@tyvm/knowhow 0.0.128 → 0.0.129

package/src/agents/tools/execCommand.ts

@@ -259,7 +259,7 @@ const execWithTimeout = async (
 // Public tool
 export async function execCommand(
   command: string,
-  timeout?: number,
+  timeout?: number | string,
   continueInBackground?: boolean,
   logFileName?: string
 ): Promise<string> {
@@ -272,17 +272,37 @@ export async function execCommand(
     this instanceof ToolsService ? this : services().Tools
   ) as ToolsService;
 
+  // Coerce timeout to a number. Agents (and some tool-call serializers) often
+  // pass numeric arguments as strings (e.g. "120000"). Node's exec()/setTimeout
+  // reject non-integer values with "The value of \"timeout\" is out of range.
+  // It must be an unsigned integer", so normalize here before any use.
+  let normalizedTimeout: number | undefined;
+  const rawTimeout = typeof timeout === "string" ? Number(timeout.trim()) : timeout;
+  if (typeof rawTimeout !== "number" || !Number.isFinite(rawTimeout)) {
+    normalizedTimeout = undefined;
+  } else if (rawTimeout === -1) {
+    // -1 is the sentinel for "wait indefinitely" — leave as-is.
+    normalizedTimeout = -1;
+  } else {
+    // Clamp to a non-negative integer so the underlying APIs never throw.
+    normalizedTimeout = Math.max(0, Math.floor(rawTimeout));
+  }
+
   // Detect suspiciously small timeout values — agents often pass seconds instead of ms.
   // Any value < 100 is almost certainly a mistake (100ms is still very fast for a shell cmd).
   // We auto-correct to seconds and include a warning in the output so the agent learns.
-  let correctedTimeout = timeout;
+  let correctedTimeout = normalizedTimeout;
   let timeoutWarning = "";
-  if (timeout !== undefined && timeout !== -1 && timeout < 100) {
-    correctedTimeout = timeout * 1000;
+  if (
+    normalizedTimeout !== undefined &&
+    normalizedTimeout !== -1 &&
+    normalizedTimeout < 100
+  ) {
+    correctedTimeout = normalizedTimeout * 1000;
     timeoutWarning =
-      `⚠️  Warning: timeout was ${timeout}ms which is extremely small and likely a mistake. ` +
+      `⚠️  Warning: timeout was ${normalizedTimeout}ms which is extremely small and likely a mistake. ` +
       `The timeout unit is milliseconds, not seconds. ` +
-      `Auto-corrected to ${correctedTimeout}ms (${timeout}s).\n\n`;
+      `Auto-corrected to ${correctedTimeout}ms (${normalizedTimeout}s).\n\n`;
   }
 
   const context = toolService.getContext();

package/src/agents/tools/minimal/callTool.ts

@@ -0,0 +1,45 @@
+import { minimatch } from "minimatch";
+import { MinimalToolsService } from "../../../services/MinimalToolsService";
+import { ToolsService } from "../../../services/Tools";
+
+/**
+ * Calls any tool by name with the given args.
+ * This is the escape hatch that lets agents call tools outside the base tool set
+ * without changing the `tools:` array sent to the AI (preserving cache stability).
+ */
+export async function callTool(
+  this: ToolsService,
+  args: { name: string; args: Record<string, any> }
+): Promise<any> {
+  const service = this as unknown as MinimalToolsService;
+  const { name, args: toolArgs } = args;
+
+  const allTools = service.getAllTools ? service.getAllTools() : service.getTools();
+
+  // Fuzzy match: exact, endsWith, or glob
+  const matchedTool =
+    allTools.find((t) => t.function.name === name) ||
+    allTools.find((t) => t.function.name.endsWith(name)) ||
+    allTools.find((t) => minimatch(t.function.name, name, { nocase: true }));
+
+  if (!matchedTool) {
+    const available = allTools.map((t) => t.function.name).join(", ");
+    return `Error: Tool "${name}" not found. Use inspectTools() to browse available tools. Available: ${available}`;
+  }
+
+  const toolName = matchedTool.function.name;
+
+  // Build a synthetic ToolCall object and delegate to the base callTool
+  const syntheticToolCall = {
+    id: `call_minimal_${Date.now()}_${Math.random().toString(36).substr(2, 6)}`,
+    type: "function" as const,
+    function: {
+      name: toolName,
+      arguments: JSON.stringify(toolArgs),
+    },
+  };
+
+  // Use base ToolsService.callTool with the matched tool name allowed
+  const result = await service.callToolByName(syntheticToolCall, toolName);
+  return result;
+}

package/src/agents/tools/minimal/definitions.ts

@@ -0,0 +1,50 @@
+import { Tool } from "../../../clients/types";
+
+export const callToolDefinition: Tool = {
+  type: "function",
+  function: {
+    name: "callTool",
+    description:
+      "Call any available tool by name with the given arguments. Use listStoredToolResponses + inspectTools to discover available tools and their argument schemas before calling. This is the primary way to call tools that are not in the base tool set.",
+    parameters: {
+      type: "object",
+      positional: false,
+      properties: {
+        name: {
+          type: "string",
+          description: "The name of the tool to call.",
+        },
+        args: {
+          type: "object",
+          description:
+            "The arguments to pass to the tool. Should match the tool's parameter schema.",
+        },
+      },
+      required: ["name", "args"],
+    },
+  },
+};
+
+export const inspectToolsDefinition: Tool = {
+  type: "function",
+  function: {
+    name: "inspectTools",
+    description:
+      "Get the full schema/definition for tools matching the given glob patterns. Use this to discover what arguments a tool needs before calling it via callTool. Example patterns: ['read*'] finds tools starting with 'read', ['*File'] finds tools ending with 'File'.",
+    parameters: {
+      type: "object",
+      positional: true,
+      properties: {
+        patterns: {
+          type: "array",
+          items: { type: "string" },
+          description:
+            "Array of glob patterns to filter tool names. If omitted, returns all available tools.",
+        },
+      },
+      required: [],
+    },
+  },
+};
+
+export const minimalDefinitions = [callToolDefinition, inspectToolsDefinition];

package/src/agents/tools/minimal/index.ts

@@ -0,0 +1,3 @@
+export * from "./callTool";
+export * from "./inspectTools";
+export * from "./definitions";

package/src/agents/tools/minimal/inspectTools.ts

@@ -0,0 +1,27 @@
+import { minimatch } from "minimatch";
+import { MinimalToolsService } from "../../../services/MinimalToolsService";
+import { ToolsService } from "../../../services/Tools";
+
+/**
+ * Returns the full tool definitions (schema) for all tools matching the given patterns.
+ * Used by agents to discover argument shapes before calling via callTool.
+ */
+export function inspectTools(
+  this: ToolsService,
+  patterns?: string[]
+): object[] {
+  const service = this as unknown as MinimalToolsService;
+  const allTools = service.getAllTools ? service.getAllTools() : service.getTools();
+
+  if (!patterns || patterns.length === 0) {
+    return allTools.map((t) => t.function);
+  }
+
+  return allTools
+    .filter((t) =>
+      patterns.some((pattern) =>
+        minimatch(t.function.name, pattern, { nocase: true })
+      )
+    )
+    .map((t) => t.function);
+}

package/src/chat/modules/AgentModule.ts

@@ -28,6 +28,7 @@ import {
   XmlToolCallProcessor,
   HarmonyToolProcessor,
   Base64ImageProcessor,
+  MinimalToolsMessageProcessor,
 } from "../../processors/index";
 import { TaskInfo } from "../types";
 import { createAgent, agentConstructors, AgentName } from "../../agents";
@@ -124,13 +125,16 @@ export class AgentModule extends BaseChatModule {
       },
       {
         name: "poke",
-        description: "Interrupt the agent's current tool call or AI completion, so it can continue with the next step",
+        description:
+          "Interrupt the agent's current tool call or AI completion, so it can continue with the next step",
         modes: ["agent:attached"],
         handler: async (args: string[]): Promise<void> => {
           if (this.attachedAgent) {
             const message = args.length > 0 ? args.join(" ") : undefined;
             this.attachedAgent.interrupt(message);
-            console.log("Agent interrupted — it will continue with the next step.");
+            console.log(
+              "Agent interrupted — it will continue with the next step."
+            );
           }
         },
       },
@@ -531,10 +535,14 @@ export class AgentModule extends BaseChatModule {
       const threads = session.threads || [];
       // Guard against sessions saved with a flat Message[] instead of Message[][]
       // (a bug where threadUpdate emitted a single thread instead of all threads)
-      const normalizedThreads: Message[][] = threads.length > 0 && !Array.isArray(threads[0])
-        ? [threads as unknown as Message[]]
-        : threads as Message[][];
-      const lastThread = normalizedThreads.length > 0 ? normalizedThreads[normalizedThreads.length - 1] : [];
+      const normalizedThreads: Message[][] =
+        threads.length > 0 && !Array.isArray(threads[0])
+          ? [threads as unknown as Message[]]
+          : (threads as Message[][]);
+      const lastThread =
+        normalizedThreads.length > 0
+          ? normalizedThreads[normalizedThreads.length - 1]
+          : [];
       const resumeMessages = [...lastThread];
 
       // Append the resume prompt to the last user message (or add a new one)
@@ -548,7 +556,9 @@ export class AgentModule extends BaseChatModule {
         const actualIndex = lastThread.length - 1 - reversedIndex;
         resumeMessages[actualIndex] = {
           ...resumeMessages[actualIndex],
-          content: resumeMessages[actualIndex].content + `\n\n<Workflow>[RESUME CONTEXT]: ${resumePrompt}</Workflow>`,
+          content:
+            resumeMessages[actualIndex].content +
+            `\n\n<Workflow>[RESUME CONTEXT]: ${resumePrompt}</Workflow>`,
         };
       }
 
@@ -584,7 +594,12 @@ export class AgentModule extends BaseChatModule {
         }
       });
 
-      await this.attachedAgentChatLoop(taskId, agent, resumePrompt, resumeMessages);
+      await this.attachedAgentChatLoop(
+        taskId,
+        agent,
+        resumePrompt,
+        resumeMessages
+      );
     } catch (error) {
       console.error(
         `Failed to resume session ${sessionId}:`,
@@ -781,6 +796,10 @@ export class AgentModule extends BaseChatModule {
         new HarmonyToolProcessor().createProcessor(),
       ]);
 
+      agent.messageProcessor.setProcessors("pre_tools", [
+        new MinimalToolsMessageProcessor().createProcessor(),
+      ]);
+
       agent.messageProcessor.setProcessors("post_tools", [
         new Base64ImageProcessor(agent.tools).createProcessor(),
         ...caching,

package/src/clients/contextLimits.ts

@@ -10,6 +10,9 @@ import { Models, EmbeddingModels } from "../types";
  */
 export const ContextLimits: Record<string, number> = {
   // ─── OpenAI ───────────────────────────────────────────────────────────────
+  [Models.openai.GPT_56_Sol]: 1_000_000,
+  [Models.openai.GPT_56_Terra]: 1_000_000,
+  [Models.openai.GPT_56_Luna]: 1_000_000,
   [Models.openai.GPT_55]: 1_000_000,
   [Models.openai.GPT_54]: 1_000_000,
   [Models.openai.GPT_54_Mini]: 400_000,
@@ -83,6 +86,11 @@ export const ContextLimits: Record<string, number> = {
   // ─── Anthropic ────────────────────────────────────────────────────────────
   [Models.anthropic.Opus4_8Fast]: 1_000_000,
   [Models.anthropic.Opus4_8]: 1_000_000,
+  // Claude 5 generation (launched 2026-06-09, 1M context window)
+  [Models.anthropic.Fable5]: 1_000_000,
+  [Models.anthropic.Sonnet5]: 1_000_000,
+  [Models.anthropic.Mythos5]: 1_000_000,
+  [Models.anthropic.MythosPreview]: 1_000_000,
   [Models.anthropic.Opus4_7]: 1_000_000,
   [Models.anthropic.Opus4_6]: 1_000_000,
   [Models.anthropic.Opus4_6Fast]: 1_000_000,
@@ -166,6 +174,7 @@ export const ContextLimits: Record<string, number> = {
   [Models.xai.GrokBeta]: 131_072,
   [Models.xai.GrokVisionBeta]: 131_072,
   [Models.xai.Grok4_1_Fast]: 2_000_000,
+  [Models.xai.Grok4_5]: 500_000,
   [Models.xai.Grok4Fast]: 2_000_000,
   [Models.xai.Grok4FastNonReasoning]: 2_000_000,
   // xAI image/video models — no text context window

package/src/clients/index.ts

@@ -41,6 +41,7 @@ import {
   XaiImagePricing,
   XaiVideoPricing,
 } from "./pricing/xai";
+import { MetaTextPricing } from "./pricing/meta";
 import type {
   ModelPricing,
   ModelType,
@@ -56,6 +57,7 @@ import { GenericMistralClient } from "./mistral";
 import { GitHubCopilotClient } from "./copilot";
 import { GenericLlamaClient } from "./llama";
 import { GenericFireworksClient } from "./fireworks";
+import { GenericMetaClient } from "./meta";
 export {
   OpenAiTextPricing,
   AnthropicTextPricing,
@@ -100,6 +102,7 @@ const BUILT_IN_PROVIDER_REGISTRY: Record<string, ProviderRegistryEntry> = {
   "github-copilot": { clientClass: GitHubCopilotClient },
   llama: { clientClass: GenericLlamaClient },
   fireworks: { clientClass: GenericFireworksClient },
+  meta: { clientClass: GenericMetaClient },
   knowhow: {
     createClient: (entry: ModelProvider) => {
       const jwt = loadKnowhowJwt();
@@ -130,6 +133,7 @@ const DEFAULT_PROVIDERS: ModelProvider[] = [
   { provider: "github-copilot", envKey: "GITHUB_COPILOT_TOKEN" },
   { provider: "llama", envKey: "LLAMA_API_KEY" },
   { provider: "fireworks", envKey: "FIREWORKS_API_KEY" },
+  { provider: "meta", envKey: "META_API_KEY" },
 ];
 
 export class AIClient {
@@ -1021,6 +1025,7 @@ export class AIClient {
       ...GeminiPricing,
       ...XaiTextPricing,
       ...FireworksTextPricing,
+      ...MetaTextPricing,
     };
     const allImagePricing: Record<string, ModelPricing> = {
       ...XaiImagePricing,
@@ -1066,3 +1071,4 @@ export * from "./mistral";
 export * from "./llama";
 export * from "./copilot";
 export * from "./fireworks";
+export * from "./meta";

package/src/clients/meta.ts

@@ -0,0 +1,18 @@
+import { HttpClient } from "./http";
+import { MetaTextPricing } from "./pricing/meta";
+
+/**
+ * Meta Model API client — OpenAI-compatible API
+ * https://api.meta.ai
+ * Drop-in compatible with the OpenAI SDK (Responses + Chat Completions) and
+ * the Anthropic SDK (Messages API). Serves Meta's Muse Spark model with a
+ * 1,048,576 token context window.
+ * Set env var META_API_KEY to enable.
+ */
+export class GenericMetaClient extends HttpClient {
+  constructor(apiKey = process.env.META_API_KEY) {
+    super("https://api.meta.ai");
+    if (apiKey) this.setJwt(apiKey);
+    this.setPrices(MetaTextPricing);
+  }
+}

package/src/clients/pricing/anthropic.ts

@@ -9,6 +9,10 @@ import { completions, ModelCatalogEntry, ModelPricing } from "./types";
 
 export const AnthropicModels = {
   // Active models
+  Fable5:    "claude-fable-5",
+  Sonnet5:   "claude-sonnet-5",
+  Mythos5:   "claude-mythos-5",
+  MythosPreview: "claude-mythos-preview",
   Opus4_8:   "claude-opus-4-8",
   Opus4_8Fast: "claude-opus-4-8-fast",
   Opus4_7:   "claude-opus-4-7",
@@ -39,6 +43,8 @@ export const AnthropicModels = {
 
 export const AnthropicTextModels: string[] = [
   AnthropicModels.Opus4_8,
+  AnthropicModels.Fable5,
+  AnthropicModels.Sonnet5,
   AnthropicModels.Opus4_7,
   AnthropicModels.Opus4_6,
   AnthropicModels.Sonnet4_6,
@@ -51,6 +57,8 @@ export const AnthropicTextModels: string[] = [
 // Models in our catalog but not yet publicly available
 export const AnthropicLimitedAvailabilityModels: string[] = [
   AnthropicModels.Opus4_8Fast, // fast mode pricing
+  AnthropicModels.Mythos5,     // invitation-only (Project Glasswing)
+  AnthropicModels.MythosPreview, // invitation-only (defensive cybersecurity)
   AnthropicModels.Opus4_6Fast, // 404 – not publicly available yet
 ];
 
@@ -80,6 +88,11 @@ const AnthropicAllModels: string[] = [
 export const AnthropicTextPricing: Record<string, ModelPricing> = {
   [AnthropicModels.Opus4_8]:   { input: 5.0,  cache_write: 6.25,   cache_hit: 0.5,  output: 25.0 },
   [AnthropicModels.Opus4_8Fast]: { input: 10.0, cache_write: 12.5,  cache_hit: 1.0,  output: 50.0, limitedAvailability: true },
+  // Claude 5 generation (launched 2026-06-09)
+  [AnthropicModels.Fable5]:    { input: 10.0, cache_write: 12.5,   cache_hit: 1.0,  output: 50.0 },
+  [AnthropicModels.Sonnet5]:   { input: 3.0,  cache_write: 3.75,   cache_hit: 0.3,  output: 15.0 },
+  [AnthropicModels.Mythos5]:   { input: 10.0, cache_write: 12.5,   cache_hit: 1.0,  output: 50.0, limitedAvailability: true },
+  [AnthropicModels.MythosPreview]: { input: 10.0, cache_write: 12.5, cache_hit: 1.0, output: 50.0, limitedAvailability: true },
   [AnthropicModels.Opus4_7]:   { input: 5.0,  cache_write: 6.25,   cache_hit: 0.5,  output: 25.0 },
   [AnthropicModels.Opus4_6]:   { input: 5.0,  cache_write: 6.25,   cache_hit: 0.5,  output: 25.0 },
   [AnthropicModels.Opus4_6Fast]: { input: 30.0, cache_write: 37.5, cache_hit: 3.0,  output: 150.0, limitedAvailability: true },

package/src/clients/pricing/fireworks.ts

@@ -1,8 +1,10 @@
+import { ModelPricing } from "./types";
+
 /**
  * Fireworks AI pricing (USD per 1M tokens)
  * Source: https://fireworks.ai/pricing
  */
-export const FireworksTextPricing: Record<string, { input: number; output: number; cache_hit?: number }> = {
+export const FireworksTextPricing: Record<string, ModelPricing> = {
   // Moonshot AI
   "accounts/fireworks/models/kimi-k2p7-code": { input: 0.95, cache_hit: 0.19, output: 4.0 },
   "accounts/fireworks/models/kimi-k2-6": { input: 0.95, cache_hit: 0.16, output: 4.0 },
@@ -14,22 +16,26 @@ export const FireworksTextPricing: Record<string, { input: number; output: numbe
   "accounts/fireworks/models/minimax-m2-5": { input: 0.30, cache_hit: 0.03, output: 1.20 },
 
   // Qwen
-  "accounts/fireworks/models/qwen3-6-plus": { input: 0.50, cache_hit: 0.10, output: 3.0 },
-  "accounts/fireworks/models/qwen3p6-plus": { input: 0.50, cache_hit: 0.10, output: 3.0 },
-  "accounts/fireworks/models/qwen3-vl-30b-a3b-thinking": { input: 0.15, cache_hit: 0.08, output: 0.60 },
-  "accounts/fireworks/models/qwen3-vl-30b-a3b-instruct": { input: 0.15, cache_hit: 0.08, output: 0.60 },
-  "accounts/fireworks/models/qwen3-8b": { input: 0.20, cache_hit: 0.10, output: 0.20 },
+  "accounts/fireworks/models/qwen3p7-plus": { input: 0.40, cache_hit: 0.08, output: 1.60 },
+
+  // Deprecated QWEN Models
+  "accounts/fireworks/models/qwen3-6-plus": { input: 0.50, cache_hit: 0.10, output: 3.0, deprecated: true},
+  "accounts/fireworks/models/qwen3p6-plus": { input: 0.50, cache_hit: 0.10, output: 3.0, deprecated: true},
+  "accounts/fireworks/models/qwen3-vl-30b-a3b-thinking": { input: 0.15, cache_hit: 0.08, output: 0.60, deprecated: true},
+  "accounts/fireworks/models/qwen3-vl-30b-a3b-instruct": { input: 0.15, cache_hit: 0.08, output: 0.60, deprecated: true},
+  "accounts/fireworks/models/qwen3-8b": { input: 0.20, cache_hit: 0.10, output: 0.20, deprecated: true},
 
   // Z.ai
   "accounts/fireworks/models/glm-5p2": { input: 1.40, cache_hit: 0.26, output: 4.40 },
-  "accounts/fireworks/models/glm-5-2": { input: 1.40, cache_hit: 0.26, output: 4.40 },
-  "accounts/fireworks/models/glm-5-1": { input: 1.40, cache_hit: 0.26, output: 4.40 },
-  "accounts/fireworks/models/glm-5": { input: 1.00, cache_hit: 0.20, output: 3.20 },
-  "accounts/fireworks/models/glm-4-7": { input: 0.60, cache_hit: 0.30, output: 2.20 },
+  "accounts/fireworks/models/glm-5p1": { input: 1.40, cache_hit: 0.26, output: 4.40 },
+  "accounts/fireworks/models/glm-5": { input: 1.00, cache_hit: 0.20, output: 3.20, deprecated: true },
+  "accounts/fireworks/models/glm-4-7": { input: 0.60, cache_hit: 0.30, output: 2.20, deprecated: true},
 
   // DeepSeek AI
-  "accounts/fireworks/models/deepseek-v3-2": { input: 0.56, cache_hit: 0.28, output: 1.68 },
-  "accounts/fireworks/models/deepseek-v3-1": { input: 0.56, cache_hit: 0.28, output: 1.68 },
+  "accounts/fireworks/models/deepseek-v4-pro": { input: 1.74, cache_hit: 0.15, output: 3.48 },
+  "accounts/fireworks/models/deepseek-v4-flash": { input: 0.14, cache_hit: 0.03, output: 0.28 },
+  "accounts/fireworks/models/deepseek-v3-2": { input: 0.56, cache_hit: 0.28, output: 1.68, deprecated: true },
+  "accounts/fireworks/models/deepseek-v3-1": { input: 0.56, cache_hit: 0.28, output: 1.68, deprecated: true },
 
   // OpenAI OSS
   "accounts/fireworks/models/gpt-oss-120b": { input: 0.15, cache_hit: 0.01, output: 0.60 },

package/src/clients/pricing/index.ts

@@ -36,6 +36,7 @@ export { OpenRouterTextPricing } from "./openrouter";
 export { LlamaTextPricing } from "./llama";
 export { CopilotTextPricing, CopilotModelMultipliers } from "./copilot";
 export { FireworksTextPricing } from "./fireworks";
+export { MetaTextPricing, MetaWebSearchPricingPer1000Queries } from "./meta";
 
 // ─── Combined catalog ─────────────────────────────────────────────────────────
 export { ALL_MODEL_CATALOG, USAGE_MARKUP_PERCENT } from "./models";

package/src/clients/pricing/meta.ts

@@ -0,0 +1,12 @@
+/**
+ * Meta Model API pricing (USD per 1M tokens)
+ * Source: Meta Model API docs — https://api.meta.ai
+ * New accounts start with $20 in free credits.
+ */
+export const MetaTextPricing: Record<string, { input: number; output: number; cached_input?: number }> = {
+  "muse-spark-1.1": { input: 1.25, output: 4.25, cached_input: 0.15 },
+};
+
+// Web search grounding is billed per 1,000 search queries ($2.50/1,000),
+// not per-token, so it isn't represented in the per-token pricing map above.
+export const MetaWebSearchPricingPer1000Queries = 2.5;

package/src/clients/pricing/openai.ts

@@ -16,6 +16,9 @@ import {
 // ─── Model IDs ────────────────────────────────────────────────────────────────
 
 export const OpenAiModels = {
+  GPT_56_Sol: "gpt-5.6-sol",
+  GPT_56_Terra: "gpt-5.6-terra",
+  GPT_56_Luna: "gpt-5.6-luna",
   GPT_55: "gpt-5.5",
   GPT_55_Pro: "gpt-5.5-pro",
   GPT_54: "gpt-5.4",
@@ -95,6 +98,9 @@ export const OpenAiEmbeddingModels: Record<string, string> = {
 
 export const OpenAiReasoningModels: string[] = [
   OpenAiModels.GPT_5,
+  OpenAiModels.GPT_56_Sol,
+  OpenAiModels.GPT_56_Terra,
+  OpenAiModels.GPT_56_Luna,
   OpenAiModels.GPT_5_Pro,
   OpenAiModels.GPT_5_Mini,
   OpenAiModels.GPT_5_Nano,
@@ -125,6 +131,9 @@ export const OpenAiChatModels: string[] = [
   OpenAiModels.GPT_41,
   OpenAiModels.GPT_41_Mini,
   OpenAiModels.GPT_41_Nano,
+  OpenAiModels.GPT_56_Sol,
+  OpenAiModels.GPT_56_Terra,
+  OpenAiModels.GPT_56_Luna,
   OpenAiModels.GPT_54,
   OpenAiModels.GPT_54_Mini,
   OpenAiModels.GPT_55,
@@ -218,6 +227,10 @@ export const OpenAiSearchModels: string[] = [
 // ─── Pricing (USD per 1M tokens) ──────────────────────────────────────────────
 
 export const OpenAiTextPricing: Record<string, ModelPricing> = {
+  // gpt-5.6 series — short context / long context pricing
+  [OpenAiModels.GPT_56_Sol]:   { input: 5.0, cached_input: 0.5,  cache_write: 6.25,  output: 30.0, input_gt_200k: 10.0, output_gt_200k: 45.0 },
+  [OpenAiModels.GPT_56_Terra]: { input: 2.5, cached_input: 0.25, cache_write: 3.125, output: 15.0, input_gt_200k: 5.0,  output_gt_200k: 22.5 },
+  [OpenAiModels.GPT_56_Luna]:  { input: 1.0, cached_input: 0.10, cache_write: 1.25,  output: 6.0,  input_gt_200k: 2.0,  output_gt_200k: 9.0  },
   [OpenAiModels.GPT_55]: { input: 5.0, cached_input: 0.5, output: 30.0 },
   [OpenAiModels.GPT_55_Pro]: { input: 30.0, cached_input: 0, output: 180.0, reasoningLevels: ["medium", "high", "xhigh"] },
   [OpenAiModels.GPT_54]: { input: 2.5, cached_input: 0.25, output: 15.0 },

package/src/clients/pricing/xai.ts

@@ -10,6 +10,7 @@ export const XaiModels = {
   Grok_4_20_Reasoning: "grok-4.20-0309-reasoning",
   Grok_4_20_NonReasoning: "grok-4.20-0309-non-reasoning",
   Grok_4_20_MultiAgent: "grok-4.20-multi-agent-0309",
+  Grok4_5: "grok-4.5",
   Grok4_1_Fast_Reasoning: "grok-4-1-fast-reasoning",
   Grok4_1_Fast_NonReasoning: "grok-4-1-fast-non-reasoning",
   GrokCodeFast: "grok-code-fast-1",
@@ -44,6 +45,7 @@ export const XaiTextModels: string[] = [
   XaiModels.Grok_4_20_Reasoning, XaiModels.Grok_4_20_NonReasoning,
   XaiModels.Grok_4_20_MultiAgent,
   XaiModels.Grok4_1_Fast_Reasoning, XaiModels.Grok4_1_Fast_NonReasoning,
+  XaiModels.Grok4_5,
   XaiModels.GrokCodeFast, XaiModels.Grok4,
   XaiModels.Grok3Beta, XaiModels.Grok3MiniBeta, XaiModels.Grok3FastBeta, XaiModels.Grok3MiniFastBeta,
 ];
@@ -87,6 +89,7 @@ export const XaiTextPricing: Record<string, ModelPricing> = {
   [XaiModels.Grok_4_20_MultiAgent]:      { input: 2.0, cache_hit: 0.20, output: 6.0, reasoningLevels: ["low", "medium", "high", "xhigh"], useResponsesApi: true },
   [XaiModels.Grok4_1_Fast_Reasoning]:    { input: 0.2, cache_hit: 0.05, output: 0.5, useResponsesApi: true },
   [XaiModels.Grok4_1_Fast_NonReasoning]: { input: 0.2, cache_hit: 0.05, output: 0.5, useResponsesApi: true },
+  [XaiModels.Grok4_5]:                   { input: 2.0, cache_hit: 0.50, output: 6.0 },
   [XaiModels.GrokCodeFast]:              { input: 0.2, cache_hit: 0.02, output: 1.5 },
   [XaiModels.Grok4]:                     { input: 3.0, output: 15.0 },
   [XaiModels.Grok3Beta]:                 { input: 3.0, output: 15.0 },

package/src/commands/services.ts

@@ -1,6 +1,6 @@
 import { includedTools } from "../agents/tools/list";
 import * as allTools from "../agents/tools";
-import { LazyToolsService, services } from "../services";
+import { LazyToolsService, services, MinimalToolsService } from "../services";
 import { agents } from "../agents";
 import { ModulesService } from "../services/modules";
 

package/src/processors/MinimalToolsMessageProcessor.ts

@@ -0,0 +1,113 @@
+import { Message, ToolCall } from "../clients/types";
+import { MessageProcessorFunction } from "../services/MessageProcessor";
+
+/**
+ * MinimalToolsMessageProcessor
+ *
+ * Works with MinimalToolsService to unwrap `callTool(name, args)` invocations
+ * in the message history, rewriting them as if the real tool had been called
+ * directly.
+ *
+ * This is registered on the `pre_tools` lifecycle so it runs right before tool
+ * calls are dispatched. It rewrites the last assistant message's tool_calls so
+ * that:
+ *   callTool({ name: "readFile", args: { filePath: "..." } })
+ * becomes:
+ *   readFile({ filePath: "..." })
+ *
+ * This means:
+ * 1. The `tools:` array sent to the AI never needs to change (cache stable).
+ * 2. The agent terminates correctly when callTool("finalAnswer", ...) is called,
+ *    because after transformation the required tool name matches.
+ * 3. The conversation history looks clean — the model sees native tool calls in
+ *    subsequent turns rather than nested callTool wrappers.
+ */
+export class MinimalToolsMessageProcessor {
+  /**
+   * Unwraps a single callTool ToolCall into the real underlying ToolCall.
+   * Returns null if this is not a callTool invocation.
+   */
+  private unwrapCallTool(toolCall: ToolCall): ToolCall | null {
+    if (toolCall.function.name !== "callTool") {
+      return null;
+    }
+
+    let parsed: { name?: string; args?: Record<string, any> };
+    try {
+      parsed =
+        typeof toolCall.function.arguments === "string"
+          ? JSON.parse(toolCall.function.arguments)
+          : toolCall.function.arguments;
+    } catch {
+      return null;
+    }
+
+    const { name, args } = parsed;
+    if (!name || typeof name !== "string") {
+      return null;
+    }
+
+    return {
+      id: toolCall.id,
+      type: "function",
+      function: {
+        name,
+        arguments: JSON.stringify(args ?? {}),
+      },
+    };
+  }
+
+  /**
+   * Processes the modified messages array in-place:
+   * - Finds assistant messages whose tool_calls contain callTool invocations.
+   * - Replaces those callTool entries with the real underlying tool calls.
+   *
+   * Only the most-recent assistant message (the one just added by the AI) is
+   * rewritten — older messages are left untouched so caching is not affected.
+   */
+  private processMessages(
+    _originalMessages: Message[],
+    modifiedMessages: Message[]
+  ): void {
+    // Walk backwards to find the last assistant message with tool_calls
+    for (let i = modifiedMessages.length - 1; i >= 0; i--) {
+      const msg = modifiedMessages[i];
+      if (msg.role !== "assistant" || !msg.tool_calls?.length) {
+        continue;
+      }
+
+      let didUnwrap = false;
+      const newToolCalls: ToolCall[] = msg.tool_calls.map((tc) => {
+        const unwrapped = this.unwrapCallTool(tc);
+        if (unwrapped) {
+          didUnwrap = true;
+          return unwrapped;
+        }
+        return tc;
+      });
+
+      if (didUnwrap) {
+        msg.tool_calls = newToolCalls;
+      }
+
+      // Only process the last assistant message
+      break;
+    }
+  }
+
+  /**
+   * Returns a MessageProcessorFunction suitable for registering on the
+   * `pre_tools` lifecycle of a MessageProcessor.
+   */
+  createProcessor(): MessageProcessorFunction {
+    return (
+      originalMessages: Message[],
+      modifiedMessages: Message[]
+    ): void => {
+      this.processMessages(originalMessages, modifiedMessages);
+    };
+  }
+}
+
+export const globalMinimalToolsMessageProcessor =
+  new MinimalToolsMessageProcessor();