@kenkaiiii/gg-core 5.55.1 → 5.56.1

package/dist/index.d.cts

@@ -15,8 +15,8 @@ declare function getNextThinkingLevel(provider: Provider, model: string, current
  * provider in gg-ai's stream.ts); the only per-server difference is where the
  * *capabilities* come from, because `GET /v1/models` reports nothing useful:
  *
- *   - Ollama    → `POST /api/show`      → `capabilities[]` + `model_info["<arch>.context_length"]`
- *   - LM Studio → `GET /api/v0/models`  → `type`, `state`, `max_context_length`
+ *   - Ollama    → `/api/show` + `/api/ps` → capabilities + running context
+ *   - LM Studio → `/api/v1/models` (v0 fallback) → capabilities + loaded context
  *   - llama.cpp → `GET /props`          → `default_generation_settings.n_ctx`
  *   - vLLM/other → nothing; `max_model_len` sometimes rides the model object.
  *
@@ -70,7 +70,7 @@ declare const DEFAULT_LOCAL_ENDPOINTS: readonly LocalEndpoint[];
  * conservative: over-guessing means the provider 400s mid-run at a point
  * auto-compaction already sailed past, while under-guessing only compacts early.
  */
-declare const FALLBACK_CONTEXT_WINDOW = 8192;
+declare const FALLBACK_CONTEXT_WINDOW = 4096;
 /** Placeholder token for endpoints with no key — these servers ignore it. */
 declare const LOCAL_API_KEY_PLACEHOLDER = "local";
 /**

package/dist/index.d.ts

@@ -15,8 +15,8 @@ declare function getNextThinkingLevel(provider: Provider, model: string, current
  * provider in gg-ai's stream.ts); the only per-server difference is where the
  * *capabilities* come from, because `GET /v1/models` reports nothing useful:
  *
- *   - Ollama    → `POST /api/show`      → `capabilities[]` + `model_info["<arch>.context_length"]`
- *   - LM Studio → `GET /api/v0/models`  → `type`, `state`, `max_context_length`
+ *   - Ollama    → `/api/show` + `/api/ps` → capabilities + running context
+ *   - LM Studio → `/api/v1/models` (v0 fallback) → capabilities + loaded context
  *   - llama.cpp → `GET /props`          → `default_generation_settings.n_ctx`
  *   - vLLM/other → nothing; `max_model_len` sometimes rides the model object.
  *
@@ -70,7 +70,7 @@ declare const DEFAULT_LOCAL_ENDPOINTS: readonly LocalEndpoint[];
  * conservative: over-guessing means the provider 400s mid-run at a point
  * auto-compaction already sailed past, while under-guessing only compacts early.
  */
-declare const FALLBACK_CONTEXT_WINDOW = 8192;
+declare const FALLBACK_CONTEXT_WINDOW = 4096;
 /** Placeholder token for endpoints with no key — these servers ignore it. */
 declare const LOCAL_API_KEY_PLACEHOLDER = "local";
 /**

package/dist/index.js

@@ -54,7 +54,7 @@ import {
   registerRuntimeModels,
   usesOpenAICodexTransport,
   withFileLock
-} from "./chunk-WPX6Z5NG.js";
+} from "./chunk-PSLSIL3V.js";
 import {
   getAppPaths
 } from "./chunk-EAIPT76S.js";
@@ -69,7 +69,8 @@ var OPENAI_GPT_56_THINKING_LEVELS = [
   "max",
   "ultra"
 ];
-var SAKANA_THINKING_LEVELS = ["high", "xhigh"];
+var SAKANA_THINKING_LEVELS = ["high", "xhigh", "max"];
+var DEEPSEEK_THINKING_LEVELS = ["low", "high", "max"];
 var XAI_THINKING_LEVELS = ["low", "medium", "high", "xhigh"];
 var ANTHROPIC_XHIGH_THINKING_LEVELS = [
   "low",
@@ -133,6 +134,7 @@ function getSupportedThinkingLevels(provider, model) {
     return XAI_THINKING_LEVELS.slice(0, maxIndex2 + 1);
   }
   if (isMoonshotK3Model(provider, model)) return MOONSHOT_K3_THINKING_LEVELS;
+  if (provider === "deepseek") return DEEPSEEK_THINKING_LEVELS;
   if (isGlmModel(provider)) {
     const maxIndex2 = GLM_THINKING_LEVELS.indexOf(maxLevel);
     if (maxIndex2 === -1) return GLM_THINKING_LEVELS;
@@ -149,7 +151,7 @@ function isThinkingLevelSupported(provider, model, level) {
 }
 function getNextThinkingLevel(provider, model, current) {
   const supportedLevels = getSupportedThinkingLevels(provider, model);
-  const shouldCycleLevels = isOpenAIGptModel(provider, model) || isAnthropicAdaptiveModel(provider, model) || isSakanaModel(provider) || isXaiModel(provider) || isMoonshotK3Model(provider, model) || isGlmModel(provider) || // Local servers take a real effort level, not just on/off: Ollama accepts
+  const shouldCycleLevels = isOpenAIGptModel(provider, model) || isAnthropicAdaptiveModel(provider, model) || isSakanaModel(provider) || isXaiModel(provider) || isMoonshotK3Model(provider, model) || isGlmModel(provider) || provider === "deepseek" || // Local servers take a real effort level, not just on/off: Ollama accepts
   // low/medium/high on `reasoning_effort` (verified against 0.32) and the
   // other OpenAI-compatible servers use the same three. A model that can't
   // reason at all already has no supported levels, so it never gets here.
@@ -167,7 +169,7 @@ function getNextThinkingLevel(provider, model, current) {
 var DEFAULT_LOCAL_ENDPOINTS = [
   { id: "ollama", label: "Ollama", baseUrl: "http://127.0.0.1:11434/v1", kind: "ollama" }
 ];
-var FALLBACK_CONTEXT_WINDOW = 8192;
+var FALLBACK_CONTEXT_WINDOW = 4096;
 var LOCAL_API_KEY_PLACEHOLDER = "local";
 var DEFAULT_PROBE_TIMEOUT_MS = 1200;
 var ENRICH_CONCURRENCY = 6;
@@ -260,7 +262,7 @@ async function enrich(endpoint, entries, options) {
   return entries.map((entry) => genericModel(endpoint, entry));
 }
 function genericModel(endpoint, entry) {
-  const declared = typeof entry.max_model_len === "number" ? entry.max_model_len : void 0;
+  const declared = contextLength(entry.max_model_len);
   return {
     rawId: entry.id,
     endpointId: endpoint.id,
@@ -273,16 +275,23 @@ function genericModel(endpoint, entry) {
 }
 async function enrichOllama(endpoint, entries, options) {
   const showUrl = `${endpointRoot(endpoint.baseUrl)}/api/show`;
+  const running = await fetchJson(
+    `${endpointRoot(endpoint.baseUrl)}/api/ps`,
+    endpoint,
+    options
+  );
+  const runningModels = Array.isArray(running?.models) ? running.models : [];
   const enriched = await mapLimited(entries, ENRICH_CONCURRENCY, async (entry) => {
     const show = await fetchJson(showUrl, endpoint, {
       ...options,
       method: "POST",
       body: { model: entry.id }
     });
-    if (!show) return genericModel(endpoint, entry);
-    const caps = show.capabilities ?? [];
+    const caps = Array.isArray(show?.capabilities) ? show.capabilities : [];
     if (caps.includes("embedding") && !caps.includes("completion")) return void 0;
-    const ctx = ollamaContextLength(show.model_info);
+    const ctx = contextLength(
+      runningModels.find((model) => model?.name === entry.id || model?.model === entry.id)?.context_length
+    );
     return {
       rawId: entry.id,
       endpointId: endpoint.id,
@@ -290,38 +299,61 @@ async function enrichOllama(endpoint, entries, options) {
       contextWindowKnown: ctx !== void 0,
       // Ollama reports capabilities honestly, so trust it here rather than
       // using the optimistic generic default.
-      supportsTools: caps.includes("tools"),
+      supportsTools: show ? caps.includes("tools") : true,
       supportsImages: caps.includes("vision"),
       supportsThinking: caps.includes("thinking")
     };
   });
   return enriched.filter((model) => model !== void 0);
 }
-function ollamaContextLength(info) {
-  if (!info) return void 0;
-  for (const [key, value] of Object.entries(info)) {
-    if (key.endsWith(".context_length") && typeof value === "number" && value > 0) return value;
-  }
-  return void 0;
+function contextLength(value) {
+  return typeof value === "number" && Number.isSafeInteger(value) && value > 0 ? value : void 0;
 }
 async function enrichLmStudio(endpoint, entries, options) {
-  const detail = await fetchJson(
-    `${endpointRoot(endpoint.baseUrl)}/api/v0/models`,
-    endpoint,
-    options
-  );
-  if (!detail?.data) return entries.map((entry) => genericModel(endpoint, entry));
-  const byId = new Map(detail.data.filter((m) => m.id).map((m) => [m.id, m]));
+  const root = endpointRoot(endpoint.baseUrl);
+  const detail = await fetchJson(`${root}/api/v1/models`, endpoint, options);
+  const byId = /* @__PURE__ */ new Map();
+  if (Array.isArray(detail?.models)) {
+    for (const model of detail.models) {
+      if (typeof model?.key !== "string") continue;
+      const instances = Array.isArray(model.loaded_instances) ? model.loaded_instances : [];
+      const contexts = instances.map((instance) => contextLength(instance?.config?.context_length));
+      const ctx = contexts.length && contexts.every((value) => value !== void 0) ? contexts.reduce((min, value) => Math.min(min, value)) : void 0;
+      const info = {
+        id: model.key,
+        type: model.type === "llm" && model.capabilities?.vision === true ? "vlm" : model.type,
+        state: instances.length ? "loaded" : "not-loaded",
+        loaded_context_length: ctx
+      };
+      byId.set(model.key, info);
+      for (const instance of instances) {
+        if (typeof instance?.id === "string" && instance.id !== model.key) {
+          byId.set(instance.id, {
+            ...info,
+            id: instance.id,
+            loaded_context_length: contextLength(instance.config?.context_length)
+          });
+        }
+      }
+    }
+  } else {
+    const legacy = await fetchJson(`${root}/api/v0/models`, endpoint, options);
+    if (Array.isArray(legacy?.data)) {
+      for (const model of legacy.data) {
+        if (typeof model?.id === "string") byId.set(model.id, model);
+      }
+    }
+  }
   const models = [];
   for (const entry of entries) {
     const info = byId.get(entry.id);
     if (info && info.type !== "llm" && info.type !== "vlm") continue;
-    const ctx = info?.max_context_length;
+    const ctx = info?.state === "loaded" ? contextLength(info.loaded_context_length) : void 0;
     models.push({
       rawId: entry.id,
       endpointId: endpoint.id,
-      contextWindow: typeof ctx === "number" && ctx > 0 ? ctx : FALLBACK_CONTEXT_WINDOW,
-      contextWindowKnown: typeof ctx === "number" && ctx > 0,
+      contextWindow: ctx ?? FALLBACK_CONTEXT_WINDOW,
+      contextWindowKnown: ctx !== void 0,
       // LM Studio doesn't report tool support; it gates per-model at request time.
       supportsTools: true,
       supportsImages: info?.type === "vlm",
@@ -337,8 +369,8 @@ async function enrichLlamaCpp(endpoint, entries, options) {
     endpoint,
     options
   );
-  const nCtx = props?.default_generation_settings?.n_ctx;
-  const known = typeof nCtx === "number" && nCtx > 0;
+  const nCtx = contextLength(props?.default_generation_settings?.n_ctx);
+  const known = nCtx !== void 0;
   return entries.map((entry) => ({
     ...genericModel(endpoint, entry),
     contextWindow: known ? nCtx : FALLBACK_CONTEXT_WINDOW,