prism-mcp-server 13.0.1 → 13.1.1

package/README.md

@@ -19,33 +19,46 @@
 
 ### Downloads
 
-| Platform | Package | Size |
-|----------|---------|------|
-| **Windows** | [Prism Coder-1.0.0-Setup.exe](https://github.com/dcostenco/prism-coder/releases/download/v1.0.0/Prism.Coder-1.0.0-Setup.exe) | 99 MB |
-| **macOS (Apple Silicon)** | [Prism Coder-1.0.0-arm64.dmg](https://github.com/dcostenco/prism-coder/releases/download/v1.0.0/Prism.Coder-1.0.0-arm64.dmg) | 113 MB |
-| **Linux** | [Prism Coder-1.0.0.AppImage](https://github.com/dcostenco/prism-coder/releases/download/v1.0.0/Prism.Coder-1.0.0.AppImage) | 119 MB |
-| **npm (MCP Server)** | `npx -y prism-mcp-server` | — |
-
-| What You Get | Time Saved vs. Traditional |
+Latest IDE release: **v2.1.1** ([release notes](https://github.com/dcostenco/prism-coder/releases/tag/v2.1.1-ide)) — Merge conflict editor (Accept Current/Incoming/Both + auto-stage), enhanced Builder/Deploy/Workflow panels, all v2.1.0 features intact.
+
+| Platform | Package | Version | Size |
+|----------|---------|:-------:|:----:|
+| **macOS (Apple Silicon)** | [Prism Coder-2.1.1-arm64.dmg](https://github.com/dcostenco/prism-coder/releases/download/v2.1.1-ide/Prism.Coder-2.1.1-arm64.dmg) | **v2.1.1** | 115 MB |
+| **macOS (Apple Silicon, zip)** | [Prism Coder-2.1.1-arm64.zip](https://github.com/dcostenco/prism-coder/releases/download/v2.1.1-ide/Prism.Coder-2.1.1-arm64.zip) | **v2.1.1** | 110 MB |
+| **Windows (signed)** | [Prism Coder-2.1.1-Setup.exe](https://github.com/dcostenco/prism-coder/releases/download/v2.1.1-ide/Prism.Coder-2.1.1-Setup.exe) | **v2.1.1** | 103 MB |
+| **Linux (arm64)** | [Prism Coder-2.1.1.AppImage](https://github.com/dcostenco/prism-coder/releases/download/v2.1.1-ide/Prism.Coder-2.1.1.AppImage) | **v2.1.1** | 121 MB |
+| **npm (MCP Server)** | `npx -y prism-mcp-server` | latest | — |
+
+> **macOS:** Signed with Developer ID. First launch: right-click → **Open** if Gatekeeper warns.
+>
+> **Windows:** Authenticode-signed (DigiCert-timestamped). SmartScreen may warn — click **More info → Run anyway**.
+
+| What You Get | Status |
 |---|:---:|
-| 🤖 **Agent Mode** — autonomous multi-step task execution with diff previews | ~95% |
-| 🏗️ **Website Builder** — 6 templates, section editor, export to HTML/ZIP | ~90% |
-| 🎨 **Visual Drag & Drop** — 11 component types, canvas drop zone, live property editor | ~85% |
-| 🔑 **Auth & Database** — 6 auth providers, table CRUD, RLS, storage buckets | ~90% |
-| 🐳 **DevContainers** — 8 base images, port forwarding, resource limits, Codespaces export | ~80% |
-| 📋 **Customer Board (HIPAA)** — 12-pattern PHI scanner, moderator controls, ticket lifecycle | ~70% |
-| 🎨 **Media Studio** — AI image/video/3D generation, tier-gated quality | ~98% |
-| 🚀 **One-Click Deploy** — Vercel, Netlify, Synalux Cloud, custom server | ~98% |
-| 👥 **Real-Time Collaboration** — multiplayer editing with cursor presence | ~60% |
-| 📊 **SEO + Analytics** — 8-category audit + traffic dashboard | ~99% |
-| 🏪 **Marketplace** — 10-category extension registry, install with one click | ~90% |
-| 📋 **Workflow Engine** — natural language → structured project workflows | ~90% |
-| 🔀 **Git Integration** — branch, stage, commit, push without leaving IDE | ~60% |
-| 🌐 **12-Language i18n** — full UI translation including Arabic RTL | ~100% |
-
-**27/27 features** — more than any competitor (Cursor: 9, Windsurf: 9, Replit: 12, Bolt: 9).
-
-👉 **[Full IDE README with screenshots, architecture, and technical details →](https://github.com/dcostenco/prism-coder/releases/tag/v1.0.0)**
+| 💻 **Real Terminal** — xterm.js + node-pty, multi-tab, resize | ✅ v2.1 |
+| ✨ **Inline AI Completions** — ghost-text via local Ollama FIM, Tab to accept | ✅ v2.1 |
+| 🤖 **Agent Mode** — autonomous multi-step task execution with diff previews | ✅ |
+| 🎨 **Media Studio** — AI image/video/3D generation with prompt + tier-gated quality | ✅ v2.1 |
+| 🏗️ **Website Builder** — 6 templates, section editor, export to HTML | ✅ |
+| 🎨 **Visual Drag & Drop** — 11 component types, canvas drop zone, property editor | ✅ |
+| 📋 **Customer Board (HIPAA)** — 12-pattern PHI scanner, moderator controls, ticket lifecycle | ✅ |
+| 🔀 **Git Integration** — status, stage, commit, push, pull, branches, Monaco diff viewer | ✅ |
+| 🐙 **GitHub** — PRs, issues, actions via `gh` CLI | ✅ |
+| 📋 **Tasks Runner** — VS Code tasks.json parser, JSONC, streaming output | ✅ |
+| 🐛 **Debugger** — Node.js CDP, breakpoints, stepping, expression eval | ✅ |
+| 🔑 **Auth & Database** — 6 auth provider toggles, DB connect form, CLI guides | ✅ |
+| 🐳 **DevContainers** — 8 base images, docker run + .devcontainer.json export | ✅ |
+| 🏪 **Skill Browser** — scans ~/.agent/skills/, search, rescan | ✅ v2.1 |
+| 🚀 **Deploy** — env var management + CLI recipes (Vercel/Netlify/Custom) | ✅ |
+| 📋 **Workflow Engine** — 8-step default dev workflow, status tracking, persistence | ✅ |
+| 🌐 **12-Language i18n** — full UI translation including Arabic RTL | ✅ |
+| 📊 **SEO Audit** — URL input + Lighthouse trigger, analytics provider guide | ✅ v2.1 |
+| 🔒 **SSH Remote** — host registry, command execution, streaming output | ✅ |
+| 🎯 **Design-to-Code** — screenshot → Builder sections via vision model | ✅ |
+| 📜 **Version History** — git-backed snapshots per builder project | ✅ |
+| 👥 **Collaboration** — v2.2 (CRDT editing, presence, team chat) | 🟡 roadmap |
+
+👉 **[Full v2.1.1 release notes →](https://github.com/dcostenco/prism-coder/releases/tag/v2.1.1-ide)**
 
 ---
 
@@ -53,6 +66,8 @@
 
 **Your AI agent forgets everything between sessions. Prism fixes that — then teaches it to think.**
 
+> **🆕 May 2026 — `prism-coder:7b` v18aac released.** New on-device LLM backing the agent: AAC realigned eval **47/48 (98%)**, emergency_qa **13/13 perfect**, text_correct **15/15 perfect**, translate **8/8 perfect**, plus first-class gesture-recognition support for accessibility (smile/brow/jaw/blink/head-tilt). Built on Qwen2.5-Coder-7B-Instruct with full SFT on commercial-safe data (Apache 2.0 / CC-BY-4.0). Hybrid `prism-coder:7b-coder` (BFCL-optimized variant) coming soon for a top-rank function-calling track. See `prism/training/HYBRID_RESEARCH.md` for the dual-model deployment plan.
+
 <details>
 <summary>Cognitive Architecture Deep Dive (v12.5)</summary>
 
@@ -120,7 +135,7 @@ Prism v12.5 unifies Prism and Synalux into a **single billing architecture** wit
 - **10 Vendor Skills** — Vercel, Supabase, Stripe, Sentry, OpenAI, Addy Osmani, Garry Tan/gstack
 - **30+ Community Skills** — Installed and optimized from the Gemini ecosystem
 
-> 📦 **Packages:** [`prism-mcp-server`](https://www.npmjs.com/package/prism-mcp-server) (npm) · [Prism Coder IDE](https://github.com/dcostenco/prism-coder/releases/tag/v1.0.0) (VS Code) · `prism` CLI
+> 📦 **Packages:** [`prism-mcp-server`](https://www.npmjs.com/package/prism-mcp-server) (npm) · [Prism Coder IDE](https://github.com/dcostenco/prism-coder/releases/tag/v2.1.1-ide) (VS Code) · `prism` CLI
 
 ---
 
@@ -195,7 +210,7 @@ Prism features a cutting-edge **Zero-Search Retrieval** system for its cognitive
 | **BFCL Submission** | [PR #1332](https://github.com/ShishirPatil/gorilla/pull/1332) |
 | **Changelog** | [synalux-docs/CHANGELOG.md](https://github.com/dcostenco/synalux-docs/blob/main/CHANGELOG.md) |
 | **Roadmap** | [synalux-docs/ROADMAP.md](https://github.com/dcostenco/synalux-docs/blob/main/ROADMAP.md) |
-| **PrismAAC** | [github.com/dcostenco/prism-aac](https://github.com/dcostenco/prism-aac) — AAC web app for children with motor impairments |
+| **PrismAAC** | [github.com/dcostenco/prism-aac](https://github.com/dcostenco/prism-aac) — AAC web app for children with motor impairments. Powered by `prism-coder:7b` for offline text correction, AI Chat, translation, caregiver-note parsing, and emergency phone-call AI (13/13 on live Twilio test). Ships **all-neural offline TTS across 14 locales** via Kokoro-82M (Apache-2.0, 6 langs) + Piper (MIT, 5 langs) + Azure Neural fallback (Cantonese + emotional styles). Optional local servers `mlx-vlm` and `mlx-whisper` add scene understanding and HQ STT for Mac users. |
 
 ---
 
@@ -735,14 +750,17 @@ Prism scores coding tasks across **6 weighted heuristic signals** (keyword analy
 To achieve zero-latency, offline routing and memory compilation without cloud dependencies, Prism utilizes an internal fine-tuned ML model: **`prism-coder:7b`**.
 Built atop Qwen 2.5 Coder 7B using the MLX framework for Apple Silicon, this engine underwent aggressive Supervised Fine-Tuning (SFT) over 3,300+ session traces, then aligned using **GRPO (Group Relative Policy Optimization)** with a decomposed 4-component reward function.
 
-**Benchmark Results ([`training/benchmark.py`](https://github.com/dcostenco/prism-mcp/blob/main/training/benchmark.py), N=15 held-out):**
-- **Tool-Call Accuracy:** 100.0% — correct tool on unseen prompts (15/15)
-- **Tool Selection:** 100.0% (7/7) — perfect on all tool-call prompts
-- **Retrieval Accuracy:** 100.0% (3/3) — perfect on search/list/knowledge tasks
-- **JSON Validity:** 100.0% — every output parses as valid JSON
-- **Parameter Accuracy:** 73.3% — required params present when tool is correct
-- **Generation Speed:** 29.9 Tokens/sec (Apple M4 Max, 36GB)
-- **Avg Latency:** 2.2s per prompt
+**Benchmark Results — v12 production model:**
+- **BFCL Tool-Call Accuracy:** 100% (64/64)
+- **AAC realigned eval (production paths from prism-aac):** 89.6% overall — text-correct 80%, emergency Q&A **100% (13/13)**, translation 87.5%, caregiver-note parsing 85.7%, ask-AI 100%
+- **Live emergency phone call (Twilio):** 13/13 questions answered correctly with proactive safety guidance
+- **Generation Speed:** 130 ms / call on Apple M5 Max
+- **Quantization:** Q4_K_M GGUF, 4 GB on disk
+- **Distribution:** [`dcostenco/prism-coder-7b`](https://huggingface.co/dcostenco/prism-coder-7b) on HuggingFace; `ollama pull prism-coder:7b`
+
+**v18aac (current):** full SFT on commercial-safe data (Apache 2.0 / CC-BY-4.0). AAC eval 47/48 (98%), emergency_qa 13/13 perfect, text_correct 15/15, translate 8/8. First-class gesture-recognition support. DoRA rank=256 on Qwen2.5-Coder-7B-Instruct base.
+
+**Sister local servers (M-series Apple Silicon, optional):** when running on a Mac, prism-aac users can additionally stand up `mlx-vlm` (Qwen2.5-VL-7B for camera/scene understanding, port 8001) and `mlx-whisper` (large-v3-turbo for high-accuracy STT, port 8002) for fully-offline multimodal AI.
 
 **Integration**: Run via Ollama natively to power autonomous file operations and session routing entirely within the local host environment.
 
@@ -1043,8 +1061,9 @@ The Generator strips the `console.log`, resubmits, and the next `EVALUATE` retur
 
 ## <a name="whats-new"></a>🆕 What's New
 
-> **Current release: v12.5 — Unified Billing & Agent Skill Ecosystem**
+> **Current release: v12.6 — Prism Coder 14B sibling + tier-aware local routing**
 
+- 🤖 **v12.6 — Prism Coder 14B sibling + tier-aware local routing:** `prism-coder:7b` re-trained from clean Qwen2.5-Coder-7B base. **BFCL 88.1%** (3-run StdDev 0%, +40.9pp over previous prod), AAC realigned **47/48 (97.9%)**, caregiver targeted **20/20**, emergency_qa **13/13**, text_correct **15/15**, translate **8/8**. New **`prism-coder:14b`** sibling tag (BFCL 85.9%, AAC 46/48, 32K context) auto-routed for paid-tier medium-length queries via Synalux portal — keeps inference local, $0 marginal cost vs Claude/Gemini. Rollback: `ollama cp prism-coder:7b-prev-20260504-1325 prism-coder:7b`. Coordinated with synalux-private v0.14.4. → [Changelog](CHANGELOG.md)
 - 💳 **v12.5 — Unified Billing & Agent Skill Ecosystem:** Aligned Prism and Synalux tiers (Standard $19, Advanced $49, Enterprise $99), 14-day free trial, 54 agent skills, BSL-1.1 license. → [Changelog](CHANGELOG.md#1200)
 - 🏗️ **v11.6.0 — Agent Infrastructure Resilience:** Production-grade concurrent agent execution with serialized queue (Python `fcntl`), memory guardian daemon, queue watchdog, and unified status dashboard. 115/115 tests verified across 5 suites. → [Changelog](CHANGELOG.md#1160)
 - 🧠 **v11.5.1 — Structural GRPO Alignment:** GRPO-aligned local engine with held-out benchmark suite (N=15). **100% tool-call accuracy**, 100% JSON validity, 100% retrieval accuracy, 100% reasoning abstention. → [Changelog](CHANGELOG.md#1150)
@@ -1778,10 +1797,11 @@ Prism MCP is open-source and free for individual developers. For teams and enter
 
 ## <a name="milestones-roadmap"></a>📦 Milestones & Roadmap
 
-> **Current: v12.5** — Unified Billing & Agent Skill Ecosystem ([CHANGELOG](CHANGELOG.md))
+> **Current: v13.0** — The Adaptive Release ([CHANGELOG](CHANGELOG.md#1300))
 
 | Release | Headline |
 |---------|----------|
+| **v13.0** | 🧬 **The Adaptive Release** — BCBA behavioral brain (universal skill), project-aware skill auto-loading (23 skills), adaptive context injection, 83 security findings fixed, encrypted sync IV fix, timing-safe checksums. [Details →](CHANGELOG.md#1300) |
 | **v12.5** | 💳 **Unified Billing & Agent Skill Ecosystem** — Synalux-priced tiers ($19/$49/$99), 14-day trial, 54 skills, BSL-1.1 license. |
 | **v11.6.0** | 🏗️ **Agent Infrastructure Resilience** — Production-grade serialized queue, memory guardian, queue watchdog, status dashboard. 115/115 tests. |
 | **v11.5.1** | 🧠 **Structural GRPO Alignment** — Perfect 100% accuracy cross-validated on Synalux Elite platform. |
@@ -1797,8 +1817,7 @@ Prism MCP is open-source and free for individual developers. For teams and enter
 | **v7.0** | 🧬 ACT-R Activation Memory |
 
 ### Future Tracks
-- **v12.0: Distal Memory** — Semantic clustering of long-term history with Active-Prism background maintenance.
-- **v13.0: Team Handoff** — Encrypted peer-to-peer session syncing with multi-agent task routing and verifiable memory.
+- **v14.0: Prism-Coder v2** — Retrained model with adaptive behavioral alignment (training in progress).
 
 👉 **[Full ROADMAP.md →](ROADMAP.md)**
 

package/dist/utils/localLlm.js

@@ -22,6 +22,7 @@
  *   if (summary) { use(summary); } else { fallback to Gemini }
  */
 import { debugLog } from "./logger.js";
+import { normalizeToolCallFormat } from "./normalizeToolCallFormat.js";
 import { PRISM_LOCAL_LLM_ENABLED, PRISM_LOCAL_LLM_MODEL, PRISM_LOCAL_LLM_URL, PRISM_LOCAL_LLM_TIMEOUT_MS, } from "../config.js";
 // ─── Helpers ──────────────────────────────────────────────────────────────────
 /** Redact credentials from a URL for safe logging (strips user:pass@). */
@@ -149,7 +150,9 @@ export async function callLocalLlm(userPrompt, model = PRISM_LOCAL_LLM_MODEL, sy
         //   2. <|im_start|>...<|im_end|>          (Qwen native ChatML)
         //   3. <think>...<tool_call>              (standard format)
         // We normalize all to return just the clean content/JSON.
-        let content = rawContent;
+        // v18-clean coerces stochastic plural-wrapper / XML-attr / CJK-bracket
+        // emissions into canonical <tool_call>{json}</tool_call> first.
+        let content = normalizeToolCallFormat(rawContent);
         // Strip thinking blocks (all known formats)
         const thinkPatterns = [
             /<\|synalux_think\|>[\s\S]*?<\/\|synalux_think\|>\s*/,

package/dist/utils/normalizeToolCallFormat.js

@@ -0,0 +1,70 @@
+/**
+ * normalizeToolCallFormat — coerce stochastic Qwen2.5-Coder tool-call
+ * formats into the canonical singular-wrapper JSON form.
+ *
+ * v18-clean SFT models occasionally emit non-canonical formats despite
+ * training data being clean canonical. Three known variants in the wild:
+ *
+ *   1. plural wrapper + XML-attr params:
+ *      <tool_calls><tool_call name="X"><param name="Y" value="Z"/></tool_call></tool_calls>
+ *   2. CJK angle brackets:
+ *      〈tool_calls〉〈tool_call function_name="X"〉〈param name="Y" value="Z" /〉〈/tool_call〉〈/tool_calls〉
+ *   3. <functioncall> with stringified arguments:
+ *      <functioncall> {"name":"X","arguments":'{"Y":"Z"}'}
+ *
+ * Each is rewritten to:
+ *      <tool_call>
+ *      {"name":"X","arguments":{"Y":"Z"}}
+ *      </tool_call>
+ *
+ * Canonical inputs pass through unchanged.
+ */
+const PARAM_ATTR_RE = /<param\s+name="([^"]+)"\s+value="([^"]*)"\s*\/?>/gi;
+const XML_ATTR_TOOL_CALL_RE = /<tool_call\s+(?:name|function_name)="([^"]+)"[^>]*>([\s\S]*?)<\/tool_call>/gi;
+const PLURAL_WRAPPER_RE = /<tool_calls>\s*([\s\S]*?)\s*<\/tool_calls>/gi;
+const FUNCTION_CALL_RE = /<functioncall>\s*(\{[\s\S]*?\})(?=\s*$|\s*<|\s*\n\s*\n)/gi;
+function rewriteXmlAttrToolCalls(input) {
+    return input.replace(XML_ATTR_TOOL_CALL_RE, (_match, name, body) => {
+        const args = {};
+        PARAM_ATTR_RE.lastIndex = 0;
+        let pm;
+        while ((pm = PARAM_ATTR_RE.exec(body)) !== null) {
+            args[pm[1]] = pm[2];
+        }
+        return `<tool_call>\n${JSON.stringify({ name, arguments: args })}\n</tool_call>`;
+    });
+}
+export function normalizeToolCallFormat(text) {
+    if (!text)
+        return text;
+    let s = text;
+    // Step 1: replace CJK angle brackets so subsequent regex matches
+    if (s.includes("〈") || s.includes("〉")) {
+        s = s.replace(/〈/g, "<").replace(/〉/g, ">");
+    }
+    // Step 2: unwrap plural <tool_calls>...</tool_calls> and rewrite each inner call
+    s = s.replace(PLURAL_WRAPPER_RE, (_m, inner) => rewriteXmlAttrToolCalls(inner));
+    // Step 3: rewrite any remaining bare XML-attr <tool_call> outside plural wrapper
+    s = rewriteXmlAttrToolCalls(s);
+    // Step 4: rewrite <functioncall> {...} forms with stringified arguments
+    s = s.replace(FUNCTION_CALL_RE, (match, jsonish) => {
+        try {
+            // Strip single-quote wrapping that occasionally surrounds nested JSON args
+            const cleaned = jsonish
+                .replace(/"arguments":\s*'(\{[\s\S]*?\})'/g, '"arguments": $1')
+                .replace(/"arguments":\s*'([^']*)'/g, '"arguments": "$1"');
+            const obj = JSON.parse(cleaned);
+            if (obj && typeof obj.name === "string") {
+                const args = typeof obj.arguments === "string"
+                    ? JSON.parse(obj.arguments)
+                    : (obj.arguments ?? {});
+                return `<tool_call>\n${JSON.stringify({ name: obj.name, arguments: args })}\n</tool_call>`;
+            }
+        }
+        catch {
+            /* fall through — leave original */
+        }
+        return match;
+    });
+    return s;
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prism-mcp-server",
-  "version": "13.0.1",
+  "version": "13.1.1",
   "mcpName": "io.github.dcostenco/prism-coder",
   "description": "Prism-Coder v12.5: The world's first O(1) Cognitive Memory Architecture for AI Agents. 100% Tool-Call Accuracy (BFCL Gold Certified), 54 Agent Skills, Zero-Search Retrieval (HDC/HRR), HIPAA-hardened local-first storage, and SLERP-optimized GRPO alignment.",
   "module": "index.ts",