feat: cap Claude Code auto-compact at 196608-token self-hosted window

src/lib/configure.ts

@@ -11,7 +11,12 @@ import {
 import { homedir } from "node:os";
 import { dirname, join } from "node:path";
 import TOML from "@iarna/toml";
-import { AI_GATEWAY_OPENAI_URL, AI_GATEWAY_URL } from "@/lib/const.js";
+import {
+	AI_GATEWAY_OPENAI_URL,
+	AI_GATEWAY_URL,
+	CLAUDE_AUTO_COMPACT_WINDOW,
+	CLAUDE_AUTOCOMPACT_PCT,
+} from "@/lib/const.js";
 
 export type Tool =
 	| "claude-code"
@@ -93,6 +98,8 @@ const CLAUDE_K = {
 	sonnet: atob("QU5USFJPUElDX0RFRkFVTFRfU09OTkVUX01PREVM"),
 	haiku: atob("QU5USFJPUElDX0RFRkFVTFRfSEFJS1VfTU9ERUw="),
 	agentTeams: atob("Q0xBVURFX0NPREVfRVhQRVJJTUVOVEFMX0FHRU5UX1RFQU1T"),
+	autoCompactWindow: atob("Q0xBVURFX0NPREVfQVVUT19DT01QQUNUX1dJTkRPVw=="),
+	autoCompactPct: atob("Q0xBVURFX0FVVE9DT01QQUNUX1BDVF9PVkVSUklERQ=="),
 };
 
 const CODEX_K = {
@@ -395,6 +402,8 @@ export function configureClaudeCode(creds: Credentials): ConfigureResult[] {
 			[CLAUDE_K.sonnet]: model,
 			[CLAUDE_K.haiku]: model,
 			[CLAUDE_K.agentTeams]: "1",
+			[CLAUDE_K.autoCompactWindow]: CLAUDE_AUTO_COMPACT_WINDOW,
+			[CLAUDE_K.autoCompactPct]: CLAUDE_AUTOCOMPACT_PCT,
 		},
 	});
 

src/lib/const.ts

@@ -9,6 +9,12 @@ export const SSO_URL = `${BASE_URL}/sso-wrapper`;
 export const AI_GATEWAY_URL = `${BASE_URL}/gateway`;
 export const AI_GATEWAY_OPENAI_URL = `${AI_GATEWAY_URL}/v1`;
 
+// Self-hosted gateway model has a 196608-token window. Telling Claude Code
+// to treat that as its effective window and to fire auto-compaction at 85%
+// of it (≈167K) keeps compaction well below the hard limit.
+export const CLAUDE_AUTO_COMPACT_WINDOW = "196608";
+export const CLAUDE_AUTOCOMPACT_PCT = "85";
+
 export const VERSION: string = pkg.version;
 
 export const HELP_HINT = "Run `codev --help` to see all commands.";

tests/UpdateApp.test.tsx

@@ -51,6 +51,13 @@ function allFrames(frames: string[]): string {
 	return frames.join("\n");
 }
 
+// Windows CI is 2-3× slower than Linux/macOS and vi.waitFor's default ~1s
+// timeout isn't enough for the full UpdateApp render pipeline (detect →
+// updating → TaskList settle → parent setPhase → render Happy coding).
+// 10s matches the slack we give Ink renders elsewhere (see tests/
+// InstallApp.test.tsx#waitForFrame).
+const WAIT_OPTS = { timeout: 10_000, interval: 50 } as const;
+
 afterEach(() => {
 	cleanup();
 });
@@ -70,7 +77,10 @@ describe("UpdateApp", () => {
 			);
 
 		const { frames } = render(<UpdateApp />);
-		await vi.waitFor(() => expect(allFrames(frames)).toContain("Happy coding"));
+		await vi.waitFor(
+			() => expect(allFrames(frames)).toContain("Happy coding"),
+			WAIT_OPTS,
+		);
 
 		const history = allFrames(frames);
 		expect(history).toContain("Updated opencode-ai");
@@ -86,7 +96,10 @@ describe("UpdateApp", () => {
 		const existsSpy = vi.mocked(fs.existsSync).mockReturnValue(false);
 
 		const { frames } = render(<UpdateApp />);
-		await vi.waitFor(() => expect(allFrames(frames)).toContain("Happy coding"));
+		await vi.waitFor(
+			() => expect(allFrames(frames)).toContain("Happy coding"),
+			WAIT_OPTS,
+		);
 
 		const history = allFrames(frames);
 		expect(history).toContain("Nothing to update");
@@ -109,8 +122,9 @@ describe("UpdateApp", () => {
 			);
 
 		const { frames } = render(<UpdateApp />);
-		await vi.waitFor(() =>
-			expect(allFrames(frames)).toContain("Failed to update opencode-ai"),
+		await vi.waitFor(
+			() => expect(allFrames(frames)).toContain("Failed to update opencode-ai"),
+			WAIT_OPTS,
 		);
 
 		const history = allFrames(frames);

tests/lib/configure.test.ts

@@ -152,6 +152,8 @@ describe("configureClaudeCode", () => {
 			ANTHROPIC_DEFAULT_SONNET_MODEL: "chosen-model",
 			ANTHROPIC_DEFAULT_HAIKU_MODEL: "chosen-model",
 			CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS: "1",
+			CLAUDE_CODE_AUTO_COMPACT_WINDOW: "196608",
+			CLAUDE_AUTOCOMPACT_PCT_OVERRIDE: "85",
 		});
 	});