feat: add numPredict, numCtx, maxCompletionTokens to model config pipeline

Fixes premature AI API response truncation by propagating inference
parameters through the entire probe → storage → runtime → API call chain.

Root cause: Ollama defaults num_predict to 128 tokens and num_ctx to
4096, silently truncating output and context. We never overrode these.

Changes:
- IAiModelSettings: add numPredict, maxCompletionTokens fields
- IDroneModelConfig: moved from gadget-drone to @gadget/api (shared),
  expanded with numPredict, numCtx, maxCompletionTokens params
- IAiModelConfig.params: add numPredict, numCtx, maxCompletionTokens
- IAiModelProbeResult.settings: add numPredict, maxCompletionTokens
- AiModelSettingsSchema (Mongoose): add numPredict, maxCompletionTokens
- Ollama extractSettings(): extract num_predict from model parameters
- Ollama generate()/chat(): pass options: { num_ctx, num_predict }
- OpenAI all three create() calls: add max_completion_tokens
- web-cli.ts onProviderProbe(): compute numPredict (-1 for Ollama)
  and maxCompletionTokens (contextWindow for OpenAI) during probe
- agent.ts main + subagent loops: read model settings from provider
  cached models, build IDroneModelConfig with stored params
- ai.ts: remove local IDroneModelConfig, import from @gadget/api
- chat-session.ts: add new params to title generation call
- Tests: update all fixtures with new params, all 19 tests pass

Defaults when model settings unavailable:
- numPredict: -1 (Ollama unlimited - generate until natural stop)
- numCtx: 131072 (128k - covers most modern models)
- maxCompletionTokens: 16384 (16k - reasonable OpenAI default)

gadget-code/src/models/ai-provider.ts

@@ -29,6 +29,8 @@ export const AiModelSettingsSchema = new Schema<IAiModelSettings>(
     topP: { type: Number },
     topK: { type: Number },
     numCtx: { type: Number },
+    numPredict: { type: Number },
+    maxCompletionTokens: { type: Number },
   },
   { _id: false },
 );

gadget-code/src/services/chat-session.ts

@@ -439,6 +439,9 @@ class ChatSessionService extends DtpService {
           temperature: 1.0,
           topK: 0.6,
           topP: 0.4,
+          numPredict: -1,
+          numCtx: 131072,
+          maxCompletionTokens: 256,
         },
       },
       {

gadget-code/src/web-cli.ts

@@ -586,6 +586,22 @@ class DtpWebCli extends DtpProcess {
       try {
         const probeResult = await api.probeModel(modelInfo.id);
 
+        // Compute provider-specific inference settings
+        const settings: IAiModelSettings = {
+          ...(probeResult.settings as IAiModelSettings | undefined),
+        };
+
+        if (provider.apiType === 'ollama') {
+          // Ollama: always override numPredict to -1 (unlimited) for agentic workflows
+          // The model must generate until its natural stop token or context limit
+          settings.numPredict = -1;
+          // numCtx is already populated by probeResult.settings from extractSettings()
+        } else if (provider.apiType === 'openai') {
+          // OpenAI-compatible: set maxCompletionTokens to model's context window
+          // This prevents compatible providers (Gab AI, etc.) from imposing low defaults
+          settings.maxCompletionTokens = modelInfo.contextWindow || 16384;
+        }
+
         const model: IAiModel = {
           id: modelInfo.id,
           name: modelInfo.name,
@@ -593,7 +609,7 @@ class DtpWebCli extends DtpProcess {
           parameterLabel: modelInfo.parameterLabel,
           contextWindow: modelInfo.contextWindow,
           capabilities: probeResult.capabilities as IAiModelCapabilities,
-          settings: probeResult.settings as IAiModelSettings | undefined,
+          settings,
         };
 
         models.push(model);

gadget-drone/src/services/agent.ts

@@ -17,11 +17,14 @@ import {
 } from "@gadget/ai";
 import {
   type IAiProvider,
+  type IAiModelSettings,
+  type IDroneModelConfig,
   type IChatSession,
   type IChatTurn,
   type IChatSubagentProcess,
   type IChatToolCall,
   type IUser,
+  type GadgetId,
   type ServerToClientEvents,
   type ClientToServerEvents,
   ChatSessionMode,
@@ -209,10 +212,7 @@ class AgentService extends GadgetService {
           try {
             response = await AiService.chat(
               turn.provider,
-              {
+              this.buildDroneModelConfig(turn.provider, turn.llm, currentReasoning),
-                modelId: turn.llm,
-                params: { reasoning: currentReasoning, temperature: 0.8, topP: 0.9, topK: 40 },
-              },
               chatOptions,
               this.makeStreamHandler(socket),
             );
@@ -410,6 +410,46 @@ class AgentService extends GadgetService {
     };
   }
 
+  /**
+   * Builds an IDroneModelConfig by looking up the model's stored settings
+   * from the provider's cached models array. Falls back to safe defaults
+   * when stored settings aren't available (e.g., model not yet probed).
+   */
+  private buildDroneModelConfig(
+    provider: IAiProvider | GadgetId,
+    modelId: string,
+    reasoning: boolean | "low" | "medium" | "high",
+  ): Omit<IDroneModelConfig, "provider"> {
+    const settings = this.getModelSettings(provider, modelId);
+    return {
+      modelId,
+      params: {
+        reasoning,
+        temperature: settings?.temperature ?? 0.8,
+        topP: settings?.topP ?? 0.9,
+        topK: settings?.topK ?? 40,
+        numPredict: settings?.numPredict ?? -1,         // -1 = unlimited (Ollama)
+        numCtx: settings?.numCtx ?? 131072,             // 128k fallback
+        maxCompletionTokens: settings?.maxCompletionTokens ?? 16384, // 16k fallback
+      },
+    };
+  }
+
+  /**
+   * Looks up the stored IAiModelSettings for a model from the provider's
+   * cached models array. Returns undefined if the provider or model isn't found.
+   */
+  private getModelSettings(
+    provider: IAiProvider | GadgetId,
+    modelId: string,
+  ): IAiModelSettings | undefined {
+    if (typeof provider === "string") return undefined;
+    const modelRecord = (provider as IAiProvider).models?.find(
+      (m) => m.id === modelId,
+    );
+    return modelRecord?.settings;
+  }
+
   private async executeTool(name: string, argsJson: string): Promise<string> {
     const tool = this.toolbox.getTool(name);
     if (!tool) {
@@ -528,10 +568,7 @@ class AgentService extends GadgetService {
 
         const response = await AiService.chat(
           provider,
-          {
+          this.buildDroneModelConfig(provider, turn.llm, false),
-            modelId: turn.llm,
-            params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40 },
-          },
           chatOptions,
           streamHandler,
         );

gadget-drone/src/services/ai.ts

@@ -15,7 +15,7 @@ const aiEnv: IAiEnvironment = {
   },
 };
 
-import { IAiProvider as DbAiProvider, GadgetId } from "@gadget/api";
+import { IAiProvider as DbAiProvider, GadgetId, type IDroneModelConfig } from "@gadget/api";
 import { GadgetService } from "../lib/service.js";
 import {
   type IAiChatOptions,
@@ -29,20 +29,6 @@ import {
   IAiEnvironment,
 } from "@gadget/ai";
 
-/**
- * Drone-specific model config that accepts the database provider type.
- */
-export interface IDroneModelConfig {
-  provider: DbAiProvider | GadgetId;
-  modelId: string;
-  params: {
-    reasoning: boolean | "low" | "medium" | "high";
-    temperature: number;
-    topP: number;
-    topK: number;
-  };
-}
-
 /**
  * An abstraction of the backend AI APIs (Ollama, OpenAI) that provides one
  * common interface and contract for working with different AI APIs.

packages/ai/src/api.ts

@@ -22,6 +22,12 @@ export interface IAiModelConfig {
     temperature: number;
     topP: number;
     topK: number;
+    /** Ollama: -1 = unlimited (generate until natural stop or context limit) */
+    numPredict: number;
+    /** Context window size (input + output tokens); Ollama passes as num_ctx */
+    numCtx: number;
+    /** OpenAI-compatible: maximum completion tokens the model can generate */
+    maxCompletionTokens: number;
   };
 }
 
@@ -146,6 +152,10 @@ export interface IAiModelProbeResult {
     topP?: number;
     topK?: number;
     numCtx?: number;
+    /** Ollama: discovered num_predict from model parameters (informational; overridden to -1 at inference time) */
+    numPredict?: number;
+    /** OpenAI-compatible: discovered from model info */
+    maxCompletionTokens?: number;
   };
 }
 

packages/ai/src/ollama.test.ts

@@ -87,7 +87,7 @@ describe('OllamaAiApi', () => {
         {
           provider: mockProvider as any,
           modelId: 'test-model',
-          params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40 },
+          params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40, numPredict: -1, numCtx: 131072, maxCompletionTokens: 16384 },
         },
         {
           userPrompt: 'Test prompt',
@@ -150,7 +150,7 @@ describe('OllamaAiApi', () => {
         {
           provider: mockProvider as any,
           modelId: 'test-model',
-          params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40 },
+          params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40, numPredict: -1, numCtx: 131072, maxCompletionTokens: 16384 },
         },
         {
           userPrompt: 'Test prompt',
@@ -201,7 +201,7 @@ describe('OllamaAiApi', () => {
         {
           provider: mockProvider as any,
           modelId: 'test-model',
-          params: { reasoning: true, temperature: 0.8, topP: 0.9, topK: 40 },
+          params: { reasoning: true, temperature: 0.8, topP: 0.9, topK: 40, numPredict: -1, numCtx: 131072, maxCompletionTokens: 16384 },
         },
         {
           userPrompt: 'What is the answer?',
@@ -243,7 +243,7 @@ describe('OllamaAiApi', () => {
         {
           provider: mockProvider as any,
           modelId: 'test-model',
-          params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40 },
+          params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40, numPredict: -1, numCtx: 131072, maxCompletionTokens: 16384 },
         },
         {
           userPrompt: 'Test prompt',

packages/ai/src/ollama.ts

@@ -126,6 +126,11 @@ export class OllamaAiApi extends AiApi {
       settings.numCtx = parseInt(numCtxMatch[1], 10);
     }
 
+    const numPredictMatch = parameters.match(/num_predict\s+(-?\d+)/i);
+    if (numPredictMatch) {
+      settings.numPredict = parseInt(numPredictMatch[1], 10);
+    }
+
     return Object.keys(settings).length > 0 ? settings : undefined;
   }
 
@@ -144,6 +149,10 @@ export class OllamaAiApi extends AiApi {
       prompt: options.prompt,
       system: options.systemPrompt,
       stream: true,
+      options: {
+        num_ctx: model.params.numCtx,
+        num_predict: model.params.numPredict,
+      },
     });
 
     const content = {
@@ -259,6 +268,10 @@ export class OllamaAiApi extends AiApi {
       stream: true,
       think: model.params.reasoning,
       tools: ollamaTools,
+      options: {
+        num_ctx: model.params.numCtx,
+        num_predict: model.params.numPredict,
+      },
     });
 
     let lastChunk;

packages/ai/src/openai.test.ts

@@ -65,7 +65,7 @@ describe("OpenAiApi", () => {
       {
         provider: mockProvider as any,
         modelId: "test-model",
-        params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40 },
+        params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40, numPredict: -1, numCtx: 131072, maxCompletionTokens: 16384 },
       },
       { userPrompt: "Hello", context: [], tools: [] },
       vi.fn(),
@@ -105,7 +105,7 @@ describe("OpenAiApi", () => {
       {
         provider: mockProvider as any,
         modelId: "test-model",
-        params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40 },
+        params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40, numPredict: -1, numCtx: 131072, maxCompletionTokens: 16384 },
       },
       { userPrompt: "Read index.html", context: [], tools: [] },
       streamCallback,
@@ -133,7 +133,7 @@ describe("OpenAiApi", () => {
       {
         provider: mockProvider as any,
         modelId: "test-model",
-        params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40 },
+        params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40, numPredict: -1, numCtx: 131072, maxCompletionTokens: 16384 },
       },
       { userPrompt: "Hello", context: [], tools: [] },
       streamCallback,
@@ -153,7 +153,7 @@ describe("OpenAiApi", () => {
       {
         provider: mockProvider as any,
         modelId: "test-model",
-        params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40 },
+        params: { reasoning: false, temperature: 0.8, topP: 0.9, topK: 40, numPredict: -1, numCtx: 131072, maxCompletionTokens: 16384 },
       },
       {
         userPrompt: "Don't edit code. Just talk to me.",

packages/ai/src/openai.ts

@@ -202,6 +202,9 @@ export class OpenAiApi extends AiApi {
         { role: "user" as const, content: options.prompt },
       ],
       stream: true,
+      ...(model.params.maxCompletionTokens
+        ? { max_completion_tokens: model.params.maxCompletionTokens }
+        : {}),
       ...(typeof model.params.reasoning === "string"
         ? {
             reasoning_effort: model.params.reasoning as
@@ -373,6 +376,9 @@ export class OpenAiApi extends AiApi {
       messages,
       tools,
       stream: true,
+      ...(model.params.maxCompletionTokens
+        ? { max_completion_tokens: model.params.maxCompletionTokens }
+        : {}),
       ...(typeof model.params.reasoning === "string"
         ? {
             reasoning_effort: model.params.reasoning as
@@ -439,6 +445,9 @@ export class OpenAiApi extends AiApi {
       messages,
       tools,
       stream: false,
+      ...(model.params.maxCompletionTokens
+        ? { max_completion_tokens: model.params.maxCompletionTokens }
+        : {}),
       ...(typeof model.params.reasoning === "string"
         ? {
             reasoning_effort: model.params.reasoning as

packages/api/src/interfaces/ai-provider.ts

@@ -17,6 +17,10 @@ export interface IAiModelSettings {
   topP?: number;
   topK?: number;
   numCtx?: number;
+  /** Ollama: maximum number of tokens to predict (-1 = unlimited, generate until natural stop or context limit) */
+  numPredict?: number;
+  /** OpenAI-compatible: maximum number of completion tokens the model can generate */
+  maxCompletionTokens?: number;
 }
 
 export interface IAiModelCapabilities {
@@ -61,3 +65,26 @@ export interface IAiProvider {
 }
 
 export type AiProviderDocument = HydratedDocument<IAiProvider>;
+
+/**
+ * Drone-side model configuration that accepts the database IAiProvider type.
+ * This is the canonical interface used by the drone to configure AI API calls,
+ * carrying the model's stored inference settings (numCtx, numPredict, etc.)
+ * from the provider's cached model data through to the AI package.
+ */
+export interface IDroneModelConfig {
+  provider: IAiProvider | GadgetId;
+  modelId: string;
+  params: {
+    reasoning: boolean | "low" | "medium" | "high";
+    temperature: number;
+    topP: number;
+    topK: number;
+    /** Ollama: -1 = unlimited (generate until natural stop or context limit) */
+    numPredict: number;
+    /** Context window size (input + output tokens); Ollama passes as num_ctx */
+    numCtx: number;
+    /** OpenAI-compatible: maximum completion tokens the model can generate */
+    maxCompletionTokens: number;
+  };
+}