gsd-pi 2.82.0-dev.ed17d078d → 3.0.0-dev.1b44e695b

package/packages/pi-ai/dist/providers/simple-options.test.js

@@ -0,0 +1,50 @@
+import { describe, test } from "node:test";
+import assert from "node:assert/strict";
+import { buildBaseOptions, defaultMaxTokens } from "./simple-options.js";
+function makeModel(overrides = {}) {
+    return {
+        id: "test-model",
+        name: "Test Model",
+        api: "bedrock-converse-stream",
+        provider: "amazon-bedrock",
+        baseUrl: "",
+        reasoning: false,
+        input: ["text"],
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+        contextWindow: 16384,
+        maxTokens: 16384,
+        ...overrides,
+    };
+}
+describe("defaultMaxTokens", () => {
+    test("leaves prompt room when a non-Anthropic model declares output equal to context", () => {
+        const model = makeModel({
+            id: "qwen.qwen3-32b-v1:0",
+            contextWindow: 16384,
+            maxTokens: 16384,
+        });
+        assert.equal(defaultMaxTokens(model), 8192);
+    });
+    test("preserves smaller declared output windows", () => {
+        const model = makeModel({
+            contextWindow: 32000,
+            maxTokens: 8192,
+        });
+        assert.equal(defaultMaxTokens(model), 8192);
+    });
+    test("keeps the native Anthropic 32k ceiling within the context cap", () => {
+        const model = makeModel({
+            api: "anthropic-messages",
+            provider: "anthropic",
+            contextWindow: 200000,
+            maxTokens: 64000,
+        });
+        assert.equal(defaultMaxTokens(model), 32000);
+    });
+    test("honors explicit maxTokens", () => {
+        const model = makeModel();
+        const options = buildBaseOptions(model, { maxTokens: 12000 });
+        assert.equal(options.maxTokens, 12000);
+    });
+});
+//# sourceMappingURL=simple-options.test.js.map

package/packages/pi-ai/dist/providers/simple-options.test.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"simple-options.test.js","sourceRoot":"","sources":["../../src/providers/simple-options.test.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAC3C,OAAO,MAAM,MAAM,oBAAoB,CAAC;AAExC,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AAGzE,SAAS,SAAS,CAAC,YAAiC,EAAE;IACrD,OAAO;QACN,EAAE,EAAE,YAAY;QAChB,IAAI,EAAE,YAAY;QAClB,GAAG,EAAE,yBAAyB;QAC9B,QAAQ,EAAE,gBAAgB;QAC1B,OAAO,EAAE,EAAE;QACX,SAAS,EAAE,KAAK;QAChB,KAAK,EAAE,CAAC,MAAM,CAAC;QACf,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE;QAC1D,aAAa,EAAE,KAAK;QACpB,SAAS,EAAE,KAAK;QAChB,GAAG,SAAS;KACZ,CAAC;AACH,CAAC;AAED,QAAQ,CAAC,kBAAkB,EAAE,GAAG,EAAE;IACjC,IAAI,CAAC,gFAAgF,EAAE,GAAG,EAAE;QAC3F,MAAM,KAAK,GAAG,SAAS,CAAC;YACvB,EAAE,EAAE,qBAAqB;YACzB,aAAa,EAAE,KAAK;YACpB,SAAS,EAAE,KAAK;SAChB,CAAC,CAAC;QAEH,MAAM,CAAC,KAAK,CAAC,gBAAgB,CAAC,KAAK,CAAC,EAAE,IAAI,CAAC,CAAC;IAC7C,CAAC,CAAC,CAAC;IAEH,IAAI,CAAC,2CAA2C,EAAE,GAAG,EAAE;QACtD,MAAM,KAAK,GAAG,SAAS,CAAC;YACvB,aAAa,EAAE,KAAK;YACpB,SAAS,EAAE,IAAI;SACf,CAAC,CAAC;QAEH,MAAM,CAAC,KAAK,CAAC,gBAAgB,CAAC,KAAK,CAAC,EAAE,IAAI,CAAC,CAAC;IAC7C,CAAC,CAAC,CAAC;IAEH,IAAI,CAAC,+DAA+D,EAAE,GAAG,EAAE;QAC1E,MAAM,KAAK,GAAG,SAAS,CAAC;YACvB,GAAG,EAAE,oBAAoB;YACzB,QAAQ,EAAE,WAAW;YACrB,aAAa,EAAE,MAAM;YACrB,SAAS,EAAE,KAAK;SAChB,CAAC,CAAC;QAEH,MAAM,CAAC,KAAK,CAAC,gBAAgB,CAAC,KAAK,CAAC,EAAE,KAAK,CAAC,CAAC;IAC9C,CAAC,CAAC,CAAC;IAEH,IAAI,CAAC,2BAA2B,EAAE,GAAG,EAAE;QACtC,MAAM,KAAK,GAAG,SAAS,EAAE,CAAC;QAC1B,MAAM,OAAO,GAAG,gBAAgB,CAAC,KAAK,EAAE,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC,CAAC;QAE9D,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,SAAS,EAAE,KAAK,CAAC,CAAC;IACxC,CAAC,CAAC,CAAC;AACJ,CAAC,CAAC,CAAC","sourcesContent":["import { describe, test } from \"node:test\";\nimport assert from \"node:assert/strict\";\n\nimport { buildBaseOptions, defaultMaxTokens } from \"./simple-options.js\";\nimport type { Api, Model } from \"../types.js\";\n\nfunction makeModel(overrides: Partial<Model<Api>> = {}): Model<Api> {\n\treturn {\n\t\tid: \"test-model\",\n\t\tname: \"Test Model\",\n\t\tapi: \"bedrock-converse-stream\",\n\t\tprovider: \"amazon-bedrock\",\n\t\tbaseUrl: \"\",\n\t\treasoning: false,\n\t\tinput: [\"text\"],\n\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },\n\t\tcontextWindow: 16384,\n\t\tmaxTokens: 16384,\n\t\t...overrides,\n\t};\n}\n\ndescribe(\"defaultMaxTokens\", () => {\n\ttest(\"leaves prompt room when a non-Anthropic model declares output equal to context\", () => {\n\t\tconst model = makeModel({\n\t\t\tid: \"qwen.qwen3-32b-v1:0\",\n\t\t\tcontextWindow: 16384,\n\t\t\tmaxTokens: 16384,\n\t\t});\n\n\t\tassert.equal(defaultMaxTokens(model), 8192);\n\t});\n\n\ttest(\"preserves smaller declared output windows\", () => {\n\t\tconst model = makeModel({\n\t\t\tcontextWindow: 32000,\n\t\t\tmaxTokens: 8192,\n\t\t});\n\n\t\tassert.equal(defaultMaxTokens(model), 8192);\n\t});\n\n\ttest(\"keeps the native Anthropic 32k ceiling within the context cap\", () => {\n\t\tconst model = makeModel({\n\t\t\tapi: \"anthropic-messages\",\n\t\t\tprovider: \"anthropic\",\n\t\t\tcontextWindow: 200000,\n\t\t\tmaxTokens: 64000,\n\t\t});\n\n\t\tassert.equal(defaultMaxTokens(model), 32000);\n\t});\n\n\ttest(\"honors explicit maxTokens\", () => {\n\t\tconst model = makeModel();\n\t\tconst options = buildBaseOptions(model, { maxTokens: 12000 });\n\n\t\tassert.equal(options.maxTokens, 12000);\n\t});\n});\n"]}

package/packages/pi-ai/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gsd/pi-ai",
-  "version": "2.82.0",
+  "version": "3.0.0",
   "description": "Unified LLM API (vendored from pi-mono)",
   "type": "module",
   "gsd": {

package/packages/pi-ai/src/providers/google-gemini-cli.test.ts

@@ -0,0 +1,49 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import type { Context, Model } from "../types.js";
+import { streamGoogleGeminiCli } from "./google-gemini-cli.js";
+
+function antigravityModel(id: string): Model<"google-gemini-cli"> {
+	return {
+		id,
+		name: id,
+		api: "google-gemini-cli",
+		provider: "google-antigravity",
+		baseUrl: "https://antigravity.example.test",
+		reasoning: true,
+		input: ["text"],
+		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+		contextWindow: 200000,
+		maxTokens: 32000,
+	};
+}
+
+test("antigravity 404 names Antigravity instead of Cloud Code Assist (#4606)", async (t) => {
+	const originalFetch = globalThis.fetch;
+	const originalSetTimeout = globalThis.setTimeout;
+
+	t.after(() => {
+		globalThis.fetch = originalFetch;
+		globalThis.setTimeout = originalSetTimeout;
+	});
+
+	globalThis.fetch = async () =>
+		new Response(JSON.stringify({ error: { message: "model not found" } }), { status: 404 });
+	globalThis.setTimeout = ((callback: (...args: unknown[]) => void, _ms?: number, ...args: unknown[]) => {
+		queueMicrotask(() => callback(...args));
+		return 0;
+	}) as unknown as typeof setTimeout;
+
+	const context: Context = {
+		messages: [{ role: "user", content: "hello", timestamp: Date.now() }],
+	};
+
+	const stream = streamGoogleGeminiCli(antigravityModel("removed-antigravity-model"), context, {
+		apiKey: JSON.stringify({ token: "token", projectId: "project" }),
+	});
+	const result = await stream.result();
+
+	assert.equal(result.stopReason, "error");
+	assert.match(result.errorMessage ?? "", /Antigravity API error \(404\)/);
+	assert.doesNotMatch(result.errorMessage ?? "", /Cloud Code Assist API error/);
+});

package/packages/pi-ai/src/providers/google-gemini-cli.ts

@@ -447,6 +447,13 @@ export const streamGoogleGeminiCli: StreamFunction<"google-gemini-cli", GoogleGe
 
 					// Not retryable or max retries exceeded
 					if (response.status === 404) {
+						if (isAntigravity) {
+							throw new Error(
+								`Antigravity API error (404): Model "${model.id}" was not found. ` +
+								`This model may have been removed or renamed in the Antigravity backend. ` +
+								`Please switch to a supported model (e.g., claude-opus-4-6-thinking or gemini-3.1-pro-high).`,
+							);
+						}
 						throw new Error(
 							`Cloud Code Assist API error (404): Model "${model.id}" was not found. ` +
 							`This model may not be available via Cloud Code Assist. ` +

package/packages/pi-ai/src/providers/openai-codex-responses.test.ts

@@ -0,0 +1,63 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import type { Context, Model } from "../types.js";
+import { streamOpenAICodexResponses } from "./openai-codex-responses.js";
+
+function codexModel(): Model<"openai-codex-responses"> {
+	return {
+		id: "gpt-5.5",
+		name: "GPT-5.5",
+		api: "openai-codex-responses",
+		provider: "openai-codex",
+		baseUrl: "https://chatgpt.com/backend-api",
+		reasoning: true,
+		input: ["text"],
+		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+		contextWindow: 400000,
+		maxTokens: 128000,
+	};
+}
+
+function fakeCodexToken(): string {
+	const payload = Buffer.from(
+		JSON.stringify({ "https://api.openai.com/auth": { chatgpt_account_id: "account-123" } }),
+	).toString("base64");
+	return `header.${payload}.signature`;
+}
+
+test("openai codex responses uses Retry-After as the minimum 429 retry delay (#5677)", async (t) => {
+	const originalFetch = globalThis.fetch;
+	const originalSetTimeout = globalThis.setTimeout;
+	const delays: number[] = [];
+	let requests = 0;
+
+	t.after(() => {
+		globalThis.fetch = originalFetch;
+		globalThis.setTimeout = originalSetTimeout;
+	});
+
+	globalThis.fetch = async () => {
+		requests++;
+		return new Response(JSON.stringify({ error: { code: "rate_limit_exceeded", message: "rate limited" } }), {
+			status: 429,
+			headers: { "Retry-After": "60" },
+		});
+	};
+
+	globalThis.setTimeout = ((callback: (...args: unknown[]) => void, ms?: number, ...args: unknown[]) => {
+		delays.push(Number(ms));
+		queueMicrotask(() => callback(...args));
+		return 0;
+	}) as unknown as typeof setTimeout;
+
+	const context: Context = {
+		messages: [{ role: "user", content: "hello", timestamp: Date.now() }],
+	};
+
+	const stream = streamOpenAICodexResponses(codexModel(), context, { apiKey: fakeCodexToken() });
+	const result = await stream.result();
+
+	assert.equal(requests, 4);
+	assert.deepEqual(delays, [60000, 60000, 60000]);
+	assert.equal(result.stopReason, "error");
+});

package/packages/pi-ai/src/providers/openai-codex-responses.ts

@@ -89,6 +89,94 @@ function isRetryableError(status: number, errorText: string): boolean {
 	return /rate.?limit|overloaded|service.?unavailable|upstream.?connect|connection.?refused/i.test(errorText);
 }
 
+function extractRetryDelay(errorText: string, response?: Response | Headers): number | undefined {
+	const normalizeDelay = (ms: number): number | undefined => (ms > 0 ? Math.ceil(ms) : undefined);
+
+	const headers = response instanceof Headers ? response : response?.headers;
+	if (headers) {
+		const retryAfter = headers.get("retry-after");
+		if (retryAfter) {
+			const retryAfterSeconds = Number(retryAfter);
+			if (Number.isFinite(retryAfterSeconds)) {
+				const delay = normalizeDelay(retryAfterSeconds * 1000);
+				if (delay !== undefined) {
+					return delay;
+				}
+			}
+
+			const retryAfterDate = new Date(retryAfter).getTime();
+			if (!Number.isNaN(retryAfterDate)) {
+				const delay = normalizeDelay(retryAfterDate - Date.now());
+				if (delay !== undefined) {
+					return delay;
+				}
+			}
+		}
+
+		const rateLimitReset = headers.get("x-ratelimit-reset");
+		if (rateLimitReset) {
+			const resetSeconds = Number.parseInt(rateLimitReset, 10);
+			if (!Number.isNaN(resetSeconds)) {
+				const delay = normalizeDelay(resetSeconds * 1000 - Date.now());
+				if (delay !== undefined) {
+					return delay;
+				}
+			}
+		}
+
+		const rateLimitResetAfter = headers.get("x-ratelimit-reset-after");
+		if (rateLimitResetAfter) {
+			const resetAfterSeconds = Number(rateLimitResetAfter);
+			if (Number.isFinite(resetAfterSeconds)) {
+				const delay = normalizeDelay(resetAfterSeconds * 1000);
+				if (delay !== undefined) {
+					return delay;
+				}
+			}
+		}
+	}
+
+	const durationMatch = errorText.match(/reset after (?:(\d+)h)?(?:(\d+)m)?(\d+(?:\.\d+)?)s/i);
+	if (durationMatch) {
+		const hours = durationMatch[1] ? parseInt(durationMatch[1], 10) : 0;
+		const minutes = durationMatch[2] ? parseInt(durationMatch[2], 10) : 0;
+		const seconds = parseFloat(durationMatch[3]);
+		if (!Number.isNaN(seconds)) {
+			const totalMs = ((hours * 60 + minutes) * 60 + seconds) * 1000;
+			const delay = normalizeDelay(totalMs);
+			if (delay !== undefined) {
+				return delay;
+			}
+		}
+	}
+
+	const retryInMatch = errorText.match(/Please retry in ([0-9.]+)(ms|s)/i);
+	if (retryInMatch?.[1]) {
+		const value = parseFloat(retryInMatch[1]);
+		if (!Number.isNaN(value) && value > 0) {
+			const ms = retryInMatch[2].toLowerCase() === "ms" ? value : value * 1000;
+			const delay = normalizeDelay(ms);
+			if (delay !== undefined) {
+				return delay;
+			}
+		}
+	}
+
+	const retryDelayMatch = errorText.match(/"retryDelay":\s*"([0-9.]+)(ms|s)"/i);
+	if (retryDelayMatch?.[1]) {
+		const value = parseFloat(retryDelayMatch[1]);
+		if (!Number.isNaN(value) && value > 0) {
+			const ms = retryDelayMatch[2].toLowerCase() === "ms" ? value : value * 1000;
+			const delay = normalizeDelay(ms);
+			if (delay !== undefined) {
+				return delay;
+			}
+		}
+	}
+
+	return undefined;
+}
+
 function sleep(ms: number, signal?: AbortSignal): Promise<void> {
 	return new Promise((resolve, reject) => {
 		if (signal?.aborted) {
@@ -205,7 +293,9 @@ export const streamOpenAICodexResponses: StreamFunction<"openai-codex-responses"
 
 					const errorText = await response.text();
 					if (attempt < MAX_RETRIES && isRetryableError(response.status, errorText)) {
-						const delayMs = BASE_DELAY_MS * 2 ** attempt;
+						const backoffMs = BASE_DELAY_MS * 2 ** attempt;
+						const serverDelayMs = extractRetryDelay(errorText, response);
+						const delayMs = Math.max(backoffMs, serverDelayMs ?? 0);
 						await sleep(delayMs, options?.signal);
 						continue;
 					}

package/packages/pi-ai/src/providers/simple-options.test.ts

@@ -0,0 +1,60 @@
+import { describe, test } from "node:test";
+import assert from "node:assert/strict";
+
+import { buildBaseOptions, defaultMaxTokens } from "./simple-options.js";
+import type { Api, Model } from "../types.js";
+
+function makeModel(overrides: Partial<Model<Api>> = {}): Model<Api> {
+	return {
+		id: "test-model",
+		name: "Test Model",
+		api: "bedrock-converse-stream",
+		provider: "amazon-bedrock",
+		baseUrl: "",
+		reasoning: false,
+		input: ["text"],
+		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+		contextWindow: 16384,
+		maxTokens: 16384,
+		...overrides,
+	};
+}
+
+describe("defaultMaxTokens", () => {
+	test("leaves prompt room when a non-Anthropic model declares output equal to context", () => {
+		const model = makeModel({
+			id: "qwen.qwen3-32b-v1:0",
+			contextWindow: 16384,
+			maxTokens: 16384,
+		});
+
+		assert.equal(defaultMaxTokens(model), 8192);
+	});
+
+	test("preserves smaller declared output windows", () => {
+		const model = makeModel({
+			contextWindow: 32000,
+			maxTokens: 8192,
+		});
+
+		assert.equal(defaultMaxTokens(model), 8192);
+	});
+
+	test("keeps the native Anthropic 32k ceiling within the context cap", () => {
+		const model = makeModel({
+			api: "anthropic-messages",
+			provider: "anthropic",
+			contextWindow: 200000,
+			maxTokens: 64000,
+		});
+
+		assert.equal(defaultMaxTokens(model), 32000);
+	});
+
+	test("honors explicit maxTokens", () => {
+		const model = makeModel();
+		const options = buildBaseOptions(model, { maxTokens: 12000 });
+
+		assert.equal(options.maxTokens, 12000);
+	});
+});

package/packages/pi-ai/src/providers/simple-options.ts

@@ -4,16 +4,15 @@ import type { Api, Model, SimpleStreamOptions, StreamOptions, ThinkingBudgets, T
  * Compute the default maxTokens for a model when no explicit value is provided.
  *
  * The 32 k cap is retained only for native Anthropic models (api === "anthropic-messages")
- * where the Anthropic API historically rejected higher values. Custom and
- * Anthropic-compatible models (e.g. OpenAI-completions, Vertex, etc.) use their
- * declared model.maxTokens directly so that providers with larger output windows
- * (131 072 tokens, etc.) are not silently capped.
+ * where the Anthropic API historically rejected higher values. Defaults also
+ * leave at least half the context window available for prompts.
  */
 export function defaultMaxTokens(model: Model<Api>): number {
+	const contextCappedMaxTokens = Math.max(1, Math.floor(model.contextWindow / 2));
 	if (model.api === "anthropic-messages") {
-		return Math.min(model.maxTokens, 32000);
+		return Math.min(model.maxTokens, 32000, contextCappedMaxTokens);
 	}
-	return model.maxTokens;
+	return Math.min(model.maxTokens, contextCappedMaxTokens);
 }
 
 export function buildBaseOptions(model: Model<Api>, options?: SimpleStreamOptions, apiKey?: string): StreamOptions {