@bastani/pi-ai 0.9.20-alpha.5 → 0.9.20-alpha.7

package/dist/api/anthropic-messages.js

@@ -43,13 +43,13 @@ function getCacheControl(model, cacheRetention, env) {
 // Stealth mode: Mimic Claude Code's tool naming exactly
 //
 // Anthropic gates newer models on the `claude-cli/<version>` user agent alone and rejects an
-// older one with `claude_code_version_too_old`. `claude-fable-5-1` requires >= 2.1.251, bisected
-// against the live API: 2.1.250 -> 400, 2.1.251 -> 200. This is pinned to that exact published
-// minimum rather than the newest release, and it is a strict superset of the previous 2.1.75 --
-// every model this provider ships answers 200 at 2.1.251. Raise it only when a model rejects
-// this value; a caller can override it for one client through the `headers` option, whose
-// lowercase `user-agent` key is merged last.
-const claudeCodeVersion = "2.1.251";
+// older one with `claude_code_version_too_old`. `claude-fable-5-1` requires >= 2.1.251 (bisected
+// against the live API: 2.1.250 -> 400, 2.1.251 -> 200) and Claude Opus 5.5 requires >= 2.1.280,
+// the floor the API names in its rejection. This is pinned to the highest published minimum
+// among shipped models rather than the newest release; every model this provider ships answers
+// 200 at 2.1.280. Raise it only when a model rejects this value; a caller can override it for
+// one client through the `headers` option, whose lowercase `user-agent` key is merged last.
+const claudeCodeVersion = "2.1.280";
 // 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
@@ -129,6 +129,61 @@ const SERVER_SIDE_FALLBACK_BETA = "server-side-fallback-2026-07-01";
 const THINKING_BINDING_CONTROLS_BETA = "thinking-binding-controls-2026-08-01";
 const MID_CONVERSATION_OUTPUT_CONFIG_BETA = "mid-conversation-output-config-2026-07-01";
 const MID_CONVERSATION_TOOL_CHANGES_BETA = "mid-conversation-tool-changes-2026-07-01";
+const FAST_MODE_BETA = "fast-mode-2026-02-01";
+/**
+ * Fast mode bills every token category at twice the standard rate for each supported model, across
+ * the full context window, and prompt-caching multipliers stack on top.
+ * https://platform.claude.com/docs/en/build-with-claude/fast-mode#pricing
+ */
+const FAST_MODE_PRICE_MULTIPLIER = 2;
+function requestedSpeed(model) {
+    return model.fastRoute?.speed;
+}
+function upstreamModelId(model) {
+    return model.fastRoute?.upstreamModelId ?? model.id;
+}
+function reportedSpeed(usage) {
+    const speed = usage?.speed;
+    return speed === "fast" || speed === "standard" ? speed : undefined;
+}
+function applyFastModePricing(usage) {
+    usage.cost.input *= FAST_MODE_PRICE_MULTIPLIER;
+    usage.cost.output *= FAST_MODE_PRICE_MULTIPLIER;
+    usage.cost.cacheRead *= FAST_MODE_PRICE_MULTIPLIER;
+    usage.cost.cacheWrite *= FAST_MODE_PRICE_MULTIPLIER;
+    usage.cost.total = usage.cost.input + usage.cost.output + usage.cost.cacheRead + usage.cost.cacheWrite;
+}
+/**
+ * Reject a payload hook that rewrote a field the model's fast route owns.
+ *
+ * A fast variant is recorded, persisted, and billed under its own `-fast` identity, so a hook that
+ * changes the upstream model or drops `speed: "fast"` would send a different request under that
+ * identity. Refusing loudly surfaces the misconfigured hook instead of silently restoring the route.
+ */
+function assertPayloadPreservesFastRoute(model, payload) {
+    const fastRoute = model.fastRoute;
+    if (fastRoute?.speed === undefined)
+        return;
+    const remedy = `A fast model variant must send its route's upstream model and speed, because it is recorded, persisted, ` +
+        `and billed under its own identity. Select the normal model "${model.provider}/${fastRoute.baseModelId}" ` +
+        `if the request needs different routing`;
+    if (typeof payload !== "object" || payload === null || Array.isArray(payload)) {
+        throw new Error(`A request hook changed payload for fast model "${model.provider}/${model.id}" to a value that cannot carry ` +
+            `the route's upstream model and speed. ${remedy}, or return an object that preserves those fields.`);
+    }
+    const record = payload;
+    const conflicts = [];
+    if (record.model !== fastRoute.upstreamModelId) {
+        conflicts.push(`model (expected "${fastRoute.upstreamModelId}", got ${JSON.stringify(record.model)})`);
+    }
+    if (record.speed !== fastRoute.speed) {
+        conflicts.push(`speed (expected ${JSON.stringify(fastRoute.speed)}, got ${JSON.stringify(record.speed)})`);
+    }
+    if (conflicts.length === 0)
+        return;
+    throw new Error(`A request hook changed ${conflicts.join(" and ")} for fast model "${model.provider}/${model.id}". ` +
+        `${remedy}, or stop rewriting those fields.`);
+}
 /**
  * Stable deferred tool declared whenever native tool changes are in use. Anthropic adds
  * hidden prompt scaffolding as soon as any tool has `defer_loading`; declaring this
@@ -220,7 +275,7 @@ function addEarlierAttemptCosts(output, model, servingModel, event) {
         if (!iteration.output_tokens)
             continue;
         // Price at the rates of the model that ran the attempt, which is not the serving model.
-        const attemptCost = iteration.model === model.id
+        const attemptCost = iteration.model === upstreamModelId(model)
             ? model.cost
             : iteration.model === servingModel.id
                 ? servingModel.cost
@@ -485,6 +540,7 @@ export const stream = (model, context, options) => {
             let client;
             let isOAuth;
             let usageModel = model;
+            let servedSpeed = requestedSpeed(model);
             if (options?.client) {
                 client = options.client;
                 isOAuth = false;
@@ -515,6 +571,7 @@ export const stream = (model, context, options) => {
             if (nextParams !== undefined) {
                 params = nextParams;
             }
+            assertPayloadPreservesFastRoute(model, params);
             const requestOptions = {
                 ...(streamDeadline.signal ? { signal: streamDeadline.signal } : {}),
                 ...(options?.timeoutMs !== undefined ? { timeout: options.timeoutMs } : {}),
@@ -532,9 +589,9 @@ export const stream = (model, context, options) => {
                 if (event.type === "message_start") {
                     output.responseId = event.message.id;
                     const responseModel = event.message.model;
-                    if (responseModel !== model.id)
+                    if (responseModel !== upstreamModelId(model))
                         output.responseModel = responseModel;
-                    const fallbackCost = responseModel === model.id
+                    const fallbackCost = responseModel === upstreamModelId(model)
                         ? undefined
                         : model.compat?.allowedFallbackModels?.find((fallback) => fallback.provider === model.provider && fallback.model === responseModel)?.cost;
                     usageModel = fallbackCost ? { ...model, id: responseModel, cost: fallbackCost } : model;
@@ -549,6 +606,9 @@ export const stream = (model, context, options) => {
                     output.usage.totalTokens =
                         output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
                     calculateCost(usageModel, output.usage);
+                    servedSpeed = reportedSpeed(event.message.usage) ?? servedSpeed;
+                    if (servedSpeed === "fast")
+                        applyFastModePricing(output.usage);
                     inputTransformations = getInputTransformations(event.message) ?? inputTransformations;
                 }
                 else if (event.type === "content_block_start") {
@@ -744,6 +804,9 @@ export const stream = (model, context, options) => {
                     output.usage.totalTokens =
                         output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
                     calculateCost(usageModel, output.usage);
+                    servedSpeed = reportedSpeed(event.usage) ?? servedSpeed;
+                    if (servedSpeed === "fast")
+                        applyFastModePricing(output.usage);
                     // A non-empty final fallback report supersedes the request-start report with
                     // the serving model's transformations. An empty report must not erase drops
                     // already reported at message_start.
@@ -868,6 +931,8 @@ function createClient(model, apiKey, interleavedThinking, useFineGrainedToolStre
     }
     if (useNativeToolChanges)
         betaFeatures.push(MID_CONVERSATION_TOOL_CHANGES_BETA);
+    if (requestedSpeed(model) === "fast")
+        betaFeatures.push(FAST_MODE_BETA);
     const uniqueBetaFeatures = [...new Set(betaFeatures)];
     // Copilot: Bearer auth, selective betas.
     if (model.provider === "github-copilot") {
@@ -959,6 +1024,8 @@ function getNativeToolChangeBetas(model, context, isOAuth, options) {
         features.push(THINKING_BINDING_CONTROLS_BETA);
     if (supportsMidConvoEffort(model))
         features.push(MID_CONVERSATION_OUTPUT_CONFIG_BETA, THINKING_BINDING_CONTROLS_BETA);
+    if (requestedSpeed(model) === "fast")
+        features.push(FAST_MODE_BETA);
     return [...new Set(features)];
 }
 function buildParams(model, context, isOAuthToken, options) {
@@ -979,7 +1046,7 @@ function buildParams(model, context, isOAuthToken, options) {
     const converted = convertMessages(conversationMessages, isOAuthToken, cacheControl, compat.allowEmptySignature, supportsMidConvoEffort(model) ? model.provider : undefined, nativeToolChanges);
     const activeEffort = options?.effort ?? "high";
     const params = {
-        model: model.id,
+        model: upstreamModelId(model),
         messages: (supportsMidConvoEffort(model)
             ? insertThinkingLevelMessages(converted, activeEffort)
             : converted.messages),
@@ -988,6 +1055,9 @@ function buildParams(model, context, isOAuthToken, options) {
     };
     if (nativeToolChanges)
         params.betas = getNativeToolChangeBetas(model, context, isOAuthToken, options);
+    const speed = requestedSpeed(model);
+    if (speed !== undefined)
+        params.speed = speed;
     // For OAuth tokens, we MUST include Claude Code identity
     if (isOAuthToken) {
         params.system = [