@code-yeongyu/senpi-ai 2026.10.10-6 → 2026.10.10-8

package/dist/api/anthropic-messages.js

@@ -71,7 +71,7 @@ function getCacheControl(model, cacheRetention, env) {
 // The bundled Claude Code version and the floor of the advertised `claude-cli/<version>`
 // (see utils/claude-code-version.ts). Keep this exact declaration: a downstream installer
 // (oh-my-openagent) rewrites it byte-for-byte in the installed dist and bundle.
-const claudeCodeVersion = "2.1.289";
+const claudeCodeVersion = "2.1.292";
 // Claude Code 2.x tool names (canonical casing)
 // Source: https://cchistory.mariozechner.at/data/prompts-2.1.11.md
 // To update: https://github.com/badlogic/cchistory
@@ -164,6 +164,7 @@ const ADAPTIVE_THINKING_MODEL_MARKERS = [
     "opus-5",
     "sonnet-4-6",
     "sonnet-5",
+    "haiku-5-5",
     "fable-5",
     "mythos-5",
 ];
@@ -176,14 +177,20 @@ const NATIVE_XHIGH_EFFORT_MODEL_MARKERS = [
     "opus-4-8",
     "opus-5",
     "sonnet-5",
+    "haiku-5-5",
     "fable-5",
     "mythos-5",
 ];
 /**
- * Adaptive families that reject `thinking: {type: "disabled"}` outright (verified 400:
- * `"thinking.type.disabled" is not supported for this model`). The generated catalog also encodes
- * this as `compat.supportsDisabledThinking: false`, but `models.json` entries and third-party
- * gateway rows carry no generated compat, so the family fact has to live here as well.
+ * Families senpi never sends `thinking: {type: "disabled"}` to. Fable 5, Opus 5.5 and Sonnet 5.5 reject it
+ * outright (verified live 400: `"thinking.type.disabled" is not supported for this model`); Mythos 5 is
+ * listed by family with Fable. Claude Haiku 5.5 is listed by choice, not because it rejects `disabled`: its
+ * effort docs accept `disabled` at effort `high` or below (400 only at `xhigh` / `max`) and reject a
+ * differing per-message effort while disabled, so a real thinking-off needs marker handling first
+ * (senpi#2927). Until then a thinking-off turn sends no `thinking` and effort `low`, which is valid on all of
+ * them. The generated catalog also encodes this as `compat.supportsDisabledThinking: false`, but
+ * `models.json` entries and third-party gateway rows carry no generated compat, so the family fact has to
+ * live here as well.
  */
 const DISABLED_THINKING_REJECTING_MODEL_MARKERS = [
     "fable-5",
@@ -192,6 +199,8 @@ const DISABLED_THINKING_REJECTING_MODEL_MARKERS = [
     "opus-5.5",
     "sonnet-5-5",
     "sonnet-5.5",
+    "haiku-5-5",
+    "haiku-5.5",
 ];
 const UNSUPPORTED_NATIVE_COMPUTER_TOOL_MODEL_MARKERS = [
     "opus-4-6",
@@ -922,7 +931,7 @@ export const stream = (model, context, options) => {
     const normalizedContext = resolveTranscript(context, supportsMidConvoSystemMessages(model));
     const currentTools = getCurrentTools(normalizedContext.messages);
     (async () => {
-        const providerThinkingLevel = model.compat?.supportsMidConvoEffort ? (options?.effort ?? "high") : undefined;
+        const providerThinkingLevel = managedEffortForRequest(model, options);
         const output = {
             role: "assistant",
             content: [],
@@ -1459,6 +1468,25 @@ function cannotDisableThinking(model, compat) {
         return true;
     return matchesModelMarker(model, DISABLED_THINKING_REJECTING_MODEL_MARKERS);
 }
+/**
+ * Effort a per-message-effort request runs at, which is also the active marker and the level recorded
+ * on the reply: the caller's effort with thinking on, `low` for a thinking-off turn on a family that
+ * cannot disable thinking, and none when thinking is actually disabled (senpi#2912).
+ */
+function managedEffortForRequest(model, options) {
+    if (model.compat?.supportsMidConvoEffort !== true)
+        return undefined;
+    if (options?.thinkingEnabled === false) {
+        return cannotDisableThinking(model, getAnthropicCompat(model)) ? "low" : undefined;
+    }
+    return options?.effort ?? "high";
+}
+function isEffortMarker(message) {
+    return (message.role === "system" &&
+        Array.isArray(message.content) &&
+        message.content.length === 0 &&
+        message.output_config?.effort !== undefined);
+}
 function disableThinkingForRequest(params, model, compat) {
     // A degraded/disabled turn must not retain the caller's higher effort.
     delete params.output_config;
@@ -1470,6 +1498,8 @@ function disableThinkingForRequest(params, model, compat) {
         return;
     }
     params.thinking = { type: "disabled" };
+    // Per-message effort is rejected alongside disabled thinking (senpi#1399).
+    params.messages = params.messages.filter((message) => !isEffortMarker(message));
 }
 function supportsAdaptiveThinking(model) {
     if (model.compat?.forceAdaptiveThinking !== undefined) {
@@ -1741,7 +1771,7 @@ function buildParams(model, context, isOAuthToken, options, unsignedThinkingRepl
         deferredTools = [];
     }
     const deferredToolNames = new Set(deferredTools.map((tool) => normalizeToolName(tool.name)));
-    const activeEffort = options?.effort ?? "high";
+    const activeEffort = managedEffortForRequest(model, options) ?? "high";
     const betaFeatures = getBetaFeatures(model, context, isOAuthToken, nativeToolChanges, options);
     const convertedMessages = convertMessages(conversationMessages, model, isOAuthToken, cacheControl, unsignedThinkingReplay, deferredToolNames, normalizeToolName, model.compat?.supportsMidConvoEffort === true ? model.provider : undefined, nativeToolChanges);
     const messages = model.compat?.supportsMidConvoEffort === true
@@ -1826,8 +1856,9 @@ function buildParams(model, context, isOAuthToken, options, unsignedThinkingRepl
         ];
     }
     // Managed effort models always use adaptive thinking so prefix mismatches can
-    // be dropped instead of surfacing as persistent 400 responses. Thinking-off is
-    // handled before this managed branch because these models accept `disabled`.
+    // be dropped instead of surfacing as persistent 400 responses. A thinking-off turn
+    // skips this branch: `disableThinkingForRequest` sends `disabled` where the family
+    // accepts it and pins effort `low` where it does not (Opus/Sonnet/Haiku 5.5, Fable, Mythos).
     if (model.compat?.supportsMidConvoEffort === true && options?.thinkingEnabled !== false) {
         params.thinking = {
             type: "adaptive",

package/dist/api/anthropic-tool-references.js

@@ -73,6 +73,27 @@ export function demoteUnavailableToolReferences(params) {
     }
     let changed = false;
     const availableToolNames = [...available];
+    // A native search pair whose every reference stopped resolving is demoted as a
+    // unit: the result decides, and its `server_tool_use` follows. The pair can span
+    // two assistant messages (a deferred server tool resumes in the continuation), so
+    // the decision is collected over the whole request before any message is rewritten.
+    const droppedSearchUseIds = new Set();
+    const droppedSearchNames = new Map();
+    for (const message of messages) {
+        if (message.role !== "assistant" || !Array.isArray(message.content))
+            continue;
+        for (const block of message.content) {
+            if (!isNativeToolSearchResultBlock(block))
+                continue;
+            const names = block.content.tool_references
+                .filter((item) => isRecord(item) && item.type === "tool_reference")
+                .map((item) => (typeof item.tool_name === "string" ? item.tool_name : ""));
+            if (names.length > 0 && names.every((name) => resolve(name) === undefined)) {
+                droppedSearchUseIds.add(block.tool_use_id);
+                droppedSearchNames.set(block.tool_use_id, names);
+            }
+        }
+    }
     const seenDemotedCallNames = new Set();
     const rewrittenMessages = [];
     for (const message of messages) {
@@ -81,23 +102,6 @@ export function demoteUnavailableToolReferences(params) {
             continue;
         }
         let messageChanged = false;
-        // A native search pair whose every reference stopped resolving is demoted
-        // as a unit: the result decides, and its `server_tool_use` follows.
-        const droppedSearchUseIds = new Set();
-        const droppedSearchNames = new Map();
-        if (message.role === "assistant") {
-            for (const block of message.content) {
-                if (!isNativeToolSearchResultBlock(block))
-                    continue;
-                const names = block.content.tool_references
-                    .filter((item) => isRecord(item) && item.type === "tool_reference")
-                    .map((item) => (typeof item.tool_name === "string" ? item.tool_name : ""));
-                if (names.length > 0 && names.every((name) => resolve(name) === undefined)) {
-                    droppedSearchUseIds.add(block.tool_use_id);
-                    droppedSearchNames.set(block.tool_use_id, names);
-                }
-            }
-        }
         const content = [];
         for (const block of message.content) {
             if (message.role === "assistant" && isRecord(block) && block.type === "tool_use") {
@@ -168,7 +172,9 @@ export function demoteUnavailableToolReferences(params) {
             }
             content.push(block);
         }
-        if (content.length === 0) {
+        // Only a message this pass emptied is dropped; a message that arrived empty (a per-message
+        // effort marker carries `content: []` by design) passes through untouched (senpi#2912).
+        if (messageChanged && content.length === 0) {
             changed = true;
             continue;
         }

package/dist/api/bedrock-converse-stream.js

@@ -590,6 +590,7 @@ function supportsAdaptiveThinking(modelId, modelName) {
         s.includes("opus-5") ||
         s.includes("sonnet-4-6") ||
         s.includes("sonnet-5") ||
+        s.includes("haiku-5-5") ||
         s.includes("fable-5") ||
         s.includes("mythos-5"));
 }
@@ -599,6 +600,7 @@ function supportsNativeXhighEffort(model) {
         s.includes("opus-4-8") ||
         s.includes("opus-5") ||
         s.includes("sonnet-5") ||
+        s.includes("haiku-5-5") ||
         s.includes("fable-5") ||
         s.includes("mythos-5"));
 }
@@ -616,7 +618,9 @@ function rejectsDisabledThinking(model) {
         s.includes("opus-5-5") ||
         s.includes("opus-5.5") ||
         s.includes("sonnet-5-5") ||
-        s.includes("sonnet-5.5"));
+        s.includes("sonnet-5.5") ||
+        s.includes("haiku-5-5") ||
+        s.includes("haiku-5.5"));
 }
 function mapThinkingLevelToEffort(model, level) {
     if (level === "xhigh" && supportsNativeXhighEffort(model))

package/dist/api/lazy.js

@@ -1,5 +1,7 @@
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
+import { isOAuthRefreshUnavailableError } from "../utils/oauth-refresh-error.js";
 import { describeReplacedInstall } from "../utils/provider-failure-description.js";
+import { OAUTH_REFRESH_UNAVAILABLE_DIAGNOSTIC } from "../utils/retry.js";
 function createSetupErrorMessage(model, error) {
     return {
         role: "assistant",
@@ -17,6 +19,17 @@ function createSetupErrorMessage(model, error) {
         },
         stopReason: "error",
         errorMessage: describeReplacedInstall(error) ?? (error instanceof Error ? error.message : String(error)),
+        ...(isOAuthRefreshUnavailableError(error)
+            ? {
+                diagnostics: [
+                    {
+                        type: OAUTH_REFRESH_UNAVAILABLE_DIAGNOSTIC,
+                        timestamp: Date.now(),
+                        details: { provider: model.provider },
+                    },
+                ],
+            }
+            : {}),
         timestamp: Date.now(),
     };
 }

package/dist/api/mistral-conversations.js

@@ -612,6 +612,21 @@ function stripSymbolKeys(value) {
     }
     return value;
 }
+/**
+ * Consecutive user turns become one user message, as the OpenAI Chat converter does for
+ * non-OpenAI endpoints: a word message rebuilt after tool results (senpi#2920) is often followed
+ * by a steer or the next prompt, and Mistral's role-order validation is not documented to accept
+ * two user messages in a row.
+ */
+function appendUserMessage(result, content) {
+    const previous = result[result.length - 1];
+    if (previous?.role !== "user") {
+        result.push({ role: "user", content });
+        return;
+    }
+    const chunks = (value) => typeof value === "string" ? [{ type: "text", text: value }] : (value ?? []);
+    previous.content = [...chunks(previous.content), ...chunks(content)];
+}
 function toChatMessages(messages, supportsImages) {
     const result = [];
     for (const [index, msg] of messages.entries()) {
@@ -625,7 +640,7 @@ function toChatMessages(messages, supportsImages) {
         }
         if (msg.role === "user") {
             if (typeof msg.content === "string") {
-                result.push({ role: "user", content: sanitizeSurrogates(msg.content) });
+                appendUserMessage(result, sanitizeSurrogates(msg.content));
                 continue;
             }
             const hadImages = msg.content.some((item) => item.type === "image");
@@ -637,11 +652,11 @@ function toChatMessages(messages, supportsImages) {
                 return { type: "image_url", imageUrl: `data:${item.mimeType};base64,${item.data}` };
             });
             if (content.length > 0) {
-                result.push({ role: "user", content });
+                appendUserMessage(result, content);
                 continue;
             }
             if (hadImages && !supportsImages) {
-                result.push({ role: "user", content: "(image omitted: model does not support images)" });
+                appendUserMessage(result, "(image omitted: model does not support images)");
             }
             continue;
         }

package/dist/auth/oauth/anthropic.js

@@ -4,6 +4,7 @@
  * NOTE: This module uses Node.js http.createServer for the OAuth callback server.
  * It is only intended for CLI use, not browser environments.
  */
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 import { getProviderEnvValue } from "../../utils/provider-env.js";
 import { PREFERRED_CALLBACK_PORT, startCallbackListener, } from "./anthropic-callback-listener.js";
 import { parseAuthorizationInput } from "./authorization-input.js";
@@ -57,7 +58,7 @@ async function postJson(url, body, signal) {
     });
     const responseBody = await response.text();
     if (!response.ok) {
-        throw new Error(`HTTP request failed. status=${response.status}; url=${url}; body=${responseBody}`);
+        throw new OAuthTokenEndpointError(`HTTP request failed. status=${response.status}; url=${url}; body=${responseBody}`, response.status);
     }
     return responseBody;
 }
@@ -74,7 +75,7 @@ async function exchangeAuthorizationCode(code, state, verifier, redirectUri, sig
         }, signal);
     }
     catch (error) {
-        throw new Error(`Token exchange request failed. url=${TOKEN_URL}; redirect_uri=${redirectUri}; response_type=authorization_code; details=${formatErrorDetails(error)}`);
+        throw new Error(`Token exchange request failed. url=${TOKEN_URL}; redirect_uri=${redirectUri}; response_type=authorization_code; details=${formatErrorDetails(error)}`, { cause: error });
     }
     let tokenData;
     try {
@@ -240,7 +241,7 @@ async function refreshAnthropicToken(refreshToken, signal) {
         }, signal);
     }
     catch (error) {
-        throw new Error(`Anthropic token refresh request failed. url=${TOKEN_URL}; details=${formatErrorDetails(error)}`);
+        throw new Error(`Anthropic token refresh request failed. url=${TOKEN_URL}; details=${formatErrorDetails(error)}`, { cause: error });
     }
     let data;
     try {

package/dist/auth/oauth/chatgpt-subscription.js

@@ -12,6 +12,7 @@ if (typeof process !== "undefined" && (process.versions?.node || process.version
     });
 }
 import { extractChatGptSubscriptionAccountId } from "../../utils/chatgpt-subscription-auth.js";
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 import { getProviderEnvValue } from "../../utils/provider-env.js";
 import { getWireIdentity } from "../../wire-identity.js";
 import { startOAuthCallbackServer, waitForCallbackOrManualInput } from "./callback-server.js";
@@ -80,7 +81,7 @@ async function fetchWithLoginCancellation(input, init) {
 async function readTokenResponse(response, operation) {
     if (!response.ok) {
         const text = await response.text().catch(() => "");
-        throw new Error(`ChatGPT Subscription token ${operation} failed (${response.status}): ${text || response.statusText}`);
+        throw new OAuthTokenEndpointError(`ChatGPT Subscription token ${operation} failed (${response.status}): ${text || response.statusText}`, response.status);
     }
     const rawJson = await response.json();
     const json = rawJson;
@@ -123,7 +124,7 @@ async function refreshAccessToken(refreshToken, signal) {
         });
     }
     catch (error) {
-        throw new Error(`ChatGPT Subscription token refresh error: ${error instanceof Error ? error.message : String(error)}`);
+        throw new Error(`ChatGPT Subscription token refresh error: ${error instanceof Error ? error.message : String(error)}`, { cause: error });
     }
     return readTokenResponse(response, "refresh");
 }

package/dist/auth/oauth/cursor.js

@@ -10,6 +10,7 @@
  * (or a dashboard user API key) at `auth/exchange_user_api_key` for a fresh
  * session JWT.
  */
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 import { generatePKCE } from "./pkce.js";
 const CURSOR_LOGIN_URL = "https://cursor.com/loginDeepControl";
 const CURSOR_POLL_URL = "https://api2.cursor.sh/auth/poll";
@@ -205,12 +206,12 @@ async function refreshCursorCredential(credential, signal) {
     }
     catch (error) {
         if (signal.aborted)
-            throw new Error("Cursor token refresh cancelled");
+            throw new Error("Cursor token refresh cancelled", { cause: signal.reason });
         throw error;
     }
     const body = await readJsonBody(response);
     if (!response.ok) {
-        throw new Error(`Cursor token refresh failed (${describeFailure(response.status, body)})`);
+        throw new OAuthTokenEndpointError(`Cursor token refresh failed (${describeFailure(response.status, body)})`, response.status);
     }
     const access = requiredString(body ?? {}, "accessToken");
     const rotated = body?.refreshToken;

package/dist/auth/oauth/devin-token.js

@@ -7,6 +7,7 @@
  * "token" that serves as both the access and the refresh credential. Expiry is
  * read from the JWT itself because the response carries no expires_in.
  */
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 export const DEVIN_TOKEN_URL = "https://api.devin.ai/auth/cli/token";
 export const DEVIN_API_ENDPOINT = "https://api.devin.ai";
 /** Devin issues long-lived CLI tokens; used when the JWT carries no usable exp. */
@@ -91,7 +92,7 @@ export async function exchangeDevinAuthorizationCode(code, verifier, signal) {
     }
     if (!response.ok) {
         const detail = errorDetail(body);
-        throw new Error(`Devin OAuth token exchange failed (HTTP ${response.status})${detail ? `: ${detail}` : ""}`);
+        throw new OAuthTokenEndpointError(`Devin OAuth token exchange failed (HTTP ${response.status})${detail ? `: ${detail}` : ""}`, response.status);
     }
     if (typeof body.token !== "string" || body.token.length === 0) {
         throw new Error('Devin OAuth response carries no "token"');

package/dist/auth/oauth/github-copilot.js

@@ -3,6 +3,7 @@
  */
 import { GITHUB_COPILOT_INDIVIDUAL_BASE_URL, githubCopilotTokenField, parseGitHubCopilotApiEndpoint, resolveGitHubCopilotBaseUrl, } from "../../api/github-copilot-endpoint.js";
 import { GITHUB_COPILOT_REJECTED_TOKEN_STATUSES } from "../../api/github-copilot-headers.js";
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 import { sleep } from "../../utils/sleep.js";
 import { pollOAuthDeviceCodeFlow } from "./device-code.js";
 import { parseGitHubCopilotModelCatalog } from "./github-copilot-model-catalog.js";
@@ -79,7 +80,7 @@ async function fetchGitHubCopilotModels(copilotToken, enterpriseDomain, signal,
         },
     }, signal, retryPolicy);
     if (!response.ok) {
-        throw new Error(`${response.status} ${response.statusText}: ${await response.text()}`);
+        throw new OAuthTokenEndpointError(`${response.status} ${response.statusText}: ${await response.text()}`, response.status);
     }
     return parseGitHubCopilotModelCatalog(await response.json(), allowPolicyFallback);
 }
@@ -87,7 +88,7 @@ async function fetchJson(url, init) {
     const response = await fetch(url, init);
     if (!response.ok) {
         const text = await response.text();
-        throw new Error(`${response.status} ${response.statusText}: ${text}`);
+        throw new OAuthTokenEndpointError(`${response.status} ${response.statusText}: ${text}`, response.status);
     }
     return response.json();
 }

package/dist/auth/oauth/kimi-coding.js

@@ -6,6 +6,7 @@
  * access token authenticates requests to that region's coding API as an
  * `Authorization: Bearer` header; see kimi-region.ts for the host table.
  */
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 import { sleep } from "../../utils/sleep.js";
 import { pollOAuthDeviceCodeFlow } from "./device-code.js";
 import { kimiCodeIdentityHeaders } from "./kimi-identity.js";
@@ -178,7 +179,7 @@ async function refreshToken(oauthHost, refreshTokenValue, signal) {
             await sleep(1000 * 2 ** (attempt - 1), signal);
         }
         if (signal.aborted) {
-            throw new Error("Kimi Code token refresh aborted");
+            throw new Error("Kimi Code token refresh aborted", { cause: signal.reason });
         }
         let response;
         try {
@@ -208,14 +209,14 @@ async function refreshToken(oauthHost, refreshTokenValue, signal) {
         // Unauthorized: the stored credential is dead; Models clears it and prompts re-login.
         if (response.status === 401 || response.status === 403 || json?.error === "invalid_grant") {
             const description = typeof json?.error_description === "string" ? `: ${json.error_description}` : "";
-            throw new Error(`Kimi Code token refresh unauthorized (status ${response.status})${description}`);
+            throw new OAuthTokenEndpointError(`Kimi Code token refresh unauthorized (status ${response.status})${description}`, response.status);
         }
         if (isRetryableRefreshFailure(response) && attempt < REFRESH_MAX_RETRIES) {
-            lastError = new Error(`Kimi Code token refresh failed with status ${response.status}`);
+            lastError = new OAuthTokenEndpointError(`Kimi Code token refresh failed with status ${response.status}`, response.status);
             continue;
         }
         const text = JSON.stringify(json);
-        throw new Error(`Kimi Code token refresh failed with status ${response.status}${text ? `: ${text}` : ""}`);
+        throw new OAuthTokenEndpointError(`Kimi Code token refresh failed with status ${response.status}${text ? `: ${text}` : ""}`, response.status);
     }
     throw lastError ?? new Error("Kimi Code token refresh failed");
 }

package/dist/auth/oauth/openai-chatgpt.js

@@ -7,6 +7,7 @@
 import { randomBytes } from "node:crypto";
 import { createServer } from "node:http";
 import { oauthErrorHtml, oauthSuccessHtml } from "../../utils/oauth-page.js";
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 import { getProviderEnvValue } from "../../utils/provider-env.js";
 import { generatePKCE } from "./pkce.js";
 // every login registers a new client with this ID; OpenAI returns the issued client ID in the callback
@@ -119,7 +120,7 @@ async function requestToken(body, signal) {
     });
     if (!response.ok) {
         const responseBody = await response.text().catch(() => "");
-        throw new Error(`OpenAI OAuth token request failed (${response.status}): ${responseBody || response.statusText}`);
+        throw new OAuthTokenEndpointError(`OpenAI OAuth token request failed (${response.status}): ${responseBody || response.statusText}`, response.status);
     }
     const data = await response.json();
     if (typeof data !== "object" || data === null || Array.isArray(data)) {

package/dist/auth/oauth/openrouter.js

@@ -10,6 +10,7 @@
  * NOTE: This module uses node:http (via callback-server.ts) for the OAuth callback server.
  * It is only intended for CLI use, not browser environments.
  */
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 import { getProviderEnvValue } from "../../utils/provider-env.js";
 import { startOAuthCallbackServer, waitForCallbackOrManualInput } from "./callback-server.js";
 import { generatePKCE } from "./pkce.js";
@@ -88,7 +89,7 @@ async function exchangeAuthorizationCode(code, verifier, signal) {
     }
     if (!response.ok) {
         const detail = errorDetail(body);
-        throw new Error(`OpenRouter OAuth key exchange failed (HTTP ${response.status})${detail ? `: ${detail}` : ""}`);
+        throw new OAuthTokenEndpointError(`OpenRouter OAuth key exchange failed (HTTP ${response.status})${detail ? `: ${detail}` : ""}`, response.status);
     }
     if (typeof body.key !== "string" || body.key.length === 0) {
         throw new Error('OpenRouter OAuth response carries no "key"');

package/dist/auth/oauth/radius.js

@@ -9,6 +9,7 @@
  * It is only intended for CLI use, not browser environments.
  */
 import { normalizeRadiusGatewayUrl } from "../../providers/radius-config.js";
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 import { startOAuthCallbackServer } from "./callback-server.js";
 import { pollOAuthDeviceCodeFlow } from "./device-code.js";
 import { generatePKCE } from "./pkce.js";
@@ -36,15 +37,14 @@ async function loadRadiusOAuthDiscovery(gateway, signal) {
     }
     return { authorizationEndpoint: discovery.authorizationEndpoint };
 }
-class OAuthResponseError extends Error {
+class OAuthResponseError extends OAuthTokenEndpointError {
     constructor(status, oauthError, description, message) {
         const detail = oauthError
             ? description
                 ? `${oauthError}: ${description}`
                 : oauthError
             : description || String(status);
-        super(`${message}: ${detail}`);
-        this.status = status;
+        super(`${message}: ${detail}`, status);
         this.oauthError = oauthError;
     }
 }
@@ -76,7 +76,7 @@ async function requestOAuthToken(gateway, body, signal) {
     }
     catch (error) {
         if (signal.aborted) {
-            throw new Error("Login cancelled");
+            throw new Error("Login cancelled", { cause: signal.reason });
         }
         throw error;
     }
@@ -149,7 +149,7 @@ async function requestDeviceAuthorization(gateway, signal) {
     }
     catch (error) {
         if (signal.aborted) {
-            throw new Error("Login cancelled");
+            throw new Error("Login cancelled", { cause: signal.reason });
         }
         throw error;
     }

package/dist/auth/oauth/xai.js

@@ -1,6 +1,7 @@
 /**
  * xAI OAuth device-code flow.
  */
+import { OAuthTokenEndpointError } from "../../utils/oauth-refresh-error.js";
 import { pollOAuthDeviceCodeFlow } from "./device-code.js";
 const XAI_CLIENT_ID = "b1a00492-073a-47ea-816f-4c329264a828";
 const XAI_SCOPE = "openid profile email offline_access grok-cli:access api:access";
@@ -53,7 +54,7 @@ async function postForm(url, fields, signal) {
     }
     catch (error) {
         if (signal.aborted) {
-            throw new Error("Login cancelled");
+            throw new Error("Login cancelled", { cause: signal.reason });
         }
         throw error;
     }
@@ -64,9 +65,11 @@ async function postForm(url, fields, signal) {
     }
     catch {
         if (signal.aborted) {
-            throw new Error("Login cancelled");
+            throw new Error("Login cancelled", { cause: signal.reason });
         }
-        throw new Error(`xAI OAuth returned invalid JSON (HTTP ${response.status})`);
+        const message = `xAI OAuth returned invalid JSON (HTTP ${response.status})`;
+        // An error page (an HTML 503 from a proxy, say) still carries the status the retry decision needs.
+        throw response.ok ? new Error(message) : new OAuthTokenEndpointError(message, response.status);
     }
     return {
         ok: response.ok,
@@ -78,7 +81,7 @@ function requestFailure(action, response) {
     const error = typeof response.body.error === "string" ? response.body.error : undefined;
     const description = typeof response.body.error_description === "string" ? response.body.error_description : undefined;
     const detail = [error, description].filter(Boolean).join(": ");
-    return new Error(`xAI OAuth ${action} failed (HTTP ${response.status})${detail ? `: ${detail}` : ""}`);
+    return new OAuthTokenEndpointError(`xAI OAuth ${action} failed (HTTP ${response.status})${detail ? `: ${detail}` : ""}`, response.status);
 }
 function parseDeviceCode(body) {
     // RFC 8628 allows interval 0 (no minimum wait); fall back to the poller's

package/dist/auth/oauth-refresh.js

@@ -15,6 +15,7 @@
  * superseded catalog refresh stops waiting but never aborts the exchange.
  */
 import { raceWithAbortSignal } from "../utils/abort.js";
+import { oauthRefreshFailureCause } from "../utils/oauth-refresh-error.js";
 import { mergeRefreshed, mergeRefreshedSlot, projectSlot } from "./pool/slots.js";
 export const DEFAULT_OAUTH_REFRESH_TIMEOUT_MS = 15_000;
 /** The provider rejected or timed out the token exchange. */
@@ -103,6 +104,15 @@ async function exchangeAndStore(request, signal) {
         signal.throwIfAborted();
     }
     catch (error) {
+        signal.throwIfAborted();
+        const causeClass = oauthRefreshFailureCause(error);
+        if (causeClass !== undefined) {
+            console.warn("OAuth refresh unavailable", JSON.stringify({
+                provider: providerId,
+                ...(slotName ? { slot: slotName } : {}),
+                cause: causeClass,
+            }));
+        }
         throw new OAuthRefreshExchangeError(error);
     }
     try {

package/dist/auth/resolve.js

@@ -1,5 +1,6 @@
 import { operationSignal, raceWithAbortSignal } from "../utils/abort.js";
 import { ModelsError } from "../utils/models-error.js";
+import { classifyOAuthRefreshFailure, OAuthRefreshUnavailableError } from "../utils/oauth-refresh-error.js";
 import { OAuthRefreshExchangeError, OAuthRefreshStoreError, projectOAuthSlot, refreshOAuthCredential, } from "./oauth-refresh.js";
 import { projectSlot } from "./pool/slots.js";
 export { ModelsError } from "../utils/models-error.js";
@@ -82,6 +83,9 @@ export function oauthRefreshModelsError(error, providerId) {
     if (error instanceof ModelsError)
         return error;
     if (error instanceof OAuthRefreshExchangeError) {
+        if (classifyOAuthRefreshFailure(error.cause) === "transient") {
+            return new OAuthRefreshUnavailableError(providerId, error.cause);
+        }
         return new ModelsError("oauth", `OAuth refresh failed for ${providerId}`, { cause: error.cause });
     }
     const cause = error instanceof OAuthRefreshStoreError ? error.cause : error;
@@ -112,6 +116,7 @@ async function resolveStoredOAuth(credentials, providerId, oauth, stored, authCo
             });
         }
         catch (error) {
+            signal.throwIfAborted();
             throw oauthRefreshModelsError(error, providerId);
         }
         if (post?.type !== "oauth")

package/dist/providers/all.d.ts

@@ -50,7 +50,7 @@ declare const FORK_OWNED_CATALOGS: {
                     readonly input: 3;
                     readonly output: 15;
                     readonly cacheRead: 0.3;
-                    readonly cacheWrite: 0;
+                    readonly cacheWrite: 3;
                 };
                 readonly contextWindow: 1048576;
                 readonly maxTokens: 131072;