@pellux/goodvibes-sdk 0.34.0 → 0.34.1

package/dist/platform/providers/anthropic-sse-assembler.js

@@ -0,0 +1,146 @@
+import { mapAnthropicStopReason } from './stop-reason-maps.js';
+import { fromAnthropicContent, parseToolCallArguments } from './tool-formats.js';
+import { logger } from '../utils/logger.js';
+import { SseLineBuffer } from './sse-line-buffer.js';
+/** Create a fresh zero-valued SSE accumulator. */
+export function createAnthropicSSEState() {
+    return {
+        responseText: '',
+        inputTokens: 0,
+        outputTokens: 0,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0,
+        rawStopReason: undefined,
+        stopReason: 'unknown',
+        toolBlocks: new Map(),
+    };
+}
+/**
+ * Dispatch one Anthropic SSE event into the accumulator.
+ * Mutates state in place and calls onDelta for real-time delivery.
+ */
+export function processAnthropicSSEEvent(event, state, onDelta) {
+    if (event.type === 'content_block_start' && event.content_block?.type === 'tool_use') {
+        const idx = event.index ?? 0;
+        state.toolBlocks.set(idx, {
+            id: event.content_block.id ?? '',
+            name: event.content_block.name ?? '',
+            args: '',
+        });
+        if (onDelta) {
+            onDelta({ toolCalls: [{ index: idx, id: event.content_block.id, name: event.content_block.name }] });
+        }
+    }
+    else if (event.type === 'content_block_delta') {
+        const idx = event.index ?? 0;
+        if (event.delta?.type === 'text_delta' && event.delta.text) {
+            state.responseText += event.delta.text;
+            if (onDelta)
+                onDelta({ content: event.delta.text });
+        }
+        else if (event.delta?.type === 'thinking_delta' && event.delta.thinking) {
+            if (onDelta)
+                onDelta({ reasoning: event.delta.thinking });
+        }
+        else if (event.delta?.type === 'input_json_delta' && event.delta.partial_json) {
+            const block = state.toolBlocks.get(idx);
+            if (block)
+                block.args += event.delta.partial_json;
+            if (onDelta) {
+                onDelta({ toolCalls: [{ index: idx, arguments: event.delta.partial_json }] });
+            }
+        }
+    }
+    else if (event.type === 'message_delta') {
+        if (event.delta?.stop_reason) {
+            state.rawStopReason = event.delta.stop_reason;
+            state.stopReason = mapAnthropicStopReason(state.rawStopReason);
+        }
+        if (event.usage?.output_tokens)
+            state.outputTokens = event.usage.output_tokens;
+        if (event.usage?.cache_read_input_tokens != null)
+            state.cacheReadTokens = event.usage.cache_read_input_tokens;
+        if (event.usage?.cache_creation_input_tokens != null)
+            state.cacheWriteTokens = event.usage.cache_creation_input_tokens;
+    }
+    else if (event.type === 'message_start') {
+        if (event.message?.usage) {
+            state.inputTokens = event.message.usage.input_tokens;
+            state.outputTokens = event.message.usage.output_tokens;
+            state.cacheReadTokens = event.message.usage.cache_read_input_tokens ?? 0;
+            state.cacheWriteTokens = event.message.usage.cache_creation_input_tokens ?? 0;
+        }
+    }
+}
+/**
+ * Read an Anthropic SSE stream to completion, dispatching each event through
+ * processAnthropicSSEEvent. Uses SseLineBuffer for correct CRLF handling.
+ *
+ * @param reader       ReadableStream reader from the fetch response body.
+ * @param state        Accumulator to mutate (created with createAnthropicSSEState).
+ * @param onDelta      Optional streaming delta callback.
+ * @param providerLabel Used in warn logs, e.g. 'Anthropic' or 'AnthropicCompat(my-proxy)'.
+ */
+export async function readAnthropicSSEStream(reader, state, onDelta, providerLabel) {
+    const sseBuffer = new SseLineBuffer();
+    try {
+        while (true) {
+            const { done, value } = await reader.read();
+            if (done)
+                break;
+            for (const line of sseBuffer.feed(value)) {
+                if (!line.startsWith('data: '))
+                    continue;
+                const data = line.slice(6).trim();
+                if (!data || data === '[DONE]')
+                    continue;
+                let event;
+                try {
+                    event = JSON.parse(data);
+                }
+                catch {
+                    logger.warn(`${providerLabel} SSE: failed to parse JSON chunk`, {
+                        chunkPreview: data.slice(0, 200),
+                        chunkLength: data.length,
+                    });
+                    continue;
+                }
+                processAnthropicSSEEvent(event, state, onDelta);
+            }
+        }
+    }
+    finally {
+        reader.releaseLock();
+    }
+}
+/**
+ * Assemble the final content + toolCalls from accumulated streaming state.
+ * Sorts tool blocks by stream index, parses JSON arguments, and calls
+ * fromAnthropicContent to produce the canonical response shape.
+ *
+ * @param fallbackBlocks Used by the SDK provider: if no streaming content was
+ *   accumulated, fall back to finalMessage.content instead of returning empty.
+ */
+export function assembleAnthropicContentBlocks(toolBlocks, responseText, providerName, fallbackBlocks) {
+    const contentBlocks = [];
+    if (responseText) {
+        contentBlocks.push({ type: 'text', text: responseText });
+    }
+    for (const [, block] of [...toolBlocks.entries()].sort(([a], [b]) => a - b)) {
+        const parsedInput = parseToolCallArguments(block.args, {
+            provider: providerName,
+            toolName: block.name,
+            callId: block.id,
+        });
+        if (parsedInput === undefined)
+            continue;
+        contentBlocks.push({
+            type: 'tool_use',
+            id: block.id,
+            name: block.name,
+            input: parsedInput,
+        });
+    }
+    const target = contentBlocks.length > 0 ? contentBlocks : (fallbackBlocks ?? []);
+    return fromAnthropicContent(target);
+}

package/dist/platform/providers/anthropic-stream.d.ts

@@ -0,0 +1,30 @@
+/**
+ * Shared Anthropic streaming helpers.
+ *
+ * Extracted to avoid duplicating thinking-budget logic across the three
+ * Anthropic-family providers (anthropic.ts, anthropic-compat.ts,
+ * anthropic-sdk-provider.ts).
+ */
+/**
+ * Extra output tokens added above the thinking budget to satisfy the
+ * Anthropic API invariant: max_tokens > budget_tokens.
+ */
+export declare const ANTHROPIC_THINKING_HEADROOM = 4096;
+/**
+ * Apply Anthropic extended-thinking parameters to a mutable request body.
+ *
+ * Sets `body.thinking` and adjusts `body.max_tokens` so that:
+ *   - `max_tokens > budget_tokens` (Anthropic API requirement), and
+ *   - `max_tokens <= modelOutputCap` (avoids over-cap requests).
+ *
+ * If the budget meets or exceeds the model's output cap, thinking is skipped
+ * entirely — it is impossible to satisfy `max_tokens > budget_tokens` within
+ * the cap. Pass `Infinity` for compat/SDK providers that have no per-model cap;
+ * this preserves the original `budget + HEADROOM` behaviour for those paths.
+ *
+ * @param body           Mutable request body (mutated in place).
+ * @param reasoningEffort Effort label ('low'|'medium'|'high'|'instant') or falsy.
+ * @param modelOutputCap  Maximum output tokens the model accepts.
+ */
+export declare function applyAnthropicThinking(body: Record<string, unknown>, reasoningEffort: string | null | undefined, modelOutputCap: number): void;
+//# sourceMappingURL=anthropic-stream.d.ts.map

package/dist/platform/providers/anthropic-stream.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"anthropic-stream.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/anthropic-stream.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAIH;;;GAGG;AACH,eAAO,MAAM,2BAA2B,OAAO,CAAC;AAEhD;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,sBAAsB,CACpC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,EAC7B,eAAe,EAAE,MAAM,GAAG,IAAI,GAAG,SAAS,EAC1C,cAAc,EAAE,MAAM,GACrB,IAAI,CAWN"}

package/dist/platform/providers/anthropic-stream.js

@@ -0,0 +1,44 @@
+/**
+ * Shared Anthropic streaming helpers.
+ *
+ * Extracted to avoid duplicating thinking-budget logic across the three
+ * Anthropic-family providers (anthropic.ts, anthropic-compat.ts,
+ * anthropic-sdk-provider.ts).
+ */
+import { REASONING_BUDGET_MAP } from './interface.js';
+/**
+ * Extra output tokens added above the thinking budget to satisfy the
+ * Anthropic API invariant: max_tokens > budget_tokens.
+ */
+export const ANTHROPIC_THINKING_HEADROOM = 4096;
+/**
+ * Apply Anthropic extended-thinking parameters to a mutable request body.
+ *
+ * Sets `body.thinking` and adjusts `body.max_tokens` so that:
+ *   - `max_tokens > budget_tokens` (Anthropic API requirement), and
+ *   - `max_tokens <= modelOutputCap` (avoids over-cap requests).
+ *
+ * If the budget meets or exceeds the model's output cap, thinking is skipped
+ * entirely — it is impossible to satisfy `max_tokens > budget_tokens` within
+ * the cap. Pass `Infinity` for compat/SDK providers that have no per-model cap;
+ * this preserves the original `budget + HEADROOM` behaviour for those paths.
+ *
+ * @param body           Mutable request body (mutated in place).
+ * @param reasoningEffort Effort label ('low'|'medium'|'high'|'instant') or falsy.
+ * @param modelOutputCap  Maximum output tokens the model accepts.
+ */
+export function applyAnthropicThinking(body, reasoningEffort, modelOutputCap) {
+    if (!reasoningEffort || reasoningEffort === 'instant')
+        return;
+    const budget = REASONING_BUDGET_MAP[reasoningEffort];
+    if (typeof budget !== 'number' || budget <= 0)
+        return;
+    // Cannot satisfy API invariant max_tokens > budget_tokens within the model cap.
+    if (budget >= modelOutputCap)
+        return;
+    body['thinking'] = { type: 'enabled', budget_tokens: budget };
+    const currentMax = body['max_tokens'] ?? modelOutputCap;
+    if (currentMax <= budget) {
+        body['max_tokens'] = Math.min(budget + ANTHROPIC_THINKING_HEADROOM, modelOutputCap);
+    }
+}

package/dist/platform/providers/anthropic-vertex.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"anthropic-vertex.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/anthropic-vertex.ts"],"names":[],"mappings":"AAGA,OAAO,EAAE,oBAAoB,EAAE,MAAM,6BAA6B,CAAC;AAoMnE,qBAAa,uBAAwB,SAAQ,oBAAoB;;CAmChE"}
+{"version":3,"file":"anthropic-vertex.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/anthropic-vertex.ts"],"names":[],"mappings":"AAGA,OAAO,EAAE,oBAAoB,EAAE,MAAM,6BAA6B,CAAC;AAkMnE,qBAAa,uBAAwB,SAAQ,oBAAoB;;CAmChE"}

package/dist/platform/providers/anthropic-vertex.js

@@ -2,6 +2,7 @@ import { BaseAnthropic } from '@anthropic-ai/sdk/client';
 import * as Resources from '@anthropic-ai/sdk/resources/index';
 import { AuthClient, GoogleAuth } from 'google-auth-library';
 import { AnthropicSdkProvider } from './anthropic-sdk-provider.js';
+import { isRecord } from '../utils/record-coerce.js';
 const DEFAULT_VERSION = 'vertex-2023-10-16';
 const MODEL_ENDPOINTS = new Set(['/v1/messages', '/v1/messages?beta=true']);
 const VERTEX_MODELS = [
@@ -12,9 +13,6 @@ const VERTEX_MODELS = [
 function readEnv(name) {
     return process.env[name] ?? null;
 }
-function isRecord(value) {
-    return typeof value === 'object' && value !== null && !Array.isArray(value);
-}
 function resolveVertexBaseUrl(region) {
     if (region === 'global')
         return 'https://aiplatform.googleapis.com/v1';

package/dist/platform/providers/anthropic.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"anthropic.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/anthropic.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,WAAW,EACX,WAAW,EACX,YAAY,EAEZ,oBAAoB,EAKpB,uBAAuB,EACvB,2BAA2B,EAC5B,MAAM,gBAAgB,CAAC;AAKxB,OAAO,KAAK,EAAgB,eAAe,EAAE,MAAM,qBAAqB,CAAC;AA4EzE;;;;;GAKG;AACH,qBAAa,iBAAkB,YAAW,WAAW;IACnD,QAAQ,CAAC,IAAI,eAAe;IAC5B,QAAQ,CAAC,MAAM,EAAE,MAAM,EAAE,CAAM;IAC/B,QAAQ,CAAC,KAAK,EAAE,oBAAoB,CAAC;IAErC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAS;IAChC,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAqD;gBAGnF,MAAM,EAAE,MAAM,EACd,eAAe,GAAE,IAAI,CAAC,eAAe,EAAE,YAAY,GAAG,YAAY,CAA0B;IAcxF,IAAI,CAAC,MAAM,EAAE,WAAW,GAAG,OAAO,CAAC,YAAY,CAAC;IAwUhD,eAAe,CAAC,IAAI,EAAE,2BAA2B,GAAG,OAAO,CAAC,uBAAuB,CAAC;IA4C1F,YAAY,IAAI,OAAO;YAIT,eAAe;YAmBf,aAAa;YAcb,WAAW;YAcX,eAAe;IAuC7B,OAAO,CAAC,YAAY;IASpB,OAAO,CAAC,6BAA6B;IAwBrC,OAAO,CAAC,uBAAuB;YAUjB,gBAAgB;IAsB9B,OAAO,CAAC,yBAAyB;IA2BjC,OAAO,CAAC,mCAAmC;CAoB5C"}
+{"version":3,"file":"anthropic.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/anthropic.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,WAAW,EACX,WAAW,EACX,YAAY,EAEZ,oBAAoB,EAKpB,uBAAuB,EACvB,2BAA2B,EAC5B,MAAM,gBAAgB,CAAC;AAKxB,OAAO,KAAK,EAAgB,eAAe,EAAE,MAAM,qBAAqB,CAAC;AAgDzE;;;;;GAKG;AACH,qBAAa,iBAAkB,YAAW,WAAW;IACnD,QAAQ,CAAC,IAAI,eAAe;IAC5B,QAAQ,CAAC,MAAM,EAAE,MAAM,EAAE,CAAM;IAC/B,QAAQ,CAAC,KAAK,EAAE,oBAAoB,CAAC;IAErC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAS;IAChC,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAqD;gBAGnF,MAAM,EAAE,MAAM,EACd,eAAe,GAAE,IAAI,CAAC,eAAe,EAAE,YAAY,GAAG,YAAY,CAA0B;IAcxF,IAAI,CAAC,MAAM,EAAE,WAAW,GAAG,OAAO,CAAC,YAAY,CAAC;IAwNhD,eAAe,CAAC,IAAI,EAAE,2BAA2B,GAAG,OAAO,CAAC,uBAAuB,CAAC;IA4C1F,YAAY,IAAI,OAAO;YAIT,eAAe;YAmBf,aAAa;YAcb,WAAW;YAcX,eAAe;IAuC7B,OAAO,CAAC,YAAY;IASpB,OAAO,CAAC,6BAA6B;IAiBrC,OAAO,CAAC,uBAAuB;YAUjB,gBAAgB;IAsB9B,OAAO,CAAC,yBAAyB;IA2BjC,OAAO,CAAC,mCAAmC;CAoB5C"}

package/dist/platform/providers/anthropic.js

@@ -1,12 +1,13 @@
-import { REASONING_BUDGET_MAP } from './interface.js';
+import { applyAnthropicThinking } from './anthropic-stream.js';
 import { getCacheCapability } from './cache-capability.js';
 import { mapAnthropicStopReason } from './stop-reason-maps.js';
 import { getDefaultStrategy } from './cache-strategy.js';
 import { ProviderError } from '../types/errors.js';
 import { withRetry } from '../utils/retry.js';
-import { logger } from '../utils/logger.js';
 import { instrumentedLlmCall } from '../runtime/llm-observability.js';
-import { toAnthropicTools, toAnthropicMessages, fromAnthropicContent, parseToolCallArguments, } from './tool-formats.js';
+import { toAnthropicTools, toAnthropicMessages, fromAnthropicContent, normalizeAnthropicModel, } from './tool-formats.js';
+import { createAnthropicSSEState, readAnthropicSSEStream, assembleAnthropicContentBlocks } from './anthropic-sse-assembler.js';
+import { resolveCompletedStopReason, withProviderStopReason } from './provider-stop-reason.js';
 import { summarizeError, toProviderError } from '../utils/error-display.js';
 import { instrumentedFetch } from '../utils/fetch-with-timeout.js';
 import { toRecord } from '../utils/record-coerce.js';
@@ -24,13 +25,6 @@ const NOOP_CACHE_HIT_TRACKER = {
     getHitRate: () => 0,
     recordTurn: () => { },
 };
-function normalizeAnthropicModel(model) {
-    if (model.startsWith('anthropic:'))
-        return model.slice('anthropic:'.length);
-    if (model.startsWith('anthropic/'))
-        return model.slice('anthropic/'.length);
-    return model;
-}
 /** Clamp max_tokens to the model's known limit. */
 function clampMaxTokens(model, requested) {
     for (const { match, cap } of ANTHROPIC_MAX_OUTPUT) {
@@ -180,17 +174,7 @@ export class AnthropicProvider {
                 }
             }
             body['messages'] = anthropicMessages;
-            if (reasoningEffort && reasoningEffort !== 'instant') {
-                const budget = REASONING_BUDGET_MAP[reasoningEffort];
-                if (budget !== undefined && budget > 0) {
-                    body['thinking'] = { type: 'enabled', budget_tokens: budget };
-                    // max_tokens must be strictly greater than thinking.budget_tokens
-                    const currentMax = body['max_tokens'] ?? 8192;
-                    if (currentMax <= budget) {
-                        body['max_tokens'] = budget + 4096;
-                    }
-                }
-            }
+            applyAnthropicThinking(body, reasoningEffort, clampMaxTokens(resolvedModel, Infinity));
             const headers = {
                 'Content-Type': 'application/json',
                 'anthropic-version': ANTHROPIC_API_VERSION,
@@ -235,15 +219,7 @@ export class AnthropicProvider {
                 });
             }
             // Parse SSE stream
-            let responseText = '';
-            let inputTokens = 0;
-            let outputTokens = 0;
-            let cacheReadTokens = 0;
-            let cacheWriteTokens = 0;
-            let rawStopReason;
-            let stopReason = 'unknown';
-            // Accumulate tool use blocks by index
-            const toolBlocks = new Map();
+            const state = createAnthropicSSEState();
             const reader = res.body?.getReader();
             if (!reader) {
                 throw new ProviderError('Anthropic chat returned no response body.', {
@@ -253,125 +229,21 @@ export class AnthropicProvider {
                     phase: 'response',
                 });
             }
-            const decoder = new TextDecoder();
-            let buffer = '';
-            try {
-                while (true) {
-                    const { done, value } = await reader.read();
-                    if (done)
-                        break;
-                    buffer += decoder.decode(value, { stream: true });
-                    const lines = buffer.split('\n');
-                    // Keep the last (potentially incomplete) line in buffer
-                    buffer = lines.pop() ?? '';
-                    for (const line of lines) {
-                        if (!line.startsWith('data: '))
-                            continue;
-                        const data = line.slice(6).trim();
-                        if (!data || data === '[DONE]')
-                            continue;
-                        let event;
-                        try {
-                            event = JSON.parse(data);
-                        }
-                        catch {
-                            logger.warn('Anthropic SSE: failed to parse JSON chunk', {
-                                chunkPreview: data.slice(0, 200),
-                                chunkLength: data.length,
-                            });
-                            continue;
-                        }
-                        if (event.type === 'content_block_start' && event.content_block?.type === 'tool_use') {
-                            const idx = event.index ?? 0;
-                            toolBlocks.set(idx, {
-                                id: event.content_block.id ?? '',
-                                name: event.content_block.name ?? '',
-                                args: '',
-                            });
-                            if (onDelta) {
-                                onDelta({ toolCalls: [{ index: idx, id: event.content_block.id, name: event.content_block.name }] });
-                            }
-                        }
-                        else if (event.type === 'content_block_delta') {
-                            const idx = event.index ?? 0;
-                            if (event.delta?.type === 'text_delta' && event.delta.text) {
-                                responseText += event.delta.text;
-                                if (onDelta)
-                                    onDelta({ content: event.delta.text });
-                            }
-                            else if (event.delta?.type === 'thinking_delta' && event.delta.thinking) {
-                                if (onDelta)
-                                    onDelta({ reasoning: event.delta.thinking });
-                            }
-                            else if (event.delta?.type === 'input_json_delta' && event.delta.partial_json) {
-                                const block = toolBlocks.get(idx);
-                                if (block)
-                                    block.args += event.delta.partial_json;
-                                if (onDelta) {
-                                    onDelta({ toolCalls: [{ index: idx, arguments: event.delta.partial_json }] });
-                                }
-                            }
-                        }
-                        else if (event.type === 'message_delta') {
-                            if (event.delta?.stop_reason) {
-                                rawStopReason = event.delta.stop_reason;
-                                stopReason = mapAnthropicStopReason(rawStopReason);
-                            }
-                            if (event.usage?.output_tokens)
-                                outputTokens = event.usage.output_tokens;
-                            if (event.usage?.cache_read_input_tokens != null)
-                                cacheReadTokens = event.usage.cache_read_input_tokens;
-                            if (event.usage?.cache_creation_input_tokens != null)
-                                cacheWriteTokens = event.usage.cache_creation_input_tokens;
-                        }
-                        else if (event.type === 'message_start') {
-                            if (event.message?.usage) {
-                                inputTokens = event.message.usage.input_tokens;
-                                outputTokens = event.message.usage.output_tokens;
-                                cacheReadTokens = event.message.usage.cache_read_input_tokens ?? 0;
-                                cacheWriteTokens = event.message.usage.cache_creation_input_tokens ?? 0;
-                            }
-                        }
-                    }
-                }
-            }
-            finally {
-                reader.releaseLock();
-            }
-            // Build the content blocks for fromAnthropicContent
-            const contentBlocks = [];
-            if (responseText) {
-                contentBlocks.push({ type: 'text', text: responseText });
-            }
-            for (const [, block] of [...toolBlocks.entries()].sort(([a], [b]) => a - b)) {
-                const parsedInput = parseToolCallArguments(block.args, {
-                    provider: this.name,
-                    toolName: block.name,
-                    callId: block.id,
-                });
-                if (parsedInput === undefined)
-                    continue;
-                contentBlocks.push({
-                    type: 'tool_use',
-                    id: block.id,
-                    name: block.name,
-                    input: parsedInput,
-                });
-            }
-            const { text, toolCalls } = fromAnthropicContent(contentBlocks);
+            await readAnthropicSSEStream(reader, state, onDelta, 'Anthropic');
+            const { text, toolCalls } = assembleAnthropicContentBlocks(state.toolBlocks, state.responseText, this.name);
             // Record cache metrics for strategy adaptation.
-            this.cacheHitTracker.recordTurn({ inputTokens, cacheReadTokens, cacheWriteTokens });
+            this.cacheHitTracker.recordTurn({ inputTokens: state.inputTokens, cacheReadTokens: state.cacheReadTokens, cacheWriteTokens: state.cacheWriteTokens });
             const cap = getCacheCapability('anthropic');
             // Exclude write tokens from the denominator: writes are a one-time cost and inflate the
             // apparent miss rate on the first request. Read rate = reads / (billed input + reads).
-            const hitRateDenom = inputTokens + cacheReadTokens;
-            const hitRate = hitRateDenom > 0 ? cacheReadTokens / hitRateDenom : undefined;
+            const hitRateDenom = state.inputTokens + state.cacheReadTokens;
+            const hitRate = hitRateDenom > 0 ? state.cacheReadTokens / hitRateDenom : undefined;
             return {
                 content: text,
                 toolCalls,
-                usage: { inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens },
-                stopReason: stopReason === 'unknown' && text ? 'completed' : stopReason,
-                ...(rawStopReason !== undefined ? { providerStopReason: rawStopReason } : {}),
+                usage: { inputTokens: state.inputTokens, outputTokens: state.outputTokens, cacheReadTokens: state.cacheReadTokens, cacheWriteTokens: state.cacheWriteTokens },
+                stopReason: resolveCompletedStopReason(state.stopReason, text),
+                ...withProviderStopReason(state.rawStopReason),
                 cacheMetrics: {
                     strategy: cap.type === 'explicit' ? `explicit-${cap.maxBreakpoints}bp` : cap.type,
                     breakpointsPlaced,
@@ -531,15 +403,7 @@ export class AnthropicProvider {
         if (params.tools && params.tools.length > 0) {
             body['tools'] = toAnthropicTools(params.tools);
         }
-        if (params.reasoningEffort && params.reasoningEffort !== 'instant') {
-            const budget = REASONING_BUDGET_MAP[params.reasoningEffort];
-            if (budget !== undefined && budget > 0) {
-                body['thinking'] = { type: 'enabled', budget_tokens: budget };
-                const currentMax = body['max_tokens'] ?? 8192;
-                if (currentMax <= budget)
-                    body['max_tokens'] = budget + 4096;
-            }
-        }
+        applyAnthropicThinking(body, params.reasoningEffort, clampMaxTokens(resolvedModel, Infinity));
         return body;
     }
     mapAnthropicBatchStatus(body) {

package/dist/platform/providers/capabilities.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"capabilities.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/capabilities.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAsB,KAAK,SAAS,EAAE,MAAM,uBAAuB,CAAC;AAC3E,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,gBAAgB,CAAC;AAMlD;;;;GAIG;AACH,MAAM,WAAW,kBAAkB;IACjC,4DAA4D;IAC5D,SAAS,EAAE,OAAO,CAAC;IACnB,uEAAuE;IACvE,WAAW,EAAE,OAAO,CAAC;IACrB,qEAAqE;IACrE,aAAa,EAAE,OAAO,CAAC;IACvB,mEAAmE;IACnE,QAAQ,EAAE,OAAO,CAAC;IAClB,oEAAoE;IACpE,iBAAiB,EAAE,OAAO,CAAC;IAC3B,6DAA6D;IAC7D,gBAAgB,EAAE,MAAM,CAAC;IACzB,6DAA6D;IAC7D,eAAe,EAAE,MAAM,CAAC;IACxB,sDAAsD;IACtD,SAAS,EAAE,MAAM,CAAC;IAClB,0DAA0D;IAC1D,OAAO,EAAE,SAAS,CAAC;CACpB;AAED;;;GAGG;AACH,MAAM,WAAW,cAAc;IAC7B,qDAAqD;IACrD,iBAAiB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IACxC,oDAAoD;IACpD,mBAAmB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IAC1C,2DAA2D;IAC3D,qBAAqB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IAC5C,4EAA4E;IAC5E,gBAAgB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IACvC,yEAAyE;IACzE,yBAAyB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IAChD,iEAAiE;IACjE,gBAAgB,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IACtC,6DAA6D;IAC7D,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;CACtC;AAMD;;;GAGG;AACH,eAAO,MAAM,kBAAkB;;;;;;;;CAQrB,CAAC;AAEX,MAAM,MAAM,kBAAkB,GAAG,CAAC,OAAO,kBAAkB,CAAC,CAAC,MAAM,OAAO,kBAAkB,CAAC,CAAC;AAE9F,wDAAwD;AACxD,MAAM,WAAW,oBAAoB;IACnC,uCAAuC;IACvC,IAAI,EAAE,kBAAkB,CAAC;IACzB,iEAAiE;IACjE,MAAM,EAAE,MAAM,CAAC;IACf,+CAA+C;IAC/C,MAAM,EAAE,OAAO,GAAG,MAAM,GAAG,MAAM,CAAC;IAClC,+CAA+C;IAC/C,QAAQ,EAAE,OAAO,GAAG,MAAM,GAAG,MAAM,CAAC;CACrC;AAED,mFAAmF;AACnF,MAAM,MAAM,gBAAgB,GACxB;IACE,QAAQ,EAAE,IAAI,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,2DAA2D;IAC3D,OAAO,EAAE,MAAM,CAAC;IAChB,6DAA6D;IAC7D,UAAU,EAAE,kBAAkB,CAAC;CAChC,GACD;IACE,QAAQ,EAAE,KAAK,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,6DAA6D;IAC7D,OAAO,EAAE,MAAM,CAAC;IAChB,0EAA0E;IAC1E,UAAU,EAAE,oBAAoB,EAAE,CAAC;IACnC,6DAA6D;IAC7D,UAAU,EAAE,kBAAkB,CAAC;CAChC,CAAC;AAwRN;;;;;;;;;;;;;;;;GAgBG;AACH,qBAAa,0BAA0B;IACrC,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAyC;IAE/D;;;;;;;OAOG;IACH,aAAa,CACX,UAAU,EAAE,MAAM,EAClB,OAAO,EAAE,MAAM,EACf,QAAQ,CAAC,EAAE,IAAI,CAAC,WAAW,EAAE,cAAc,CAAC,GAC3C,kBAAkB;IAUrB;;;OAGG;IACH,UAAU,IAAI,IAAI;IAIlB;;;;;;OAMG;IACH,SAAS,CAAC,UAAU,EAAE,kBAAkB,EAAE,OAAO,EAAE,cAAc,GAAG,OAAO;IAI3E;;;;;;;;;OASG;IACH,mBAAmB,CACjB,UAAU,EAAE,MAAM,EAClB,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,cAAc,EACvB,QAAQ,CAAC,EAAE,IAAI,CAAC,WAAW,EAAE,cAAc,CAAC,GAC3C,gBAAgB;IA6BnB,OAAO,CAAC,QAAQ;IAwBhB,OAAO,CAAC,kBAAkB;CA6E3B"}
+{"version":3,"file":"capabilities.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/capabilities.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAsB,KAAK,SAAS,EAAE,MAAM,uBAAuB,CAAC;AAC3E,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,gBAAgB,CAAC;AAMlD;;;;GAIG;AACH,MAAM,WAAW,kBAAkB;IACjC,4DAA4D;IAC5D,SAAS,EAAE,OAAO,CAAC;IACnB,uEAAuE;IACvE,WAAW,EAAE,OAAO,CAAC;IACrB,qEAAqE;IACrE,aAAa,EAAE,OAAO,CAAC;IACvB,mEAAmE;IACnE,QAAQ,EAAE,OAAO,CAAC;IAClB,oEAAoE;IACpE,iBAAiB,EAAE,OAAO,CAAC;IAC3B,6DAA6D;IAC7D,gBAAgB,EAAE,MAAM,CAAC;IACzB,6DAA6D;IAC7D,eAAe,EAAE,MAAM,CAAC;IACxB,sDAAsD;IACtD,SAAS,EAAE,MAAM,CAAC;IAClB,0DAA0D;IAC1D,OAAO,EAAE,SAAS,CAAC;CACpB;AAED;;;GAGG;AACH,MAAM,WAAW,cAAc;IAC7B,qDAAqD;IACrD,iBAAiB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IACxC,oDAAoD;IACpD,mBAAmB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IAC1C,2DAA2D;IAC3D,qBAAqB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IAC5C,4EAA4E;IAC5E,gBAAgB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IACvC,yEAAyE;IACzE,yBAAyB,CAAC,EAAE,OAAO,GAAG,SAAS,CAAC;IAChD,iEAAiE;IACjE,gBAAgB,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IACtC,6DAA6D;IAC7D,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;CACtC;AAMD;;;GAGG;AACH,eAAO,MAAM,kBAAkB;;;;;;;;CAQrB,CAAC;AAEX,MAAM,MAAM,kBAAkB,GAAG,CAAC,OAAO,kBAAkB,CAAC,CAAC,MAAM,OAAO,kBAAkB,CAAC,CAAC;AAE9F,wDAAwD;AACxD,MAAM,WAAW,oBAAoB;IACnC,uCAAuC;IACvC,IAAI,EAAE,kBAAkB,CAAC;IACzB,iEAAiE;IACjE,MAAM,EAAE,MAAM,CAAC;IACf,+CAA+C;IAC/C,MAAM,EAAE,OAAO,GAAG,MAAM,GAAG,MAAM,CAAC;IAClC,+CAA+C;IAC/C,QAAQ,EAAE,OAAO,GAAG,MAAM,GAAG,MAAM,CAAC;CACrC;AAED,mFAAmF;AACnF,MAAM,MAAM,gBAAgB,GACxB;IACE,QAAQ,EAAE,IAAI,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,2DAA2D;IAC3D,OAAO,EAAE,MAAM,CAAC;IAChB,6DAA6D;IAC7D,UAAU,EAAE,kBAAkB,CAAC;CAChC,GACD;IACE,QAAQ,EAAE,KAAK,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,6DAA6D;IAC7D,OAAO,EAAE,MAAM,CAAC;IAChB,0EAA0E;IAC1E,UAAU,EAAE,oBAAoB,EAAE,CAAC;IACnC,6DAA6D;IAC7D,UAAU,EAAE,kBAAkB,CAAC;CAChC,CAAC;AA4RN;;;;;;;;;;;;;;;;GAgBG;AACH,qBAAa,0BAA0B;IACrC,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAyC;IAE/D;;;;;;;OAOG;IACH,aAAa,CACX,UAAU,EAAE,MAAM,EAClB,OAAO,EAAE,MAAM,EACf,QAAQ,CAAC,EAAE,IAAI,CAAC,WAAW,EAAE,cAAc,CAAC,GAC3C,kBAAkB;IAUrB;;;OAGG;IACH,UAAU,IAAI,IAAI;IAIlB;;;;;;OAMG;IACH,SAAS,CAAC,UAAU,EAAE,kBAAkB,EAAE,OAAO,EAAE,cAAc,GAAG,OAAO;IAI3E;;;;;;;;;OASG;IACH,mBAAmB,CACjB,UAAU,EAAE,MAAM,EAClB,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,cAAc,EACvB,QAAQ,CAAC,EAAE,IAAI,CAAC,WAAW,EAAE,cAAc,CAAC,GAC3C,gBAAgB;IA6BnB,OAAO,CAAC,QAAQ;IAwBhB,OAAO,CAAC,kBAAkB;CA6E3B"}

package/dist/platform/providers/capabilities.js

@@ -240,7 +240,7 @@ const PROVIDER_DEFAULTS = {
         parallelTools: true,
         jsonMode: true,
         reasoningControls: false,
-        maxContextTokens: 131_072,
+        maxContextTokens: 256_000,
         maxOutputTokens: 8_192,
         timeoutMs: 120_000,
     },
@@ -274,12 +274,16 @@ const MODEL_OVERRIDES = {
     'claude-3-5-sonnet-20241022': { maxOutputTokens: 8_192 },
     'claude-3-7-sonnet-20250219': { reasoningControls: true, maxOutputTokens: 64_000 },
     // OpenAI reasoning models
-    'o1': { reasoningControls: true, jsonMode: false, maxOutputTokens: 32_768 },
-    'o1-mini': { reasoningControls: true, jsonMode: false, maxOutputTokens: 65_536 },
-    'o1-preview': { reasoningControls: true, jsonMode: false, maxOutputTokens: 32_768 },
-    'o3': { reasoningControls: true, maxOutputTokens: 100_000 },
-    'o3-mini': { reasoningControls: true, maxOutputTokens: 65_536 },
-    'o4-mini': { reasoningControls: true, maxOutputTokens: 65_536 },
+    'o1': { reasoningControls: true, jsonMode: false, maxContextTokens: 200_000, maxOutputTokens: 32_768 },
+    'o1-mini': { reasoningControls: true, jsonMode: false, maxContextTokens: 200_000, maxOutputTokens: 65_536 },
+    'o1-preview': { reasoningControls: true, jsonMode: false, maxContextTokens: 200_000, maxOutputTokens: 32_768 },
+    'o3': { reasoningControls: true, maxContextTokens: 200_000, maxOutputTokens: 100_000 },
+    'o3-mini': { reasoningControls: true, maxContextTokens: 200_000, maxOutputTokens: 65_536 },
+    'o4-mini': { reasoningControls: true, maxContextTokens: 200_000, maxOutputTokens: 65_536 },
+    // OpenAI large-context models
+    'gpt-5': { maxContextTokens: 400_000 },
+    'gpt-4.1': { maxContextTokens: 400_000 },
+    'gpt-4.1-mini': { maxContextTokens: 400_000 },
     // Mercury models (InceptionLabs)
     'mercury-2': { toolCalling: false, parallelTools: false, reasoningControls: true, maxOutputTokens: 32_000 },
     'mercury-edit': { toolCalling: false, parallelTools: false, reasoningControls: false, maxOutputTokens: 32_000 },

package/dist/platform/providers/context-discovery.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"context-discovery.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/context-discovery.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAaH,gFAAgF;AAChF,eAAO,MAAM,qBAAqB,eAAe,CAAC;AA8WlD;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,wBAAsB,sBAAsB,CAC1C,OAAO,EAAE,MAAM,EACf,MAAM,CAAC,EAAE,MAAM,GAAG,SAAS,GAC1B,OAAO,CAAC,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CA0C9B"}
+{"version":3,"file":"context-discovery.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/context-discovery.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAaH,gFAAgF;AAChF,eAAO,MAAM,qBAAqB,eAAe,CAAC;AAuWlD;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,wBAAsB,sBAAsB,CAC1C,OAAO,EAAE,MAAM,EACf,MAAM,CAAC,EAAE,MAAM,GAAG,SAAS,GAC1B,OAAO,CAAC,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CA0C9B"}

package/dist/platform/providers/context-discovery.js

@@ -18,7 +18,7 @@
  */
 import { logger } from '../utils/logger.js';
 import { summarizeError } from '../utils/error-display.js';
-import { instrumentedFetch } from '../utils/fetch-with-timeout.js';
+import { instrumentedFetch, fetchWithTimeout } from '../utils/fetch-with-timeout.js';
 // ---------------------------------------------------------------------------
 // Constants
 // ---------------------------------------------------------------------------
@@ -48,21 +48,17 @@ function extractOrigin(baseURL) {
  * Returns the Response on success, or null on any error (timeout, network, 4xx/5xx).
  */
 async function probe(url, options = {}) {
-    const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(), PROBE_TIMEOUT_MS);
-    timer.unref?.();
+    const headers = { Accept: 'application/json' };
+    if (options.apiKey)
+        headers['Authorization'] = `Bearer ${options.apiKey}`;
+    if (options.body)
+        headers['Content-Type'] = 'application/json';
     try {
-        const headers = { Accept: 'application/json' };
-        if (options.apiKey)
-            headers['Authorization'] = `Bearer ${options.apiKey}`;
-        if (options.body)
-            headers['Content-Type'] = 'application/json';
-        const response = await instrumentedFetch(url, {
+        const response = await fetchWithTimeout(url, {
             method: options.method ?? 'GET',
             headers,
             ...(options.body !== undefined ? { body: options.body } : {}),
-            signal: controller.signal,
-        });
+        }, PROBE_TIMEOUT_MS, instrumentedFetch);
         if (!response.ok) {
             logger.debug('[context-discovery] Non-OK probe response', { url, status: response.status });
             return null;
@@ -71,7 +67,7 @@ async function probe(url, options = {}) {
     }
     catch (err) {
         const name = err?.name;
-        if (name === 'AbortError') {
+        if (name === 'AbortError' || name === 'TimeoutError') {
             logger.debug('[context-discovery] Probe timed out', { url });
         }
         else {
@@ -79,9 +75,6 @@ async function probe(url, options = {}) {
         }
         return null;
     }
-    finally {
-        clearTimeout(timer);
-    }
 }
 /**
  * Safely parse JSON from a Response. Returns null on parse error.

package/dist/platform/providers/context-window-fallback.d.ts

@@ -0,0 +1,24 @@
+/**
+ * Provider/model-family-aware context-window fallback.
+ *
+ * Used only when neither the live catalog (models.dev) nor the provider API
+ * reports a context window for a model — e.g. a model newer than the models.dev
+ * snapshot (the "brand-new gpt-5.5" case). A single flat default badly mis-sizes
+ * the window and drives auto-compaction to fire far too early (small default) or
+ * too late (large default, risking a provider "context length exceeded" error).
+ * Infer a sane family default from the provider and model id instead.
+ *
+ * The real window always wins whenever the catalog / OpenRouter / provider_api
+ * supplies one; this is the last line of defense, not a primary source.
+ */
+/** Last-resort default when neither provider nor model id is recognised. */
+export declare const FALLBACK_CONTEXT_WINDOW = 128000;
+/**
+ * Best-effort context window (in tokens) inferred from provider + model id.
+ *
+ * Over-estimating is the dangerous direction (it risks context-overflow errors),
+ * so only known-large families get a large window; older/unknown families stay at
+ * the conservative {@link FALLBACK_CONTEXT_WINDOW}.
+ */
+export declare function inferFallbackContextWindow(provider: string, modelId?: string): number;
+//# sourceMappingURL=context-window-fallback.d.ts.map

package/dist/platform/providers/context-window-fallback.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"context-window-fallback.d.ts","sourceRoot":"","sources":["../../../src/platform/providers/context-window-fallback.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAEH,4EAA4E;AAC5E,eAAO,MAAM,uBAAuB,SAAU,CAAC;AAE/C;;;;;;GAMG;AACH,wBAAgB,0BAA0B,CAAC,QAAQ,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,MAAM,GAAG,MAAM,CAwBrF"}