientInitiator : "user";
+  }
+
   async refreshCopilotToken(githubAccessToken, log) {
     try {
       const response = await fetch("https://api.github.com/copilot_internal/v2/token", {

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/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");