From 5291b2f5a07f81e64a8312f00c7c7e2f04a1e02a Mon Sep 17 00:00:00 2001
From: roboomp <omp@can.ac>
Date: Tue, 23 Jun 2026 01:42:53 +0000
Subject: [PATCH] fix(ai): skipped context management for injected clients

Injected Anthropic clients bypass buildAnthropicClientOptions, so this package cannot add the context-management beta header their SDK instance would need before accepting context_management.clear_thinking_20251015.

Omit context_management for options.client requests while preserving thinking itself, and add regression coverage for injected-client payload shaping.

Fixes #3288
---
 packages/ai/CHANGELOG.md                      |  2 +-
 packages/ai/src/providers/anthropic.ts        | 13 ++++----
 .../ai/test/anthropic-stream-envelope.test.ts | 30 +++++++++++++++++++
 3 files changed, 39 insertions(+), 6 deletions(-)

diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index f167fe2d3..bd038ba1a 100644
--- a/packages/ai/CHANGELOG.md
+++ b/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. 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))
+- 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. Injected SDK clients and GitHub Copilot's Anthropic proxy are excluded because this code path cannot add the beta to caller-owned clients, while Copilot 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
 
diff --git a/packages/ai/src/providers/anthropic.ts b/packages/ai/src/providers/anthropic.ts
index 7543d0594..8300316a2 100644
--- a/packages/ai/src/providers/anthropic.ts
+++ b/packages/ai/src/providers/anthropic.ts
@@ -2966,12 +2966,15 @@ function buildParams(
 	// 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. 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.
+	// injected clients because this code cannot add the required
+	// `context-management-2025-06-27` beta to caller-owned SDK clients. Skip
+	// Copilot because its proxy strips Anthropic betas and demotes thinking
+	// blocks to text upstream, so `keep: "all"` is a no-op that risks proxy
+	// rejection of an unrecognized field.
 	const shouldKeepThinkingContext =
-		model.provider !== "github-copilot" && (thinking?.type === "adaptive" || thinking?.type === "enabled");
+		!options?.client &&
+		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;
diff --git a/packages/ai/test/anthropic-stream-envelope.test.ts b/packages/ai/test/anthropic-stream-envelope.test.ts
index f73748009..27cee8e5c 100644
--- a/packages/ai/test/anthropic-stream-envelope.test.ts
+++ b/packages/ai/test/anthropic-stream-envelope.test.ts
@@ -576,6 +576,36 @@ describe("anthropic stream envelope handling", () => {
 		expect(capturedParams?.tools?.map(tool => tool.name)).toEqual(["web_search"]);
 		expect(capturedOptions?.headers).toEqual({ "X-Umans-Websearch-Provider": "exa" });
 	});
+
+	it("does not send context_management through injected clients", async () => {
+		type CapturedPayload = {
+			thinking?: { type?: string };
+			context_management?: unknown;
+		};
+		let capturedParams: CapturedPayload | undefined;
+		const client: AnthropicMessagesClientLike = {
+			messages: {
+				create(params) {
+					capturedParams = params as CapturedPayload;
+					return createMockRequest(createTextSuccessEvents("done"));
+				},
+			},
+		};
+
+		const stream = streamAnthropic(model, context, {
+			client,
+			thinkingEnabled: true,
+		});
+		const events: AssistantMessageEvent[] = [];
+		for await (const event of stream) {
+			events.push(event);
+		}
+		const result = await stream.result();
+
+		expect(result.content).toEqual([{ type: "text", text: "done" }]);
+		expect(capturedParams?.thinking?.type).toBe("enabled");
+		expect(capturedParams?.context_management).toBeUndefined();
+	});
 	it("unwraps thinking blocks that Anthropic streams with literal thinking tags", async () => {
 		const wrappedThinking =
 			"<thinking>\n<thinking>\nCheck logs before accepting container health.\n</thinking></thinking>";