@code-yeongyu/senpi-ai 2026.10.8 → 2026.10.10-10

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: [],
@@ -972,7 +981,7 @@ export const stream = (model, context, options) => {
                 const cacheRetention = resolveCacheRetention(options?.cacheRetention ?? model.cacheRetention, options?.env, "short");
                 const cacheSessionId = cacheRetention === "none" ? undefined : options?.sessionId;
                 openClient = () => {
-                    const created = createClient(model, apiKey, options?.interleavedThinking ?? true, shouldUseFineGrainedToolStreamingBeta(model, normalizedContext), options?.refusalFallbacks !== undefined, optionsHeaders, options?.fetch, copilotDynamicHeaders, cacheSessionId, options?.env, federation);
+                    const created = createClient(model, apiKey, options?.interleavedThinking ?? true, shouldUseFineGrainedToolStreamingBeta(model, normalizedContext), options?.refusalFallbacks !== undefined && configuredBetaHeader(model, options) !== null, optionsHeaders, options?.fetch, copilotDynamicHeaders, cacheSessionId, options?.env, federation);
                     client = created.client;
                     isOAuth = created.isOAuthToken;
                     advertisedClaudeCodeVersion = created.claudeCodeVersion;
@@ -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) {
@@ -1665,27 +1695,46 @@ export function buildAnthropicWarmPromptCacheParams(model, context, options) {
         max_tokens: 0,
     });
 }
-function getBetaFeatures(model, context, isOAuthToken, nativeToolChanges, options) {
-    let configuredFeatures;
+function configuredBetaHeader(model, options) {
+    let configured;
     for (const headers of [model.headers, options?.headers]) {
         for (const [name, value] of Object.entries(headers ?? {})) {
             if (name.toLowerCase() === "anthropic-beta")
-                configuredFeatures = value;
+                configured = value;
         }
     }
-    if (configuredFeatures === null)
+    return configured;
+}
+function getBetaFeatures(model, context, isOAuthToken, shape, options) {
+    const configured = configuredBetaHeader(model, options);
+    if (configured === null)
         return [];
-    if (configuredFeatures !== undefined) {
-        return [
-            ...new Set(configuredFeatures
-                .split(",")
-                .map((feature) => feature.trim())
-                .filter(Boolean)),
-        ];
-    }
-    const features = [];
+    // Betas the request body depends on: the API rejects the request without them.
+    const required = [];
     if (isOAuthToken)
-        features.push("claude-code-20250219", "oauth-2025-04-20");
+        required.push("claude-code-20250219", "oauth-2025-04-20");
+    if (shape.midConvoEffort)
+        required.push(MID_CONVERSATION_OUTPUT_CONFIG_BETA);
+    if (shape.thinkingBinding)
+        required.push(THINKING_BINDING_CONTROLS_BETA);
+    if (shape.nativeToolChanges)
+        required.push(MID_CONVERSATION_TOOL_CHANGES_BETA);
+    if (shape.refusalFallbacks)
+        required.push(SERVER_SIDE_FALLBACK_BETA);
+    // A configured header owns the optional betas and is merged with the required ones, never replacing them: a
+    // proxy route's own header used to drop the effort beta, and Claude 5.5 models then 400ed (senpi#2957).
+    // Interleaved thinking is stripped on adaptive models here too, as it is from the client headers.
+    const optional = configured === undefined
+        ? defaultOptionalBetas(model, context, options)
+        : configured
+            .split(",")
+            .map((feature) => feature.trim())
+            .filter(Boolean)
+            .filter((feature) => !(feature === INTERLEAVED_THINKING_BETA && supportsAdaptiveThinking(model)));
+    return [...new Set([...optional, ...required])];
+}
+function defaultOptionalBetas(model, context, options) {
+    const features = [];
     if (shouldUseFineGrainedToolStreamingBeta(model, context))
         features.push(FINE_GRAINED_TOOL_STREAMING_BETA);
     if (model.reasoning &&
@@ -1696,12 +1745,7 @@ function getBetaFeatures(model, context, isOAuthToken, nativeToolChanges, option
     }
     if (shouldUseServerSideFallbackBeta(model))
         features.push(SERVER_SIDE_FALLBACK_BETA);
-    if (model.compat?.supportsMidConvoEffort === true) {
-        features.push(MID_CONVERSATION_OUTPUT_CONFIG_BETA, THINKING_BINDING_CONTROLS_BETA);
-    }
-    if (nativeToolChanges)
-        features.push(MID_CONVERSATION_TOOL_CHANGES_BETA);
-    return [...new Set(features)];
+    return features;
 }
 function buildParams(model, context, isOAuthToken, options, unsignedThinkingReplay = getAnthropicCompat(model).unsignedThinkingReplay) {
     const compat = getAnthropicCompat(model);
@@ -1718,7 +1762,16 @@ function buildParams(model, context, isOAuthToken, options, unsignedThinkingRepl
     // and Anthropic rejects a tool list where every tool is deferred, so there must be an
     // initial active tool to anchor the deferred ones. Otherwise the current tool list is sent.
     const initialTools = initialSystemMessage?.toolsAdded ?? [];
-    const nativeToolChanges = supportsMidConvoSystemMessages(model) &&
+    // A configured `anthropic-beta: null` asks for no betas at all, so the request takes a shape that needs none:
+    // the current tool list instead of native tool changes, top-level effort instead of per-message markers, and
+    // no server-side fallbacks (senpi#2957). OAuth still needs its identity betas; getBetaFeatures names them.
+    const betasSuppressed = configuredBetaHeader(model, options) === null;
+    if (betasSuppressed && isOAuthToken) {
+        throw new Error(`The anthropic-beta header is set to null, but ${model.provider}/${model.id} is called with an OAuth token, which needs the claude-code-20250219 and oauth-2025-04-20 betas. Remove the null, or list the betas you want instead; senpi adds the ones a request needs.`);
+    }
+    const midConvoEffort = model.compat?.supportsMidConvoEffort === true && !betasSuppressed;
+    const nativeToolChanges = !betasSuppressed &&
+        supportsMidConvoSystemMessages(model) &&
         supportsMidConvoToolChanges(model) &&
         initialTools.length > 0 &&
         !hasToolRedefinitions(context.messages);
@@ -1741,15 +1794,21 @@ function buildParams(model, context, isOAuthToken, options, unsignedThinkingRepl
         deferredTools = [];
     }
     const deferredToolNames = new Set(deferredTools.map((tool) => normalizeToolName(tool.name)));
-    const activeEffort = options?.effort ?? "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
+    const activeEffort = managedEffortForRequest(model, options) ?? "high";
+    const sendFallbacks = options?.refusalFallbacks !== undefined && !betasSuppressed;
+    const betaFeatures = getBetaFeatures(model, context, isOAuthToken, {
+        midConvoEffort,
+        thinkingBinding: midConvoEffort && options?.thinkingEnabled !== false,
+        nativeToolChanges,
+        refusalFallbacks: sendFallbacks,
+    }, options);
+    const convertedMessages = convertMessages(conversationMessages, model, isOAuthToken, cacheControl, unsignedThinkingReplay, deferredToolNames, normalizeToolName, midConvoEffort ? model.provider : undefined, nativeToolChanges);
+    const messages = midConvoEffort
         ? insertThinkingLevelMessages(convertedMessages, activeEffort)
         : convertedMessages.messages;
     // Effort markers are appended after history; re-run the fork's final-user cache checkpoint
     // against the actual last user/tool-result message so the marker cannot displace cache_control.
-    if (model.compat?.supportsMidConvoEffort === true && cacheControl) {
+    if (midConvoEffort && cacheControl) {
         for (let index = messages.length - 1; index >= 0; index--) {
             const message = messages[index];
             // A native system update after the last turn already holds the checkpoint;
@@ -1826,9 +1885,10 @@ 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`.
-    if (model.compat?.supportsMidConvoEffort === true && options?.thinkingEnabled !== false) {
+    // 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 (midConvoEffort && options?.thinkingEnabled !== false) {
         params.thinking = {
             type: "adaptive",
             display: options?.thinkingDisplay ?? "summarized",
@@ -1897,7 +1957,7 @@ function buildParams(model, context, isOAuthToken, options, unsignedThinkingRepl
         }
     }
     applyExtraBodyToAnthropicParams(params, options?.extraBody);
-    if (options?.refusalFallbacks !== undefined) {
+    if (sendFallbacks && options?.refusalFallbacks !== undefined) {
         // AnthropicRefusalFallback models a readonly list; the SDK's BetaFallbacksParam
         // is mutable, so copy rather than widen the fork's public type.
         params.fallbacks = options.refusalFallbacks === "default" ? "default" : [...options.refusalFallbacks];
@@ -2070,7 +2130,6 @@ function convertMessages(transformedMessages, model, isOAuthToken, cacheControl,
             }
         }
         else if (msg.role === "assistant") {
-            flushPendingSystemMessages();
             const blocks = [];
             const isSameModel = isSameAnthropicModel(msg, model);
             // Blocks before the final fallback marker are the declined attempt; the
@@ -2156,8 +2215,12 @@ function convertMessages(transformedMessages, model, isOAuthToken, cacheControl,
                     }
                 }
             }
+            // Pending system messages are emitted only in front of an assistant param that is actually sent: a reply
+            // that converts to no blocks is skipped, and flushing for it would leave a `system` param directly before
+            // the next `user` turn, which Anthropic rejects on every later request (senpi#2864).
             if (blocks.length === 0)
                 continue;
+            flushPendingSystemMessages();
             const messageIndex = params.length;
             params.push({
                 role: "assistant",

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/api/openai-codex-responses.js

@@ -450,9 +450,9 @@ function buildRequestBody(model, context, options, cacheSessionId, grammarToolIn
 function getServiceTierCostMultiplier(model, serviceTier) {
     switch (serviceTier) {
         case "ultrafast":
-            // OpenAI publishes an Ultrafast price for GPT-6 Astra only: 6x Standard on every
+            // OpenAI prices GPT-6 Astra and GPT-6.1 Sol Ultrafast at 6x Standard on every
             // token class and context tier. Any other model keeps its base rate.
-            return model.id === "gpt-6-astra" ? 6 : 1;
+            return ["gpt-6-astra", "gpt-6.1-sol"].includes(model.upstreamModelId ?? model.id) ? 6 : 1;
         case "flex":
             return 0.5;
         case "priority":

package/dist/api/openai-completions.js

@@ -14,7 +14,7 @@ import { getProviderEnvValue } from "../utils/provider-env.js";
 import { retryProviderStreamRequest } from "../utils/provider-retry.js";
 import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
 import { getSystemMessageText, renderSystemMessageUpdate } from "../utils/text.js";
-import { sendWithForcedToolChoiceFallback } from "../utils/tool-choice-fallback.js";
+import { primeStreamUntilContent, sendWithForcedToolChoiceFallback } from "../utils/tool-choice-fallback.js";
 import { normalizeToolParametersForMoonshot, normalizeToolParametersForOpenAICompat, } from "../utils/tool-schema-compat.js";
 import { getCurrentTools, getDeclaredTools, resolveTranscript, resolveTranscriptTools, } from "../utils/transcript.js";
 import { appendGrammarToolInputJsonDelta, createGrammarToolInputProperties, getGrammarToolInput, getJsonSchemaToolParameters, resolveGrammarConstrainedSampling, resolveJsonSchemaStrictSampling, } from "./constrained-sampling.js";
@@ -169,6 +169,23 @@ function getOpenRouterRawMetadata(error) {
 function isForcedOpenAICompletionsToolChoice(toolChoice) {
     return toolChoice !== undefined && toolChoice !== "auto" && toolChoice !== "none";
 }
+/** A chunk that carries model output or a finish: text, reasoning, a tool call, or a finish_reason. */
+function chatChunkHasContent(chunk) {
+    if (!isRecord(chunk) || !Array.isArray(chunk.choices))
+        return false;
+    return chunk.choices.some((choice) => {
+        if (!isRecord(choice))
+            return false;
+        if (typeof choice.finish_reason === "string" && choice.finish_reason.length > 0)
+            return true;
+        const delta = choice.delta;
+        if (!isRecord(delta))
+            return false;
+        if (Array.isArray(delta.tool_calls) && delta.tool_calls.length > 0)
+            return true;
+        return ["content", "reasoning_content", "reasoning", "reasoning_text"].some((field) => typeof delta[field] === "string" && delta[field].length > 0);
+    });
+}
 function isReasoningDetailObject(detail) {
     return typeof detail === "object" && detail !== null && !Array.isArray(detail);
 }
@@ -362,7 +379,14 @@ export const stream = (model, context, options) => {
                     params,
                     acceptsForcedToolChoice: compat.supportsForcedToolChoice !== false,
                     isForced: isForcedOpenAICompletionsToolChoice,
-                    send: createStream,
+                    // A keepalive gateway answers 200 and refuses a forced choice in-band (senpi#2801): a forced
+                    // request is read up to its first content so that refusal can still be retried.
+                    send: async (sentParams) => {
+                        const created = await createStream(sentParams);
+                        if (!isForcedOpenAICompletionsToolChoice(sentParams.tool_choice))
+                            return created;
+                        return { ...created, data: await primeStreamUntilContent(created.data, chatChunkHasContent) };
+                    },
                 });
                 params = sent.params;
                 return sent.result;

package/dist/api/openai-responses.js

@@ -861,9 +861,9 @@ function buildWebSocketHeaders(model, context, apiKey, optionsHeaders, sessionId
 function getServiceTierCostMultiplier(model, serviceTier) {
     switch (serviceTier) {
         case "ultrafast":
-            // OpenAI publishes an Ultrafast price for GPT-6 Astra only: 6x Standard on every
+            // OpenAI prices GPT-6 Astra and GPT-6.1 Sol Ultrafast at 6x Standard on every
             // token class and context tier. Any other model keeps its base rate.
-            return model.id === "gpt-6-astra" ? 6 : 1;
+            return ["gpt-6-astra", "gpt-6.1-sol"].includes(model.upstreamModelId ?? model.id) ? 6 : 1;
         case "flex":
             return 0.5;
         case "priority":

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();
 }