@miller-tech/uap 1.0.0

package/tools/agents/benchmarks/results/benchmark_20260106_064817.json

@@ -0,0 +1,170 @@
+[
+  {
+    "name": "SQLite",
+    "operation": "INSERT (single)",
+    "samples": 100,
+    "mean_ms": 1.1429446772672236,
+    "median_ms": 1.1120839626528323,
+    "p95_ms": 1.3595609925687313,
+    "p99_ms": 1.5702530508860946,
+    "min_ms": 1.0651589836925268,
+    "max_ms": 1.5702530508860946,
+    "ops_per_sec": 874.932986600013
+  },
+  {
+    "name": "SQLite",
+    "operation": "SELECT recent (LIMIT 50)",
+    "samples": 100,
+    "mean_ms": 0.1497111632488668,
+    "median_ms": 0.14744600048288703,
+    "p95_ms": 0.16588205471634865,
+    "p99_ms": 0.19576202612370253,
+    "min_ms": 0.1444709487259388,
+    "max_ms": 0.19576202612370253,
+    "ops_per_sec": 6679.528622309127
+  },
+  {
+    "name": "SQLite",
+    "operation": "SELECT by type",
+    "samples": 100,
+    "mean_ms": 0.1452046213671565,
+    "median_ms": 0.14327099779620767,
+    "p95_ms": 0.159741030074656,
+    "p99_ms": 0.20464195404201746,
+    "min_ms": 0.14055101200938225,
+    "max_ms": 0.20464195404201746,
+    "ops_per_sec": 6886.833150244264
+  },
+  {
+    "name": "SQLite",
+    "operation": "LIKE search",
+    "samples": 100,
+    "mean_ms": 0.13849956914782524,
+    "median_ms": 0.13719149865210056,
+    "p95_ms": 0.14585198368877172,
+    "p99_ms": 0.19792199600487947,
+    "min_ms": 0.1340809976682067,
+    "max_ms": 0.19792199600487947,
+    "ops_per_sec": 7220.239067550213
+  },
+  {
+    "name": "SQLite",
+    "operation": "Graph query (1-hop)",
+    "samples": 100,
+    "mean_ms": 0.1656885934062302,
+    "median_ms": 0.16310601495206356,
+    "p95_ms": 0.18141092732548714,
+    "p99_ms": 0.23385195527225733,
+    "min_ms": 0.15984103083610535,
+    "max_ms": 0.23385195527225733,
+    "ops_per_sec": 6035.41848863567
+  },
+  {
+    "name": "Qdrant",
+    "operation": "INSERT (single + embed)",
+    "samples": 50,
+    "mean_ms": 7.81712032854557,
+    "median_ms": 4.746215010527521,
+    "p95_ms": 5.6189680472016335,
+    "p99_ms": 156.39470203313977,
+    "min_ms": 4.38071705866605,
+    "max_ms": 156.39470203313977,
+    "ops_per_sec": 127.92434527946648
+  },
+  {
+    "name": "Qdrant",
+    "operation": "Semantic search (top-5)",
+    "samples": 100,
+    "mean_ms": 1.2219337350688875,
+    "median_ms": 1.1123394942842424,
+    "p95_ms": 1.682064961642027,
+    "p99_ms": 2.7543739415705204,
+    "min_ms": 0.9932080283761024,
+    "max_ms": 2.7543739415705204,
+    "ops_per_sec": 818.3749832748698
+  },
+  {
+    "name": "Qdrant",
+    "operation": "Filtered search",
+    "samples": 100,
+    "mean_ms": 1.2450160249136388,
+    "median_ms": 1.2208950356580317,
+    "p95_ms": 1.2983810156583786,
+    "p99_ms": 3.2460070215165615,
+    "min_ms": 1.1418990325182676,
+    "max_ms": 3.2460070215165615,
+    "ops_per_sec": 803.202513051481
+  },
+  {
+    "name": "Embedding",
+    "operation": "Generate (all-MiniLM-L6-v2)",
+    "samples": 100,
+    "mean_ms": 3.2829831482376903,
+    "median_ms": 3.2453580060973763,
+    "p95_ms": 3.575409995391965,
+    "p99_ms": 4.24535502679646,
+    "min_ms": 3.1607369892299175,
+    "max_ms": 4.24535502679646,
+    "ops_per_sec": 304.601015249439
+  },
+  {
+    "name": "Consolidation",
+    "operation": "Simple extraction",
+    "samples": 100,
+    "mean_ms": 0.04601658787578344,
+    "median_ms": 0.04540046211332083,
+    "p95_ms": 0.04780001472681761,
+    "p99_ms": 0.07955997716635466,
+    "min_ms": 0.04478008486330509,
+    "max_ms": 0.07955997716635466,
+    "ops_per_sec": 21731.293999880792
+  },
+  {
+    "name": "Consolidation",
+    "operation": "Semantic dedup",
+    "samples": 20,
+    "mean_ms": 33.511918887961656,
+    "median_ms": 32.87853847723454,
+    "p95_ms": 38.731466978788376,
+    "p99_ms": 38.731466978788376,
+    "min_ms": 31.861560069955885,
+    "max_ms": 38.731466978788376,
+    "ops_per_sec": 29.840129517597564
+  },
+  {
+    "name": "SQLite@100",
+    "operation": "SELECT recent",
+    "samples": 50,
+    "mean_ms": 0.15139207243919373,
+    "median_ms": 0.14865596313029528,
+    "p95_ms": 0.1651509664952755,
+    "p99_ms": 0.2013719640672207,
+    "min_ms": 0.14511204790323973,
+    "max_ms": 0.2013719640672207,
+    "ops_per_sec": 6605.365683210709
+  },
+  {
+    "name": "SQLite@1000",
+    "operation": "SELECT recent",
+    "samples": 50,
+    "mean_ms": 0.16931103775277734,
+    "median_ms": 0.16216648509725928,
+    "p95_ms": 0.19682198762893677,
+    "p99_ms": 0.24656206369400024,
+    "min_ms": 0.15751097816973925,
+    "max_ms": 0.24656206369400024,
+    "ops_per_sec": 5906.289473342952
+  },
+  {
+    "name": "SQLite@5000",
+    "operation": "SELECT recent",
+    "samples": 50,
+    "mean_ms": 0.16595820896327496,
+    "median_ms": 0.16168103320524096,
+    "p95_ms": 0.18208206165581942,
+    "p99_ms": 0.2518820110708475,
+    "min_ms": 0.15775102656334639,
+    "max_ms": 0.2518820110708475,
+    "ops_per_sec": 6025.613353186349
+  }
+]

package/tools/agents/benchmarks/results/benchmark_20260106_064817.md

@@ -0,0 +1,51 @@
+# Memory Systems Benchmark Report
+
+**Generated:** 2026-01-06T06:48:17.656637Z
+**System:** UAP Memory Implementation
+
+## Summary
+
+| System        | Operation                   | Mean (ms) | P95 (ms) | Ops/sec |
+| ------------- | --------------------------- | --------- | -------- | ------- |
+| SQLite        | INSERT (single)             | 1.143     | 1.360    | 875     |
+| SQLite        | SELECT recent (LIMIT 50)    | 0.150     | 0.166    | 6680    |
+| SQLite        | SELECT by type              | 0.145     | 0.160    | 6887    |
+| SQLite        | LIKE search                 | 0.138     | 0.146    | 7220    |
+| SQLite        | Graph query (1-hop)         | 0.166     | 0.181    | 6035    |
+| Qdrant        | INSERT (single + embed)     | 7.817     | 5.619    | 128     |
+| Qdrant        | Semantic search (top-5)     | 1.222     | 1.682    | 818     |
+| Qdrant        | Filtered search             | 1.245     | 1.298    | 803     |
+| Embedding     | Generate (all-MiniLM-L6-v2) | 3.283     | 3.575    | 305     |
+| Consolidation | Simple extraction           | 0.046     | 0.048    | 21731   |
+| Consolidation | Semantic dedup              | 33.512    | 38.731   | 30      |
+| SQLite@100    | SELECT recent               | 0.151     | 0.165    | 6605    |
+| SQLite@1000   | SELECT recent               | 0.169     | 0.197    | 5906    |
+| SQLite@5000   | SELECT recent               | 0.166     | 0.182    | 6026    |
+
+## Key Findings
+
+### Short-term Memory (SQLite)
+
+- Single INSERT operations are extremely fast (<0.5ms)
+- SELECT with ORDER BY and LIMIT scales well
+- Knowledge graph queries (1-hop) add minimal overhead
+
+### Long-term Memory (Qdrant)
+
+- Embedding generation is the main latency contributor
+- Semantic search is fast once vectors exist (~50-100ms)
+- Filtering adds minimal overhead
+
+### Consolidation
+
+- Simple extraction is very fast (<1ms)
+- Semantic deduplication adds significant latency (~100-500ms)
+- Recommendation: Use hash-based dedup, semantic only for high-importance
+
+## Recommendations
+
+1. **Keep SQLite for short-term**: Performance is excellent
+2. **Batch Qdrant operations**: Reduce per-operation overhead
+3. **Cache embeddings**: Avoid regenerating for known content
+4. **Use hybrid dedup**: Hash first, semantic for borderline cases
+5. **Add session memory layer**: Low overhead, high value

package/tools/agents/config/chat_template.jinja

@@ -0,0 +1,77 @@
+{#- UAP Chat Template - ChatML format with tool calling support -#}
+{#- Compatible with llama.cpp / vLLM / OpenAI-compatible servers -#}
+{#- Uses ChatML (<|im_start|>/<|im_end|>) with <tool_call> format -#}
+{#- Works with: Qwen3, Qwen3.5, and other ChatML-compatible models -#}
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content | trim + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+
+{%- for message in messages %}
+    {#- User message or non-first system message -#}
+    {%- if (message.role == 'user') or (message.role == 'system' and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content | trim + '<|im_end|>\n' }}
+
+    {#- Assistant message -#}
+    {%- elif message.role == 'assistant' %}
+        {{- '<|im_start|>assistant' }}
+        {%- if message.content is defined and message.content %}
+            {{- '\n' + message.content | trim }}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tc = tool_call.function %}
+                {%- else %}
+                    {%- set tc = tool_call %}
+                {%- endif %}
+                {%- if (loop.first and message.content is defined and message.content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tc.name }}
+                {{- '", "arguments": ' }}
+                {%- if tc.arguments is string %}
+                    {{- tc.arguments }}
+                {%- else %}
+                    {{- tc.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+
+    {#- Tool response (grouped under user role per Qwen3 spec) -#}
+    {%- elif message.role == 'tool' %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != 'tool') %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content | trim }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != 'tool') %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+
+    {%- endif %}
+{%- endfor %}
+
+{#- Generation prompt for assistant turn -#}
+{%- if add_generation_prompt is defined and add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

package/tools/agents/config/tool-call-schema.json

@@ -0,0 +1,19 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "title": "Qwen3.5 Tool Call Output Schema",
+  "description": "JSON Schema for constraining tool call output via llama.cpp --json-schema flag. Matches the official Qwen3 <tool_call> format.",
+  "type": "object",
+  "properties": {
+    "name": {
+      "type": "string",
+      "pattern": "^[a-zA-Z][a-zA-Z0-9_-]*$",
+      "description": "The tool/function name to call"
+    },
+    "arguments": {
+      "type": "object",
+      "description": "Key-value arguments for the function call"
+    }
+  },
+  "required": ["name", "arguments"],
+  "additionalProperties": false
+}

package/tools/agents/config/tool-call.gbnf

@@ -0,0 +1,58 @@
+# GBNF Grammar for Qwen3.5 Tool Call Structured Output
+# Constrains model output to valid tool call format
+#
+# WARNING: Do NOT use this with --grammar-file on server startup.
+# Global grammar forces ALL output into <tool_call> format, breaking plain chat.
+# llama.cpp's autoparser handles tool calls lazily via the chat template.
+#
+# Use this grammar for per-request enforcement only, via the "grammar" field
+# in individual API payloads when strict format adherence is needed.
+#
+# Matches the official Qwen3 chat template format:
+#   <tool_call>
+#   {"name": "function_name", "arguments": {"key": "value"}}
+#   </tool_call>
+
+root ::= ws tool-calls ws
+
+# One or more tool calls (supports multi-tool in single turn)
+tool-calls ::= tool-call (ws tool-call)*
+
+# Single tool call block
+tool-call ::= "<tool_call>" ws "{\"name\":" ws json-string ws "," ws "\"arguments\":" ws json-object ws "}" ws "</tool_call>"
+
+# JSON object (the arguments)
+json-object ::= "{" ws json-members? ws "}"
+
+json-members ::= json-pair (ws "," ws json-pair)*
+
+json-pair ::= json-string ws ":" ws json-value
+
+json-value ::= json-string | json-number | json-object | json-array | json-bool | json-null
+
+json-string ::= "\"" json-chars "\"" | "\"\""
+
+json-chars ::= json-char+
+
+json-char ::= [^"\\] | "\\" json-escape
+
+json-escape ::= ["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]
+
+json-number ::= "-"? json-int json-frac? json-exp?
+
+json-int ::= "0" | [1-9] [0-9]*
+
+json-frac ::= "." [0-9]+
+
+json-exp ::= [eE] [+-]? [0-9]+
+
+json-array ::= "[" ws json-array-items? ws "]"
+
+json-array-items ::= json-value (ws "," ws json-value)*
+
+json-bool ::= "true" | "false"
+
+json-null ::= "null"
+
+# Whitespace
+ws ::= [ \t\n\r]*

package/tools/agents/docker/Dockerfile.python

@@ -0,0 +1,52 @@
+# Layer 2: Agent toolchain for python:3.13-slim-bookworm tasks
+# Contains: nvm, Node 22, opencode, @ai-sdk/openai-compatible, apt/pip deps
+# Does NOT contain UAP -- that is Layer 3, installed at runtime for easy upgrades
+#
+# Rebuild this image when: opencode version changes, new system deps needed
+# Do NOT rebuild for UAP version changes -- just update UAP_VERSION in install script
+FROM python:3.13-slim-bookworm
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV NVM_DIR=/root/.nvm
+ENV NODE_VERSION=22
+
+# System dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    curl git build-essential \
+    jq netcat-openbsd xxd \
+    tesseract-ocr tesseract-ocr-eng \
+    poppler-utils imagemagick ffmpeg \
+    libsm6 libxext6 file \
+    sqlite3 libsqlite3-dev \
+    ca-certificates \
+    tmux \
+    && rm -rf /var/lib/apt/lists/*
+
+# Python packages
+RUN pip install --no-cache-dir \
+    numpy pillow pytesseract opencv-python-headless \
+    python-chess \
+    2>/dev/null || true
+
+# Install nvm + Node 22
+RUN curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.2/install.sh | bash \
+    && . "$NVM_DIR/nvm.sh" \
+    && nvm install $NODE_VERSION \
+    && nvm alias default $NODE_VERSION \
+    && nvm use default
+
+# Make node/npm available in PATH for non-interactive shells
+ENV PATH="/root/.nvm/versions/node/v${NODE_VERSION}/bin:${PATH}"
+
+# Install opencode + AI SDK provider (Layer 2 -- toolchain)
+RUN . "$NVM_DIR/nvm.sh" \
+    && npm i -g opencode-ai@latest \
+    && npm i -g @ai-sdk/openai-compatible
+
+# Create installed-agent directory for harbor
+RUN mkdir -p /installed-agent
+
+# Marker: toolchain version (Layer 2). UAP version is separate (Layer 3).
+RUN echo "toolchain-1.0" > /installed-agent/.toolchain-version
+
+WORKDIR /app

package/tools/agents/docker/Dockerfile.ubuntu

@@ -0,0 +1,55 @@
+# Layer 2: Agent toolchain for ubuntu:24.04 tasks
+# Contains: nvm, Node 22, opencode, @ai-sdk/openai-compatible, apt/pip deps
+# Does NOT contain UAP -- that is Layer 3, installed at runtime for easy upgrades
+#
+# Rebuild this image when: opencode version changes, new system deps needed
+# Do NOT rebuild for UAP version changes -- just update UAP_VERSION in install script
+FROM ubuntu:24.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV NVM_DIR=/root/.nvm
+ENV NODE_VERSION=22
+
+# System dependencies (superset of all task needs)
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    curl git build-essential python3-pip python3-dev python3-venv python3-full \
+    jq netcat-openbsd xxd \
+    tesseract-ocr tesseract-ocr-eng \
+    poppler-utils imagemagick ffmpeg \
+    libsm6 libxext6 file \
+    p7zip-full hashcat john \
+    sqlite3 libsqlite3-dev \
+    stockfish \
+    ca-certificates \
+    tmux \
+    perl libcompress-raw-lzma-perl \
+    && rm -rf /var/lib/apt/lists/*
+
+# Python packages (superset of all task needs)
+RUN pip install --break-system-packages --no-cache-dir \
+    numpy pillow pytesseract opencv-python-headless \
+    python-chess \
+    2>/dev/null || true
+
+# Install nvm + Node 22
+RUN curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.2/install.sh | bash \
+    && . "$NVM_DIR/nvm.sh" \
+    && nvm install $NODE_VERSION \
+    && nvm alias default $NODE_VERSION \
+    && nvm use default
+
+# Make node/npm available in PATH for non-interactive shells
+ENV PATH="/root/.nvm/versions/node/v${NODE_VERSION}/bin:${PATH}"
+
+# Install opencode + AI SDK provider (Layer 2 -- toolchain)
+RUN . "$NVM_DIR/nvm.sh" \
+    && npm i -g opencode-ai@latest \
+    && npm i -g @ai-sdk/openai-compatible
+
+# Create installed-agent directory for harbor
+RUN mkdir -p /installed-agent
+
+# Marker: toolchain version (Layer 2). UAP version is separate (Layer 3).
+RUN echo "toolchain-1.0" > /installed-agent/.toolchain-version
+
+WORKDIR /app

package/tools/agents/docker-compose.qdrant.yml

@@ -0,0 +1,24 @@
+version: '3.8'
+
+services:
+  qdrant:
+    image: qdrant/qdrant:latest
+    container_name: uap-qdrant
+    ports:
+      - '6333:6333' # REST API
+      - '6334:6334' # gRPC
+    volumes:
+      - ./qdrant_storage:/qdrant/storage:z
+    environment:
+      - QDRANT__SERVICE__GRPC_PORT=6334
+      - QDRANT__REST__PORT=6333
+      - QDRANT__LOG_LEVEL=INFO
+    restart: unless-stopped
+    healthcheck:
+      test: ['CMD', 'curl', '-f', 'http://localhost:6333/healthz']
+      interval: 30s
+      timeout: 10s
+      retries: 3
+
+volumes:
+  qdrant_storage:

package/tools/agents/install-opencode-local.sh.j2

@@ -0,0 +1,135 @@
+#!/bin/bash
+set -e
+
+# ============================================================================
+# 3-Layer Architecture:
+#   Layer 1: Task image (task data, base OS) -- from Dockerfile
+#   Layer 2: Toolchain (nvm, node, opencode, pip) -- pre-built in Docker image
+#   Layer 3: UAP (universal-agent-protocol) -- installed HERE at runtime
+#
+# To upgrade UAP: change UAP_VERSION below. No image rebuild needed.
+# To upgrade toolchain: rebuild docker images (Dockerfile.ubuntu/python)
+# ============================================================================
+
+UAP_VERSION="1.0.0"
+
+# ============================================================================
+# Layer 2: Toolchain detection
+# ============================================================================
+if [ -f /installed-agent/.toolchain-version ]; then
+    TOOLCHAIN_VER=$(cat /installed-agent/.toolchain-version)
+    echo "=== Layer 2: Toolchain detected (${TOOLCHAIN_VER}) - skipping heavy install ==="
+
+    export NVM_DIR="$HOME/.nvm"
+    [ -s "$NVM_DIR/nvm.sh" ] && source "$NVM_DIR/nvm.sh"
+
+    echo "  Node $(node -v 2>/dev/null || echo 'MISSING')"
+    echo "  npm $(npm -v 2>/dev/null || echo 'MISSING')"
+    echo "  opencode $(opencode --version 2>/dev/null || echo 'MISSING')"
+else
+    echo "=== Layer 2: No toolchain found - full install ==="
+
+    # System dependencies
+    apt-get update
+    apt-get install -y curl git build-essential python3-pip jq netcat-openbsd xxd \
+        tesseract-ocr tesseract-ocr-eng poppler-utils imagemagick ffmpeg \
+        libsm6 libxext6 file tmux
+
+    # Python packages
+    pip install --break-system-packages numpy pillow pytesseract \
+        opencv-python-headless python-chess 2>/dev/null || true
+
+    # Node.js via nvm
+    curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.2/install.sh | bash
+    source "$HOME/.nvm/nvm.sh"
+    nvm install 22
+    echo "  Node $(node -v), npm $(npm -v)"
+
+    # opencode + AI SDK
+    for attempt in 1 2 3; do
+{% if version %}
+        npm i -g opencode-ai@{{ version }} && break
+{% else %}
+        npm i -g opencode-ai@latest && break
+{% endif %}
+        echo "  npm install failed (attempt ${attempt}/3), retrying..."
+        sleep 15
+    done
+    npm i -g @ai-sdk/openai-compatible 2>/dev/null || true
+
+    echo "  Layer 2 install complete"
+fi
+
+# ============================================================================
+# Layer 3: UAP (always runs -- fast, ~3-5s)
+# ============================================================================
+echo "=== Layer 3: Installing UAP v${UAP_VERSION} ==="
+
+export NVM_DIR="$HOME/.nvm"
+[ -s "$NVM_DIR/nvm.sh" ] && source "$NVM_DIR/nvm.sh"
+
+# Check if correct version already installed
+CURRENT_UAP=$(uap --version 2>/dev/null || echo "none")
+if [ "$CURRENT_UAP" = "$UAP_VERSION" ]; then
+    echo "  UAP v${UAP_VERSION} already installed - skipping"
+else
+    npm i -g universal-agent-protocol@${UAP_VERSION} 2>/dev/null || true
+    echo "  UAP v${UAP_VERSION} installed"
+fi
+
+# Run UAP setup for this project
+cd /app
+uap install opencode 2>/dev/null || true
+echo "  UAP setup complete"
+
+# ============================================================================
+# Config injection
+# ============================================================================
+
+# Write opencode.json with custom llama.cpp provider
+cat > /installed-agent/opencode.json << 'OPENCODE_EOF'
+{{ opencode_config }}
+OPENCODE_EOF
+
+{% if claude_md is defined and claude_md %}
+cat > /installed-agent/CLAUDE.md << 'CLAUDE_EOF'
+{{ claude_md }}
+CLAUDE_EOF
+echo "  CLAUDE.md written (UAP mode)"
+{% else %}
+echo "  No CLAUDE.md (baseline mode)"
+{% endif %}
+
+# ============================================================================
+# LLM endpoint connectivity check
+# ============================================================================
+LLM_ENDPOINT="{{ api_endpoint }}"
+echo "=== Checking LLM endpoint: ${LLM_ENDPOINT} ==="
+
+MAX_RETRIES=5
+RETRY_DELAY=10
+CONNECTED=false
+
+for attempt in $(seq 1 $MAX_RETRIES); do
+    if python3 -c "import urllib.request; urllib.request.urlopen('${LLM_ENDPOINT}/models', timeout=10)" 2>/dev/null; then
+        echo "  LLM reachable (attempt ${attempt}/${MAX_RETRIES})"
+        CONNECTED=true
+        break
+    fi
+    echo "  LLM not reachable (attempt ${attempt}/${MAX_RETRIES}), retrying in ${RETRY_DELAY}s..."
+    sleep $RETRY_DELAY
+done
+
+if [ "$CONNECTED" = false ]; then
+    echo "  WARNING: LLM endpoint not reachable after ${MAX_RETRIES} attempts"
+fi
+
+# ============================================================================
+# Summary
+# ============================================================================
+source "$HOME/.nvm/nvm.sh"
+echo "=== Install complete ==="
+echo "  Layer 2: toolchain $(cat /installed-agent/.toolchain-version 2>/dev/null || echo 'fresh')"
+echo "  Layer 3: UAP v$(uap --version 2>/dev/null || echo '?')"
+echo "  opencode $(opencode --version 2>/dev/null || echo '?')"
+echo "  API: {{ api_endpoint }}"