koneck 2.25.66 → 2.25.67

package/dist/ink-chat.js

@@ -2,12 +2,12 @@ import { jsx as _jsx, Fragment as _Fragment, jsxs as _jsxs } from "react/jsx-run
 import React, { useState, useEffect, useRef } from 'react';
 import { Box, Static, Text, render, useApp, useInput, useStdin } from 'ink';
 import { execa } from 'execa';
-import { createAgentSession, buildClient, resolveAgentConcurrency, toolsUnsupportedNotice, isLocalRuntime } from './engine.js';
+import { createAgentSession, buildClient, resolveAgentConcurrency, toolsUnsupportedNotice } from './engine.js';
 import { estimateCost, formatCost } from './pricing.js';
 import { generateSessionId, saveSession, listSessions, loadSession, relativeAge, renameSession, forkSession, archiveSession, findSession, } from './session.js';
 import { loadMemory } from './memory.js';
 import { loadKoneckConfig, saveKoneckConfig, setConfigKey, configKeyDescriptions, CONFIG_FILE, CONFIG_SCHEMA, stepValue, displayValue } from './config-store.js';
-import { PROVIDERS, resolveProvider, getApiKey, addUserProvider, userProvidersPath, needsNoCredential } from './providers.js';
+import { PROVIDERS, resolveProvider, getApiKey, addUserProvider, userProvidersPath, needsNoCredential, servesLocally } from './providers.js';
 import { resolveEndpoint, normaliseEndpoint, withEndpoint, savedEndpoints, shortSource, resolveModelChoice, withModel } from './endpoints.js';
 import { loadProjectConfig, loadGlobalConfig } from './config.js';
 import { listCheckpoints, revertCheckpoint, snapshotTree, filesTouchedBy } from './checkpoint.js';
@@ -810,6 +810,8 @@ function App({ config: initialConfig, clearFrame }) {
     const incompleteRef = useRef(null);
     /** The model already reported as unable to call tools, so the notice is not repeated. */
     const toolsWarnedRef = useRef(null);
+    /** The crowded-window notice is about the setup, so it is said once per session. */
+    const crowdedWarnedRef = useRef(false);
     /**
      * The id to save under, mirrored in a ref.
      *
@@ -1031,16 +1033,25 @@ function App({ config: initialConfig, clearFrame }) {
             toolsWarnedRef.current = model;
             addSystem(toolsUnsupportedNotice(model));
         },
+        onCrowdedWindow: (_fixed, _limit, advice) => {
+            // Once per session: it is a property of the setup, not of the turn.
+            if (crowdedWarnedRef.current)
+                return;
+            crowdedWarnedRef.current = true;
+            addSystem(advice);
+        },
         onLooping: (copies) => {
             // The loop itself is the symptom. On a local model the cause is almost always the window:
             // Ollama serves 4096 tokens unless told otherwise, whatever the model was trained for, and
             // a conversation that no longer fits is exactly what makes a model start repeating.
-            const local = isLocalRuntime(cfg.provider);
+            // Judged by the endpoint, not the provider's name: a LAN Ollama reached through a slot
+            // called "omni" is still Ollama, and asking about the name missed exactly that setup.
+            const local = servesLocally(cfg.provider, endpointFor(cfg.provider).url);
             addSystem(`${cfg.model} repeated the same text ${copies} times, so the reply was cut short.` +
                 (local
-                    ? `\n\nThis is usually the context window rather than the model. Ollama serves 4096 ` +
-                        `tokens by default however large the model is — set ` +
-                        `\`OLLAMA_CONTEXT_LENGTH=32768\` and restart it, then \`/context\` to confirm.`
+                    ? `\n\nThis is usually the context window rather than the model. Ollama serves 4,096 ` +
+                        `tokens by default however large the model is — set \`OLLAMA_CONTEXT_LENGTH=32768\` ` +
+                        `where Ollama runs, restart it, then \`/context\` to confirm.`
                     : ''));
         },
         onThinking: (text, totalChars) => {