infinicode 2.3.5 → 2.4.0

package/dist/kernel/free-providers.js

@@ -101,6 +101,89 @@ export const FREE_PROVIDERS = [
         ],
         notes: 'NVIDIA build.nvidia.com offers free NIM endpoints. 1000 req/month free tier.',
     },
+    {
+        id: 'cerebras',
+        name: 'Cerebras',
+        baseURL: 'https://api.cerebras.ai/v1',
+        keyUrl: 'https://cloud.cerebras.ai',
+        free: true,
+        recommended: {
+            coding: ['qwen-3-235b-a22b-instruct-2507', 'gpt-oss-120b'],
+            architecture: ['qwen-3-235b-a22b-instruct-2507'],
+            research: ['gpt-oss-120b'],
+            review: ['qwen-3-235b-a22b-instruct-2507'],
+            reasoning: ['qwen-3-235b-a22b-instruct-2507'],
+        },
+        knownModels: [
+            { id: 'qwen-3-235b-a22b-instruct-2507', name: 'Qwen3 235B', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'gpt-oss-120b', name: 'GPT-OSS 120B', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'qwen-3-32b', name: 'Qwen3 32B', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'llama-3.3-70b', name: 'Llama 3.3 70B', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+        ],
+        notes: 'Fastest inference (wafer-scale). Free: ~1M tokens/day, 30 RPM. Frontier open models.',
+    },
+    {
+        id: 'chutes',
+        name: 'Chutes',
+        baseURL: 'https://llm.chutes.ai/v1',
+        keyUrl: 'https://chutes.ai',
+        free: true,
+        recommended: {
+            coding: ['zai-org/GLM-4.6', 'deepseek-ai/DeepSeek-V3.1', 'moonshotai/Kimi-K2-Instruct'],
+            architecture: ['deepseek-ai/DeepSeek-V3.1', 'zai-org/GLM-4.6'],
+            research: ['deepseek-ai/DeepSeek-V3.1'],
+            reasoning: ['deepseek-ai/DeepSeek-R1'],
+            review: ['zai-org/GLM-4.6'],
+        },
+        knownModels: [
+            { id: 'zai-org/GLM-4.6', name: 'GLM-4.6', contextLength: 200_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'moonshotai/Kimi-K2-Instruct', name: 'Kimi K2', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'deepseek-ai/DeepSeek-V3.1', name: 'DeepSeek V3.1', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'Qwen/Qwen3-235B-A22B', name: 'Qwen3 235B', contextLength: 256_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'deepseek-ai/DeepSeek-R1', name: 'DeepSeek R1', contextLength: 128_000, capabilities: ['reasoning'] },
+        ],
+        notes: 'Huge decentralized catalog — GLM, Kimi, DeepSeek, Qwen frontier models. Generous free tier.',
+    },
+    {
+        id: 'sambanova',
+        name: 'SambaNova',
+        baseURL: 'https://api.sambanova.ai/v1',
+        keyUrl: 'https://cloud.sambanova.ai',
+        free: true,
+        recommended: {
+            coding: ['DeepSeek-V3.1', 'Meta-Llama-3.3-70B-Instruct'],
+            architecture: ['DeepSeek-V3.1'],
+            research: ['Meta-Llama-3.3-70B-Instruct'],
+            reasoning: ['DeepSeek-R1'],
+        },
+        knownModels: [
+            { id: 'DeepSeek-V3.1', name: 'DeepSeek V3.1', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'Meta-Llama-3.3-70B-Instruct', name: 'Llama 3.3 70B', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'gpt-oss-120b', name: 'GPT-OSS 120B', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'Qwen3-32B', name: 'Qwen3 32B', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+        ],
+        notes: 'Very fast (RDU). Permanent free tier (20 RPM) + $5 trial credits. No card.',
+    },
+    {
+        id: 'mistral',
+        name: 'Mistral',
+        baseURL: 'https://api.mistral.ai/v1',
+        keyUrl: 'https://console.mistral.ai/api-keys',
+        free: true,
+        recommended: {
+            coding: ['codestral-latest', 'mistral-large-latest'],
+            architecture: ['mistral-large-latest'],
+            research: ['mistral-large-latest'],
+            reasoning: ['magistral-medium-latest'],
+        },
+        knownModels: [
+            { id: 'mistral-large-latest', name: 'Mistral Large', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'codestral-latest', name: 'Codestral', contextLength: 256_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+            { id: 'magistral-medium-latest', name: 'Magistral Medium', contextLength: 128_000, capabilities: ['reasoning'], supportsFunctionCalling: true },
+            { id: 'mistral-small-latest', name: 'Mistral Small', contextLength: 128_000, capabilities: ['coding', 'reasoning'], supportsFunctionCalling: true },
+        ],
+        notes: 'La Plateforme free (Experiment) tier ≈ 1B tokens/month. Codestral is a strong coder.',
+    },
     {
         id: 'hf',
         name: 'Hugging Face Inference',

package/dist/kernel/gateway/openai-gateway.d.ts

@@ -45,6 +45,16 @@ export interface KernelGatewayOptions {
     executor?: ProviderExecutor;
     /** Capabilities to route interactive chat under. */
     capabilities?: Capability[];
+    /**
+     * Soft benchmark floor for `auto` routing. Biases interactive coding toward
+     * capable models (weak ones become fallbacks, not first picks) so demanding
+     * work doesn't open on a lazy model — without hard-pinning anything.
+     */
+    minBenchmark?: number;
+    /** Stall guards (ms). Override the defaults (mainly for tests). */
+    headersTimeoutMs?: number;
+    firstChunkTimeoutMs?: number;
+    idleTimeoutMs?: number;
 }
 export declare class KernelGateway {
     private opts;
@@ -56,15 +66,55 @@ export declare class KernelGateway {
     constructor(opts: KernelGatewayOptions);
     get port(): number;
     get url(): string;
-    /** Start listening. Scans ports [port .. port+10] to avoid collisions. */
+    /**
+     * Start listening. Scans ports [port .. port+20] then falls back to an OS-chosen
+     * ephemeral port. Retries past BOTH EADDRINUSE (a lingering session holds it) and
+     * EACCES (a Windows excluded/reserved range) — never dies on a single bad port,
+     * since the whole point is that the gateway must come up so the TUI gets routing.
+     */
     start(): Promise<void>;
     private listen;
     stop(): Promise<void>;
     private handle;
     private listModels;
     private handleChat;
-    /** Stream bytes through untouched, or relay the JSON body for non-streaming. */
-    private relay;
+    /**
+     * Build a continuation request: the original messages plus the partial answer as
+     * an assistant prefix so a fresh model resumes it instead of restarting. `prefix:
+     * true` is the continue-this-message convention (DeepSeek/Mistral/Zhipu/Qwen);
+     * providers that ignore it at worst restart — still better than a dead task.
+     */
+    private buildContinuation;
+    /** Close an already-open SSE stream with a synthetic clean finish so the client
+     *  gets a proper end-of-turn instead of hanging on a half-open connection. */
+    private endStream;
+    /**
+     * Run one model. Returns forwarded:true only after the response is validated
+     * and sent to the client; otherwise nothing is written to `res` and the caller
+     * swaps to the next model. Catches BOTH transport failures (non-2xx) AND bad
+     * tool calls the model streams back on a 200 (a weak model naming a tool that
+     * isn't in request.tools — e.g. `glob={...}` — which downstream would reject).
+     */
+    private attempt;
+    /**
+     * Relay an SSE stream, but with a small look-ahead: buffer frames until the
+     * first tool-call name or content delta. If that first tool call names a tool
+     * not in request.tools (or an error frame arrives), fail WITHOUT writing to the
+     * client so the caller can swap models. Once a valid start is seen, commit and
+     * stream the rest live (so large generations still stream token-by-token).
+     */
+    private streamValidated;
+    /**
+     * Classify one SSE frame: does it error, is it a safe commit point, is it a
+     * TERMINAL frame (finish_reason / `[DONE]` — held back so a truncation can be
+     * continued), and what visible content does it carry (accumulated so a fresh
+     * model can continue the answer from where a dead provider left off)?
+     */
+    private inspectFrame;
+    /** Validate a non-streamed response body for unknown tool calls / errors. */
+    private validateResponse;
+    /** The set of tool names the request declared (OpenAI tools[].function.name). */
+    private allowedTools;
     /** Resolve the initial target: honor an explicit provider/model, else route. */
     private resolveTarget;
     /** Ask the kernel router for the best untried, executable model. */
@@ -75,6 +125,8 @@ export declare class KernelGateway {
      * escalate to a stronger model via the kernel ladder, then fall back to rank.
      */
     private nextTarget;
+    /** First configured provider/model not in `tried` (optionally skipping a provider). */
+    private firstUntried;
     private toTarget;
     private rank;
     private routingRequest;