contextos-mcp 0.4.0

package/dist/prompts/orchestrator.d.ts

@@ -0,0 +1,6 @@
+/**
+ * Orchestrator system prompt — teaches the LLM how to orchestrate swarm tasks
+ * via declarative, typed JSON actions.
+ */
+import type { SwarmConfig } from "../core/types.js";
+export declare function buildSwarmSystemPrompt(config: SwarmConfig, agentDescriptions?: string, language?: "javascript" | "python" | "json"): string;

package/dist/prompts/orchestrator.js

@@ -0,0 +1,114 @@
+/**
+ * Orchestrator system prompt — teaches the LLM how to orchestrate swarm tasks
+ * via declarative, typed JSON actions.
+ */
+export function buildSwarmSystemPrompt(config, agentDescriptions, language) {
+    if (language === "python") {
+        return `You are a text analysis agent assisting with codebase decomposition. Analyze the provided context and return structured findings.`;
+    }
+    return `You are a Swarm Orchestrator — an AI agent coordinating parallel coding agents across isolated git worktrees using declarative JSON actions.
+
+## Action Protocol
+
+You do not write or execute arbitrary scripts. Every turn you MUST respond with a single JSON action object conforming to version 1 of the action schema.
+
+Available actions:
+1. \`spawn\` — Spawn a subagent thread in an isolated git worktree with an explicit write scope.
+2. \`wait\` — Await completion of specified thread IDs.
+3. \`inspect_diff\` — Inspect the git diff produced by a thread.
+4. \`review\` — Mark a thread as reviewed after verifying changes.
+5. \`merge\` — Merge a thread's branch into the workspace.
+6. \`finish\` — Conclude orchestration and return the final report.
+
+## Action Schemas
+
+### 1. \`spawn\`
+\`\`\`json
+{
+  "version": 1,
+  "action": "spawn",
+  "task": "Precise description of the task for the subagent",
+  "writeScope": ["src/module.ts", "tests/module.test.ts"],
+  "model": "${config.default_model}"
+}
+\`\`\`
+- \`writeScope\`: Array of repository-relative file paths the agent is permitted to touch.
+- \`model\`: Optional model override.
+
+### 2. \`wait\`
+\`\`\`json
+{
+  "version": 1,
+  "action": "wait",
+  "threadIds": ["thread_id_1", "thread_id_2"]
+}
+\`\`\`
+
+### 3. \`inspect_diff\`
+\`\`\`json
+{
+  "version": 1,
+  "action": "inspect_diff",
+  "threadId": "thread_id_1"
+}
+\`\`\`
+
+### 4. \`review\`
+\`\`\`json
+{
+  "version": 1,
+  "action": "review",
+  "threadId": "thread_id_1"
+}
+\`\`\`
+
+### 5. \`merge\`
+\`\`\`json
+{
+  "version": 1,
+  "action": "merge",
+  "threadId": "thread_id_1"
+}
+\`\`\`
+
+### 6. \`finish\`
+\`\`\`json
+{
+  "version": 1,
+  "action": "finish",
+  "summary": "Comprehensive summary of changes, verification results, and next steps."
+}
+\`\`\`
+
+${agentDescriptions
+        ? `## Available Agents
+
+${agentDescriptions}
+
+`
+        : ""}## Orchestration Rules
+
+1. Decompose the objective into independent tasks with explicit, isolated \`writeScope\` file lists.
+2. Spawn tasks using \`spawn\`.
+3. Use \`wait\` to check completion of running threads.
+4. Use \`inspect_diff\` to inspect the exact diff produced by each thread.
+5. Review the changes with \`review\`.
+6. Merge valid completed threads with \`merge\`.
+7. Conclude the session with \`finish\`.
+8. Never output executable code, scripts, or placeholders.
+9. Always respond with a single JSON object matching the action schema.
+
+## Output format
+
+Respond with ONLY a JSON code block. No explanations before or after.
+
+\`\`\`json
+{
+  "version": 1,
+  "action": "spawn",
+  "task": "Example task",
+  "writeScope": ["src/example.ts"]
+}
+\`\`\``;
+}
+//# sourceMappingURL=orchestrator.js.map

package/dist/prompts/orchestrator.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"orchestrator.js","sourceRoot":"","sources":["../../src/prompts/orchestrator.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAIH,MAAM,UAAU,sBAAsB,CACrC,MAAmB,EACnB,iBAA0B,EAC1B,QAA2C;IAE3C,IAAI,QAAQ,KAAK,QAAQ,EAAE,CAAC;QAC3B,OAAO,mIAAmI,CAAC;IAC5I,CAAC;IAED,OAAO;;;;;;;;;;;;;;;;;;;;;;;cAuBM,MAAM,CAAC,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EAoDjC,iBAAiB;QAChB,CAAC,CAAC;;EAEF,iBAAiB;;CAElB;QACC,CAAC,CAAC,EACJ;;;;;;;;;;;;;;;;;;;;;;;OAuBO,CAAC;AACR,CAAC"}

package/dist/routing/model-resolver.d.ts

@@ -0,0 +1,27 @@
+/**
+ * Model resolver — resolves model IDs to pi-ai Model objects.
+ *
+ * Handles three cases:
+ *   1. Standard pi-ai models (anthropic, openai, google) — looked up from pi-ai registry
+ *   2. Ollama models (ollama/*) — creates synthetic Model<"openai-completions"> pointing at localhost:11434
+ *   3. OpenRouter models (openrouter/*) — creates synthetic Model<"openai-completions"> pointing at openrouter.ai
+ *
+ * This preserves the RLM loop for all backends — the orchestrator always uses pi-ai's completeSimple().
+ */
+import { type Model } from "@mariozechner/pi-ai";
+type Api = "openai-completions" | "openai-responses";
+export interface ResolvedModel {
+    model: Model<Api>;
+    provider: string;
+}
+/**
+ * Resolve a model ID to a pi-ai Model object.
+ *
+ * Supports:
+ *   - "ollama/deepseek-coder-v2" → Ollama local model
+ *   - "openrouter/auto" → OpenRouter cloud model
+ *   - "claude-sonnet-4-6" → standard pi-ai model lookup
+ *   - Falls back to any available provider's default model
+ */
+export declare function resolveModel(modelId: string, warnFn?: (msg: string) => void): ResolvedModel | null;
+export {};

package/dist/routing/model-resolver.js

@@ -0,0 +1,265 @@
+/**
+ * Model resolver — resolves model IDs to pi-ai Model objects.
+ *
+ * Handles three cases:
+ *   1. Standard pi-ai models (anthropic, openai, google) — looked up from pi-ai registry
+ *   2. Ollama models (ollama/*) — creates synthetic Model<"openai-completions"> pointing at localhost:11434
+ *   3. OpenRouter models (openrouter/*) — creates synthetic Model<"openai-completions"> pointing at openrouter.ai
+ *
+ * This preserves the RLM loop for all backends — the orchestrator always uses pi-ai's completeSimple().
+ */
+import { getEnvApiKey, getModels, getProviders } from "@mariozechner/pi-ai";
+// Patch global fetch to bypass WAFs that block Node.js by default
+const originalFetch = globalThis.fetch;
+globalThis.fetch = async function (url, options) {
+    if (url.toString().includes("air-outer.com") || url.toString().includes("agentrouter.org")) {
+        options = options || {};
+        options.headers = {
+            ...options.headers,
+            "User-Agent": "OpenCode",
+        };
+    }
+    return originalFetch.apply(this, [url, options]);
+};
+const DEFAULT_MODELS = {
+    anthropic: "claude-sonnet-4-6",
+    openai: "gpt-4o",
+    google: "gemini-2.5-flash",
+};
+/**
+ * Create a synthetic pi-ai Model for Ollama (OpenAI-compatible API at localhost:11434).
+ *
+ * Uses provider "openai" so pi-ai's API key lookup resolves to OPENAI_API_KEY.
+ * We set a dummy OPENAI_API_KEY if none exists — Ollama ignores auth headers.
+ */
+function createOllamaModel(modelId) {
+    const shortId = modelId.replace("ollama/", "");
+    // pi-ai requires an API key for the "openai" provider. Ollama doesn't need one,
+    // but we must satisfy pi-ai's check. Set a dummy key if no real one exists.
+    if (!process.env.OPENAI_API_KEY) {
+        process.env.OPENAI_API_KEY = "ollama-local";
+    }
+    return {
+        id: shortId,
+        name: shortId,
+        api: "openai-completions",
+        provider: "openai",
+        baseUrl: "http://localhost:11434/v1",
+        reasoning: false,
+        input: ["text"],
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+        contextWindow: 32768,
+        maxTokens: 4096,
+        compat: {
+            supportsStore: false,
+            supportsDeveloperRole: false,
+            supportsReasoningEffort: false,
+            supportsUsageInStreaming: false,
+            maxTokensField: "max_tokens",
+            requiresToolResultName: false,
+            requiresAssistantAfterToolResult: false,
+            requiresThinkingAsText: false,
+            requiresMistralToolIds: false,
+            thinkingFormat: "openai",
+            supportsStrictMode: false,
+        },
+    };
+}
+/**
+ * Create a synthetic pi-ai Model for OpenRouter (OpenAI-compatible API).
+ */
+function createOpenRouterModel(modelId) {
+    const shortId = modelId.replace("openrouter/", "");
+    const apiKey = process.env.OPENROUTER_API_KEY || "";
+    return {
+        id: shortId,
+        name: shortId,
+        api: "openai-completions",
+        provider: "openrouter",
+        baseUrl: "https://openrouter.ai/api/v1",
+        reasoning: false,
+        input: ["text"],
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+        contextWindow: 128000,
+        maxTokens: 4096,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+            "HTTP-Referer": "https://github.com/kok-o/contextos-agents",
+            "X-Title": "contextos-agents",
+        },
+        compat: {
+            supportsStore: false,
+            supportsDeveloperRole: false,
+            supportsReasoningEffort: false,
+            supportsUsageInStreaming: true,
+            maxTokensField: "max_tokens",
+            requiresToolResultName: false,
+            requiresAssistantAfterToolResult: false,
+            requiresThinkingAsText: false,
+            requiresMistralToolIds: false,
+            thinkingFormat: "openai",
+            supportsStrictMode: false,
+            openRouterRouting: undefined,
+        },
+    };
+}
+/**
+ * Create a synthetic pi-ai Model for AgentRouter (OpenAI-compatible API).
+ */
+function createAgentRouterModel(modelId) {
+    const shortId = modelId.replace("agentrouter/", "");
+    const apiKey = process.env.OPENAI_API_KEY || "";
+    return {
+        id: shortId,
+        name: shortId,
+        api: "openai-completions",
+        provider: "openai",
+        baseUrl: "https://agentrouter.org/v1",
+        reasoning: false,
+        input: ["text"],
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+        contextWindow: 128000,
+        maxTokens: 4096,
+        headers: {
+            Authorization: `Bearer ${apiKey}`,
+        },
+        compat: {
+            supportsStore: false,
+            supportsDeveloperRole: false,
+            supportsReasoningEffort: false,
+            supportsUsageInStreaming: true,
+            maxTokensField: "max_tokens",
+            requiresToolResultName: false,
+            requiresAssistantAfterToolResult: false,
+            requiresThinkingAsText: false,
+            requiresMistralToolIds: false,
+            thinkingFormat: "openai",
+            supportsStrictMode: false,
+        },
+    };
+}
+/**
+ * Resolve a model ID to a pi-ai Model object.
+ *
+ * Supports:
+ *   - "ollama/deepseek-coder-v2" → Ollama local model
+ *   - "openrouter/auto" → OpenRouter cloud model
+ *   - "claude-sonnet-4-6" → standard pi-ai model lookup
+ *   - Falls back to any available provider's default model
+ */
+export function resolveModel(modelId, warnFn) {
+    // Ollama models — create synthetic model
+    if (modelId.startsWith("ollama/")) {
+        return { model: createOllamaModel(modelId), provider: "ollama" };
+    }
+    // OpenRouter models — create synthetic model
+    if (modelId.startsWith("openrouter/")) {
+        return { model: createOpenRouterModel(modelId), provider: "openrouter" };
+    }
+    // AgentRouter models — create synthetic model
+    if (modelId.startsWith("agentrouter/")) {
+        return { model: createAgentRouterModel(modelId), provider: "agentrouter" };
+    }
+    // Transparent proxy via OPENAI_BASE_URL
+    if (process.env.OPENAI_BASE_URL && !modelId.includes("/")) {
+        const apiKey = process.env.OPENAI_API_KEY || "";
+        const syntheticModel = {
+            id: modelId,
+            name: modelId,
+            api: "openai-completions",
+            provider: "openai",
+            baseUrl: process.env.OPENAI_BASE_URL,
+            reasoning: false,
+            input: ["text"],
+            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+            contextWindow: 128000,
+            maxTokens: 4096,
+            headers: apiKey
+                ? {
+                    Authorization: `Bearer ${apiKey}`,
+                    "User-Agent": "OpenCode",
+                }
+                : {
+                    "User-Agent": "OpenCode",
+                },
+            compat: {
+                supportsStore: false,
+                supportsDeveloperRole: false,
+                supportsReasoningEffort: false,
+                supportsUsageInStreaming: true,
+                maxTokensField: "max_tokens",
+                requiresToolResultName: false,
+                requiresAssistantAfterToolResult: false,
+                requiresThinkingAsText: false,
+                requiresMistralToolIds: false,
+                thinkingFormat: "openai",
+                supportsStrictMode: false,
+            },
+        };
+        return { model: syntheticModel, provider: "openai" };
+    }
+    // Standard pi-ai model lookup
+    const knownProviders = new Set(["anthropic", "openai", "google", "groq", "cerebras", "xai", "openrouter", "mistral"]);
+    let model;
+    let resolvedProvider = "";
+    // Try known providers with API keys first
+    for (const provider of getProviders()) {
+        if (!knownProviders.has(provider))
+            continue;
+        if (!getEnvApiKey(provider))
+            continue;
+        for (const m of getModels(provider)) {
+            if (m.id === modelId) {
+                model = m;
+                resolvedProvider = provider;
+                break;
+            }
+        }
+        if (model)
+            break;
+    }
+    // Try unknown providers
+    if (!model) {
+        for (const provider of getProviders()) {
+            if (knownProviders.has(provider))
+                continue;
+            for (const m of getModels(provider)) {
+                if (m.id === modelId) {
+                    model = m;
+                    resolvedProvider = provider;
+                    break;
+                }
+            }
+            if (model)
+                break;
+        }
+    }
+    // Fallback: try default model for any provider that has a key
+    if (!model) {
+        for (const [prov, fallbackId] of Object.entries(DEFAULT_MODELS)) {
+            if (!getEnvApiKey(prov))
+                continue;
+            for (const p of getProviders()) {
+                if (p !== prov)
+                    continue;
+                for (const m of getModels(p)) {
+                    if (m.id === fallbackId) {
+                        model = m;
+                        resolvedProvider = prov;
+                        if (warnFn)
+                            warnFn(`Using ${fallbackId} (${prov}) — model "${modelId}" not found`);
+                        break;
+                    }
+                }
+                if (model)
+                    break;
+            }
+            if (model)
+                break;
+        }
+    }
+    if (!model)
+        return null;
+    return { model, provider: resolvedProvider };
+}
+//# sourceMappingURL=model-resolver.js.map

package/dist/routing/model-resolver.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-resolver.js","sourceRoot":"","sources":["../../src/routing/model-resolver.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,EAAE,YAAY,EAAE,SAAS,EAAE,YAAY,EAAc,MAAM,qBAAqB,CAAC;AAExF,kEAAkE;AAClE,MAAM,aAAa,GAAG,UAAU,CAAC,KAAK,CAAC;AACvC,UAAU,CAAC,KAAK,GAAG,KAAK,WAAW,GAAQ,EAAE,OAAqB;IACjE,IAAI,GAAG,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAAC,eAAe,CAAC,IAAI,GAAG,CAAC,QAAQ,EAAE,CAAC,QAAQ,CAAC,iBAAiB,CAAC,EAAE,CAAC;QAC5F,OAAO,GAAG,OAAO,IAAI,EAAE,CAAC;QACxB,OAAO,CAAC,OAAO,GAAG;YACjB,GAAG,OAAO,CAAC,OAAO;YAClB,YAAY,EAAE,UAAU;SACxB,CAAC;IACH,CAAC;IACD,OAAO,aAAa,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC,CAAC;AAClD,CAAC,CAAC;AAIF,MAAM,cAAc,GAA2B;IAC9C,SAAS,EAAE,mBAAmB;IAC9B,MAAM,EAAE,QAAQ;IAChB,MAAM,EAAE,kBAAkB;CAC1B,CAAC;AAEF;;;;;GAKG;AACH,SAAS,iBAAiB,CAAC,OAAe;IACzC,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC;IAE/C,gFAAgF;IAChF,4EAA4E;IAC5E,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,cAAc,EAAE,CAAC;QACjC,OAAO,CAAC,GAAG,CAAC,cAAc,GAAG,cAAc,CAAC;IAC7C,CAAC;IAED,OAAO;QACN,EAAE,EAAE,OAAO;QACX,IAAI,EAAE,OAAO;QACb,GAAG,EAAE,oBAAoB;QACzB,QAAQ,EAAE,QAAQ;QAClB,OAAO,EAAE,2BAA2B;QACpC,SAAS,EAAE,KAAK;QAChB,KAAK,EAAE,CAAC,MAAM,CAAC;QACf,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE;QAC1D,aAAa,EAAE,KAAK;QACpB,SAAS,EAAE,IAAI;QACf,MAAM,EAAE;YACP,aAAa,EAAE,KAAK;YACpB,qBAAqB,EAAE,KAAK;YAC5B,uBAAuB,EAAE,KAAK;YAC9B,wBAAwB,EAAE,KAAK;YAC/B,cAAc,EAAE,YAAY;YAC5B,sBAAsB,EAAE,KAAK;YAC7B,gCAAgC,EAAE,KAAK;YACvC,sBAAsB,EAAE,KAAK;YAC7B,sBAAsB,EAAE,KAAK;YAC7B,cAAc,EAAE,QAAQ;YACxB,kBAAkB,EAAE,KAAK;SACzB;KACD,CAAC;AACH,CAAC;AAED;;GAEG;AACH,SAAS,qBAAqB,CAAC,OAAe;IAC7C,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,aAAa,EAAE,EAAE,CAAC,CAAC;IACnD,MAAM,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,kBAAkB,IAAI,EAAE,CAAC;IACpD,OAAO;QACN,EAAE,EAAE,OAAO;QACX,IAAI,EAAE,OAAO;QACb,GAAG,EAAE,oBAAoB;QACzB,QAAQ,EAAE,YAAY;QACtB,OAAO,EAAE,8BAA8B;QACvC,SAAS,EAAE,KAAK;QAChB,KAAK,EAAE,CAAC,MAAM,CAAC;QACf,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE;QAC1D,aAAa,EAAE,MAAM;QACrB,SAAS,EAAE,IAAI;QACf,OAAO,EAAE;YACR,aAAa,EAAE,UAAU,MAAM,EAAE;YACjC,cAAc,EAAE,2CAA2C;YAC3D,SAAS,EAAE,kBAAkB;SAC7B;QACD,MAAM,EAAE;YACP,aAAa,EAAE,KAAK;YACpB,qBAAqB,EAAE,KAAK;YAC5B,uBAAuB,EAAE,KAAK;YAC9B,wBAAwB,EAAE,IAAI;YAC9B,cAAc,EAAE,YAAY;YAC5B,sBAAsB,EAAE,KAAK;YAC7B,gCAAgC,EAAE,KAAK;YACvC,sBAAsB,EAAE,KAAK;YAC7B,sBAAsB,EAAE,KAAK;YAC7B,cAAc,EAAE,QAAQ;YACxB,kBAAkB,EAAE,KAAK;YACzB,iBAAiB,EAAE,SAAS;SAC5B;KACD,CAAC;AACH,CAAC;AAED;;GAEG;AACH,SAAS,sBAAsB,CAAC,OAAe;IAC9C,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,cAAc,EAAE,EAAE,CAAC,CAAC;IACpD,MAAM,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,cAAc,IAAI,EAAE,CAAC;IAChD,OAAO;QACN,EAAE,EAAE,OAAO;QACX,IAAI,EAAE,OAAO;QACb,GAAG,EAAE,oBAAoB;QACzB,QAAQ,EAAE,QAAQ;QAClB,OAAO,EAAE,4BAA4B;QACrC,SAAS,EAAE,KAAK;QAChB,KAAK,EAAE,CAAC,MAAM,CAAC;QACf,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE;QAC1D,aAAa,EAAE,MAAM;QACrB,SAAS,EAAE,IAAI;QACf,OAAO,EAAE;YACR,aAAa,EAAE,UAAU,MAAM,EAAE;SACjC;QACD,MAAM,EAAE;YACP,aAAa,EAAE,KAAK;YACpB,qBAAqB,EAAE,KAAK;YAC5B,uBAAuB,EAAE,KAAK;YAC9B,wBAAwB,EAAE,IAAI;YAC9B,cAAc,EAAE,YAAY;YAC5B,sBAAsB,EAAE,KAAK;YAC7B,gCAAgC,EAAE,KAAK;YACvC,sBAAsB,EAAE,KAAK;YAC7B,sBAAsB,EAAE,KAAK;YAC7B,cAAc,EAAE,QAAQ;YACxB,kBAAkB,EAAE,KAAK;SACzB;KACD,CAAC;AACH,CAAC;AAOD;;;;;;;;GAQG;AACH,MAAM,UAAU,YAAY,CAAC,OAAe,EAAE,MAA8B;IAC3E,yCAAyC;IACzC,IAAI,OAAO,CAAC,UAAU,CAAC,SAAS,CAAC,EAAE,CAAC;QACnC,OAAO,EAAE,KAAK,EAAE,iBAAiB,CAAC,OAAO,CAAC,EAAE,QAAQ,EAAE,QAAQ,EAAE,CAAC;IAClE,CAAC;IAED,6CAA6C;IAC7C,IAAI,OAAO,CAAC,UAAU,CAAC,aAAa,CAAC,EAAE,CAAC;QACvC,OAAO,EAAE,KAAK,EAAE,qBAAqB,CAAC,OAAO,CAAC,EAAE,QAAQ,EAAE,YAAY,EAAE,CAAC;IAC1E,CAAC;IAED,8CAA8C;IAC9C,IAAI,OAAO,CAAC,UAAU,CAAC,cAAc,CAAC,EAAE,CAAC;QACxC,OAAO,EAAE,KAAK,EAAE,sBAAsB,CAAC,OAAO,CAAC,EAAE,QAAQ,EAAE,aAAa,EAAE,CAAC;IAC5E,CAAC;IAED,wCAAwC;IACxC,IAAI,OAAO,CAAC,GAAG,CAAC,eAAe,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAC,EAAE,CAAC;QAC3D,MAAM,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,cAAc,IAAI,EAAE,CAAC;QAChD,MAAM,cAAc,GAAgC;YACnD,EAAE,EAAE,OAAO;YACX,IAAI,EAAE,OAAO;YACb,GAAG,EAAE,oBAAoB;YACzB,QAAQ,EAAE,QAAQ;YAClB,OAAO,EAAE,OAAO,CAAC,GAAG,CAAC,eAAe;YACpC,SAAS,EAAE,KAAK;YAChB,KAAK,EAAE,CAAC,MAAM,CAAC;YACf,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE;YAC1D,aAAa,EAAE,MAAM;YACrB,SAAS,EAAE,IAAI;YACf,OAAO,EAAE,MAAM;gBACd,CAAC,CAAC;oBACA,aAAa,EAAE,UAAU,MAAM,EAAE;oBACjC,YAAY,EAAE,UAAU;iBACxB;gBACF,CAAC,CAAC;oBACA,YAAY,EAAE,UAAU;iBACxB;YACH,MAAM,EAAE;gBACP,aAAa,EAAE,KAAK;gBACpB,qBAAqB,EAAE,KAAK;gBAC5B,uBAAuB,EAAE,KAAK;gBAC9B,wBAAwB,EAAE,IAAI;gBAC9B,cAAc,EAAE,YAAY;gBAC5B,sBAAsB,EAAE,KAAK;gBAC7B,gCAAgC,EAAE,KAAK;gBACvC,sBAAsB,EAAE,KAAK;gBAC7B,sBAAsB,EAAE,KAAK;gBAC7B,cAAc,EAAE,QAAQ;gBACxB,kBAAkB,EAAE,KAAK;aACzB;SACD,CAAC;QACF,OAAO,EAAE,KAAK,EAAE,cAAc,EAAE,QAAQ,EAAE,QAAQ,EAAE,CAAC;IACtD,CAAC;IAED,8BAA8B;IAC9B,MAAM,cAAc,GAAG,IAAI,GAAG,CAAC,CAAC,WAAW,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,UAAU,EAAE,KAAK,EAAE,YAAY,EAAE,SAAS,CAAC,CAAC,CAAC;IACtH,IAAI,KAA6B,CAAC;IAClC,IAAI,gBAAgB,GAAG,EAAE,CAAC;IAE1B,0CAA0C;IAC1C,KAAK,MAAM,QAAQ,IAAI,YAAY,EAAE,EAAE,CAAC;QACvC,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,QAAQ,CAAC;YAAE,SAAS;QAC5C,IAAI,CAAC,YAAY,CAAC,QAAQ,CAAC;YAAE,SAAS;QACtC,KAAK,MAAM,CAAC,IAAI,SAAS,CAAC,QAAQ,CAAC,EAAE,CAAC;YACrC,IAAI,CAAC,CAAC,EAAE,KAAK,OAAO,EAAE,CAAC;gBACtB,KAAK,GAAG,CAAC,CAAC;gBACV,gBAAgB,GAAG,QAAQ,CAAC;gBAC5B,MAAM;YACP,CAAC;QACF,CAAC;QACD,IAAI,KAAK;YAAE,MAAM;IAClB,CAAC;IAED,wBAAwB;IACxB,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,KAAK,MAAM,QAAQ,IAAI,YAAY,EAAE,EAAE,CAAC;YACvC,IAAI,cAAc,CAAC,GAAG,CAAC,QAAQ,CAAC;gBAAE,SAAS;YAC3C,KAAK,MAAM,CAAC,IAAI,SAAS,CAAC,QAAQ,CAAC,EAAE,CAAC;gBACrC,IAAI,CAAC,CAAC,EAAE,KAAK,OAAO,EAAE,CAAC;oBACtB,KAAK,GAAG,CAAC,CAAC;oBACV,gBAAgB,GAAG,QAAQ,CAAC;oBAC5B,MAAM;gBACP,CAAC;YACF,CAAC;YACD,IAAI,KAAK;gBAAE,MAAM;QAClB,CAAC;IACF,CAAC;IAED,8DAA8D;IAC9D,IAAI,CAAC,KAAK,EAAE,CAAC;QACZ,KAAK,MAAM,CAAC,IAAI,EAAE,UAAU,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,cAAc,CAAC,EAAE,CAAC;YACjE,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC;gBAAE,SAAS;YAClC,KAAK,MAAM,CAAC,IAAI,YAAY,EAAE,EAAE,CAAC;gBAChC,IAAI,CAAC,KAAK,IAAI;oBAAE,SAAS;gBACzB,KAAK,MAAM,CAAC,IAAI,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC;oBAC9B,IAAI,CAAC,CAAC,EAAE,KAAK,UAAU,EAAE,CAAC;wBACzB,KAAK,GAAG,CAAC,CAAC;wBACV,gBAAgB,GAAG,IAAI,CAAC;wBACxB,IAAI,MAAM;4BAAE,MAAM,CAAC,SAAS,UAAU,KAAK,IAAI,cAAc,OAAO,aAAa,CAAC,CAAC;wBACnF,MAAM;oBACP,CAAC;gBACF,CAAC;gBACD,IAAI,KAAK;oBAAE,MAAM;YAClB,CAAC;YACD,IAAI,KAAK;gBAAE,MAAM;QAClB,CAAC;IACF,CAAC;IAED,IAAI,CAAC,KAAK;QAAE,OAAO,IAAI,CAAC;IACxB,OAAO,EAAE,KAAK,EAAE,QAAQ,EAAE,gBAAgB,EAAE,CAAC;AAC9C,CAAC"}

package/dist/routing/model-router.d.ts

@@ -0,0 +1,130 @@
+/**
+ * Model router — auto-selects the best agent + model combination per task.
+ *
+ * Inspired by Slate's approach: the orchestrator LLM naturally selects agents
+ * and models based on task complexity. This module provides a fallback
+ * rule-based router for when auto_model_selection is enabled, ensuring
+ * cost-efficient defaults even when the orchestrator doesn't explicitly choose.
+ *
+ * Two routing dimensions:
+ *   1. Task complexity (simple/medium/complex) → picks model tier
+ *   2. Task slot (execution/search/reasoning/planning) → picks agent + model specialty
+ *
+ * Two routing modes:
+ *   1. Orchestrator-driven (default): The orchestrator prompt teaches the LLM
+ *      about agent strengths, and the LLM passes agent/model in thread() calls.
+ *   2. Auto-routing (auto_model_selection=true): This router overrides the
+ *      orchestrator's choice with a cost-optimal selection based on task analysis.
+ *
+ * Enhanced with:
+ *   - FailureTracker: session-level failure tracking with decay weighting
+ *   - Success rate weighting: penalizes agents with high failure rates
+ *   - File pattern matching: boosts agents that historically handle specific file types
+ *   - Aggregate episodic stats: fallback to best-performing agent per slot
+ */
+import type { ModelSlots } from "../config.js";
+import type { SwarmConfig } from "../core/types.js";
+import type { EpisodicMemory } from "../labs/episodic-memory.js";
+export interface AgentCapability {
+    name: string;
+    /** Cost tier: 1 = cheapest, 5 = most expensive */
+    costTier: number;
+    /** Speed tier: 1 = fastest, 5 = slowest */
+    speedTier: number;
+    /** What this agent excels at */
+    strengths: string[];
+    /** Best model for this agent (provider/model-id format) */
+    defaultModel: string;
+    /** Cheaper model for simple tasks */
+    cheapModel: string;
+    /** Premium model for complex tasks */
+    premiumModel: string;
+}
+export declare const AGENT_CAPABILITIES: Record<string, AgentCapability>;
+export type TaskComplexity = "simple" | "medium" | "complex";
+/** Simple keyword-based complexity classifier. */
+export declare function classifyTaskComplexity(task: string): TaskComplexity;
+/**
+ * Task slots — named categories that map to specialized model/agent combos.
+ * Inspired by Slate's model slots (main, subagent, search, reasoning).
+ */
+export type TaskSlot = "execution" | "search" | "reasoning" | "planning";
+export declare const DEFAULT_MODEL_SLOTS: ModelSlots;
+/** Classify a task into a named slot based on keyword analysis. */
+export declare function classifyTaskSlot(task: string): TaskSlot;
+export interface FailureRecord {
+    agent: string;
+    model: string;
+    task: string;
+    error: string;
+    timestamp: number;
+    /** Transient errors (rate limit, timeout) decay faster than permanent ones. */
+    isTransient: boolean;
+}
+/**
+ * Tracks agent/model failures within a session to inform routing decisions.
+ *
+ * Failure records decay over time — recent failures weigh more heavily.
+ * Transient errors (rate limits, timeouts) decay faster than permanent ones.
+ */
+export declare class FailureTracker {
+    private failures;
+    /** Half-life for permanent failure decay (ms). Failures lose half their weight after this. */
+    private readonly permanentHalfLifeMs;
+    /** Half-life for transient failure decay (ms). Shorter — transient issues resolve quickly. */
+    private readonly transientHalfLifeMs;
+    constructor(permanentHalfLifeMs?: number, transientHalfLifeMs?: number);
+    /**
+     * Record a failure for an agent+model pair.
+     * Classifies the error as transient or permanent based on pattern matching.
+     */
+    recordFailure(agent: string, model: string, task: string, error: string): void;
+    /**
+     * Get the weighted failure rate for an agent+model pair (0-1).
+     *
+     * Uses exponential decay so recent failures count more than old ones.
+     * The rate is capped at 1.0 (effectively: agent is completely unreliable).
+     */
+    getFailureRate(agent: string, model?: string): number;
+    /**
+     * Check if an agent has a 100% failure rate (all recent attempts failed,
+     * with no significant decay). Used to skip completely broken agents.
+     */
+    isFullyFailed(agent: string): boolean;
+    /** Get all failure records (for debugging/inspection). */
+    getFailures(): FailureRecord[];
+    /** Get the number of raw (undecayed) failures for an agent. */
+    getFailureCount(agent: string): number;
+    /** Clear all failure records. */
+    clear(): void;
+}
+/**
+ * Extract file extensions mentioned in a task description.
+ * Returns unique lowercase extensions (e.g., [".ts", ".py"]).
+ */
+export declare function extractFileExtensions(task: string): string[];
+export interface RouteResult {
+    agent: string;
+    model: string;
+    slot: TaskSlot;
+    reason: string;
+}
+/**
+ * Route a task to the best agent + model combination.
+ *
+ * Logic:
+ *   1. Classify task complexity (simple/medium/complex)
+ *   2. Classify task slot (execution/search/reasoning/planning)
+ *   3. Check for slot-specific model overrides in config
+ *   4. Score agents with slot preference bonus
+ *   5. Apply failure rate penalty (skip fully-failed agents)
+ *   6. Apply file pattern matching bonus from episodic memory
+ *   7. Fallback to aggregate episodic stats when no high-confidence match
+ *   8. Pick the highest-scoring capable option
+ */
+export declare function routeTask(task: string, config: SwarmConfig, memory?: EpisodicMemory, failureTracker?: FailureTracker): Promise<RouteResult>;
+/**
+ * Get a description of available agents and their capabilities
+ * for inclusion in the orchestrator prompt.
+ */
+export declare function describeAvailableAgents(): Promise<string>;