@gamaze/hicortex 0.16.9 → 0.16.10

package/README.md

@@ -203,7 +203,7 @@ Config at `~/.hicortex/config.json`. Created by `init`. Key options:
 | `numCtx` | Context window for ollama (default 8192, one value for all phases). Scoring uses ~850 tokens, so 2048 is ample; distill/reflect/classify need more for `detectChunkSize`'s chunk sizing. |
 | `enableThinking` | Toggle the model's internal reasoning ("thinking") stream for OpenAI-compatible endpoints (default false). Only meaningful for local chat-template-aware servers (ollama, mlx-lm); leave unset for cloud OpenAI/OpenRouter/Groq endpoints (they 400 on the unknown `chat_template_kwargs` field). |
 | `maxTokens` | Max output tokens for all phases (default 8192). A ceiling, not a target — the model stops early when done. |
-| `ollamaFlushEvery` | Flush ollama's accumulated memory every N scoring calls (default 0 = off). Workaround for ollama's memory-growth bug (~171 MB/call). N=15 caps a cycle at ~2.5 GB. Scoped to the fast tier (scoring) only. |
+| `ollamaFlushEvery` | Flush ollama's accumulated memory every N scoring calls. **Off by default (0)** — opt-in only for an **ollama** install whose runner RSS growth (~171 MB/call) swap-thrashes long consolidations on a RAM-constrained box; N=15 caps a cycle at ~2.5 GB. Gated on the provider being ollama (local **or** remote) — no effect for non-ollama providers. Only you can judge whether your ollama endpoint actually suffers the growth (a managed/cloud ollama host may not), so it stays off until you set it. |
 | `ollamaFlushWaitMs` | Milliseconds to wait after an ollama flush for the runner to exit + release memory (default 180000 = 3 min). |
 | `authToken` | Bearer token for endpoint auth. Generated on first `init` in server mode. Find the active token with `hicortex status` or in `~/.hicortex/config.json`. |
 | `corsAllowedOrigins` | Browser origins allowed to read cross-origin responses, e.g. `["https://ui.example.com"]`. **Empty by default** — the server sends no `Access-Control-Allow-Origin` and never `Allow-Credentials`, so no external web page can read its data. The bundled `/viz` and `/context/ui` pages are same-origin and need no entry. |

package/dist/config-read.d.ts

@@ -29,3 +29,10 @@ export declare function readStrictBoolean(config: Record<string, unknown>, key:
  * value (e.g. ollamaFlushEvery).
  */
 export declare function readNonNegativeConfig(config: Record<string, unknown>, key: string, def: number): number;
+/**
+ * Warn if the saved config carries keys that 0.16.8+ ignores. Call at every
+ * config read (daemon boot + nightly). The warning clears once the keys are
+ * removed and (for the model keys) the model is consolidated into
+ * llmModel/llmBaseUrl/llmProvider.
+ */
+export declare function warnIgnoredConfigKeys(savedConfig: Record<string, unknown> | null | undefined): void;

package/dist/config-read.js

@@ -16,6 +16,7 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.readPositiveConfig = readPositiveConfig;
 exports.readStrictBoolean = readStrictBoolean;
 exports.readNonNegativeConfig = readNonNegativeConfig;
+exports.warnIgnoredConfigKeys = warnIgnoredConfigKeys;
 /**
  * Read a positive finite number from a config object. Returns `def` when the
  * key is absent OR present-but-invalid (with a warn in the latter case).
@@ -58,3 +59,49 @@ function readNonNegativeConfig(config, key, def) {
     console.warn(`[hicortex] config "${key}" = ${String(v)} is not a non-negative finite number — using default ${def}.`);
     return def;
 }
+/**
+ * Config keys that ≤0.16.7 accepted and 0.16.8+ IGNORES. Two groups:
+ *  - per-stage model keys + the nested `models` block (#231): one model now
+ *    serves all phases, so distillation/reflection/classification silently run
+ *    on `llmModel` — a quality downgrade if a larger model was on a distill/
+ *    reflect tier.
+ *  - `distillFallback` (#232): removed; strict mode is default. A `"local"`
+ *    value no longer falls back — failures retry next run.
+ * Both were public (documented in the README config table). Warn loudly at the
+ * config boundary so neither change is silent.
+ *
+ * NOTE: this list is deliberately WIDER than the `HicortexConfig` type. In
+ * ≤0.16.7 most of these keys (the flat `distill*`/`reflect*` set and
+ * `distillFallback`) were NEVER declared in the interface — they were
+ * accepted-but-untyped, read loosely off the config object, with the README
+ * config table as their only public contract. So deriving this list from the
+ * type would miss them. The list is the README's documented keys; don't trim it
+ * to match the interface (a future audit that did so would silently re-introduce
+ * the ignored-key gap this warning exists to close).
+ */
+const IGNORED_CONFIG_KEYS = [
+    "distillModel", "distillBaseUrl", "distillApiKey", "distillProvider",
+    "reflectModel", "reflectBaseUrl", "reflectApiKey", "reflectProvider",
+    "classifyModel", "classifyBaseUrl", "classifyApiKey", "classifyProvider",
+    "distillFallback",
+];
+/**
+ * Warn if the saved config carries keys that 0.16.8+ ignores. Call at every
+ * config read (daemon boot + nightly). The warning clears once the keys are
+ * removed and (for the model keys) the model is consolidated into
+ * llmModel/llmBaseUrl/llmProvider.
+ */
+function warnIgnoredConfigKeys(savedConfig) {
+    if (!savedConfig)
+        return;
+    const present = IGNORED_CONFIG_KEYS.filter((k) => savedConfig[k] !== undefined);
+    const hasModelsBlock = savedConfig.models !== undefined;
+    if (present.length === 0 && !hasModelsBlock)
+        return;
+    const detail = [...present, ...(hasModelsBlock ? ["models"] : [])].join(", ");
+    console.warn(`[hicortex] config has keys IGNORED since 0.16.8 (${detail}). They have no effect now. ` +
+        `Per-stage model keys / the \`models\` block: one model serves all phases — set ` +
+        `llmModel/llmBaseUrl/llmProvider (+ llmApiKey) to your intended model. ` +
+        `distillFallback: removed (strict mode is default — a failed distill retries next run). ` +
+        `Remove these keys to clear this warning. See the 0.16.8 changelog.`);
+}

package/dist/mcp-server.js

@@ -57,6 +57,7 @@ const zod_1 = require("zod");
 const db_js_1 = require("./db.js");
 const llm_js_1 = require("./llm.js");
 const features_js_1 = require("./features.js");
+const config_read_js_1 = require("./config-read.js");
 const state_js_1 = require("./state.js");
 const embedder_js_1 = require("./embedder.js");
 const storage = __importStar(require("./storage.js"));
@@ -373,6 +374,9 @@ async function startServer(options = {}) {
     // If nothing is configured: start recall-only with an unmissable warning.
     // One model serves all phases (#231) — no per-tier overlay here.
     const savedConfig = readConfigFile(stateDir);
+    // 0.16.8 upgrade guard: per-stage keys are silently ignored now. Warn loudly
+    // so a carried-over distill/reflect model doesn't silently downgrade quality.
+    (0, config_read_js_1.warnIgnoredConfigKeys)(savedConfig);
     // 0.16.2 activation gap: self-heal the agentId provenance field for
     // pre-0.16.2 server installs on first boot after upgrade. The server's own
     // nightly captures its sessions to localhost:8787/distill and needs this id;

package/dist/nightly.js

@@ -219,6 +219,8 @@ async function runNightly(options = {}) {
     (0, state_js_1.migrateLegacyState)(stateDir);
     // Check mode: client or server
     const savedConfig = readNightlyConfig(stateDir);
+    // 0.16.8 upgrade guard: warn if ignored per-stage keys are still present.
+    (0, config_read_js_1.warnIgnoredConfigKeys)(savedConfig);
     // 0.16.2 activation gap: pre-0.16.2 installs never re-run init, so their
     // config has no agentId → capture sent source_agent_id: null forever (the
     // provenance feature was inert for the whole existing fleet). Self-heal on

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gamaze/hicortex",
-  "version": "0.16.9",
+  "version": "0.16.10",
   "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": {