fix(ai): advertised context-management beta for api-key thinking

Anthropic rejects context_management.clear_thinking_20251015 without the context-management-2025-06-27 beta header. OAuth requests carried it via claudeCodeAgentBetaDefaults; API-key requests pushed the new context_management field without the beta, so the server rejected them.

Push the beta into extraBetas alongside the field for every API-key thinking request, and exclude the GitHub Copilot proxy (which strips Anthropic betas and demotes thinking blocks upstream) from emitting the field.

Fixes #3288

packages/ai/CHANGELOG.md

@@ -4,7 +4,7 @@
 
 ### Fixed
 
-- Fixed Anthropic-compatible thinking requests sending replayed thinking blocks without `context_management.keep: "all"`, preserving multi-turn reasoning context for API-key providers. ([#3288](https://github.com/can1357/oh-my-pi/issues/3288))
+- Fixed Anthropic-compatible thinking requests sending replayed thinking blocks without `context_management.keep: "all"`, preserving multi-turn reasoning context for API-key providers. API-key requests now also advertise the required `context-management-2025-06-27` beta header so the field is honored instead of rejected. GitHub Copilot's Anthropic proxy is excluded because it strips Anthropic betas and demotes thinking blocks to text upstream. ([#3288](https://github.com/can1357/oh-my-pi/issues/3288))
 
 ## [16.1.15] - 2026-06-22
 

packages/ai/src/providers/anthropic.ts

@@ -120,10 +120,11 @@ export function buildBetaHeader(baseBetas: readonly string[], extraBetas: readon
 }
 
 const midConversationSystemBeta = "mid-conversation-system-2026-04-07";
+const contextManagementBeta = "context-management-2025-06-27";
 const claudeCodeUtilityBetaDefaults = [
 	"oauth-2025-04-20",
 	"interleaved-thinking-2025-05-14",
-	"context-management-2025-06-27",
+	contextManagementBeta,
 	"prompt-caching-scope-2026-01-05",
 	"structured-outputs-2025-12-15",
 ] as const;
@@ -131,7 +132,7 @@ const claudeCodeAgentBetaDefaults = [
 	"claude-code-20250219",
 	"oauth-2025-04-20",
 	"interleaved-thinking-2025-05-14",
-	"context-management-2025-06-27",
+	contextManagementBeta,
 	"prompt-caching-scope-2026-01-05",
 	midConversationSystemBeta,
 	"advanced-tool-use-2025-11-20",
@@ -1680,6 +1681,20 @@ const streamAnthropicOnce = (
 					// carry it in the Claude Code list).
 					extraBetas.push(midConversationSystemBeta);
 				}
+				// `context_management.clear_thinking_20251015` requires this beta. OAuth
+				// requests carry it in `claudeCodeAgentBetaDefaults`; API-key requests
+				// need it added explicitly so the field is honored instead of rejected
+				// (#3288). Skip Copilot — its proxy strips Anthropic betas and the
+				// upstream compat flag demotes thinking blocks to text, so there is
+				// nothing for `keep: "all"` to preserve.
+				if (
+					model.reasoning &&
+					options?.thinkingEnabled &&
+					model.provider !== "github-copilot" &&
+					!extraBetas.includes(contextManagementBeta)
+				) {
+					extraBetas.push(contextManagementBeta);
+				}
 
 				const created = createClient(model, {
 					model,
@@ -2950,8 +2965,13 @@ function buildParams(
 	// strip the replayed thinking blocks `replayUnsignedThinking` puts back
 	// on the wire, so the model loses the prior reasoning chain across turns
 	// and the KV cache misses every turn (#3288). Narrowing this guard back
-	// to `isOAuthToken` regresses every API-key thinking provider.
-	const shouldKeepThinkingContext = thinking?.type === "adaptive" || thinking?.type === "enabled";
+	// to `isOAuthToken` regresses every API-key thinking provider. Skip
+	// Copilot — its proxy strips Anthropic betas (so the required
+	// `context-management-2025-06-27` header never lands) and the compat
+	// flag demotes thinking blocks to text, so `keep: "all"` is a no-op
+	// that risks the proxy rejecting an unrecognized field.
+	const shouldKeepThinkingContext =
+		model.provider !== "github-copilot" && (thinking?.type === "adaptive" || thinking?.type === "enabled");
 	const contextManagement = shouldKeepThinkingContext
 		? { edits: [{ type: "clear_thinking_20251015" as const, keep: "all" as const }] }
 		: undefined;

packages/ai/test/anthropic-alignment.test.ts

@@ -406,6 +406,41 @@ describe("Anthropic request fingerprint alignment", () => {
 		expect(capturedBeta).toContain("mid-conversation-system-2026-04-07");
 	});
 
+	it("adds the context-management beta to API-key thinking requests", async () => {
+		let capturedBeta: string | undefined;
+		const fetchMock = (async (_input: string | URL | Request, init?: RequestInit) => {
+			capturedBeta = (init?.headers as Record<string, string> | undefined)?.["anthropic-beta"];
+			return new Response(
+				JSON.stringify({ type: "error", error: { type: "invalid_request_error", message: "captured" } }),
+				{ status: 400, headers: { "Content-Type": "application/json" } },
+			);
+		}) as typeof fetch;
+
+		// `context_management.clear_thinking_20251015` is rejected without
+		// the `context-management-2025-06-27` beta. OAuth requests carry it
+		// via `claudeCodeAgentBetaDefaults`; API-key requests must add it
+		// explicitly whenever thinking is enabled so the field is honored
+		// instead of dropped on the floor (#3288).
+		await streamAnthropic(
+			ANTHROPIC_MODEL,
+			{ systemPrompt: ["Stay concise."], messages: [{ role: "user", content: "Hi", timestamp: Date.now() }] },
+			{ apiKey: "sk-ant-api-test", thinkingEnabled: true, fetch: fetchMock },
+		).result();
+
+		expect(capturedBeta).toContain("context-management-2025-06-27");
+
+		capturedBeta = undefined;
+		await streamAnthropic(
+			ANTHROPIC_MODEL,
+			{ systemPrompt: ["Stay concise."], messages: [{ role: "user", content: "Hi", timestamp: Date.now() }] },
+			{ apiKey: "sk-ant-api-test", thinkingEnabled: false, fetch: fetchMock },
+		).result();
+
+		// No context_management field is sent when thinking is disabled, so the
+		// beta MUST NOT be advertised either.
+		expect(capturedBeta ?? "").not.toContain("context-management-2025-06-27");
+	});
+
 	it("billing-header fingerprint uses first user message, not leading developer message", async () => {
 		const userText = "Hello from user with enough chars padding here";
 

packages/ai/test/github-copilot-reasoning.test.ts

@@ -56,9 +56,15 @@ describe("GitHub Copilot reasoning request construction", () => {
 		const payload = (await captureAnthropicPayload(model)) as {
 			thinking?: { type?: string };
 			output_config?: { effort?: string };
+			context_management?: unknown;
 		};
 
 		expect(payload.thinking).toEqual({ type: "adaptive" });
 		expect(payload.output_config).toEqual({ effort: "high" });
+		// The Copilot Anthropic proxy strips Anthropic betas and demotes
+		// thinking blocks to text upstream — the `context_management` field
+		// would have no replayed thinking to keep and risks proxy rejection
+		// of an unrecognized field. The field MUST NOT be sent (#3288).
+		expect(payload.context_management).toBeUndefined();
 	});
 });