titan-agent 7.0.0 → 7.2.0

package/dist/providers/moa.js

@@ -0,0 +1,151 @@
+#!/usr/bin/env node
+import { LLMProvider } from "./base.js";
+import { loadConfig } from "../config/config.js";
+import logger from "../utils/logger.js";
+const COMPONENT = "MoA";
+const DEFAULT_REFERENCE_TIMEOUT_MS = 45e3;
+const DEFAULT_REFERENCE_MAX_TOKENS = 1024;
+const TOOL_RESULT_TRIM = 4e3;
+const ADVISOR_SYSTEM_PROMPT = [
+  "You are a reference advisor inside a Mixture-of-Agents process. You are NOT the acting agent:",
+  "you cannot call tools, and the user never sees your words. Another model (the aggregator) acts",
+  "and answers \u2014 your output is private guidance handed to it. Give concise, concrete advice:",
+  "the likely best next steps, tool-use strategy, risks, and anything the aggregator might miss.",
+  "If the conversation references tools or systems you cannot access, advise from reasoning alone \u2014",
+  "never refuse for lack of access."
+].join(" ");
+function trimForReference(messages) {
+  const out = [{ role: "system", content: ADVISOR_SYSTEM_PROMPT }];
+  for (const m of messages) {
+    if (m.role === "system") continue;
+    if (m.role === "tool") {
+      const body = (m.content || "").slice(0, TOOL_RESULT_TRIM);
+      out.push({ role: "user", content: `[Tool result]
+${body}` });
+      continue;
+    }
+    if (m.role === "assistant" && m.toolCalls?.length) {
+      const calls = m.toolCalls.map((tc) => `${tc.function.name}(${(tc.function.arguments || "").slice(0, 300)})`).join("; ");
+      out.push({ role: "assistant", content: `${m.content || ""}
+[Called tools: ${calls}]`.trim() });
+      continue;
+    }
+    out.push({ role: m.role, content: m.content });
+  }
+  return out;
+}
+function buildGuidanceBlock(results) {
+  const blocks = results.map((r, i) => {
+    if (r.content) return `Reference ${i + 1} \u2014 ${r.model}:
+${r.content}`;
+    return `Reference ${i + 1} \u2014 ${r.model}: (unavailable: ${r.error || "no response"})`;
+  });
+  return [
+    "[Mixture of Agents context \u2014 private guidance from reference advisors. The user cannot see this.",
+    "Weigh it critically; you are the acting agent and the only voice.]",
+    "",
+    ...blocks
+  ].join("\n");
+}
+function getMoaPresets() {
+  try {
+    const cfg = loadConfig();
+    return cfg.moa?.presets || {};
+  } catch {
+    return {};
+  }
+}
+class MoaProvider extends LLMProvider {
+  name = "moa";
+  displayName = "Mixture of Agents";
+  async healthCheck() {
+    return Object.keys(getMoaPresets()).length > 0;
+  }
+  async listModels() {
+    return Object.keys(getMoaPresets()).map((p) => `moa/${p}`);
+  }
+  resolvePreset(model) {
+    const presets = getMoaPresets();
+    const name = (model || "").replace(/^moa\//, "") || "default";
+    const preset = presets[name];
+    if (!preset) throw new Error(`MoA preset "${name}" not found. Configure moa.presets.${name} in titan.json`);
+    if (preset.references.some((r) => r.model.startsWith("moa/"))) {
+      throw new Error("MoA presets cannot reference other MoA presets (recursion blocked)");
+    }
+    if (preset.aggregator.startsWith("moa/")) {
+      throw new Error("MoA aggregator cannot itself be a MoA preset (recursion blocked)");
+    }
+    return { name, preset };
+  }
+  /** Fan out to advisors; never throws — failures become guidance footnotes. */
+  async gatherGuidance(preset, messages, signal) {
+    const { chat } = await import("./router.js");
+    const trimmed = trimForReference(messages);
+    const tasks = preset.references.map(async (ref) => {
+      const started = Date.now();
+      const timeoutMs = ref.timeoutMs ?? DEFAULT_REFERENCE_TIMEOUT_MS;
+      const controller = new AbortController();
+      const timer = setTimeout(() => controller.abort(), timeoutMs);
+      const onOuter = () => controller.abort();
+      signal?.addEventListener("abort", onOuter, { once: true });
+      try {
+        const res = await chat({
+          model: ref.model,
+          messages: trimmed,
+          maxTokens: ref.maxTokens ?? DEFAULT_REFERENCE_MAX_TOKENS,
+          temperature: ref.temperature,
+          signal: controller.signal,
+          noFallback: true
+        });
+        return { model: ref.model, content: res.content || null, ms: Date.now() - started };
+      } catch (e) {
+        return { model: ref.model, content: null, error: e.message.slice(0, 160), ms: Date.now() - started };
+      } finally {
+        clearTimeout(timer);
+        signal?.removeEventListener("abort", onOuter);
+      }
+    });
+    return Promise.all(tasks);
+  }
+  async prepare(options) {
+    const { name, preset } = this.resolvePreset(options.model);
+    if (preset.enabled === false) {
+      return { preset, name, aggregatorOptions: { ...options, model: preset.aggregator } };
+    }
+    const t0 = Date.now();
+    const guidance = await this.gatherGuidance(preset, options.messages, options.signal);
+    const ok = guidance.filter((g) => g.content).length;
+    logger.info(COMPONENT, `[${name}] ${ok}/${preset.references.length} advisors answered in ${Date.now() - t0}ms (${guidance.map((g) => `${g.model.split("/").pop()}:${g.ms}ms`).join(", ")})`);
+    try {
+      const { titanEvents } = await import("../agent/daemon.js");
+      titanEvents.emit("moa:guidance", { preset: name, advisors: guidance.map((g) => ({ model: g.model, ok: !!g.content, ms: g.ms })) });
+    } catch {
+    }
+    const messages = [
+      ...options.messages,
+      { role: "user", content: buildGuidanceBlock(guidance) }
+    ];
+    return {
+      preset,
+      name,
+      aggregatorOptions: { ...options, model: preset.aggregator, messages, maxTokens: preset.maxTokens ?? options.maxTokens }
+    };
+  }
+  async chat(options) {
+    const { chat } = await import("./router.js");
+    const { aggregatorOptions } = await this.prepare(options);
+    return chat(aggregatorOptions);
+  }
+  async *chatStream(options) {
+    const { chatStream } = await import("./router.js");
+    const { aggregatorOptions } = await this.prepare(options);
+    yield* chatStream(aggregatorOptions);
+  }
+}
+export {
+  MoaProvider,
+  buildGuidanceBlock,
+  getMoaPresets,
+  trimForReference
+};
+//# sourceMappingURL=moa.js.map

package/dist/providers/moa.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/providers/moa.ts"],"sourcesContent":["/**\n * MoA — Mixture-of-Agents virtual provider (v7.1).\n *\n * A MoA preset = N reference \"advisors\" + 1 aggregator, exposed as a virtual\n * model (`moa/<preset>`) selectable anywhere a model is. On every chat call:\n *\n *   1. Advisors receive a TRIMMED conversation (no TITAN system prompt, no\n *      tool schemas, tool results truncated) plus an advisory system prompt —\n *      they advise, they never act.\n *   2. Advisors fan out in parallel with per-reference timeouts; a slow or\n *      failed advisor is dropped/noted, never awaited into oblivion.\n *   3. Their labeled guidance is appended at the END of the aggregator's\n *      prompt (prompt-cache prefix stays stable).\n *   4. The aggregator — the only voice — answers with the FULL tool schema\n *      through the normal loop.\n *\n * TITAN's edge over the same feature elsewhere: advisors default to LOCAL\n * models spread across separate machines (a \"local council\"), so the\n * mixture costs nothing and parallelizes across GPUs; and because whole\n * agents can be advisors via MCP, the mixture isn't limited to bare models.\n */\nimport { LLMProvider, type ChatOptions, type ChatResponse, type ChatStreamChunk, type ChatMessage } from './base.js';\nimport { loadConfig } from '../config/config.js';\nimport logger from '../utils/logger.js';\n\nconst COMPONENT = 'MoA';\nconst DEFAULT_REFERENCE_TIMEOUT_MS = 45_000;\nconst DEFAULT_REFERENCE_MAX_TOKENS = 1_024;\nconst TOOL_RESULT_TRIM = 4_000;\n\nexport interface MoaReference {\n    /** provider/model id, e.g. \"litellm/glm-4.7-flash\" */\n    model: string;\n    maxTokens?: number;\n    temperature?: number;\n    timeoutMs?: number;\n}\n\nexport interface MoaPreset {\n    references: MoaReference[];\n    /** provider/model id of the acting aggregator */\n    aggregator: string;\n    enabled?: boolean;\n    /** Optional cap for the aggregator's output */\n    maxTokens?: number;\n}\n\nconst ADVISOR_SYSTEM_PROMPT = [\n    'You are a reference advisor inside a Mixture-of-Agents process. You are NOT the acting agent:',\n    'you cannot call tools, and the user never sees your words. Another model (the aggregator) acts',\n    'and answers — your output is private guidance handed to it. Give concise, concrete advice:',\n    'the likely best next steps, tool-use strategy, risks, and anything the aggregator might miss.',\n    'If the conversation references tools or systems you cannot access, advise from reasoning alone —',\n    'never refuse for lack of access.',\n].join(' ');\n\n/** Trim the conversation for advisors: no system prompt, no tool schemas,\n *  tool transcripts flattened + truncated. Exported for tests. */\nexport function trimForReference(messages: ChatMessage[]): ChatMessage[] {\n    const out: ChatMessage[] = [{ role: 'system', content: ADVISOR_SYSTEM_PROMPT }];\n    for (const m of messages) {\n        if (m.role === 'system') continue; // TITAN's own harness prompt stays private\n        if (m.role === 'tool') {\n            const body = (m.content || '').slice(0, TOOL_RESULT_TRIM);\n            out.push({ role: 'user', content: `[Tool result]\\n${body}` } as ChatMessage);\n            continue;\n        }\n        if (m.role === 'assistant' && m.toolCalls?.length) {\n            const calls = m.toolCalls.map(tc => `${tc.function.name}(${(tc.function.arguments || '').slice(0, 300)})`).join('; ');\n            out.push({ role: 'assistant', content: `${m.content || ''}\\n[Called tools: ${calls}]`.trim() } as ChatMessage);\n            continue;\n        }\n        out.push({ role: m.role, content: m.content } as ChatMessage);\n    }\n    return out;\n}\n\n/** Build the guidance block appended to the aggregator's prompt. Exported for tests. */\nexport function buildGuidanceBlock(results: Array<{ model: string; content: string | null; error?: string }>): string {\n    const blocks = results.map((r, i) => {\n        if (r.content) return `Reference ${i + 1} — ${r.model}:\\n${r.content}`;\n        return `Reference ${i + 1} — ${r.model}: (unavailable: ${r.error || 'no response'})`;\n    });\n    return [\n        '[Mixture of Agents context — private guidance from reference advisors. The user cannot see this.',\n        'Weigh it critically; you are the acting agent and the only voice.]',\n        '',\n        ...blocks,\n    ].join('\\n');\n}\n\nexport function getMoaPresets(): Record<string, MoaPreset> {\n    try {\n        const cfg = loadConfig() as { moa?: { presets?: Record<string, MoaPreset> } };\n        return cfg.moa?.presets || {};\n    } catch {\n        return {};\n    }\n}\n\nexport class MoaProvider extends LLMProvider {\n    readonly name = 'moa';\n    readonly displayName = 'Mixture of Agents';\n\n    async healthCheck(): Promise<boolean> {\n        return Object.keys(getMoaPresets()).length > 0;\n    }\n\n    async listModels(): Promise<string[]> {\n        return Object.keys(getMoaPresets()).map(p => `moa/${p}`);\n    }\n\n    private resolvePreset(model?: string): { name: string; preset: MoaPreset } {\n        const presets = getMoaPresets();\n        const name = (model || '').replace(/^moa\\//, '') || 'default';\n        const preset = presets[name];\n        if (!preset) throw new Error(`MoA preset \"${name}\" not found. Configure moa.presets.${name} in titan.json`);\n        if (preset.references.some(r => r.model.startsWith('moa/'))) {\n            throw new Error('MoA presets cannot reference other MoA presets (recursion blocked)');\n        }\n        if (preset.aggregator.startsWith('moa/')) {\n            throw new Error('MoA aggregator cannot itself be a MoA preset (recursion blocked)');\n        }\n        return { name, preset };\n    }\n\n    /** Fan out to advisors; never throws — failures become guidance footnotes. */\n    private async gatherGuidance(preset: MoaPreset, messages: ChatMessage[], signal?: AbortSignal):\n        Promise<Array<{ model: string; content: string | null; error?: string; ms: number }>> {\n        const { chat } = await import('./router.js');\n        const trimmed = trimForReference(messages);\n        const tasks = preset.references.map(async ref => {\n            const started = Date.now();\n            const timeoutMs = ref.timeoutMs ?? DEFAULT_REFERENCE_TIMEOUT_MS;\n            const controller = new AbortController();\n            const timer = setTimeout(() => controller.abort(), timeoutMs);\n            const onOuter = () => controller.abort();\n            signal?.addEventListener('abort', onOuter, { once: true });\n            try {\n                const res = await chat({\n                    model: ref.model,\n                    messages: trimmed,\n                    maxTokens: ref.maxTokens ?? DEFAULT_REFERENCE_MAX_TOKENS,\n                    temperature: ref.temperature,\n                    signal: controller.signal,\n                    noFallback: true,\n                } as ChatOptions & { noFallback: boolean });\n                return { model: ref.model, content: res.content || null, ms: Date.now() - started };\n            } catch (e) {\n                return { model: ref.model, content: null, error: (e as Error).message.slice(0, 160), ms: Date.now() - started };\n            } finally {\n                clearTimeout(timer);\n                signal?.removeEventListener('abort', onOuter);\n            }\n        });\n        return Promise.all(tasks);\n    }\n\n    private async prepare(options: ChatOptions): Promise<{ preset: MoaPreset; name: string; aggregatorOptions: ChatOptions }> {\n        const { name, preset } = this.resolvePreset(options.model);\n        if (preset.enabled === false) {\n            // Preset disabled → aggregator acts alone (Hermes-compatible semantics)\n            return { preset, name, aggregatorOptions: { ...options, model: preset.aggregator } };\n        }\n        const t0 = Date.now();\n        const guidance = await this.gatherGuidance(preset, options.messages, options.signal);\n        const ok = guidance.filter(g => g.content).length;\n        logger.info(COMPONENT, `[${name}] ${ok}/${preset.references.length} advisors answered in ${Date.now() - t0}ms (${guidance.map(g => `${g.model.split('/').pop()}:${g.ms}ms`).join(', ')})`);\n\n        // Emit an observable event for the UI (mascot/Live Studio can narrate)\n        try {\n            const { titanEvents } = await import('../agent/daemon.js');\n            titanEvents.emit('moa:guidance', { preset: name, advisors: guidance.map(g => ({ model: g.model, ok: !!g.content, ms: g.ms })) });\n        } catch { /* events are best-effort */ }\n\n        // Append guidance at the END (prompt-cache prefix stays stable)\n        const messages: ChatMessage[] = [\n            ...options.messages,\n            { role: 'user', content: buildGuidanceBlock(guidance) } as ChatMessage,\n        ];\n        return {\n            preset, name,\n            aggregatorOptions: { ...options, model: preset.aggregator, messages, maxTokens: preset.maxTokens ?? options.maxTokens },\n        };\n    }\n\n    async chat(options: ChatOptions): Promise<ChatResponse> {\n        const { chat } = await import('./router.js');\n        const { aggregatorOptions } = await this.prepare(options);\n        return chat(aggregatorOptions);\n    }\n\n    async *chatStream(options: ChatOptions): AsyncGenerator<ChatStreamChunk> {\n        const { chatStream } = await import('./router.js');\n        const { aggregatorOptions } = await this.prepare(options);\n        yield* chatStream(aggregatorOptions);\n    }\n}\n"],"mappings":";AAqBA,SAAS,mBAAgG;AACzG,SAAS,kBAAkB;AAC3B,OAAO,YAAY;AAEnB,MAAM,YAAY;AAClB,MAAM,+BAA+B;AACrC,MAAM,+BAA+B;AACrC,MAAM,mBAAmB;AAmBzB,MAAM,wBAAwB;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACJ,EAAE,KAAK,GAAG;AAIH,SAAS,iBAAiB,UAAwC;AACrE,QAAM,MAAqB,CAAC,EAAE,MAAM,UAAU,SAAS,sBAAsB,CAAC;AAC9E,aAAW,KAAK,UAAU;AACtB,QAAI,EAAE,SAAS,SAAU;AACzB,QAAI,EAAE,SAAS,QAAQ;AACnB,YAAM,QAAQ,EAAE,WAAW,IAAI,MAAM,GAAG,gBAAgB;AACxD,UAAI,KAAK,EAAE,MAAM,QAAQ,SAAS;AAAA,EAAkB,IAAI,GAAG,CAAgB;AAC3E;AAAA,IACJ;AACA,QAAI,EAAE,SAAS,eAAe,EAAE,WAAW,QAAQ;AAC/C,YAAM,QAAQ,EAAE,UAAU,IAAI,QAAM,GAAG,GAAG,SAAS,IAAI,KAAK,GAAG,SAAS,aAAa,IAAI,MAAM,GAAG,GAAG,CAAC,GAAG,EAAE,KAAK,IAAI;AACpH,UAAI,KAAK,EAAE,MAAM,aAAa,SAAS,GAAG,EAAE,WAAW,EAAE;AAAA,iBAAoB,KAAK,IAAI,KAAK,EAAE,CAAgB;AAC7G;AAAA,IACJ;AACA,QAAI,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,EAAE,QAAQ,CAAgB;AAAA,EAChE;AACA,SAAO;AACX;AAGO,SAAS,mBAAmB,SAAmF;AAClH,QAAM,SAAS,QAAQ,IAAI,CAAC,GAAG,MAAM;AACjC,QAAI,EAAE,QAAS,QAAO,aAAa,IAAI,CAAC,WAAM,EAAE,KAAK;AAAA,EAAM,EAAE,OAAO;AACpE,WAAO,aAAa,IAAI,CAAC,WAAM,EAAE,KAAK,mBAAmB,EAAE,SAAS,aAAa;AAAA,EACrF,CAAC;AACD,SAAO;AAAA,IACH;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAG;AAAA,EACP,EAAE,KAAK,IAAI;AACf;AAEO,SAAS,gBAA2C;AACvD,MAAI;AACA,UAAM,MAAM,WAAW;AACvB,WAAO,IAAI,KAAK,WAAW,CAAC;AAAA,EAChC,QAAQ;AACJ,WAAO,CAAC;AAAA,EACZ;AACJ;AAEO,MAAM,oBAAoB,YAAY;AAAA,EAChC,OAAO;AAAA,EACP,cAAc;AAAA,EAEvB,MAAM,cAAgC;AAClC,WAAO,OAAO,KAAK,cAAc,CAAC,EAAE,SAAS;AAAA,EACjD;AAAA,EAEA,MAAM,aAAgC;AAClC,WAAO,OAAO,KAAK,cAAc,CAAC,EAAE,IAAI,OAAK,OAAO,CAAC,EAAE;AAAA,EAC3D;AAAA,EAEQ,cAAc,OAAqD;AACvE,UAAM,UAAU,cAAc;AAC9B,UAAM,QAAQ,SAAS,IAAI,QAAQ,UAAU,EAAE,KAAK;AACpD,UAAM,SAAS,QAAQ,IAAI;AAC3B,QAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,eAAe,IAAI,sCAAsC,IAAI,gBAAgB;AAC1G,QAAI,OAAO,WAAW,KAAK,OAAK,EAAE,MAAM,WAAW,MAAM,CAAC,GAAG;AACzD,YAAM,IAAI,MAAM,oEAAoE;AAAA,IACxF;AACA,QAAI,OAAO,WAAW,WAAW,MAAM,GAAG;AACtC,YAAM,IAAI,MAAM,kEAAkE;AAAA,IACtF;AACA,WAAO,EAAE,MAAM,OAAO;AAAA,EAC1B;AAAA;AAAA,EAGA,MAAc,eAAe,QAAmB,UAAyB,QACiB;AACtF,UAAM,EAAE,KAAK,IAAI,MAAM,OAAO,aAAa;AAC3C,UAAM,UAAU,iBAAiB,QAAQ;AACzC,UAAM,QAAQ,OAAO,WAAW,IAAI,OAAM,QAAO;AAC7C,YAAM,UAAU,KAAK,IAAI;AACzB,YAAM,YAAY,IAAI,aAAa;AACnC,YAAM,aAAa,IAAI,gBAAgB;AACvC,YAAM,QAAQ,WAAW,MAAM,WAAW,MAAM,GAAG,SAAS;AAC5D,YAAM,UAAU,MAAM,WAAW,MAAM;AACvC,cAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AACzD,UAAI;AACA,cAAM,MAAM,MAAM,KAAK;AAAA,UACnB,OAAO,IAAI;AAAA,UACX,UAAU;AAAA,UACV,WAAW,IAAI,aAAa;AAAA,UAC5B,aAAa,IAAI;AAAA,UACjB,QAAQ,WAAW;AAAA,UACnB,YAAY;AAAA,QAChB,CAA0C;AAC1C,eAAO,EAAE,OAAO,IAAI,OAAO,SAAS,IAAI,WAAW,MAAM,IAAI,KAAK,IAAI,IAAI,QAAQ;AAAA,MACtF,SAAS,GAAG;AACR,eAAO,EAAE,OAAO,IAAI,OAAO,SAAS,MAAM,OAAQ,EAAY,QAAQ,MAAM,GAAG,GAAG,GAAG,IAAI,KAAK,IAAI,IAAI,QAAQ;AAAA,MAClH,UAAE;AACE,qBAAa,KAAK;AAClB,gBAAQ,oBAAoB,SAAS,OAAO;AAAA,MAChD;AAAA,IACJ,CAAC;AACD,WAAO,QAAQ,IAAI,KAAK;AAAA,EAC5B;AAAA,EAEA,MAAc,QAAQ,SAAoG;AACtH,UAAM,EAAE,MAAM,OAAO,IAAI,KAAK,cAAc,QAAQ,KAAK;AACzD,QAAI,OAAO,YAAY,OAAO;AAE1B,aAAO,EAAE,QAAQ,MAAM,mBAAmB,EAAE,GAAG,SAAS,OAAO,OAAO,WAAW,EAAE;AAAA,IACvF;AACA,UAAM,KAAK,KAAK,IAAI;AACpB,UAAM,WAAW,MAAM,KAAK,eAAe,QAAQ,QAAQ,UAAU,QAAQ,MAAM;AACnF,UAAM,KAAK,SAAS,OAAO,OAAK,EAAE,OAAO,EAAE;AAC3C,WAAO,KAAK,WAAW,IAAI,IAAI,KAAK,EAAE,IAAI,OAAO,WAAW,MAAM,yBAAyB,KAAK,IAAI,IAAI,EAAE,OAAO,SAAS,IAAI,OAAK,GAAG,EAAE,MAAM,MAAM,GAAG,EAAE,IAAI,CAAC,IAAI,EAAE,EAAE,IAAI,EAAE,KAAK,IAAI,CAAC,GAAG;AAGzL,QAAI;AACA,YAAM,EAAE,YAAY,IAAI,MAAM,OAAO,oBAAoB;AACzD,kBAAY,KAAK,gBAAgB,EAAE,QAAQ,MAAM,UAAU,SAAS,IAAI,QAAM,EAAE,OAAO,EAAE,OAAO,IAAI,CAAC,CAAC,EAAE,SAAS,IAAI,EAAE,GAAG,EAAE,EAAE,CAAC;AAAA,IACnI,QAAQ;AAAA,IAA+B;AAGvC,UAAM,WAA0B;AAAA,MAC5B,GAAG,QAAQ;AAAA,MACX,EAAE,MAAM,QAAQ,SAAS,mBAAmB,QAAQ,EAAE;AAAA,IAC1D;AACA,WAAO;AAAA,MACH;AAAA,MAAQ;AAAA,MACR,mBAAmB,EAAE,GAAG,SAAS,OAAO,OAAO,YAAY,UAAU,WAAW,OAAO,aAAa,QAAQ,UAAU;AAAA,IAC1H;AAAA,EACJ;AAAA,EAEA,MAAM,KAAK,SAA6C;AACpD,UAAM,EAAE,KAAK,IAAI,MAAM,OAAO,aAAa;AAC3C,UAAM,EAAE,kBAAkB,IAAI,MAAM,KAAK,QAAQ,OAAO;AACxD,WAAO,KAAK,iBAAiB;AAAA,EACjC;AAAA,EAEA,OAAO,WAAW,SAAuD;AACrE,UAAM,EAAE,WAAW,IAAI,MAAM,OAAO,aAAa;AACjD,UAAM,EAAE,kBAAkB,IAAI,MAAM,KAAK,QAAQ,OAAO;AACxD,WAAO,WAAW,iBAAiB;AAAA,EACvC;AACJ;","names":[]}

package/dist/providers/modelCapabilities.js

@@ -1,5 +1,8 @@
 #!/usr/bin/env node
 import { loadConfig } from "../config/config.js";
+import { existsSync, readFileSync, writeFileSync, mkdirSync } from "fs";
+import { join } from "path";
+import { TITAN_HOME } from "../utils/constants.js";
 const STATIC_TABLE = {
   "anthropic/claude-sonnet-4-20250514": { contextWindow: 2e5, maxOutput: 64e3, supportsThinking: true },
   "anthropic/claude-opus-4": { contextWindow: 2e5, maxOutput: 32e3, supportsThinking: true },
@@ -36,6 +39,10 @@ const FAMILY_DEFAULTS = [
   { pattern: /minimax-m3/i, caps: { contextWindow: 1e6, maxOutput: 4e4 } },
   { pattern: /minimax/i, caps: { contextWindow: 2e5, maxOutput: 4e4 } },
   { pattern: /\bglm-5/i, caps: { contextWindow: 198e3, maxOutput: 128e3 } },
+  { pattern: /glm-4\.7/i, caps: { contextWindow: 198e3, maxOutput: 96e3 } },
+  { pattern: /qwen3-coder/i, caps: { contextWindow: 262144, maxOutput: 65536 } },
+  { pattern: /nemotron-nano/i, caps: { contextWindow: 131072, maxOutput: 32e3 } },
+  { pattern: /gemma-4/i, caps: { contextWindow: 131072, maxOutput: 32768 } },
   { pattern: /kimi-k2/i, caps: { contextWindow: 262144, maxOutput: 128e3 } },
   { pattern: /nemotron-3/i, caps: { contextWindow: 1e6, maxOutput: 32e3 } },
   { pattern: /^anthropic\//, caps: { contextWindow: 2e5, maxOutput: 8192 } },
@@ -44,9 +51,57 @@ const FAMILY_DEFAULTS = [
   { pattern: /^kimi\/|^moonshot\//, caps: { contextWindow: 262144, maxOutput: 128e3 } },
   { pattern: /^zhipu\/|^glm/, caps: { contextWindow: 198e3, maxOutput: 128e3 } },
   { pattern: /^gemini\//, caps: { contextWindow: 1e6, maxOutput: 8192 } },
+  { pattern: /:cloud\b/, caps: { contextWindow: 262144, maxOutput: 32768 } },
   { pattern: /^ollama\//, caps: { contextWindow: 32e3, maxOutput: 8192 } }
 ];
 const DEFAULT_FALLBACK = { contextWindow: 32e3, maxOutput: 8192 };
+const LEARNED_PATH = join(TITAN_HOME, "learned-model-limits.json");
+let learnedCache = null;
+function readLearned() {
+  if (learnedCache) return learnedCache;
+  try {
+    if (existsSync(LEARNED_PATH)) {
+      learnedCache = JSON.parse(readFileSync(LEARNED_PATH, "utf-8"));
+      return learnedCache;
+    }
+  } catch {
+  }
+  learnedCache = {};
+  return learnedCache;
+}
+function recordLearnedContextWindow(model, contextWindow) {
+  if (!Number.isFinite(contextWindow) || contextWindow < 1024) return;
+  const learned = readLearned();
+  const existing = learned[model]?.contextWindow;
+  if (existing !== void 0 && existing <= contextWindow) return;
+  learned[model] = { contextWindow, learnedAt: (/* @__PURE__ */ new Date()).toISOString() };
+  try {
+    mkdirSync(TITAN_HOME, { recursive: true });
+    writeFileSync(LEARNED_PATH, JSON.stringify(learned, null, 2));
+  } catch {
+  }
+}
+function resetLearnedLimitsCache() {
+  learnedCache = null;
+}
+function getModelCapabilitiesEx(model) {
+  if (STATIC_TABLE[model]) return { caps: STATIC_TABLE[model], source: "static" };
+  try {
+    const cfg = loadConfig();
+    const override = cfg.providers?.modelCapabilities;
+    if (override?.[model]) return { caps: override[model], source: "override" };
+  } catch {
+  }
+  const learned = readLearned()[model];
+  for (const f of FAMILY_DEFAULTS) {
+    if (f.pattern.test(model)) {
+      if (learned) return { caps: { ...f.caps, contextWindow: Math.min(learned.contextWindow, f.caps.contextWindow) }, source: "learned" };
+      return { caps: f.caps, source: "family" };
+    }
+  }
+  if (learned) return { caps: { ...DEFAULT_FALLBACK, contextWindow: Math.min(learned.contextWindow, DEFAULT_FALLBACK.contextWindow) }, source: "learned" };
+  return { caps: DEFAULT_FALLBACK, source: "fallback" };
+}
 function getModelCapabilities(model) {
   if (STATIC_TABLE[model]) return STATIC_TABLE[model];
   try {
@@ -55,9 +110,14 @@ function getModelCapabilities(model) {
     if (override?.[model]) return override[model];
   } catch {
   }
+  const learned = readLearned()[model];
   for (const f of FAMILY_DEFAULTS) {
-    if (f.pattern.test(model)) return f.caps;
+    if (f.pattern.test(model)) {
+      if (learned) return { ...f.caps, contextWindow: Math.min(learned.contextWindow, f.caps.contextWindow) };
+      return f.caps;
+    }
   }
+  if (learned) return { ...DEFAULT_FALLBACK, contextWindow: Math.min(learned.contextWindow, DEFAULT_FALLBACK.contextWindow) };
   return DEFAULT_FALLBACK;
 }
 function clampMaxTokens(model, requested) {
@@ -67,6 +127,9 @@ function clampMaxTokens(model, requested) {
 }
 export {
   clampMaxTokens,
-  getModelCapabilities
+  getModelCapabilities,
+  getModelCapabilitiesEx,
+  recordLearnedContextWindow,
+  resetLearnedLimitsCache
 };
 //# sourceMappingURL=modelCapabilities.js.map

package/dist/providers/modelCapabilities.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/providers/modelCapabilities.ts"],"sourcesContent":["/**\n * TITAN — Per-Model Capability Registry\n *\n * Single source of truth for output-token caps, context-window sizes,\n * and feature flags per model.  Providers call clampMaxTokens() so\n * DEFAULT_MAX_TOKENS can safely be a high user-preference ceiling\n * (200 K) without causing 400s on capped providers.\n */\nimport { loadConfig } from '../config/config.js';\n\nexport interface ModelCapabilities {\n    contextWindow: number;\n    maxOutput: number;\n    supportsThinking?: boolean;\n}\n\n// Exact-match table.  Keep in alphabetical order by provider / model.\nconst STATIC_TABLE: Record<string, ModelCapabilities> = {\n    'anthropic/claude-sonnet-4-20250514':   { contextWindow: 200_000, maxOutput: 64_000, supportsThinking: true },\n    'anthropic/claude-opus-4':              { contextWindow: 200_000, maxOutput: 32_000, supportsThinking: true },\n    'anthropic/claude-3-5-sonnet-20241022': { contextWindow: 200_000, maxOutput: 8_192 },\n    'anthropic/claude-3-5-haiku-20241022':  { contextWindow: 200_000, maxOutput: 8_192 },\n    'openai/gpt-4o':                        { contextWindow: 128_000, maxOutput: 16_384 },\n    'openai/gpt-4o-mini':                   { contextWindow: 128_000, maxOutput: 16_384 },\n    'openai/gpt-4.1':                       { contextWindow: 1_000_000, maxOutput: 32_768 },\n    'openai/o1':                            { contextWindow: 200_000, maxOutput: 100_000, supportsThinking: true },\n    'openai/o1-mini':                       { contextWindow: 128_000, maxOutput: 65_536, supportsThinking: true },\n    'openai/o3-mini':                       { contextWindow: 200_000, maxOutput: 100_000, supportsThinking: true },\n    'kimi/kimi-k2.6':                       { contextWindow: 262_144, maxOutput: 128_000, supportsThinking: true },\n    'moonshot/kimi-k2-0905-preview':        { contextWindow: 262_144, maxOutput: 128_000 },\n    'zhipu/glm-5.1':                        { contextWindow: 198_000, maxOutput: 128_000 },\n    'ollama/qwen3.5:cloud':                 { contextWindow: 128_000, maxOutput: 32_000 },\n    'ollama/qwen3:32b':                     { contextWindow: 32_768,  maxOutput: 8_192 },\n    'ollama/llama3.3:70b':                  { contextWindow: 128_000, maxOutput: 8_192 },\n    'ollama/deepseek-v3':                   { contextWindow: 128_000, maxOutput: 16_384 },\n    'mistral/mistral-large':                { contextWindow: 128_000, maxOutput: 8_192 },\n    'gemini/gemini-2.0-flash':              { contextWindow: 1_000_000, maxOutput: 8_192 },\n    'gemini/gemini-2.5-pro':                { contextWindow: 2_000_000, maxOutput: 65_536, supportsThinking: true },\n    'cohere/command-r-plus':                { contextWindow: 128_000, maxOutput: 4_096 },\n    'groq/llama-3.3-70b-versatile':         { contextWindow: 128_000, maxOutput: 32_768 },\n};\n\n// Family heuristic for unknown specific versions.\nconst FAMILY_DEFAULTS: Array<{ pattern: RegExp; caps: ModelCapabilities }> = [\n    // v7.0 — bare-id families for the 2026 agentic models. These match WITH OR\n    // WITHOUT a provider prefix (substring, not anchored), so models routed\n    // through the generic openai_compat provider or the local :4000 gateway\n    // (e.g. \"qwen3.6-35b-a3b\", \"deepseek-v4-pro\", \"kimi-k2.6\") get their real\n    // ceilings instead of the 32K/8K DEFAULT_FALLBACK (which truncates JSON).\n    // Placed FIRST so they win over the broad `^ollama/` catch-all below.\n    { pattern: /deepseek-v4/i,                caps: { contextWindow: 128_000, maxOutput: 64_000, supportsThinking: true } },\n    { pattern: /qwen3\\.(6|7)/i,               caps: { contextWindow: 262_144, maxOutput: 32_768, supportsThinking: true } },\n    { pattern: /minimax-m3/i,                 caps: { contextWindow: 1_000_000, maxOutput: 40_000 } },\n    { pattern: /minimax/i,                    caps: { contextWindow: 200_000, maxOutput: 40_000 } },\n    { pattern: /\\bglm-5/i,                    caps: { contextWindow: 198_000, maxOutput: 128_000 } },\n    { pattern: /kimi-k2/i,                    caps: { contextWindow: 262_144, maxOutput: 128_000 } },\n    { pattern: /nemotron-3/i,                 caps: { contextWindow: 1_000_000, maxOutput: 32_000 } },\n    { pattern: /^anthropic\\//,                caps: { contextWindow: 200_000, maxOutput: 8_192 } },\n    { pattern: /^openai\\/o\\d/,                caps: { contextWindow: 200_000, maxOutput: 100_000, supportsThinking: true } },\n    { pattern: /^openai\\//,                   caps: { contextWindow: 128_000, maxOutput: 16_384 } },\n    { pattern: /^kimi\\/|^moonshot\\//,         caps: { contextWindow: 262_144, maxOutput: 128_000 } },\n    { pattern: /^zhipu\\/|^glm/,               caps: { contextWindow: 198_000, maxOutput: 128_000 } },\n    { pattern: /^gemini\\//,                   caps: { contextWindow: 1_000_000, maxOutput: 8_192 } },\n    { pattern: /^ollama\\//,                   caps: { contextWindow: 32_000,  maxOutput: 8_192 } },\n];\n\nconst DEFAULT_FALLBACK: ModelCapabilities = { contextWindow: 32_000, maxOutput: 8_192 };\n\nexport function getModelCapabilities(model: string): ModelCapabilities {\n    if (STATIC_TABLE[model]) return STATIC_TABLE[model];\n\n    // Config override path\n    try {\n        const cfg = loadConfig();\n        const override = (cfg.providers as Record<string, unknown> | undefined)?.modelCapabilities as\n            Record<string, ModelCapabilities> | undefined;\n        if (override?.[model]) return override[model];\n    } catch { /* config not loaded yet during early boot */ }\n\n    // Family fallback\n    for (const f of FAMILY_DEFAULTS) {\n        if (f.pattern.test(model)) return f.caps;\n    }\n\n    return DEFAULT_FALLBACK;\n}\n\n/**\n * Clamp a requested maxTokens to the model's actual ceiling.\n * If requested is undefined, returns the model's own maxOutput.\n * Never returns less than 1.\n */\nexport function clampMaxTokens(model: string, requested?: number): number {\n    const caps = getModelCapabilities(model);\n    const want = requested ?? caps.maxOutput;\n    return Math.max(1, Math.min(want, caps.maxOutput));\n}\n"],"mappings":";AAQA,SAAS,kBAAkB;AAS3B,MAAM,eAAkD;AAAA,EACpD,sCAAwC,EAAE,eAAe,KAAS,WAAW,MAAQ,kBAAkB,KAAK;AAAA,EAC5G,2BAAwC,EAAE,eAAe,KAAS,WAAW,MAAQ,kBAAkB,KAAK;AAAA,EAC5G,wCAAwC,EAAE,eAAe,KAAS,WAAW,KAAM;AAAA,EACnF,uCAAwC,EAAE,eAAe,KAAS,WAAW,KAAM;AAAA,EACnF,iBAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AAAA,EACpF,sBAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AAAA,EACpF,kBAAwC,EAAE,eAAe,KAAW,WAAW,MAAO;AAAA,EACtF,aAAwC,EAAE,eAAe,KAAS,WAAW,KAAS,kBAAkB,KAAK;AAAA,EAC7G,kBAAwC,EAAE,eAAe,OAAS,WAAW,OAAQ,kBAAkB,KAAK;AAAA,EAC5G,kBAAwC,EAAE,eAAe,KAAS,WAAW,KAAS,kBAAkB,KAAK;AAAA,EAC7G,kBAAwC,EAAE,eAAe,QAAS,WAAW,OAAS,kBAAkB,KAAK;AAAA,EAC7G,iCAAwC,EAAE,eAAe,QAAS,WAAW,MAAQ;AAAA,EACrF,iBAAwC,EAAE,eAAe,OAAS,WAAW,MAAQ;AAAA,EACrF,wBAAwC,EAAE,eAAe,OAAS,WAAW,KAAO;AAAA,EACpF,oBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,uBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,sBAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AAAA,EACpF,yBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,2BAAwC,EAAE,eAAe,KAAW,WAAW,KAAM;AAAA,EACrF,yBAAwC,EAAE,eAAe,KAAW,WAAW,OAAQ,kBAAkB,KAAK;AAAA,EAC9G,yBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,gCAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AACxF;AAGA,MAAM,kBAAuE;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOzE,EAAE,SAAS,gBAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAQ,kBAAkB,KAAK,EAAE;AAAA,EACtH,EAAE,SAAS,iBAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,OAAQ,kBAAkB,KAAK,EAAE;AAAA,EACtH,EAAE,SAAS,eAA+B,MAAM,EAAE,eAAe,KAAW,WAAW,IAAO,EAAE;AAAA,EAChG,EAAE,SAAS,YAA+B,MAAM,EAAE,eAAe,KAAS,WAAW,IAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,YAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,YAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,eAA+B,MAAM,EAAE,eAAe,KAAW,WAAW,KAAO,EAAE;AAAA,EAChG,EAAE,SAAS,gBAA+B,MAAM,EAAE,eAAe,KAAS,WAAW,KAAM,EAAE;AAAA,EAC7F,EAAE,SAAS,gBAA+B,MAAM,EAAE,eAAe,KAAS,WAAW,KAAS,kBAAkB,KAAK,EAAE;AAAA,EACvH,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,uBAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,iBAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,KAAW,WAAW,KAAM,EAAE;AAAA,EAC/F,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,MAAS,WAAW,KAAM,EAAE;AACjG;AAEA,MAAM,mBAAsC,EAAE,eAAe,MAAQ,WAAW,KAAM;AAE/E,SAAS,qBAAqB,OAAkC;AACnE,MAAI,aAAa,KAAK,EAAG,QAAO,aAAa,KAAK;AAGlD,MAAI;AACA,UAAM,MAAM,WAAW;AACvB,UAAM,WAAY,IAAI,WAAmD;AAEzE,QAAI,WAAW,KAAK,EAAG,QAAO,SAAS,KAAK;AAAA,EAChD,QAAQ;AAAA,EAAgD;AAGxD,aAAW,KAAK,iBAAiB;AAC7B,QAAI,EAAE,QAAQ,KAAK,KAAK,EAAG,QAAO,EAAE;AAAA,EACxC;AAEA,SAAO;AACX;AAOO,SAAS,eAAe,OAAe,WAA4B;AACtE,QAAM,OAAO,qBAAqB,KAAK;AACvC,QAAM,OAAO,aAAa,KAAK;AAC/B,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,MAAM,KAAK,SAAS,CAAC;AACrD;","names":[]}
+{"version":3,"sources":["../../src/providers/modelCapabilities.ts"],"sourcesContent":["/**\n * TITAN — Per-Model Capability Registry\n *\n * Single source of truth for output-token caps, context-window sizes,\n * and feature flags per model.  Providers call clampMaxTokens() so\n * DEFAULT_MAX_TOKENS can safely be a high user-preference ceiling\n * (200 K) without causing 400s on capped providers.\n */\nimport { loadConfig } from '../config/config.js';\nimport { existsSync, readFileSync, writeFileSync, mkdirSync } from 'fs';\nimport { join } from 'path';\nimport { TITAN_HOME } from '../utils/constants.js';\n\nexport interface ModelCapabilities {\n    contextWindow: number;\n    maxOutput: number;\n    supportsThinking?: boolean;\n}\n\n// Exact-match table.  Keep in alphabetical order by provider / model.\nconst STATIC_TABLE: Record<string, ModelCapabilities> = {\n    'anthropic/claude-sonnet-4-20250514':   { contextWindow: 200_000, maxOutput: 64_000, supportsThinking: true },\n    'anthropic/claude-opus-4':              { contextWindow: 200_000, maxOutput: 32_000, supportsThinking: true },\n    'anthropic/claude-3-5-sonnet-20241022': { contextWindow: 200_000, maxOutput: 8_192 },\n    'anthropic/claude-3-5-haiku-20241022':  { contextWindow: 200_000, maxOutput: 8_192 },\n    'openai/gpt-4o':                        { contextWindow: 128_000, maxOutput: 16_384 },\n    'openai/gpt-4o-mini':                   { contextWindow: 128_000, maxOutput: 16_384 },\n    'openai/gpt-4.1':                       { contextWindow: 1_000_000, maxOutput: 32_768 },\n    'openai/o1':                            { contextWindow: 200_000, maxOutput: 100_000, supportsThinking: true },\n    'openai/o1-mini':                       { contextWindow: 128_000, maxOutput: 65_536, supportsThinking: true },\n    'openai/o3-mini':                       { contextWindow: 200_000, maxOutput: 100_000, supportsThinking: true },\n    'kimi/kimi-k2.6':                       { contextWindow: 262_144, maxOutput: 128_000, supportsThinking: true },\n    'moonshot/kimi-k2-0905-preview':        { contextWindow: 262_144, maxOutput: 128_000 },\n    'zhipu/glm-5.1':                        { contextWindow: 198_000, maxOutput: 128_000 },\n    'ollama/qwen3.5:cloud':                 { contextWindow: 128_000, maxOutput: 32_000 },\n    'ollama/qwen3:32b':                     { contextWindow: 32_768,  maxOutput: 8_192 },\n    'ollama/llama3.3:70b':                  { contextWindow: 128_000, maxOutput: 8_192 },\n    'ollama/deepseek-v3':                   { contextWindow: 128_000, maxOutput: 16_384 },\n    'mistral/mistral-large':                { contextWindow: 128_000, maxOutput: 8_192 },\n    'gemini/gemini-2.0-flash':              { contextWindow: 1_000_000, maxOutput: 8_192 },\n    'gemini/gemini-2.5-pro':                { contextWindow: 2_000_000, maxOutput: 65_536, supportsThinking: true },\n    'cohere/command-r-plus':                { contextWindow: 128_000, maxOutput: 4_096 },\n    'groq/llama-3.3-70b-versatile':         { contextWindow: 128_000, maxOutput: 32_768 },\n};\n\n// Family heuristic for unknown specific versions.\nconst FAMILY_DEFAULTS: Array<{ pattern: RegExp; caps: ModelCapabilities }> = [\n    // v7.0 — bare-id families for the 2026 agentic models. These match WITH OR\n    // WITHOUT a provider prefix (substring, not anchored), so models routed\n    // through the generic openai_compat provider or the local :4000 gateway\n    // (e.g. \"qwen3.6-35b-a3b\", \"deepseek-v4-pro\", \"kimi-k2.6\") get their real\n    // ceilings instead of the 32K/8K DEFAULT_FALLBACK (which truncates JSON).\n    // Placed FIRST so they win over the broad `^ollama/` catch-all below.\n    { pattern: /deepseek-v4/i,                caps: { contextWindow: 128_000, maxOutput: 64_000, supportsThinking: true } },\n    { pattern: /qwen3\\.(6|7)/i,               caps: { contextWindow: 262_144, maxOutput: 32_768, supportsThinking: true } },\n    { pattern: /minimax-m3/i,                 caps: { contextWindow: 1_000_000, maxOutput: 40_000 } },\n    { pattern: /minimax/i,                    caps: { contextWindow: 200_000, maxOutput: 40_000 } },\n    { pattern: /\\bglm-5/i,                    caps: { contextWindow: 198_000, maxOutput: 128_000 } },\n    { pattern: /glm-4\\.7/i,                   caps: { contextWindow: 198_000, maxOutput: 96_000 } },\n    { pattern: /qwen3-coder/i,                caps: { contextWindow: 262_144, maxOutput: 65_536 } },\n    { pattern: /nemotron-nano/i,              caps: { contextWindow: 131_072, maxOutput: 32_000 } },\n    { pattern: /gemma-4/i,                    caps: { contextWindow: 131_072, maxOutput: 32_768 } },\n    { pattern: /kimi-k2/i,                    caps: { contextWindow: 262_144, maxOutput: 128_000 } },\n    { pattern: /nemotron-3/i,                 caps: { contextWindow: 1_000_000, maxOutput: 32_000 } },\n    { pattern: /^anthropic\\//,                caps: { contextWindow: 200_000, maxOutput: 8_192 } },\n    { pattern: /^openai\\/o\\d/,                caps: { contextWindow: 200_000, maxOutput: 100_000, supportsThinking: true } },\n    { pattern: /^openai\\//,                   caps: { contextWindow: 128_000, maxOutput: 16_384 } },\n    { pattern: /^kimi\\/|^moonshot\\//,         caps: { contextWindow: 262_144, maxOutput: 128_000 } },\n    { pattern: /^zhipu\\/|^glm/,               caps: { contextWindow: 198_000, maxOutput: 128_000 } },\n    { pattern: /^gemini\\//,                   caps: { contextWindow: 1_000_000, maxOutput: 8_192 } },\n    { pattern: /:cloud\\b/,                    caps: { contextWindow: 262_144, maxOutput: 32_768 } },\n    { pattern: /^ollama\\//,                   caps: { contextWindow: 32_000,  maxOutput: 8_192 } },\n];\n\nconst DEFAULT_FALLBACK: ModelCapabilities = { contextWindow: 32_000, maxOutput: 8_192 };\n\n// ── Learned deployment limits (v7.1 \"truly model agnostic\") ──────────────\n// A model FAMILY may support 198K context, but the DEPLOYMENT in front of us\n// may be a vLLM serving 16K. When a backend 400 reveals its real ceiling\n// (parseMaxTokenLimit / context-overflow messages), we persist it here so\n// every later request — including tool-tier selection — fits the deployment\n// automatically. Config overrides still win; learning fills the gap.\nconst LEARNED_PATH = join(TITAN_HOME, 'learned-model-limits.json');\nlet learnedCache: Record<string, { contextWindow: number; learnedAt: string }> | null = null;\n\nfunction readLearned(): Record<string, { contextWindow: number; learnedAt: string }> {\n    if (learnedCache) return learnedCache;\n    try {\n        if (existsSync(LEARNED_PATH)) {\n            learnedCache = JSON.parse(readFileSync(LEARNED_PATH, 'utf-8'));\n            return learnedCache!;\n        }\n    } catch { /* corrupt file — relearn */ }\n    learnedCache = {};\n    return learnedCache;\n}\n\n/** Persist a deployment's real context ceiling discovered from a backend 400. */\nexport function recordLearnedContextWindow(model: string, contextWindow: number): void {\n    if (!Number.isFinite(contextWindow) || contextWindow < 1024) return;\n    const learned = readLearned();\n    const existing = learned[model]?.contextWindow;\n    // Keep the smallest observed ceiling — a deployment can't grow mid-flight,\n    // and re-serving with a bigger window is picked up by deleting the file\n    // or via the config override.\n    if (existing !== undefined && existing <= contextWindow) return;\n    learned[model] = { contextWindow, learnedAt: new Date().toISOString() };\n    try {\n        mkdirSync(TITAN_HOME, { recursive: true });\n        writeFileSync(LEARNED_PATH, JSON.stringify(learned, null, 2));\n    } catch { /* best-effort persistence */ }\n}\n\n/** Test hook: reset the learned-limits cache. */\nexport function resetLearnedLimitsCache(): void {\n    learnedCache = null;\n}\n\nexport type CapabilitySource = 'static' | 'override' | 'learned' | 'family' | 'fallback';\n\n/** Like getModelCapabilities, but reports WHERE the answer came from — a\n *  32K fallback for an unknown model is a guess, not a measurement, and\n *  context-fit tool tiering must not shrink toolsets on a guess. */\nexport function getModelCapabilitiesEx(model: string): { caps: ModelCapabilities; source: CapabilitySource } {\n    if (STATIC_TABLE[model]) return { caps: STATIC_TABLE[model], source: 'static' };\n    try {\n        const cfg = loadConfig();\n        const override = (cfg.providers as Record<string, unknown> | undefined)?.modelCapabilities as\n            Record<string, ModelCapabilities> | undefined;\n        if (override?.[model]) return { caps: override[model], source: 'override' };\n    } catch { /* early boot */ }\n    const learned = readLearned()[model];\n    for (const f of FAMILY_DEFAULTS) {\n        if (f.pattern.test(model)) {\n            if (learned) return { caps: { ...f.caps, contextWindow: Math.min(learned.contextWindow, f.caps.contextWindow) }, source: 'learned' };\n            return { caps: f.caps, source: 'family' };\n        }\n    }\n    if (learned) return { caps: { ...DEFAULT_FALLBACK, contextWindow: Math.min(learned.contextWindow, DEFAULT_FALLBACK.contextWindow) }, source: 'learned' };\n    return { caps: DEFAULT_FALLBACK, source: 'fallback' };\n}\n\nexport function getModelCapabilities(model: string): ModelCapabilities {\n    if (STATIC_TABLE[model]) return STATIC_TABLE[model];\n\n    // Config override path\n    try {\n        const cfg = loadConfig();\n        const override = (cfg.providers as Record<string, unknown> | undefined)?.modelCapabilities as\n            Record<string, ModelCapabilities> | undefined;\n        if (override?.[model]) return override[model];\n    } catch { /* config not loaded yet during early boot */ }\n\n    // Learned deployment ceiling (from real backend 400s) — beats the family\n    // guess because it reflects the endpoint actually serving this model.\n    const learned = readLearned()[model];\n\n    // Family fallback\n    for (const f of FAMILY_DEFAULTS) {\n        if (f.pattern.test(model)) {\n            if (learned) return { ...f.caps, contextWindow: Math.min(learned.contextWindow, f.caps.contextWindow) };\n            return f.caps;\n        }\n    }\n\n    if (learned) return { ...DEFAULT_FALLBACK, contextWindow: Math.min(learned.contextWindow, DEFAULT_FALLBACK.contextWindow) };\n    return DEFAULT_FALLBACK;\n}\n\n/**\n * Clamp a requested maxTokens to the model's actual ceiling.\n * If requested is undefined, returns the model's own maxOutput.\n * Never returns less than 1.\n */\nexport function clampMaxTokens(model: string, requested?: number): number {\n    const caps = getModelCapabilities(model);\n    const want = requested ?? caps.maxOutput;\n    return Math.max(1, Math.min(want, caps.maxOutput));\n}\n"],"mappings":";AAQA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,cAAc,eAAe,iBAAiB;AACnE,SAAS,YAAY;AACrB,SAAS,kBAAkB;AAS3B,MAAM,eAAkD;AAAA,EACpD,sCAAwC,EAAE,eAAe,KAAS,WAAW,MAAQ,kBAAkB,KAAK;AAAA,EAC5G,2BAAwC,EAAE,eAAe,KAAS,WAAW,MAAQ,kBAAkB,KAAK;AAAA,EAC5G,wCAAwC,EAAE,eAAe,KAAS,WAAW,KAAM;AAAA,EACnF,uCAAwC,EAAE,eAAe,KAAS,WAAW,KAAM;AAAA,EACnF,iBAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AAAA,EACpF,sBAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AAAA,EACpF,kBAAwC,EAAE,eAAe,KAAW,WAAW,MAAO;AAAA,EACtF,aAAwC,EAAE,eAAe,KAAS,WAAW,KAAS,kBAAkB,KAAK;AAAA,EAC7G,kBAAwC,EAAE,eAAe,OAAS,WAAW,OAAQ,kBAAkB,KAAK;AAAA,EAC5G,kBAAwC,EAAE,eAAe,KAAS,WAAW,KAAS,kBAAkB,KAAK;AAAA,EAC7G,kBAAwC,EAAE,eAAe,QAAS,WAAW,OAAS,kBAAkB,KAAK;AAAA,EAC7G,iCAAwC,EAAE,eAAe,QAAS,WAAW,MAAQ;AAAA,EACrF,iBAAwC,EAAE,eAAe,OAAS,WAAW,MAAQ;AAAA,EACrF,wBAAwC,EAAE,eAAe,OAAS,WAAW,KAAO;AAAA,EACpF,oBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,uBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,sBAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AAAA,EACpF,yBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,2BAAwC,EAAE,eAAe,KAAW,WAAW,KAAM;AAAA,EACrF,yBAAwC,EAAE,eAAe,KAAW,WAAW,OAAQ,kBAAkB,KAAK;AAAA,EAC9G,yBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,gCAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AACxF;AAGA,MAAM,kBAAuE;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOzE,EAAE,SAAS,gBAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAQ,kBAAkB,KAAK,EAAE;AAAA,EACtH,EAAE,SAAS,iBAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,OAAQ,kBAAkB,KAAK,EAAE;AAAA,EACtH,EAAE,SAAS,eAA+B,MAAM,EAAE,eAAe,KAAW,WAAW,IAAO,EAAE;AAAA,EAChG,EAAE,SAAS,YAA+B,MAAM,EAAE,eAAe,KAAS,WAAW,IAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,YAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,KAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,gBAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,MAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,kBAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,KAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,YAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,MAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,YAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,eAA+B,MAAM,EAAE,eAAe,KAAW,WAAW,KAAO,EAAE;AAAA,EAChG,EAAE,SAAS,gBAA+B,MAAM,EAAE,eAAe,KAAS,WAAW,KAAM,EAAE;AAAA,EAC7F,EAAE,SAAS,gBAA+B,MAAM,EAAE,eAAe,KAAS,WAAW,KAAS,kBAAkB,KAAK,EAAE;AAAA,EACvH,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,uBAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,iBAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,KAAW,WAAW,KAAM,EAAE;AAAA,EAC/F,EAAE,SAAS,YAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,MAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,MAAS,WAAW,KAAM,EAAE;AACjG;AAEA,MAAM,mBAAsC,EAAE,eAAe,MAAQ,WAAW,KAAM;AAQtF,MAAM,eAAe,KAAK,YAAY,2BAA2B;AACjE,IAAI,eAAoF;AAExF,SAAS,cAA4E;AACjF,MAAI,aAAc,QAAO;AACzB,MAAI;AACA,QAAI,WAAW,YAAY,GAAG;AAC1B,qBAAe,KAAK,MAAM,aAAa,cAAc,OAAO,CAAC;AAC7D,aAAO;AAAA,IACX;AAAA,EACJ,QAAQ;AAAA,EAA+B;AACvC,iBAAe,CAAC;AAChB,SAAO;AACX;AAGO,SAAS,2BAA2B,OAAe,eAA6B;AACnF,MAAI,CAAC,OAAO,SAAS,aAAa,KAAK,gBAAgB,KAAM;AAC7D,QAAM,UAAU,YAAY;AAC5B,QAAM,WAAW,QAAQ,KAAK,GAAG;AAIjC,MAAI,aAAa,UAAa,YAAY,cAAe;AACzD,UAAQ,KAAK,IAAI,EAAE,eAAe,YAAW,oBAAI,KAAK,GAAE,YAAY,EAAE;AACtE,MAAI;AACA,cAAU,YAAY,EAAE,WAAW,KAAK,CAAC;AACzC,kBAAc,cAAc,KAAK,UAAU,SAAS,MAAM,CAAC,CAAC;AAAA,EAChE,QAAQ;AAAA,EAAgC;AAC5C;AAGO,SAAS,0BAAgC;AAC5C,iBAAe;AACnB;AAOO,SAAS,uBAAuB,OAAsE;AACzG,MAAI,aAAa,KAAK,EAAG,QAAO,EAAE,MAAM,aAAa,KAAK,GAAG,QAAQ,SAAS;AAC9E,MAAI;AACA,UAAM,MAAM,WAAW;AACvB,UAAM,WAAY,IAAI,WAAmD;AAEzE,QAAI,WAAW,KAAK,EAAG,QAAO,EAAE,MAAM,SAAS,KAAK,GAAG,QAAQ,WAAW;AAAA,EAC9E,QAAQ;AAAA,EAAmB;AAC3B,QAAM,UAAU,YAAY,EAAE,KAAK;AACnC,aAAW,KAAK,iBAAiB;AAC7B,QAAI,EAAE,QAAQ,KAAK,KAAK,GAAG;AACvB,UAAI,QAAS,QAAO,EAAE,MAAM,EAAE,GAAG,EAAE,MAAM,eAAe,KAAK,IAAI,QAAQ,eAAe,EAAE,KAAK,aAAa,EAAE,GAAG,QAAQ,UAAU;AACnI,aAAO,EAAE,MAAM,EAAE,MAAM,QAAQ,SAAS;AAAA,IAC5C;AAAA,EACJ;AACA,MAAI,QAAS,QAAO,EAAE,MAAM,EAAE,GAAG,kBAAkB,eAAe,KAAK,IAAI,QAAQ,eAAe,iBAAiB,aAAa,EAAE,GAAG,QAAQ,UAAU;AACvJ,SAAO,EAAE,MAAM,kBAAkB,QAAQ,WAAW;AACxD;AAEO,SAAS,qBAAqB,OAAkC;AACnE,MAAI,aAAa,KAAK,EAAG,QAAO,aAAa,KAAK;AAGlD,MAAI;AACA,UAAM,MAAM,WAAW;AACvB,UAAM,WAAY,IAAI,WAAmD;AAEzE,QAAI,WAAW,KAAK,EAAG,QAAO,SAAS,KAAK;AAAA,EAChD,QAAQ;AAAA,EAAgD;AAIxD,QAAM,UAAU,YAAY,EAAE,KAAK;AAGnC,aAAW,KAAK,iBAAiB;AAC7B,QAAI,EAAE,QAAQ,KAAK,KAAK,GAAG;AACvB,UAAI,QAAS,QAAO,EAAE,GAAG,EAAE,MAAM,eAAe,KAAK,IAAI,QAAQ,eAAe,EAAE,KAAK,aAAa,EAAE;AACtG,aAAO,EAAE;AAAA,IACb;AAAA,EACJ;AAEA,MAAI,QAAS,QAAO,EAAE,GAAG,kBAAkB,eAAe,KAAK,IAAI,QAAQ,eAAe,iBAAiB,aAAa,EAAE;AAC1H,SAAO;AACX;AAOO,SAAS,eAAe,OAAe,WAA4B;AACtE,QAAM,OAAO,qBAAqB,KAAK;AACvC,QAAM,OAAO,aAAa,KAAK;AAC/B,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,MAAM,KAAK,SAAS,CAAC;AACrD;","names":[]}

package/dist/providers/ollama.js

@@ -35,6 +35,7 @@ const CLOUD_MODEL_CTX = {
   "devstral-small-2:latest": 32768,
   // Nemotron 3 Nano — 1M native, 32K practical for local
   "nemotron-3-nano": 32768,
+  "nemotron-nano": 32768,
   "nemotron-3-nano:latest": 32768,
   "nemotron-3-nano:4b": 32768,
   "nemotron-3-nano:30b": 32768,