@gajae-code/coding-agent 0.7.9 → 0.7.11

package/src/web/search/providers/gemini.ts

@@ -29,7 +29,11 @@ const ANTIGRAVITY_ENDPOINT_FALLBACKS = [ANTIGRAVITY_DAILY_ENDPOINT, ANTIGRAVITY_
 const DEFAULT_MODEL = "gemini-2.5-flash";
 const MAX_RETRIES = 3;
 const BASE_DELAY_MS = 1000;
-const RATE_LIMIT_BUDGET_MS = 5 * 60 * 1000;
+// Retry-delay ceiling for 429/5xx. The web_search chain always terminates in
+// the keyless DuckDuckGo fallback, so a rate-limited Gemini should fail
+// through in seconds instead of parking the chain behind long server-provided
+// retry delays (previously a 5-minute budget).
+const RATE_LIMIT_BUDGET_MS = 30 * 1000;
 
 const GEMINI_PROVIDERS = ["google-gemini-cli", "google-antigravity"] as const;
 type GeminiProviderId = (typeof GEMINI_PROVIDERS)[number];
@@ -230,7 +234,7 @@ async function callGeminiSearch(
 			...headers,
 		},
 		body: JSON.stringify(requestBody),
-		signal: withHardTimeout(signal),
+		signal: withHardTimeout(signal, "llm"),
 	});
 	const urlFor = (attempt: number) =>
 		`${endpoints[Math.min(attempt, endpoints.length - 1)]}/v1internal:streamGenerateContent?alt=sse`;
@@ -466,7 +470,7 @@ async function searchGeminiViaGenerativeLanguage(params: SearchParams): Promise<
 		method: "POST",
 		headers: { ...(ctx.headers ?? {}), "x-goog-api-key": apiKey, "Content-Type": "application/json" },
 		body: JSON.stringify(body),
-		signal: withHardTimeout(params.signal),
+		signal: withHardTimeout(params.signal, "llm"),
 	});
 	const text = await response.text();
 	if (!response.ok) {

package/src/web/search/providers/jina.ts

@@ -41,7 +41,7 @@ async function callJinaSearch(apiKey: string, query: string, signal?: AbortSigna
 			Accept: "application/json",
 			Authorization: `Bearer ${apiKey}`,
 		},
-		signal: withHardTimeout(signal),
+		signal: withHardTimeout(signal, "api"),
 	});
 
 	if (!response.ok) {

package/src/web/search/providers/kimi.ts

@@ -19,6 +19,11 @@ const KIMI_SEARCH_URL = "https://api.kimi.com/coding/v1/search";
 const DEFAULT_NUM_RESULTS = 10;
 const MAX_NUM_RESULTS = 20;
 const DEFAULT_TIMEOUT_SECONDS = 30;
+// Local hard ceiling aligned with the upstream `timeout_seconds` request budget
+// (plus transport headroom) instead of the generic 15s "api" class: Kimi
+// legitimately runs search + page crawling up to its advertised 30s window,
+// and aborting earlier would discard valid in-budget responses.
+const KIMI_HARD_TIMEOUT_MS = (DEFAULT_TIMEOUT_SECONDS + 5) * 1000;
 
 export interface KimiSearchParams {
 	query: string;
@@ -87,7 +92,7 @@ async function callKimiSearch(
 			enable_page_crawling: params.includeContent,
 			timeout_seconds: DEFAULT_TIMEOUT_SECONDS,
 		}),
-		signal: withHardTimeout(params.signal),
+		signal: withHardTimeout(params.signal, KIMI_HARD_TIMEOUT_MS),
 	});
 
 	if (!response.ok) {

package/src/web/search/providers/openai-compatible.ts

@@ -134,7 +134,7 @@ export class OpenAICompatibleSearchProvider extends SearchProvider {
 				method: "POST",
 				headers,
 				body: JSON.stringify(payload),
-				signal: withHardTimeout(params.signal),
+				signal: withHardTimeout(params.signal, "llm"),
 			});
 
 		// Web search is a Responses-API capability: many OpenAI-compatible

package/src/web/search/providers/parallel.ts

@@ -139,7 +139,7 @@ async function searchWithAuthStorage(
 				max_chars_per_result: 10_000,
 			},
 		}),
-		signal: withHardTimeout(params.signal),
+		signal: withHardTimeout(params.signal, "api"),
 	});
 	if (!response.ok) {
 		throw parseParallelErrorResponse(response.status, await response.text());

package/src/web/search/providers/perplexity.ts

@@ -251,7 +251,7 @@ async function callPerplexityApi(
 			"Content-Type": "application/json",
 		},
 		body: JSON.stringify(request),
-		signal: withHardTimeout(signal),
+		signal: withHardTimeout(signal, "llm"),
 	});
 
 	if (!response.ok) {
@@ -370,7 +370,7 @@ async function callPerplexityOAuth(
 				skip_search_enabled: true,
 			},
 		}),
-		signal: withHardTimeout(params.signal),
+		signal: withHardTimeout(params.signal, "llm"),
 	});
 
 	if (!response.ok) {

package/src/web/search/providers/searxng.ts

@@ -214,7 +214,7 @@ async function callSearXNGSearch(
 
 	const response = await fetch(url, {
 		headers,
-		signal: withHardTimeout(params.signal),
+		signal: withHardTimeout(params.signal, "api"),
 	});
 
 	if (!response.ok) {

package/src/web/search/providers/synthetic.ts

@@ -47,7 +47,7 @@ async function callSyntheticSearch(
 			Authorization: `Bearer ${apiKey}`,
 		},
 		body: JSON.stringify({ query }),
-		signal: withHardTimeout(signal),
+		signal: withHardTimeout(signal, "api"),
 	});
 
 	if (!response.ok) {

package/src/web/search/providers/tavily.ts

@@ -96,7 +96,7 @@ async function callTavilySearch(apiKey: string, params: TavilySearchParams): Pro
 			Authorization: `Bearer ${apiKey}`,
 		},
 		body: JSON.stringify(buildRequestBody(params)),
-		signal: withHardTimeout(params.signal),
+		signal: withHardTimeout(params.signal, "api"),
 	});
 
 	if (!response.ok) {

package/src/web/search/providers/utils.ts

@@ -44,38 +44,81 @@ export function findCredential(
 }
 
 /**
- * Default hard ceiling for a single web-search round-trip. 300s tolerates
- * legitimate slow LLM-mediated responses (anthropic web_search_20250305,
- * perplexity, gemini, OpenAI code backend) while still guaranteeing the session unfreezes
- * if Bun's `AbortSignal` fails to propagate on Windows.
- *
- * Pure search APIs (brave, exa, jina, tavily, searxng, synthetic, zai)
- * settle far faster in practice; reusing the same ceiling keeps the wiring
- * uniform without compromising correctness.
+ * Legacy hard ceiling for a single web-search round-trip, retained as the
+ * fallback for call sites that do not declare a timeout class. 300s tolerates
+ * pathological cases while still guaranteeing the session unfreezes if Bun's
+ * `AbortSignal` fails to propagate on Windows.
  */
 export const SEARCH_HARD_TIMEOUT_MS = 300_000;
 
+/**
+ * Hard ceiling for pure search APIs (brave, exa, jina, tavily, kagi,
+ * parallel, searxng, synthetic, zai, duckduckgo). These settle in ~1-3s in
+ * practice; a hung TCP/TLS connection should fall through to the next
+ * provider in seconds, not minutes. Kimi is the exception: it declares an
+ * explicit ceiling aligned with its upstream 30s request budget.
+ */
+export const SEARCH_API_TIMEOUT_MS = 15_000;
+
+/**
+ * Hard ceiling for LLM-mediated search providers (anthropic, codex, gemini,
+ * perplexity, xai, openai-compatible). These legitimately take tens of
+ * seconds because a model performs the search and synthesizes an answer.
+ */
+export const SEARCH_LLM_TIMEOUT_MS = 120_000;
+
+/** Timeout class a provider declares for its outbound round-trips. */
+export type SearchTimeoutClass = "api" | "llm";
+
+const TIMEOUT_CLASS_MS: Record<SearchTimeoutClass, number> = {
+	api: SEARCH_API_TIMEOUT_MS,
+	llm: SEARCH_LLM_TIMEOUT_MS,
+};
+
 /**
  * Runtime-configurable hard timeout, seeded from the `web_search.timeout`
- * setting via {@link setSearchHardTimeoutMs}. Falls back to
- * {@link SEARCH_HARD_TIMEOUT_MS} when unset or invalid.
+ * setting via {@link setSearchHardTimeoutMs}. When set, it overrides every
+ * class default so users keep a single knob. Unset means class defaults apply.
  */
-let configuredHardTimeoutMs = SEARCH_HARD_TIMEOUT_MS;
+let configuredHardTimeoutMs: number | undefined;
 
 /**
  * Override the hard timeout applied to every web-search round-trip.
  *
  * @param ms - Hard timeout in milliseconds. Non-finite or non-positive
- *   values reset the timeout to {@link SEARCH_HARD_TIMEOUT_MS}.
+ *   values clear the override so per-class defaults apply.
  */
 export function setSearchHardTimeoutMs(ms: number | undefined): void {
-	configuredHardTimeoutMs = typeof ms === "number" && Number.isFinite(ms) && ms > 0 ? ms : SEARCH_HARD_TIMEOUT_MS;
+	configuredHardTimeoutMs = typeof ms === "number" && Number.isFinite(ms) && ms > 0 ? ms : undefined;
+}
+
+/** Structural settings source for {@link applyConfiguredSearchTimeout}. */
+export interface SearchTimeoutSettingSource {
+	get(path: "web_search.timeout"): unknown;
+	has(path: "web_search.timeout"): boolean;
+}
+
+/**
+ * Install the `web_search.timeout` override from settings — but only when the
+ * user explicitly configured it. The settings schema ships a default value
+ * (300), and consuming that default via `get()` alone would reinstall a
+ * uniform 300s ceiling on every session, silently disabling the per-class
+ * timeout defaults. `has()` distinguishes loaded/overridden settings from
+ * schema defaults; unset clears any previous override so class defaults apply.
+ */
+export function applyConfiguredSearchTimeout(settings: SearchTimeoutSettingSource): void {
+	if (!settings.has("web_search.timeout")) {
+		setSearchHardTimeoutMs(undefined);
+		return;
+	}
+	const value = settings.get("web_search.timeout");
+	setSearchHardTimeoutMs(typeof value === "number" && Number.isFinite(value) && value > 0 ? value * 1000 : undefined);
 }
 
 /**
  * Compose a caller-supplied {@link AbortSignal} with a hard timeout so an
- * outbound `fetch()` is guaranteed to settle within `ms` even when the
- * runtime fails to propagate cancellation to the underlying transport.
+ * outbound `fetch()` is guaranteed to settle even when the runtime fails to
+ * propagate cancellation to the underlying transport.
  *
  * Bun's WinHTTP backend on Windows is known to ignore `AbortSignal` once a
  * TCP/TLS connection stalls (oven-sh/bun#15275, oven-sh/bun#18536); without
@@ -83,9 +126,16 @@ export function setSearchHardTimeoutMs(ms: number | undefined): void {
  * because the user's Esc is never delivered to the native layer.
  *
  * @param signal - Caller cancellation signal, if any.
- * @param ms - Hard timeout in milliseconds. Defaults to the configured value.
+ * @param msOrClass - Explicit timeout in milliseconds, or a
+ *   {@link SearchTimeoutClass} resolved to its class default. Omitted falls
+ *   back to the legacy {@link SEARCH_HARD_TIMEOUT_MS} ceiling. A user-set
+ *   `web_search.timeout` overrides class defaults (but not explicit ms).
  */
-export function withHardTimeout(signal: AbortSignal | undefined, ms: number = configuredHardTimeoutMs): AbortSignal {
+export function withHardTimeout(signal: AbortSignal | undefined, msOrClass?: number | SearchTimeoutClass): AbortSignal {
+	const ms =
+		typeof msOrClass === "number"
+			? msOrClass
+			: (configuredHardTimeoutMs ?? (msOrClass ? TIMEOUT_CLASS_MS[msOrClass] : SEARCH_HARD_TIMEOUT_MS));
 	const timeout = AbortSignal.timeout(ms);
 	return signal ? AbortSignal.any([signal, timeout]) : timeout;
 }

package/src/web/search/providers/xai.ts

@@ -441,7 +441,7 @@ export async function searchXai(params: XaiSearchParams): Promise<SearchResponse
 				noInlineCitations: params.no_inline_citations,
 			}),
 		),
-		signal: withHardTimeout(params.signal),
+		signal: withHardTimeout(params.signal, "llm"),
 	});
 
 	const text = await response.text();

package/src/web/search/providers/zai.ts

@@ -79,7 +79,7 @@ async function callZaiTool(apiKey: string, args: Record<string, unknown>, signal
 				arguments: args,
 			},
 		}),
-		signal: withHardTimeout(signal),
+		signal: withHardTimeout(signal, "api"),
 	});
 
 	if (!response.ok) {

package/src/web/search/render.ts

@@ -66,6 +66,7 @@ function renderFallbackText(contentText: string, expanded: boolean, theme: Theme
 export interface SearchRenderDetails {
 	response: SearchResponse;
 	error?: string;
+	warning?: string;
 }
 
 /** Render web search result with tree-based layout */
@@ -130,6 +131,8 @@ export function renderSearchResult(
 
 	const metaLines: string[] = [];
 	metaLines.push(`${theme.fg("muted", "Provider:")} ${theme.fg("text", providerLabel)}`);
+	if (details.warning)
+		metaLines.push(`${theme.fg("muted", "Fallback:")} ${theme.fg("warning", truncateToWidth(details.warning, 110))}`);
 	if (response.authMode)
 		metaLines.push(
 			`${theme.fg("muted", "Auth:")} ${theme.fg("text", response.authMode === "oauth" ? "OAuth" : response.authMode === "api_key" ? "API key" : response.authMode)}`,