prism-mcp-server 15.1.0 → 15.2.1

package/README.md

@@ -113,21 +113,31 @@ The LLM context window is treated as ephemeral scratch space. All durable state
 
 ---
 
-## Plans
+## Models
+
+All Prism Coder inference uses **only fine-tuned Prism Coder models** — no Claude, no Gemini, no OpenRouter fallbacks. Models are exclusively accessible through the Synalux router (authentication + subscription required).
+
+| Model | Where | Tier | Latency |
+|---|---|---|---|
+| **Qwen3-1.7B** (fine-tuned) | On-device — iOS CoreML / Android ONNX | Free | ~50ms offline |
+| **Qwen3-14B** (fine-tuned) | RunPod A100 via Synalux | Standard+ | ~200ms |
+| **QwQ-32B** (fine-tuned) | RunPod A100 80GB via Synalux | Pro/Enterprise | ~3–5s |
+| **Qwen3-30B-A3B** (fine-tuned MoE) | RunPod via Synalux | Enterprise | ~2–3s |
 
-| | Free (local) | Paid (Synalux portal) |
-|---|---|---|
-| Local SQLite memory | ✅ | ✅ |
-| Semantic search | ✅ (local embedding) | ✅ (cloud-backed) |
-| Cross-device sync | — | ✅ |
-| Hivemind multi-agent | ✅ local team | ✅ + cloud roster |
-| Auto-Scholar (web research → memory) | — | ✅ |
-| HRR Zero-Search retrieval | ✅ | ✅ |
-| Custom domains / SSO | — | Enterprise |
+Fine-tuned on the 3-layer corpus: AAC + BFCL tool-calling + clinical workflows. BFCL gate: ≥ 90% on all tiers before production promotion. Adapters stored at `dcostenco/prism-coder-*` (private HuggingFace).
 
-The thin-client architecture: when authenticated to Synalux, Prism Coder routes through the portal for paid features. When not authenticated (or `PRISM_FORCE_LOCAL=1`), runs purely local. Same binary.
+## Plans
+
+| | Free | Standard $19/mo | Pro $49/mo | Enterprise $99/mo |
+|---|---|---|---|---|
+| Qwen3-1.7B on-device | ✅ unlimited | ✅ | ✅ | ✅ |
+| Qwen3-14B cloud | — | ✅ 200 req/day | ✅ 2K req/day | ✅ unlimited |
+| QwQ-32B reasoning | — | — | ✅ | ✅ priority |
+| Qwen3-30B-A3B MoE | — | — | — | ✅ |
+| Custom fine-tuning | — | — | — | ✅ |
+| HIPAA BAA | — | — | — | ✅ |
 
-[Pricing →](https://synalux.ai/pricing)
+[Subscribe →](https://synalux.ai/pricing)
 
 ---
 
@@ -174,6 +184,52 @@ As of v14.0.0, Prism's algorithm exports are a **stable public contract** under
 | [PrismAAC](https://github.com/dcostenco/prism-aac) | Spreading-activation phrase ranking (recency × frequency × per-user history). Caregiver corrections auto-harvest into the personalization corpus via the audit-hooks postflight harvester. The on-device 7B model + this algorithm stack is what makes PrismAAC defensible. |
 | Synalux portal | Tier-aware model routing using experience bias on prior outcomes per fingerprint. HIPAA-compliant clinical scribe with on-device-first privacy guarantees. |
 
+## Synalux Inference Router — Architecture (v16)
+
+All Prism AAC model inference is protected behind Synalux as a mandatory router. Models are **never accessible directly** — all traffic goes through Synalux for auth, billing, and rate limiting.
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                      CLIENT LAYER                           │
+│  prism-aac (iOS/web)         │   Synalux Portal             │
+└──────────────┬──────────────────────────────────────────────┘
+               │ POST /api/v1/prism-aac/inference
+               │ Authorization: Bearer <user-JWT>
+               ▼
+┌─────────────────────────────────────────────────────────────┐
+│                   SYNALUX ROUTER                            │
+│  1. Verify JWT (no anonymous access)                        │
+│  2. Check subscription tier                                 │
+│  3. Enforce rate limit (50–2000 req/day by plan)            │
+│  4. Route to model tier by complexity                       │
+│  5. Proxy → RunPod with SECRET key (never sent to client)   │
+│  6. Log → aac_inference_log (billing audit trail)           │
+└──────────┬─────────────────────────────────────┬────────────┘
+           │ tier=fast                            │ tier=reason
+           ▼                                      ▼
+  ┌──────────────────┐               ┌───────────────────────┐
+  │  Qwen3-14B       │               │  QwQ-32B              │
+  │  RunPod A100 40G │               │  RunPod A100 80G      │
+  │  ~200ms          │               │  ~3–5s (reasoning)    │
+  │  standard/pro    │               │  pro/enterprise only  │
+  └──────────────────┘               └───────────────────────┘
+           │                                      │
+           └────────────────┬─────────────────────┘
+                            ▼
+               HuggingFace dcostenco/prism-coder-* (private)
+               RunPod pulls at pod start with server-side token
+
+On-device (free, zero latency, offline):
+  Qwen3-1.7B GGUF Q4_K_M → iOS CoreML / Android ONNX
+```
+
+| Plan | Cloud model | Daily limit | On-device |
+|---|---|---|---|
+| Free | — | unlimited local | Qwen3-1.7B |
+| Standard $5/mo | Qwen3-14B | 200 req | + cloud |
+| Pro $15/mo | QwQ-32B | 2,000 req | + reasoning |
+| Enterprise | QwQ-32B priority | unlimited | full stack |
+
 See [`docs/WOW_FEATURES.md`](docs/WOW_FEATURES.md) for the algorithm catalogue. Release notes in [`docs/releases/v14.0.0-prism-as-foundation.md`](docs/releases/v14.0.0-prism-as-foundation.md).
 
 ---

package/dist/dashboard/server.js

@@ -567,7 +567,9 @@ return false;}
                 res.writeHead(200, { "Content-Type": "application/json" });
                 return res.end(JSON.stringify({ skills }));
             }
-            // POST /api/skills → { role, content } saves skill:<role>
+            // POST /api/skills → { role, content } saves user_skill:<role> (user-local namespace).
+            // Platform skills (skill:*) are read-only — populated only by sync-skills.sh
+            // or fetched from Synalux. Users cannot overwrite platform skills from the dashboard.
             if (url.pathname === "/api/skills" && req.method === "POST") {
                 const body = await readBody(req);
                 const { role, content } = JSON.parse(body || "{}");
@@ -575,16 +577,16 @@ return false;}
                     res.writeHead(400);
                     return res.end(JSON.stringify({ error: "role required" }));
                 }
-                await setSetting(`skill:${role}`, content || "");
+                await setSetting(`user_skill:${role}`, content || "");
                 res.writeHead(200, { "Content-Type": "application/json" });
-                return res.end(JSON.stringify({ ok: true, role }));
+                return res.end(JSON.stringify({ ok: true, role, namespace: "user_skill" }));
             }
-            // DELETE /api/skills/:role → clears skill:<role>
+            // DELETE /api/skills/:role → clears user_skill:<role> (user-local only).
             if (url.pathname.startsWith("/api/skills/") && req.method === "DELETE") {
                 const role = url.pathname.replace("/api/skills/", "");
-                await setSetting(`skill:${role}`, "");
+                await setSetting(`user_skill:${role}`, "");
                 res.writeHead(200, { "Content-Type": "application/json" });
-                return res.end(JSON.stringify({ ok: true, role }));
+                return res.end(JSON.stringify({ ok: true, role, namespace: "user_skill" }));
             }
             // ─── API: Knowledge Graph (v6.2 — extracted to graphRouter.ts) ───
             if (url.pathname.startsWith("/api/graph")) {

package/dist/tools/ledgerHandlers.js

@@ -798,12 +798,19 @@ export async function sessionLoadContextHandler(args) {
         }
     }
     // ─── Project-Aware Skill Injection ──────────────────────────
-    // Routing (WHICH skills): always from Synalux /api/v1/skills/routing.
-    // Content (WHAT): paid tier → Synalux /api/v1/skills/content (batch fetch,
-    // single source of truth). Free tier / offline → local SQLite fallback.
+    // Routing (WHICH skills + user_local policy): Synalux /api/v1/skills/routing.
+    // Content (WHAT):
+    //   Platform skills  → Synalux /api/v1/skills/content (DB first, filesystem fallback)
+    //                      → local SQLite skill:<name> (free tier / offline fallback)
+    //   User-local skills → local SQLite user_skill:<name>
+    //                       ONLY when user_local.enabled=true in routing table
+    //                       OR session_load_context called with user_local=true.
+    //                       Users CANNOT write to the platform skill: namespace.
     const { resolveSkillsForProject } = await import("./skillRouting.js");
-    const skillsToLoad = await resolveSkillsForProject(project);
-    // Paid tier: batch-fetch all skill content from Synalux portal in one request.
+    const resolved = await resolveSkillsForProject(project);
+    const skillsToLoad = resolved.names;
+    const userLocalPolicy = resolved.user_local;
+    // Paid tier: batch-fetch platform skill content from Synalux in one request.
     let synaluxContent = {};
     if (SYNALUX_CONFIGURED && storage && typeof storage.fetchSkillContent === "function") {
         const missing = skillsToLoad.filter(n => !loadedSkills.includes(n));
@@ -814,19 +821,38 @@ export async function sessionLoadContextHandler(args) {
     for (const skillName of skillsToLoad) {
         if (loadedSkills.includes(skillName))
             continue;
-        // Prefer Synalux content (always up-to-date); fall back to local SQLite.
+        // Synalux (paid) → platform fallback skill:<name> (free/offline). Never user_skill:.
         const content = synaluxContent[skillName] || await getSetting(`skill:${skillName}`, "");
         if (content && content.trim()) {
-            const source = synaluxContent[skillName] ? "synalux" : "local";
+            const source = synaluxContent[skillName] ? "synalux" : "local-platform";
             skillBlock += `\n\n[📜 SKILL: ${skillName}]\n${content.trim()}`;
             loadedSkills.push(skillName);
             skillLoaded = true;
             debugLog(`[session_load_context] Skill "${skillName}" loaded (${source}) for project="${project}"`);
         }
     }
+    // ─── User-Local Skills ──────────────────────────────────────
+    // Loaded ONLY when user_local.enabled=true (set in Synalux routing table
+    // or explicitly requested). Stored under user_skill: prefix — users can
+    // write here via dashboard; they CANNOT write to the platform skill: keys.
+    if (userLocalPolicy.enabled) {
+        const prefix = userLocalPolicy.key_prefix || "user_skill:";
+        const allSettings = await storage.getAllSettings?.() || {};
+        for (const [k, v] of Object.entries(allSettings)) {
+            if (!k.startsWith(prefix) || !v)
+                continue;
+            const skillName = k.replace(prefix, "");
+            if (loadedSkills.includes(skillName))
+                continue;
+            skillBlock += `\n\n[📜 USER SKILL: ${skillName}]\n${v.trim()}`;
+            loadedSkills.push(skillName);
+            skillLoaded = true;
+            debugLog(`[session_load_context] User-local skill "${skillName}" loaded`);
+        }
+    }
     // ─── Memory-Based Skill Discovery ──────────────────────────
-    // If recent handoff/ledger mentions a skill name, auto-load it.
-    // This lets the agent's own memory drive skill activation.
+    // If recent handoff/ledger mentions a platform skill name, auto-load it.
+    // Only scans platform skill: keys — user_skill: discovery is not automatic.
     if (formattedContext.length > 0) {
         const contextText = formattedContext.toLowerCase();
         const allSkillKeys = await storage.getAllSettings?.() || {};
@@ -836,7 +862,6 @@ export async function sessionLoadContextHandler(args) {
             const skillName = k.replace("skill:", "");
             if (loadedSkills.includes(skillName))
                 continue;
-            // Only load if the skill name appears in recent context
             if (contextText.includes(skillName.replace(/-/g, " ")) || contextText.includes(skillName)) {
                 skillBlock += `\n\n[📜 CONTEXT SKILL: ${skillName}]\n${v}`;
                 loadedSkills.push(skillName);

package/dist/tools/skillRouting.js

@@ -16,12 +16,12 @@
  * Do NOT add hardcoded skill names here outside the OFFLINE_FALLBACK block
  * — that defeats the single-source-of-truth design.
  */
-// Minimal fallback when synalux is unreachable. Only the universal BCBA
-// skill — project-specific mappings need synalux to resolve.
+// Minimal fallback when synalux is unreachable.
 const OFFLINE_FALLBACK = {
     version: 1,
     universal: ['bcba_ai_assistant'],
     projects: {},
+    user_local: { enabled: false, key_prefix: 'user_skill:' },
 };
 const SYNALUX_BASE = process.env.SYNALUX_BASE_URL || 'https://synalux.ai';
 const CACHE_TTL_MS = 5 * 60 * 1000;
@@ -53,7 +53,8 @@ async function fetchOnce() {
 /**
  * Resolve the skill list for a given project (case-insensitive substring
  * match against the routing table). Always returns at least the universal
- * skills.
+ * skills. Also returns the user_local policy so callers know whether to
+ * load user_skill:* entries from local SQLite.
  */
 export async function resolveSkillsForProject(project) {
     const now = Date.now();
@@ -75,7 +76,10 @@ export async function resolveSkillsForProject(project) {
                 out.add(s);
         }
     }
-    return Array.from(out);
+    return {
+        names: Array.from(out),
+        user_local: table.user_local ?? OFFLINE_FALLBACK.user_local,
+    };
 }
 /** Force a re-fetch on the next call. Exposed for tests + admin tooling. */
 export function _invalidateRoutingCache() {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prism-mcp-server",
-  "version": "15.1.0",
+  "version": "15.2.1",
   "mcpName": "io.github.dcostenco/prism-coder",
   "description": "Prism Coder — Cognitive memory + tool-calling intelligence for AI agents. Mind Palace persistent memory (BFCL Gold Certified, 100% Tool-Call Accuracy, 54 Agent Skills, Zero-Search HDC/HRR retrieval, HIPAA-hardened local-first storage, SLERP-optimized GRPO alignment) plus the prism-coder:7b / 14b open-weights LLM fleet.",
   "module": "index.ts",