@gamaze/hicortex 0.16.4 → 0.16.5

package/dist/llm.d.ts

@@ -29,6 +29,8 @@ export interface LlmConfig {
     maxTokens?: number;
     /** Toggle thinking on the openai-compat path for the heavy phases. Absent = no kwarg sent. */
     enableThinking?: boolean;
+    /** Context window for the ollama fast tier (completeOllama). Falls back to 2048. */
+    numCtx?: number;
     /** Optional separate endpoint for reflect-tier LLM (e.g. remote Ollama with larger model). */
     reflectBaseUrl?: string;
     reflectApiKey?: string;

package/dist/llm.js

@@ -208,6 +208,9 @@ function applyTierTuningOverlay(llmConfig, savedConfig) {
     if (thinking !== undefined) {
         llmConfig.enableThinking = thinking;
     }
+    if (savedConfig.numCtx !== undefined) {
+        llmConfig.numCtx = (0, config_read_js_1.readPositiveConfig)(savedConfig, "numCtx", 2048);
+    }
 }
 /**
  * Resolve an LlmConfig from a saved ~/.hicortex/config.json object.
@@ -539,10 +542,11 @@ class LlmClient {
      */
     async completeFast(prompt, maxTokens) {
         const tokens = maxTokens ?? this.config.maxTokens ?? 2048;
-        // No `thinking` arg: scoring is excluded from the thinking toggle (it's the
-        // local fast tier; on ollama it's think:false regardless). completeFast never
-        // threads thinking, so chat_template_kwargs is never sent for scoring.
-        return this.complete(this.config.model, prompt, tokens, 600_000);
+        // Scoring is the fast tier: numCtx (default 2048, ~850-token prompts) is threaded
+        // so ONLY this tier gets the smaller context window — the heavy tiers keep 32768
+        // (preserves detectChunkSize's chunk-sizing). thinking is not threaded (scoring is
+        // excluded from the thinking toggle; ollama is think:false regardless).
+        return this.complete(this.config.model, prompt, tokens, 600_000, undefined, this.config.numCtx ?? 2048);
     }
     /**
      * Reflect-tier completion (nightly reflection, needs reasoning).
@@ -608,9 +612,14 @@ class LlmClient {
             apiKey,
             provider,
         });
+        // numCtx intentionally NOT forwarded here: only the heavy tiers route through
+        // completeWithOverride (there is no fastBaseUrl/fast-tier override), and they
+        // must land on completeOllama's 32768 default to preserve detectChunkSize. If a
+        // fast-tier override is ever added, thread numCtx here too — else the fast tier
+        // would silently revert to 32768 (the original bug, reintroduced).
         return tempClient.complete(model, prompt, maxTokens, timeoutMs, thinking);
     }
-    async complete(model, prompt, maxTokens, timeoutMs, thinking) {
+    async complete(model, prompt, maxTokens, timeoutMs, thinking, numCtx) {
         if (this.isRateLimited) {
             throw new RateLimitError(this.rateLimitedUntil - Date.now());
         }
@@ -618,7 +627,7 @@ class LlmClient {
         let lastErr;
         for (let attempt = 0; attempt <= retryDelays.length; attempt++) {
             try {
-                return await this.completeOnce(model, prompt, maxTokens, timeoutMs, thinking);
+                return await this.completeOnce(model, prompt, maxTokens, timeoutMs, thinking, numCtx);
             }
             catch (err) {
                 lastErr = err instanceof Error ? err : new Error(String(err));
@@ -635,12 +644,12 @@ class LlmClient {
         }
         throw lastErr;
     }
-    async completeOnce(model, prompt, maxTokens, timeoutMs, thinking) {
+    async completeOnce(model, prompt, maxTokens, timeoutMs, thinking, numCtx) {
         if (this.config.provider === "claude-cli") {
             return this.completeClaude(model, prompt, timeoutMs);
         }
         if (this.config.provider === "ollama") {
-            return this.completeOllama(model, prompt, maxTokens, timeoutMs);
+            return this.completeOllama(model, prompt, maxTokens, timeoutMs, numCtx);
         }
         if (this.config.provider === "anthropic") {
             return this.completeAnthropic(model, prompt, maxTokens, timeoutMs);
@@ -676,7 +685,7 @@ class LlmClient {
     /**
      * Ollama: use /api/generate with think:false (important for qwen3.5 models).
      */
-    async completeOllama(model, prompt, maxTokens, timeoutMs) {
+    async completeOllama(model, prompt, maxTokens, timeoutMs, numCtx) {
         const url = `${this.config.baseUrl.replace(/\/$/, "")}/api/generate`;
         // Ollama can take minutes to process large contexts — use streaming to avoid
         // Node.js fetch headers timeout (default ~300s kills long Ollama inferences)
@@ -688,7 +697,14 @@ class LlmClient {
                 prompt,
                 stream: true,
                 think: false,
-                options: { num_predict: maxTokens, num_ctx: 32768 },
+                // num_ctx: threaded from completeFast (the fast/scoring tier) via the numCtx
+                // param. The fast tier passes the numCtx config key (default 2048; scoring
+                // prompts are ~850 tokens) to minimize the KV/prompt-cache footprint that, at
+                // 32768, accumulated past available RAM on memory-constrained boxes during long
+                // consolidations and swap-thrashed the nightly. The heavy tiers
+                // (completeDistill/Reflect/Classify) do NOT pass numCtx → default 32768, which
+                // preserves detectChunkSize's chunk-sizing (it packs ~60% of context per chunk).
+                options: { num_predict: maxTokens, num_ctx: numCtx ?? 32768 },
             }),
             signal: AbortSignal.timeout(timeoutMs),
         });

package/dist/types.d.ts

@@ -303,6 +303,16 @@ export interface HicortexConfig {
      * See #220.
      */
     enableThinking?: boolean;
+    /**
+     * Context window for the ollama FAST tier (importance scoring via `completeFast`).
+     * Default 2048. Scoring prompts are ~850 tokens, so 2048 fits with headroom. The
+     * prior hardcoded 32768 — still the default for the heavy distill/reflect/classify
+     * tiers (which need it for `detectChunkSize`'s chunk-sizing) — made the KV-cache +
+     * prompt-cache accumulate past available RAM on memory-constrained boxes during long
+     * consolidations and swap-thrash the nightly. Applies to scoring only; raise if a
+     * scoring prompt actually needs more.
+     */
+    numCtx?: number;
 }
 /** A config-owned life-sphere domain (see HicortexConfig.domains). */
 export interface DomainDef {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gamaze/hicortex",
-  "version": "0.16.4",
+  "version": "0.16.5",
   "description": "Self-learning memory for AI agents — experience captured automatically, distilled into lessons overnight, shared across your whole fleet. Works with Hermes, OpenClaw, Claude Code, and Pi.",
   "main": "dist/index.js",
   "bin": {