@nexus-cortex/core 4.124.45 → 4.124.47

package/dist/orchestrator/transportEffortFixes.js

@@ -0,0 +1,114 @@
+/**
+ * Transport reasoning-effort fixes (2026-10-02; stack-portability map
+ * .cortex/research/stack-portability-2026-10-02.md, note N1). Each is DARK behind its own env flag; with the flag off
+ * every caller takes its pre-existing code path unchanged (byte-identical requests).
+ *
+ * - CORTEX_OPENAI_TOOLS_REASONING: the gateway forwards the request/action effort for OpenAI reasoning cards even though they
+ *   do not declare `reasoning.toggleable`, so CORTEX_ACTION_EFFORT reaches `reasoning.effort` on the RESPONSES API. Chat/completions
+ *   is unchanged: R19b's drop stays because OpenAI still rejects reasoning_effort + function tools there (live probe 2026-10-03:
+ *   gpt-5.4-nano + gpt-5.6-luna → 400 "use /v1/responses"); Responses accepts it with tools (200 on both).
+ * - CORTEX_ANTHROPIC_EFFORT: Anthropic adaptive-thinking families (claude-4.7 / 4.8 / claude-5 / claude-fable-5) got
+ *   `thinking:{type:'adaptive'}` only, so the action effort never reached the wire. On: `output_config.effort` = the
+ *   effort (low|medium|high|xhigh|max; GA, no beta header). Non-adaptive families: `max` was not in the budget map and
+ *   fell to 10000; on, `max` maps to a large budget clamped below max_tokens (budget_tokens must be < max_tokens, >= 1024).
+ * - CORTEX_GEMINI_TOOLS_THINKING: Gemini thinkingConfig was never sent with tools (the @google/genai builders skipped
+ *   it — "cannot coexist", a Gemini 2.0-era observation) and the REST builder only ever sent a card-level thinkingLevel.
+ *   On: the action effort maps to generationConfig.thinkingConfig — `thinkingLevel` on Gemini 3.x, `thinkingBudget` on
+ *   2.5 and older — on REST and SDK, with or without tools.
+ *
+ * Pure helpers — no client state. Flags are resolved per call so tests/benches can flip them.
+ */
+function flagOn(raw) {
+    const v = String(raw ?? '').trim().toLowerCase();
+    return v === 'on' || v === 'true' || v === '1';
+}
+export function isOpenAIToolsReasoningEnabled() {
+    return flagOn(process.env.CORTEX_OPENAI_TOOLS_REASONING);
+}
+export function isAnthropicEffortEnabled() {
+    return flagOn(process.env.CORTEX_ANTHROPIC_EFFORT);
+}
+export function isGeminiToolsThinkingEnabled() {
+    return flagOn(process.env.CORTEX_GEMINI_TOOLS_THINKING);
+}
+// ---------------------------------------------------------------------------------------------------------
+// Anthropic
+// ---------------------------------------------------------------------------------------------------------
+const ANTHROPIC_EFFORT_LEVELS = new Set(['low', 'medium', 'high', 'xhigh', 'max']);
+/**
+ * `output_config.effort` value for an adaptive-thinking family, or undefined (send nothing → provider default).
+ * 'none' / unknown values send nothing (adaptive families have no "off" effort).
+ */
+export function anthropicOutputEffort(reasoningEffort) {
+    const e = String(reasoningEffort ?? '').trim().toLowerCase();
+    return ANTHROPIC_EFFORT_LEVELS.has(e) ? e : undefined;
+}
+/** Target thinking budget for effort `max` on budget_tokens families (before the max_tokens clamp). */
+export const ANTHROPIC_MAX_EFFORT_BUDGET = 120000;
+/** Visible-output room kept below max_tokens when clamping the `max` budget. */
+export const ANTHROPIC_MAX_EFFORT_OUTPUT_ROOM = 1024;
+/**
+ * budget_tokens for effort `max` on a non-adaptive family: ANTHROPIC_MAX_EFFORT_BUDGET clamped to
+ * max_tokens - ANTHROPIC_MAX_EFFORT_OUTPUT_ROOM (the API requires budget_tokens < max_tokens and >= 1024).
+ * Returns undefined when max_tokens is too small to honour the floor — the caller keeps its legacy value.
+ */
+export function anthropicMaxEffortBudget(maxTokens) {
+    const mt = Number(maxTokens);
+    if (!Number.isFinite(mt) || mt <= 0)
+        return undefined;
+    const b = Math.min(ANTHROPIC_MAX_EFFORT_BUDGET, Math.floor(mt) - ANTHROPIC_MAX_EFFORT_OUTPUT_ROOM);
+    return b >= 1024 ? b : undefined;
+}
+// ---------------------------------------------------------------------------------------------------------
+// Gemini
+// ---------------------------------------------------------------------------------------------------------
+/** Gemini 3.x takes `thinkingLevel`; 2.5 and older take `thinkingBudget` (sending both is a 400). */
+export function isGeminiThinkingLevelModel(modelId) {
+    return /^(models\/)?gemini-([3-9]|\d{2,})/i.test(String(modelId ?? '').trim());
+}
+/** thinkingBudget per effort for Gemini 2.5 (Flash range 0-24576; Pro 128-32768 — every value here is valid on both). */
+export const GEMINI_BUDGET_BY_EFFORT = {
+    low: 2048,
+    medium: 8192,
+    high: 16384,
+    xhigh: 24576,
+    max: 24576,
+};
+/** thinkingLevel per effort for Gemini 3.x (levels low|medium|high; xhigh/max clamp to high). */
+export const GEMINI_LEVEL_BY_EFFORT = {
+    low: 'low',
+    medium: 'medium',
+    high: 'high',
+    xhigh: 'high',
+    max: 'high',
+};
+/**
+ * The generationConfig.thinkingConfig for an effort, or undefined (none / unknown effort → send nothing, API default).
+ */
+export function geminiThinkingConfigForEffort(modelId, reasoningEffort) {
+    const e = String(reasoningEffort ?? '').trim().toLowerCase();
+    if (isGeminiThinkingLevelModel(modelId)) {
+        const level = GEMINI_LEVEL_BY_EFFORT[e];
+        return level ? { thinkingLevel: level } : undefined;
+    }
+    const budget = GEMINI_BUDGET_BY_EFFORT[e];
+    return budget !== undefined ? { thinkingBudget: budget } : undefined;
+}
+// ---------------------------------------------------------------------------------------------------------
+// CORTEX_ACTION_EFFORT_FIRST — request 0 of each user turn
+// ---------------------------------------------------------------------------------------------------------
+/**
+ * CORTEX_ACTION_EFFORT is filled into the request params AFTER the first request of each user turn is built
+ * (CortexOrchestrator initialPrepOpts vs the `??=` in the loop setup), so request 0 ran at the card default and
+ * requests 1+ at the action effort. On DeepSeek that is harmless (card 'medium' is coerced to 'high' server-side and
+ * the anchor-lift bust lands on the same boundary); on other providers it sends a different effort on request 0 and
+ * costs one extra cache break per user turn. DARK: CORTEX_ACTION_EFFORT_FIRST=on → request 0 gets the action effort
+ * when no ramp / request param set one. Off → undefined (byte-identical).
+ */
+export function firstRequestActionEffort(env = process.env) {
+    if (!flagOn(env.CORTEX_ACTION_EFFORT_FIRST))
+        return undefined;
+    const e = String(env.CORTEX_ACTION_EFFORT ?? '').trim();
+    return e || undefined;
+}
+//# sourceMappingURL=transportEffortFixes.js.map

package/dist/orchestrator/transportEffortFixes.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"transportEffortFixes.js","sourceRoot":"","sources":["../../src/orchestrator/transportEffortFixes.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,SAAS,MAAM,CAAC,GAAuB;IACrC,MAAM,CAAC,GAAG,MAAM,CAAC,GAAG,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IACjD,OAAO,CAAC,KAAK,IAAI,IAAI,CAAC,KAAK,MAAM,IAAI,CAAC,KAAK,GAAG,CAAC;AACjD,CAAC;AAED,MAAM,UAAU,6BAA6B;IAC3C,OAAO,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,6BAA6B,CAAC,CAAC;AAC3D,CAAC;AACD,MAAM,UAAU,wBAAwB;IACtC,OAAO,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,uBAAuB,CAAC,CAAC;AACrD,CAAC;AACD,MAAM,UAAU,4BAA4B;IAC1C,OAAO,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,4BAA4B,CAAC,CAAC;AAC1D,CAAC;AAED,4GAA4G;AAC5G,YAAY;AACZ,4GAA4G;AAE5G,MAAM,uBAAuB,GAAG,IAAI,GAAG,CAAC,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,CAAC,CAAC,CAAC;AAEnF;;;GAGG;AACH,MAAM,UAAU,qBAAqB,CAAC,eAAmC;IACvE,MAAM,CAAC,GAAG,MAAM,CAAC,eAAe,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IAC7D,OAAO,uBAAuB,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;AACxD,CAAC;AAED,uGAAuG;AACvG,MAAM,CAAC,MAAM,2BAA2B,GAAG,MAAM,CAAC;AAClD,gFAAgF;AAChF,MAAM,CAAC,MAAM,gCAAgC,GAAG,IAAI,CAAC;AAErD;;;;GAIG;AACH,MAAM,UAAU,wBAAwB,CAAC,SAA6B;IACpE,MAAM,EAAE,GAAG,MAAM,CAAC,SAAS,CAAC,CAAC;IAC7B,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,EAAE,CAAC,IAAI,EAAE,IAAI,CAAC;QAAE,OAAO,SAAS,CAAC;IACtD,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,2BAA2B,EAAE,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,GAAG,gCAAgC,CAAC,CAAC;IACnG,OAAO,CAAC,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;AACnC,CAAC;AAED,4GAA4G;AAC5G,SAAS;AACT,4GAA4G;AAE5G,qGAAqG;AACrG,MAAM,UAAU,0BAA0B,CAAC,OAAe;IACxD,OAAO,oCAAoC,CAAC,IAAI,CAAC,MAAM,CAAC,OAAO,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;AACjF,CAAC;AAED,yHAAyH;AACzH,MAAM,CAAC,MAAM,uBAAuB,GAA2B;IAC7D,GAAG,EAAE,IAAI;IACT,MAAM,EAAE,IAAI;IACZ,IAAI,EAAE,KAAK;IACX,KAAK,EAAE,KAAK;IACZ,GAAG,EAAE,KAAK;CACX,CAAC;AAEF,iGAAiG;AACjG,MAAM,CAAC,MAAM,sBAAsB,GAA2B;IAC5D,GAAG,EAAE,KAAK;IACV,MAAM,EAAE,QAAQ;IAChB,IAAI,EAAE,MAAM;IACZ,KAAK,EAAE,MAAM;IACb,GAAG,EAAE,MAAM;CACZ,CAAC;AAEF;;GAEG;AACH,MAAM,UAAU,6BAA6B,CAC3C,OAAe,EACf,eAAmC;IAEnC,MAAM,CAAC,GAAG,MAAM,CAAC,eAAe,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IAC7D,IAAI,0BAA0B,CAAC,OAAO,CAAC,EAAE,CAAC;QACxC,MAAM,KAAK,GAAG,sBAAsB,CAAC,CAAC,CAAC,CAAC;QACxC,OAAO,KAAK,CAAC,CAAC,CAAC,EAAE,aAAa,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;IACtD,CAAC;IACD,MAAM,MAAM,GAAG,uBAAuB,CAAC,CAAC,CAAC,CAAC;IAC1C,OAAO,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;AACvE,CAAC;AAED,4GAA4G;AAC5G,2DAA2D;AAC3D,4GAA4G;AAE5G;;;;;;;GAOG;AACH,MAAM,UAAU,wBAAwB,CAAC,MAAyB,OAAO,CAAC,GAAG;IAC3E,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,0BAA0B,CAAC;QAAE,OAAO,SAAS,CAAC;IAC9D,MAAM,CAAC,GAAG,MAAM,CAAC,GAAG,CAAC,oBAAoB,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC;IACxD,OAAO,CAAC,IAAI,SAAS,CAAC;AACxB,CAAC"}

package/dist/orchestrator/xaiChatTransport.d.ts

@@ -0,0 +1,27 @@
+/**
+ * xAI chat/completions transport helpers (2026-10-03; stack-portability map
+ * .cortex/research/stack-portability-2026-10-02.md §"xAI CHAT/COMPLETIONS THEORY TEST").
+ *
+ * The third xAI transport, ADDITIVE beside /v1/messages (the interleaved-thinking path — untouched) and /v1/responses.
+ * Selected per card (`apiMode: 'chat'`, e.g. grok-4.3-chat / grok-build-0.1-chat) or globally for un-pinned xAI cards via
+ * XAI_API_MODE=chat. Reachable ONLY when an xAI card resolves to pattern 'chat/completions' — nothing on the shipped default
+ * path does, so every helper here is dark by construction.
+ *
+ * Pure helpers — no client state.
+ */
+import type { ModelConfig } from '../models/ModelConfig.interface.js';
+/** True when this request rides api.x.ai /v1/chat/completions. */
+export declare function isXAIChatRoute(modelConfig: Pick<ModelConfig, 'provider' | 'api'>): boolean;
+/**
+ * `reasoning_effort` for an xAI chat/completions request, or undefined (send nothing → provider default, 'high').
+ *
+ * Per-card capability: a card that declares no default effort (`reasoning.defaultEffort` / `reasoning.effort` — the
+ * XAIConfigurator `reasoningEffort` option) NEVER gets the field. grok-build-0.1 400s on it ("does not support parameter
+ * reasoningEffort", live 2026-10-02) and its cards deliberately omit `reasoningEffort`.
+ *
+ * Precedence: explicit request / action effort (CORTEX_ACTION_EFFORT, effort pulse) > card default. Mapping: the harness's
+ * `max` (the DeepSeek ladder top) → `high` (xAI has no `max`; `xhigh` is not accepted on every grok — pass it explicitly to
+ * use it); `none` / unknown → the card default (xAI reasoning cannot be disabled).
+ */
+export declare function xaiChatReasoningEffort(requested: unknown, modelConfig: Pick<ModelConfig, 'reasoning'>): string | undefined;
+//# sourceMappingURL=xaiChatTransport.d.ts.map

package/dist/orchestrator/xaiChatTransport.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"xaiChatTransport.d.ts","sourceRoot":"","sources":["../../src/orchestrator/xaiChatTransport.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AACH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oCAAoC,CAAC;AAEtE,kEAAkE;AAClE,wBAAgB,cAAc,CAAC,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,UAAU,GAAG,KAAK,CAAC,GAAG,OAAO,CAE1F;AAKD;;;;;;;;;;GAUG;AACH,wBAAgB,sBAAsB,CAAC,SAAS,EAAE,OAAO,EAAE,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,WAAW,CAAC,GAAG,MAAM,GAAG,SAAS,CAW1H"}

package/dist/orchestrator/xaiChatTransport.js

@@ -0,0 +1,33 @@
+/** True when this request rides api.x.ai /v1/chat/completions. */
+export function isXAIChatRoute(modelConfig) {
+    return modelConfig.api?.pattern === 'chat/completions' && String(modelConfig.provider ?? '').toLowerCase() === 'xai';
+}
+/** xAI reasoning_effort values (docs: model-capabilities/text/reasoning). `xhigh` = grok-4.6+; older models treat it as high. */
+const XAI_EFFORT_LEVELS = new Set(['low', 'medium', 'high', 'xhigh']);
+/**
+ * `reasoning_effort` for an xAI chat/completions request, or undefined (send nothing → provider default, 'high').
+ *
+ * Per-card capability: a card that declares no default effort (`reasoning.defaultEffort` / `reasoning.effort` — the
+ * XAIConfigurator `reasoningEffort` option) NEVER gets the field. grok-build-0.1 400s on it ("does not support parameter
+ * reasoningEffort", live 2026-10-02) and its cards deliberately omit `reasoningEffort`.
+ *
+ * Precedence: explicit request / action effort (CORTEX_ACTION_EFFORT, effort pulse) > card default. Mapping: the harness's
+ * `max` (the DeepSeek ladder top) → `high` (xAI has no `max`; `xhigh` is not accepted on every grok — pass it explicitly to
+ * use it); `none` / unknown → the card default (xAI reasoning cannot be disabled).
+ */
+export function xaiChatReasoningEffort(requested, modelConfig) {
+    const r = modelConfig.reasoning;
+    if (!r?.supported)
+        return undefined;
+    const cardDefault = r.effort ?? r.defaultEffort;
+    if (!cardDefault)
+        return undefined; // capability gate — no declared effort, no field
+    const norm = (v) => {
+        const e = String(v ?? '').trim().toLowerCase();
+        if (e === 'max')
+            return 'high';
+        return XAI_EFFORT_LEVELS.has(e) ? e : undefined;
+    };
+    return norm(requested) ?? norm(cardDefault);
+}
+//# sourceMappingURL=xaiChatTransport.js.map

package/dist/orchestrator/xaiChatTransport.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"xaiChatTransport.js","sourceRoot":"","sources":["../../src/orchestrator/xaiChatTransport.ts"],"names":[],"mappings":"AAaA,kEAAkE;AAClE,MAAM,UAAU,cAAc,CAAC,WAAkD;IAC/E,OAAO,WAAW,CAAC,GAAG,EAAE,OAAO,KAAK,kBAAkB,IAAI,MAAM,CAAC,WAAW,CAAC,QAAQ,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,KAAK,KAAK,CAAC;AACvH,CAAC;AAED,iIAAiI;AACjI,MAAM,iBAAiB,GAAG,IAAI,GAAG,CAAC,CAAC,KAAK,EAAE,QAAQ,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC;AAEtE;;;;;;;;;;GAUG;AACH,MAAM,UAAU,sBAAsB,CAAC,SAAkB,EAAE,WAA2C;IACpG,MAAM,CAAC,GAAG,WAAW,CAAC,SAAyF,CAAC;IAChH,IAAI,CAAC,CAAC,EAAE,SAAS;QAAE,OAAO,SAAS,CAAC;IACpC,MAAM,WAAW,GAAG,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,aAAa,CAAC;IAChD,IAAI,CAAC,WAAW;QAAE,OAAO,SAAS,CAAC,CAAC,iDAAiD;IACrF,MAAM,IAAI,GAAG,CAAC,CAAU,EAAsB,EAAE;QAC9C,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;QAC/C,IAAI,CAAC,KAAK,KAAK;YAAE,OAAO,MAAM,CAAC;QAC/B,OAAO,iBAAiB,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;IAClD,CAAC,CAAC;IACF,OAAO,IAAI,CAAC,SAAS,CAAC,IAAI,IAAI,CAAC,WAAW,CAAC,CAAC;AAC9C,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/core",
-  "version": "4.124.45",
+  "version": "4.124.47",
   "description": "Core orchestration library implementing Claude CLI architecture patterns",
   "type": "module",
   "main": "./dist/index.js",
@@ -47,8 +47,8 @@
     "@google/generative-ai": "^0.2.1",
     "@gradio/client": "^2.3.1",
     "@modelcontextprotocol/sdk": "^1.20.2",
-    "@nexus-cortex/executors": "4.124.45",
-    "@nexus-cortex/types": "4.124.45",
+    "@nexus-cortex/executors": "4.124.47",
+    "@nexus-cortex/types": "4.124.47",
     "@types/uuid": "^10.0.0",
     "ajv": "^8.12.0",
     "chokidar": "^3.6.0",