fix(providers): resolve untagged claude-* ids on GitHub/GHE Copilot to native targetFormat (#14575)

changelog.d/fixes/14575-copilot-claude-untagged-tool-choice.md

@@ -0,0 +1 @@
+- fix(providers): resolve untagged claude-* ids on GitHub/GHE Copilot to native targetFormat (#14575)

open-sse/config/providerModels.ts

@@ -238,6 +238,17 @@ export function getModelTargetFormat(aliasOrId: string, modelId: string): string
   // and OpenAI-shaped Mistral/Open-MaaS requests. Resource names retain enough publisher data to
   // route future dynamically-synced models without adding another pinned prefix here.
   if (alias === "vertex" || alias === "vp") return getVertexModelTargetFormat(bareModelId);
+  // #14575: GitHub/GHE Copilot's own executors (github.ts, ghe-copilot.ts) route ANY
+  // claude-named model to Copilot's Anthropic-native /v1/messages endpoint via an
+  // unconditional /claude/i name match, regardless of curated-catalog coverage. When
+  // Copilot ships a new Claude id before the catalog is updated (e.g. claude-opus-5.5),
+  // the curated lookup above misses and previously fell through to the provider's base
+  // "openai" format, leaving the request body OpenAI-shaped while it is dispatched to the
+  // Anthropic-native endpoint — which rejects it (tool_choice/tools shape mismatch).
+  // Mirror the routing heuristic here so body translation stays consistent with where the
+  // executor actually sends the request (same pattern as the openai "-pro" heuristic above,
+  // #5842).
+  if ((alias === "gh" || alias === "ghe-copilot") && /claude/i.test(bareModelId)) return "claude";
   // Model-level targetFormat is provider-scoped: a catalog entry declares how THIS
   // provider's endpoint serves the model — do NOT import another provider's tag.
   // #9994 scoped this for providers WITH a catalog; #10072 extends it to catalogless

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

@@ -87,21 +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.
+  // Use a non-Claude id so getModelTargetFormat("gh", ...) resolves to null and
+  // this stays on the /chat/completions path this test targets. Since #14575 EVERY
+  // claude-* id on the gh alias (registered or not) carries targetFormat:"claude"
+  // (native /v1/messages, which supports prefill — port of decolua/9router#2608,
+  // see github-copilot-claude-native-messages.test.ts) and intentionally skips
+  // this drop, so a Claude id can no longer exercise it.
   const body = {
-    model: "claude-sonnet-4",
+    model: "gpt-4.1",
     messages: [
       { role: "user", content: "Hi" },
       { role: "assistant", content: "Here is the answer:" },
     ],
   };
 
-  const out = executor.transformRequest("claude-sonnet-4", body, false, {});
+  const out = executor.transformRequest("gpt-4.1", body, false, {});
 
   assert.equal(out.messages.length, 1);
   assert.equal(out.messages[0].role, "user");

tests/unit/executor-github.test.ts

@@ -119,7 +119,7 @@ test("GithubExecutor.buildUrl routes gpt-5.6-sol/terra/luna to /responses (regre
   }
 });
 
-test("GithubExecutor.transformRequest injects JSON response instructions for Claude and strips reasoning fields", () => {
+test("GithubExecutor.transformRequest strips reasoning fields for Claude, and leaves response_format untouched now that any claude-named id is native (#14575)", () => {
   const executor = new GithubExecutor();
   const body = {
     response_format: {
@@ -136,16 +136,22 @@ test("GithubExecutor.transformRequest injects JSON response instructions for Cla
       // Trailing user turn: dropTrailingAssistantPrefill (9router#2143) strips a
       // conversation that ends in "assistant", which would otherwise remove the very
       // message this test inspects below. Keep the array ending in "user" so this test
-      // stays focused on response_format injection + reasoning-field stripping.
+      // stays focused on reasoning-field stripping.
       { role: "user", content: "thanks" },
     ],
   };
 
+  // #14575: getModelTargetFormat("gh", ...) now resolves "claude" for ANY claude-named
+  // id (registered or not) — mirroring buildUrl()'s own unconditional /claude/i routing
+  // to the Anthropic-native /v1/messages endpoint. So "claude-sonnet-4" is native now
+  // (see github-copilot-claude-native-messages.test.ts) and the response_format-as-
+  // system-prompt workaround (applyChatCompletionsOnlyQuirks, gated on !isClaudeNative)
+  // is unreachable for it — response_format passes through untouched. Reasoning-field
+  // stripping happens unconditionally above that gate, so it still applies.
   const result = executor.transformRequest("claude-sonnet-4", body, true, {});
 
-  assert.equal(result.response_format, undefined);
-  assert.equal(result.messages[0].role, "system");
-  assert.match(result.messages[0].content, /Respond only with valid JSON/);
+  assert.deepEqual(result.response_format, { type: "json_object" });
+  assert.equal(result.messages[0].role, "user");
   assert.equal(result.messages[2].reasoning_text, undefined);
   assert.equal(result.messages[2].reasoning_content, undefined);
 });
@@ -182,13 +188,16 @@ test("GithubExecutor.transformRequest sanitizes Anthropic-shape content parts (t
     ],
   };
 
-  // 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, {});
+  // Use a non-claude-named id so getModelTargetFormat("gh", ...) does NOT resolve
+  // "claude" and this stays on the /chat/completions path this test targets. #14575
+  // made getModelTargetFormat resolve "claude" for ANY claude-named id (registered
+  // or not, mirroring buildUrl()'s own unconditional /claude/i routing to the
+  // Anthropic-native /v1/messages endpoint), so a claude-named id — even an
+  // unregistered one like the former "claude-sonnet-4" here — is now native and
+  // intentionally skips this /chat/completions-only sanitization (see
+  // github-copilot-claude-native-messages.test.ts). The sanitization itself still
+  // matters for any other model whose client sends Anthropic-shape content parts.
+  const result = executor.transformRequest("gpt-4o", body, true, {});
 
   // user message keeps text + image_url parts untouched
   assert.equal(result.messages[0].content[0].type, "text");

tests/unit/github-copilot-claude-untagged-model-14575.test.ts

@@ -0,0 +1,71 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { GithubExecutor } from "../../open-sse/executors/github.ts";
+import { resolveChatCoreTargetFormat } from "../../open-sse/handlers/chatCore/targetFormat.ts";
+import { githubProvider } from "../../open-sse/config/providers/registry/github/index.ts";
+import { gheCopilotProvider } from "../../open-sse/config/providers/registry/ghe-copilot/index.ts";
+import { getModelTargetFormat } from "../../open-sse/config/providerModels.ts";
+
+const UNREGISTERED_CLAUDE_MODEL = "claude-opus-5.5";
+
+test("sanity: claude-opus-5.5 is not (yet) in the curated github registry", () => {
+  const entry = githubProvider.models.find((m) => m.id === UNREGISTERED_CLAUDE_MODEL);
+  assert.equal(entry, undefined);
+});
+
+test("BUG #14575: chatCore's resolved targetFormat disagrees with GithubExecutor.buildUrl()'s endpoint for an untagged claude-* model", () => {
+  const { targetFormat } = resolveChatCoreTargetFormat({
+    provider: "github",
+    resolvedModel: UNREGISTERED_CLAUDE_MODEL,
+    apiFormat: undefined,
+    sourceFormat: "openai",
+    customModelTargetFormat: undefined,
+    providerSpecificData: null,
+  });
+
+  const executor = new GithubExecutor();
+  const url = executor.buildUrl(UNREGISTERED_CLAUDE_MODEL, false, 0, { apiKey: "test-token" });
+  const routedToAnthropicNative = url === githubProvider.messagesUrl;
+
+  assert.equal(routedToAnthropicNative, true);
+  assert.equal(
+    targetFormat,
+    "claude",
+    "BUG #14575: chatCore resolved a non-Claude targetFormat " +
+      JSON.stringify(targetFormat) +
+      " for a model that GithubExecutor.buildUrl() routes to the Anthropic-native /v1/messages endpoint"
+  );
+});
+
+test("control: a CURATED claude-* model (claude-opus-5) has matching targetFormat + endpoint (no mismatch)", () => {
+  const { targetFormat } = resolveChatCoreTargetFormat({
+    provider: "github",
+    resolvedModel: "claude-opus-5",
+    apiFormat: undefined,
+    sourceFormat: "openai",
+    customModelTargetFormat: undefined,
+    providerSpecificData: null,
+  });
+  const executor = new GithubExecutor();
+  const url = executor.buildUrl("claude-opus-5", false, 0, { apiKey: "test-token" });
+  assert.equal(targetFormat, "claude");
+  assert.equal(url, githubProvider.messagesUrl);
+});
+
+test("BUG #14575 (ghe-copilot parity): an untagged claude-* id resolves targetFormat 'claude' for the ghe-copilot alias too", () => {
+  const entry = gheCopilotProvider.models.find((m) => m.id === UNREGISTERED_CLAUDE_MODEL);
+  assert.equal(
+    entry,
+    undefined,
+    "sanity: claude-opus-5.5 must not be in the curated ghe-copilot registry"
+  );
+
+  const targetFormat = getModelTargetFormat("ghe-copilot", UNREGISTERED_CLAUDE_MODEL);
+  assert.equal(
+    targetFormat,
+    "claude",
+    "ghe-copilot executor's own name-regex routes any claude-named model to the Anthropic-native endpoint " +
+      "(open-sse/executors/ghe-copilot.ts:117); getModelTargetFormat must resolve the same targetFormat"
+  );
+});

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

@@ -25,27 +25,25 @@ test.after(() => {
   globalThis.fetch = originalFetch;
 });
 
-test("T27: Claude + response_format=json_object injects system instruction and strips response_format field", () => {
+test("T27: an untagged claude-* id also skips the response_format-as-system-prompt workaround (post-#14575)", () => {
   const executor = new GithubExecutor();
   const request = {
     messages: [{ role: "user", content: "return json" }],
     response_format: { type: "json_object" },
   };
 
-  // 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.
+  // Prior to #14575, an unregistered claude-* id made getModelTargetFormat("gh", ...)
+  // return null, so this request stayed on the /chat/completions path and hit the
+  // response_format-as-system-prompt workaround below. #14575 gave getModelTargetFormat
+  // a claude-name heuristic for the gh/ghe-copilot aliases (mirroring buildUrl()'s own
+  // unconditional /claude/i routing to the Anthropic-native /v1/messages endpoint), so
+  // ANY claude-named id — registered or not — now resolves targetFormat "claude" and
+  // is treated as native, same as "claude-sonnet-4.5" in
+  // github-copilot-claude-native-messages.test.ts. The workaround below is therefore
+  // unreachable for the gh alias and response_format passes through untouched.
   const transformed = executor.transformRequest("claude-sonnet-4", request, false, {});
 
-  assert.equal(transformed.response_format, undefined);
-  assert.equal(transformed.messages[0].role, "system");
-  assert.match(
-    transformed.messages[0].content,
-    /Respond only with valid JSON\. Do not include any text/i
-  );
+  assert.deepEqual(transformed.response_format, { type: "json_object" });
 });
 
 test("T27: non-Claude models keep response_format untouched", () => {