data fixes, partial custom indicator support

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

bin/create-user

@@ -0,0 +1,140 @@
+#!/usr/bin/env bash
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+# Colors
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+NC='\033[0m'
+
+usage() {
+    echo "Usage: $0 [dev|prod]"
+    echo ""
+    echo "Create a new user and assign a license."
+    echo ""
+    echo "  dev   - Create user in dev environment (minikube)"
+    echo "  prod  - Create user in prod environment"
+    exit 1
+}
+
+ENV="${1:-dev}"
+
+if [[ "$ENV" != "dev" && "$ENV" != "prod" ]]; then
+    echo -e "${RED}Error: Environment must be 'dev' or 'prod'${NC}"
+    usage
+fi
+
+if [[ "$ENV" == "prod" ]]; then
+    KUBECTL="kubectl --context=prod"
+    BASE_URL="https://dexorder.ai"
+    MCP_URL="https://dexorder.ai/mcp"
+else
+    KUBECTL="kubectl"
+    BASE_URL="http://dexorder.local"
+    MCP_URL="http://localhost:8080/mcp"
+fi
+
+# Get postgres pod
+PG_POD=$($KUBECTL get pods -l app=postgres -o jsonpath='{.items[0].metadata.name}' 2>/dev/null)
+if [ -z "$PG_POD" ]; then
+    echo -e "${RED}No postgres pod found${NC}"
+    exit 1
+fi
+
+# Prompt for credentials
+read -p "Email: " USER_EMAIL
+read -rs -p "Password (min 8 chars): " USER_PASSWORD
+echo ""
+if [[ ${#USER_PASSWORD} -lt 8 ]]; then
+    echo -e "${RED}✗ Password must be at least 8 characters${NC}"
+    exit 1
+fi
+read -p "Display name: " USER_NAME
+read -p "License type [free|pro|enterprise] (default: pro): " LICENSE_TYPE
+LICENSE_TYPE="${LICENSE_TYPE:-pro}"
+
+# Check if user already exists
+EXISTING_ID=$($KUBECTL exec "$PG_POD" -- psql -U postgres -d iceberg -t \
+    -c "SELECT id FROM \"user\" WHERE email = '$USER_EMAIL';" \
+    2>/dev/null | tr -d ' \n')
+
+if [ -n "$EXISTING_ID" ]; then
+    echo -e "${YELLOW}⚠️  User already exists in database ($USER_EMAIL)${NC}"
+    USER_ID="$EXISTING_ID"
+else
+    echo -e "${GREEN}→${NC} Registering user via API..."
+    PAYLOAD=$(jq -n \
+        --arg email "$USER_EMAIL" \
+        --arg password "$USER_PASSWORD" \
+        --arg name "$USER_NAME" \
+        '{email: $email, password: $password, name: $name}')
+
+    HTTP_CODE=$(curl -s -o /tmp/dexorder-create-user-response.json -w "%{http_code}" \
+        -X POST "$BASE_URL/api/auth/register" \
+        -H "Content-Type: application/json" \
+        -d "$PAYLOAD")
+
+    if [[ "$HTTP_CODE" == "200" || "$HTTP_CODE" == "201" ]]; then
+        echo -e "${GREEN}✓ User registered${NC}"
+    elif [[ "$HTTP_CODE" == "400" ]]; then
+        RESPONSE=$(cat /tmp/dexorder-create-user-response.json 2>/dev/null)
+        if echo "$RESPONSE" | grep -qi "already exist\|user already\|duplicate"; then
+            echo -e "${YELLOW}⚠️  User already exists, continuing...${NC}"
+        else
+            echo -e "${RED}✗ Registration failed (400):${NC}"
+            echo "$RESPONSE"
+            rm -f /tmp/dexorder-create-user-response.json
+            exit 1
+        fi
+    else
+        echo -e "${RED}✗ API returned HTTP $HTTP_CODE${NC}"
+        cat /tmp/dexorder-create-user-response.json 2>/dev/null || true
+        rm -f /tmp/dexorder-create-user-response.json
+        exit 1
+    fi
+    rm -f /tmp/dexorder-create-user-response.json
+
+    sleep 2
+
+    USER_ID=$($KUBECTL exec "$PG_POD" -- psql -U postgres -d iceberg -t \
+        -c "SELECT id FROM \"user\" WHERE email = '$USER_EMAIL';" \
+        2>/dev/null | tr -d ' \n')
+fi
+
+if [ -z "$USER_ID" ]; then
+    echo -e "${RED}User not found in database after registration. Is the gateway running?${NC}"
+    exit 1
+fi
+echo -e "${GREEN}User ID: $USER_ID${NC}"
+
+# Build license JSON
+case "$LICENSE_TYPE" in
+    enterprise)
+        LICENSE_JSON='{"licenseType":"enterprise","features":{"maxIndicators":200,"maxStrategies":100,"maxBacktestDays":1825,"realtimeData":true,"customExecutors":true,"apiAccess":true},"resourceLimits":{"maxConcurrentSessions":20,"maxMessagesPerDay":10000,"maxTokensPerMessage":32768,"rateLimitPerMinute":300},"k8sResources":{"memoryRequest":"1Gi","memoryLimit":"4Gi","cpuRequest":"500m","cpuLimit":"4000m","storage":"50Gi","tmpSizeLimit":"1Gi","enableIdleShutdown":true,"idleTimeoutMinutes":120},"preferredModel":{"provider":"anthropic","model":"claude-opus-4-6","temperature":0.7}}'
+        ;;
+    free)
+        LICENSE_JSON='{"licenseType":"free","features":{"maxIndicators":10,"maxStrategies":3,"maxBacktestDays":30,"realtimeData":false,"customExecutors":false,"apiAccess":false},"resourceLimits":{"maxConcurrentSessions":1,"maxMessagesPerDay":100,"maxTokensPerMessage":4096,"rateLimitPerMinute":20},"k8sResources":{"memoryRequest":"256Mi","memoryLimit":"512Mi","cpuRequest":"100m","cpuLimit":"500m","storage":"2Gi","tmpSizeLimit":"128Mi","enableIdleShutdown":true,"idleTimeoutMinutes":30},"preferredModel":{"provider":"anthropic","model":"claude-haiku-4-5-20251001","temperature":0.7}}'
+        ;;
+    pro|*)
+        LICENSE_JSON='{"licenseType":"pro","features":{"maxIndicators":50,"maxStrategies":20,"maxBacktestDays":365,"realtimeData":true,"customExecutors":true,"apiAccess":true},"resourceLimits":{"maxConcurrentSessions":5,"maxMessagesPerDay":1000,"maxTokensPerMessage":8192,"rateLimitPerMinute":60},"k8sResources":{"memoryRequest":"512Mi","memoryLimit":"2Gi","cpuRequest":"250m","cpuLimit":"2000m","storage":"10Gi","tmpSizeLimit":"256Mi","enableIdleShutdown":true,"idleTimeoutMinutes":60},"preferredModel":{"provider":"anthropic","model":"claude-sonnet-4-6","temperature":0.7}}'
+        ;;
+esac
+
+echo -e "${GREEN}→${NC} Assigning $LICENSE_TYPE license..."
+$KUBECTL exec "$PG_POD" -- psql -U postgres -d iceberg -c "
+    INSERT INTO user_licenses (user_id, email, license, mcp_server_url)
+    VALUES (
+        '$USER_ID',
+        '$USER_EMAIL',
+        '$LICENSE_JSON',
+        '$MCP_URL'
+    )
+    ON CONFLICT (user_id) DO UPDATE SET
+        license = EXCLUDED.license,
+        updated_at = NOW();
+" > /dev/null
+
+echo -e "${GREEN}✓ User ready: $USER_EMAIL ($LICENSE_TYPE)${NC}"

bin/dev

@@ -217,16 +217,18 @@ deploy_services() {
     # Update configs
     echo -e "${GREEN}→${NC} Updating configs..."
 
-    # Template the gateway-config.yaml with actual image tags
-    sed -i "s/SANDBOX_TAG_PLACEHOLDER/$SANDBOX_TAG/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
-    sed -i "s/SIDECAR_TAG_PLACEHOLDER/$SIDECAR_TAG/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+    # Template gateway-config.yaml with actual image tags (backup first for safe restore)
+    local _gw_bak
+    _gw_bak=$(mktemp)
+    cp "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml" "$_gw_bak"
+    sed -i "s|sandbox_image: dexorder/ai-sandbox:.*|sandbox_image: dexorder/ai-sandbox:$SANDBOX_TAG|g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+    sed -i "s|sidecar_image: dexorder/ai-lifecycle-sidecar:.*|sidecar_image: dexorder/ai-lifecycle-sidecar:$SIDECAR_TAG|g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
 
     "$SCRIPT_DIR/config-update" dev
 
     # Create a temporary kustomization overlay with image tags
     echo -e "${GREEN}→${NC} Setting image tags in kustomization..."
     cat >> kustomization.yaml <<EOF
-
 # Image tags (added by bin/dev)
 images:
   - name: dexorder/ai-relay
@@ -258,9 +260,9 @@ EOF
     # Clean up the appended image tags from kustomization.yaml
     sed -i '/# Image tags (added by bin\/dev)/,$d' kustomization.yaml
 
-    # Restore gateway-config.yaml placeholders
-    sed -i "s/$SANDBOX_TAG/SANDBOX_TAG_PLACEHOLDER/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
-    sed -i "s/$SIDECAR_TAG/SIDECAR_TAG_PLACEHOLDER/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+    # Restore gateway-config.yaml from backup
+    cp "$_gw_bak" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+    rm "$_gw_bak"
 
     echo -e "${GREEN}✓ Services deployed${NC}"
 
@@ -583,9 +585,11 @@ deploy_service() {
         gateway)
             image_name="dexorder/ai-gateway"
             image_tag="$GATEWAY_TAG"
-            # Also need to template gateway-config.yaml
-            sed -i "s/SANDBOX_TAG_PLACEHOLDER/$SANDBOX_TAG/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
-            sed -i "s/SIDECAR_TAG_PLACEHOLDER/$SIDECAR_TAG/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+            # Also need to template gateway-config.yaml (backup for safe restore)
+            _gw_bak_single=$(mktemp)
+            cp "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml" "$_gw_bak_single"
+            sed -i "s|sandbox_image: dexorder/ai-sandbox:.*|sandbox_image: dexorder/ai-sandbox:$SANDBOX_TAG|g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+            sed -i "s|sidecar_image: dexorder/ai-lifecycle-sidecar:.*|sidecar_image: dexorder/ai-lifecycle-sidecar:$SIDECAR_TAG|g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
             "$SCRIPT_DIR/config-update" dev
             ;;
         web)
@@ -604,7 +608,6 @@ deploy_service() {
 
     # Create a temporary kustomization overlay with ONLY this service's image tag
     cat >> kustomization.yaml <<EOF
-
 # Image tags (added by bin/dev)
 images:
   - name: $image_name
@@ -616,10 +619,10 @@ EOF
     # Clean up the appended image tags from kustomization.yaml
     sed -i '/# Image tags (added by bin\/dev)/,$d' kustomization.yaml
 
-    # Restore gateway-config.yaml placeholders if we modified it
+    # Restore gateway-config.yaml from backup if we modified it
     if [ "$service" == "gateway" ]; then
-        sed -i "s/$SANDBOX_TAG/SANDBOX_TAG_PLACEHOLDER/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
-        sed -i "s/$SIDECAR_TAG/SIDECAR_TAG_PLACEHOLDER/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+        cp "$_gw_bak_single" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+        rm "$_gw_bak_single"
     fi
 
     echo -e "${GREEN}✓ $service deployed${NC}"
@@ -685,6 +688,19 @@ case "$COMMAND" in
                 fi
             done
 
+            # Sandbox restart requires gateway to redeploy with the new sandbox image tag.
+            # If gateway wasn't explicitly listed, rebuild and deploy it automatically.
+            if [ "$sandbox_requested" == "1" ]; then
+                gateway_in_list=0
+                for svc in "${deploy_services_list[@]}"; do
+                    [ "$svc" == "gateway" ] && gateway_in_list=1 && break
+                done
+                if [ "$gateway_in_list" == "0" ]; then
+                    rebuild_images "gateway"
+                    deploy_services_list+=("gateway")
+                fi
+            fi
+
             # Deploy all non-sandbox services together in one kustomize apply
             if [ ${#deploy_services_list[@]} -gt 0 ]; then
                 if [ -f "$ROOT_DIR/.dev-image-tag" ]; then
@@ -693,18 +709,20 @@ case "$COMMAND" in
 
                 cd "$ROOT_DIR/deploy/k8s/dev"
 
-                # Template gateway-config if gateway is in the list
+                # Template gateway-config if gateway is in the list (backup for safe restore)
+                _ms_gw_bak=""
                 for svc in "${deploy_services_list[@]}"; do
                     if [ "$svc" == "gateway" ]; then
-                        sed -i "s/SANDBOX_TAG_PLACEHOLDER/$SANDBOX_TAG/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
-                        sed -i "s/SIDECAR_TAG_PLACEHOLDER/$SIDECAR_TAG/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+                        _ms_gw_bak=$(mktemp)
+                        cp "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml" "$_ms_gw_bak"
+                        sed -i "s|sandbox_image: dexorder/ai-sandbox:.*|sandbox_image: dexorder/ai-sandbox:$SANDBOX_TAG|g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+                        sed -i "s|sidecar_image: dexorder/ai-lifecycle-sidecar:.*|sidecar_image: dexorder/ai-lifecycle-sidecar:$SIDECAR_TAG|g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
                         "$SCRIPT_DIR/config-update" dev
                         break
                     fi
                 done
 
                 # Build the images stanza for all services at once
-                echo "" >> kustomization.yaml
                 echo "# Image tags (added by bin/dev)" >> kustomization.yaml
                 echo "images:" >> kustomization.yaml
                 for svc in "${deploy_services_list[@]}"; do
@@ -722,18 +740,29 @@ case "$COMMAND" in
 
                 sed -i '/# Image tags (added by bin\/dev)/,$d' kustomization.yaml
 
-                # Restore gateway-config placeholders if gateway was deployed
-                for svc in "${deploy_services_list[@]}"; do
-                    if [ "$svc" == "gateway" ]; then
-                        sed -i "s/$SANDBOX_TAG/SANDBOX_TAG_PLACEHOLDER/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
-                        sed -i "s/$SIDECAR_TAG/SIDECAR_TAG_PLACEHOLDER/g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
-                        break
-                    fi
-                done
+                # Restore gateway-config from backup if we modified it
+                if [ -n "$_ms_gw_bak" ]; then
+                    cp "$_ms_gw_bak" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+                    rm "$_ms_gw_bak"
+                fi
             fi
 
             # Handle sandbox separately
             if [ "$sandbox_requested" == "1" ]; then
+                if [ -f "$ROOT_DIR/.dev-image-tag" ]; then
+                    source "$ROOT_DIR/.dev-image-tag"
+                fi
+                echo -e "${GREEN}→${NC} Updating gateway config with new sandbox image tag ($SANDBOX_TAG)..."
+                cd "$ROOT_DIR/deploy/k8s/dev"
+                _sb_bak=$(mktemp)
+                cp "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml" "$_sb_bak"
+                sed -i "s|sandbox_image: dexorder/ai-sandbox:.*|sandbox_image: dexorder/ai-sandbox:$SANDBOX_TAG|g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+                sed -i "s|sidecar_image: dexorder/ai-lifecycle-sidecar:.*|sidecar_image: dexorder/ai-lifecycle-sidecar:$SIDECAR_TAG|g" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+                "$SCRIPT_DIR/config-update" dev
+                cp "$_sb_bak" "$ROOT_DIR/deploy/k8s/dev/configs/gateway-config.yaml"
+                rm "$_sb_bak"
+                echo -e "${GREEN}→${NC} Restarting gateway to pick up new sandbox image tag..."
+                kubectl rollout restart deployment/gateway
                 echo -e "${GREEN}→${NC} Deleting user container deployments in sandbox namespace..."
                 kubectl delete deployments --all -n sandbox 2>/dev/null || true
                 echo -e "${GREEN}✓ User containers will be recreated by gateway on next login${NC}"

bin/op-setup

@@ -92,18 +92,16 @@ create_item "MinIO" \
     "secret_key[password]=REPLACE_WITH_STRONG_SECRET_KEY"
 
 # --- Gateway ---
-# Used by: ai-secrets (anthropic_api_key), gateway-secrets (all LLM keys + jwt_secret)
+# Used by: gateway-secrets (LLM keys + jwt_secret + search keys)
 # jwt_secret: used to sign user sessions — generate with: openssl rand -base64 48
-# anthropic_api_key: Anthropic Console → API Keys (https://console.anthropic.com)
-# openai_api_key: OpenAI Platform → API Keys (https://platform.openai.com)
-# google_api_key: Google AI Studio (https://aistudio.google.com)
-# openrouter_api_key: OpenRouter (https://openrouter.ai)
+# deepinfra_api_key: Deep Infra Console → API Keys (https://deepinfra.com)
+# anthropic_api_key: Anthropic Console → API Keys (https://console.anthropic.com) — kept for potential future use
+# tavily_api_key: Tavily Console → API Keys (https://app.tavily.com)
 create_item "Gateway" \
-    "anthropic_api_key[password]=sk-ant-REPLACE_ME" \
+    "deepinfra_api_key[password]=REPLACE_ME" \
     "jwt_secret[password]=REPLACE_WITH_RANDOM_64_CHAR_SECRET" \
-    "openai_api_key[password]=sk-REPLACE_ME" \
-    "google_api_key[password]=REPLACE_ME" \
-    "openrouter_api_key[password]=sk-or-REPLACE_ME"
+    "anthropic_api_key[password]=sk-ant-REPLACE_ME" \
+    "tavily_api_key[password]=tvly-REPLACE_ME"
 
 # --- Telegram ---
 # Used by: gateway-secrets (optional Telegram bot integration)

deploy/k8s/dev/configs/gateway-config.yaml

@@ -24,40 +24,40 @@ data:
 
     # Default model (if user has no preference)
     defaults:
-      model_provider: anthropic
-      model: claude-sonnet-4-6
+      model_provider: deepinfra
+      model: zai-org/GLM-5
 
     # License tier model configuration
     license_models:
       # Free tier models
       free:
-        default: claude-haiku-4-5-20251001
-        cost_optimized: claude-haiku-4-5-20251001
-        complex: claude-haiku-4-5-20251001
+        default: zai-org/GLM-5
+        cost_optimized: zai-org/GLM-5
+        complex: zai-org/GLM-5
         allowed_models:
-          - claude-haiku-4-5-20251001
+          - zai-org/GLM-5
 
       # Pro tier models
       pro:
-        default: claude-sonnet-4-6
-        cost_optimized: claude-haiku-4-5-20251001
-        complex: claude-sonnet-4-6
+        default: zai-org/GLM-5
+        cost_optimized: zai-org/GLM-5
+        complex: zai-org/GLM-5
         blocked_models:
-          - claude-opus-4-6
+          - Qwen/Qwen3-235B-A22B-Instruct-2507
 
       # Enterprise tier models
       enterprise:
-        default: claude-sonnet-4-6
-        cost_optimized: claude-haiku-4-5-20251001
-        complex: claude-opus-4-6
+        default: zai-org/GLM-5
+        cost_optimized: zai-org/GLM-5
+        complex: Qwen/Qwen3-235B-A22B-Instruct-2507
 
     # Kubernetes configuration
     kubernetes:
       namespace: sandbox
       service_namespace: default
       in_cluster: true
-      sandbox_image: dexorder/ai-sandbox:SANDBOX_TAG_PLACEHOLDER
-      sidecar_image: dexorder/ai-lifecycle-sidecar:SIDECAR_TAG_PLACEHOLDER
+      sandbox_image: dexorder/ai-sandbox:dev20260408140409
+      sidecar_image: dexorder/ai-lifecycle-sidecar:dev20260407185216
       storage_class: standard
       image_pull_policy: Never  # For minikube dev - use local images
 

deploy/k8s/dev/kustomization.yaml

@@ -43,248 +43,3 @@ secretGenerator: []
 
 generatorOptions:
   disableNameSuffixHash: true
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-

deploy/k8s/dev/secrets/ai-secrets.yaml.example

@@ -4,4 +4,4 @@ metadata:
   name: ai-secrets
 type: Opaque
 stringData:
-  anthropic-api-key: "sk-ant-YOUR_KEY_HERE"
+  deepinfra-api-key: "YOUR_DEEPINFRA_KEY_HERE"

deploy/k8s/prod/configs/gateway-config.yaml

@@ -22,8 +22,8 @@ data:
 
     # Default model (if user has no preference)
     defaults:
-      model_provider: anthropic
-      model: claude-sonnet-4-6
+      model_provider: deepinfra
+      model: zai-org/GLM-5
 
     # Kubernetes configuration
     kubernetes:

deploy/k8s/prod/secrets/ai-secrets.tpl.yaml

@@ -5,4 +5,4 @@ metadata:
   namespace: ai
 type: Opaque
 stringData:
-  anthropic-api-key: "{{ op://AI Prod/Gateway/anthropic_api_key }}"
+  deepinfra-api-key: "{{ op://AI Prod/Gateway/deepinfra_api_key }}"

deploy/k8s/prod/secrets/gateway-secrets.tpl.yaml

@@ -14,10 +14,11 @@ stringData:
 
     # LLM Provider API Keys
     llm_providers:
-      anthropic_api_key: "{{ op://AI Prod/Gateway/anthropic_api_key }}"
-      openai_api_key: "{{ op://AI Prod/Gateway/openai_api_key }}"
-      google_api_key: "{{ op://AI Prod/Gateway/google_api_key }}"
-      openrouter_api_key: "{{ op://AI Prod/Gateway/openrouter_api_key }}"
+      deepinfra_api_key: "{{ op://AI Prod/Gateway/deepinfra_api_key }}"
+
+    # Search API Keys
+    search:
+      tavily_api_key: "{{ op://AI Prod/Gateway/tavily_api_key }}"
 
     # Telegram (optional)
     telegram:

flink/src/main/java/com/dexorder/flink/iceberg/SchemaInitializer.java

@@ -80,9 +80,8 @@ public class SchemaInitializer {
      */
     // Bump this when the schema changes. Tables with a different (or missing) version
     // will be dropped and recreated. Increment by 1 for each incompatible change.
-    // v2: open/high/low/close changed from required to optional to support null gap bars
-    // v3: timestamps changed from microseconds to nanoseconds; ticker format changed to BTC/USDT.BINANCE
-    private static final String OHLC_SCHEMA_VERSION = "3";
+    // v1: open/high/low/close required; ingestor forward-fills interior gaps with previous close
+    private static final String OHLC_SCHEMA_VERSION = "1";
     private static final String SCHEMA_VERSION_PROP = "app.schema.version";
 
     private void initializeOhlcTable() {
@@ -121,7 +120,7 @@ public class SchemaInitializer {
             LOG.info("Creating OHLC table: {}", tableId);
 
             // Define the OHLC schema.
-            // timestamp is stored as BIGINT (microseconds since epoch), not a TIMESTAMP type,
+            // timestamp is stored as BIGINT (nanoseconds since epoch), not a TIMESTAMP type,
             // so that GenericRowData.setField() accepts a plain Long value.
             Schema schema = new Schema(
                 // Primary key fields
@@ -129,11 +128,11 @@ public class SchemaInitializer {
                 required(2, "period_seconds", Types.IntegerType.get(), "OHLC period in seconds"),
                 required(3, "timestamp", Types.LongType.get(), "Candle timestamp in nanoseconds since epoch"),
 
-                // OHLC price data — optional to support gap bars (null = no trades that period)
-                optional(4, "open", Types.LongType.get(), "Opening price"),
-                optional(5, "high", Types.LongType.get(), "Highest price"),
-                optional(6, "low", Types.LongType.get(), "Lowest price"),
-                optional(7, "close", Types.LongType.get(), "Closing price"),
+                // OHLC price data — required; ingestor forward-fills interior gaps with previous close
+                required(4, "open",  Types.LongType.get(), "Opening price (forward-filled for interior market gaps)"),
+                required(5, "high",  Types.LongType.get(), "Highest price"),
+                required(6, "low",   Types.LongType.get(), "Lowest price"),
+                required(7, "close", Types.LongType.get(), "Closing price"),
 
                 // Volume data
                 optional(8, "volume", Types.LongType.get(), "Total volume"),

gateway/package.json

@@ -16,22 +16,28 @@
     "@fastify/jwt": "^9.0.1",
     "@fastify/websocket": "^11.0.1",
     "@kubernetes/client-node": "^1.0.0",
-    "@langchain/anthropic": "latest",
+    "@langchain/community": "^1.1.27",
     "@langchain/core": "latest",
     "@langchain/langgraph": "latest",
+    "@langchain/openai": "^1.4.2",
     "@modelcontextprotocol/sdk": "^1.0.4",
     "@qdrant/js-client-rest": "^1.17.0",
+    "@types/pdf-parse": "^1.1.5",
     "argon2": "^0.41.1",
     "better-auth": "^1.5.3",
+    "cheerio": "^1.2.0",
     "chrono-node": "^2.7.10",
+    "duck-duck-scrape": "^2.2.7",
     "duckdb": "^1.1.3",
     "fast-json-patch": "^3.1.1",
+    "fast-xml-parser": "^5.5.10",
     "fastify": "^5.2.0",
     "gray-matter": "^4.0.3",
     "ioredis": "^5.4.2",
     "js-yaml": "^4.1.0",
     "kysely": "^0.27.3",
     "ollama": "^0.5.10",
+    "pdf-parse": "^2.4.5",
     "pg": "^8.13.1",
     "pino": "^9.6.0",
     "pino-pretty": "^13.0.0",

gateway/src/channels/websocket-handler.ts

@@ -10,6 +10,7 @@ import type { SymbolIndexService } from '../services/symbol-index-service.js';
 import type { ContainerManager } from '../k8s/container-manager.js';
 import {
   WorkspaceManager,
+  ContainerSync,
   DEFAULT_STORES,
   type ChannelAdapter,
   type ChannelCapabilities,
@@ -120,15 +121,6 @@ export class WebSocketHandler {
 
     sendStatus(socket, 'initializing', 'Starting your workspace...');
 
-    // Create workspace manager for this session
-    const workspace = new WorkspaceManager({
-      userId: authContext.userId,
-      sessionId: authContext.sessionId,
-      stores: DEFAULT_STORES,
-      // containerSync will be added when MCP client is implemented
-      logger,
-    });
-
     // Create WebSocket channel adapter
     const wsAdapter: ChannelAdapter = {
       sendSnapshot: (msg: SnapshotMessage) => {
@@ -174,31 +166,47 @@ export class WebSocketHandler {
       }),
     };
 
-    // Declare harness outside try block so it's available in catch
+    // Declare harness and workspace outside try block so they're available in catch
     let harness: AgentHarness | undefined;
+    let workspace: WorkspaceManager | undefined;
 
     try {
-      // Initialize workspace first
-      await workspace.initialize();
-      workspace.setAdapter(wsAdapter);
-      this.workspaces.set(authContext.sessionId, workspace);
-
-      // Create agent harness via factory (storage deps injected by factory)
+      // Create and connect harness first so MCP client is available for ContainerSync
       harness = this.config.createHarness({
         userId: authContext.userId,
         sessionId: authContext.sessionId,
         license: authContext.license,
         mcpServerUrl: authContext.mcpServerUrl,
         logger,
-        workspaceManager: workspace,
         channelAdapter: wsAdapter,
         channelType: authContext.channelType,
         channelUserId: authContext.channelUserId,
       });
 
       await harness.initialize();
+
+      // Wire ContainerSync now that MCP client is connected, then initialize workspace
+      const containerSync = new ContainerSync(harness.getMcpClient(), logger);
+      workspace = new WorkspaceManager({
+        userId: authContext.userId,
+        sessionId: authContext.sessionId,
+        stores: DEFAULT_STORES,
+        containerSync,
+        logger,
+      });
+
+      await workspace.initialize();
+      workspace.setAdapter(wsAdapter);
+      harness.setWorkspaceManager(workspace);
+      this.workspaces.set(authContext.sessionId, workspace);
       this.harnesses.set(authContext.sessionId, harness);
 
+      // Push all store snapshots to the client now, before 'connected'.
+      // Empty seqs force full snapshots for every store, so the browser's
+      // message queue has the current workspace state (including persistent
+      // stores loaded from the container) before TradingView initializes.
+      await workspace.handleHello({});
+
       // Register session for event system
       // Container endpoint is derived from the MCP server URL (same container, different port)
       const containerEventEndpoint = this.getContainerEventEndpoint(authContext.mcpServerUrl);
@@ -287,15 +295,18 @@ export class WebSocketHandler {
           } else if (payload.type === 'hello') {
             // Workspace sync: hello message
             logger.debug({ seqs: payload.seqs }, 'Handling workspace hello');
-            await workspace.handleHello(payload.seqs || {});
+            await workspace!.handleHello(payload.seqs || {});
           } else if (payload.type === 'patch') {
             // Workspace sync: patch message
             logger.debug({ store: payload.store, seq: payload.seq }, 'Handling workspace patch');
-            await workspace.handlePatch(payload.store, payload.seq, payload.patch || []);
+            await workspace!.handlePatch(payload.store, payload.seq, payload.patch || []);
+          } else if (payload.type === 'agent_stop') {
+            logger.info('Agent stop requested');
+            harness?.interrupt();
           } else if (this.isDatafeedMessage(payload)) {
             // Historical data request - send to OHLC service
             logger.info({ type: payload.type }, 'Routing to datafeed handler');
-            await this.handleDatafeedMessage(socket, payload, logger);
+            await this.handleDatafeedMessage(socket, payload, logger, authContext);
           } else {
             logger.warn({ type: payload.type }, 'Unknown message type received');
           }
@@ -311,8 +322,9 @@ export class WebSocketHandler {
       });
 
       // Handle disconnection
-      socket.on('close', async () => {
-        logger.info({ sessionId: authContext.sessionId }, 'WebSocket disconnected');
+      socket.on('close', async (code: number, reason: Buffer) => {
+        clearInterval(pingInterval);
+        logger.info({ sessionId: authContext.sessionId, code, reason: reason?.toString() }, 'WebSocket disconnected');
 
         // Unregister from event system
         const removedSession = this.config.sessionRegistry.unregister(authContext.sessionId);
@@ -321,7 +333,7 @@ export class WebSocketHandler {
         }
 
         // Cleanup workspace
-        await workspace.shutdown();
+        await workspace!.shutdown();
         this.workspaces.delete(authContext.sessionId);
 
         // Cleanup harness
@@ -334,11 +346,21 @@ export class WebSocketHandler {
       socket.on('error', (error: any) => {
         logger.error({ error, sessionId: authContext.sessionId }, 'WebSocket error');
       });
+
+      // Ping every 30 seconds to keep the connection alive through CloudFlare proxy.
+      // CloudFlare drops idle WebSocket connections after ~100 seconds.
+      const pingInterval = setInterval(() => {
+        if (socket.readyState === 1) { // OPEN
+          socket.ping();
+        }
+      }, 30000);
     } catch (error) {
       logger.error({ error }, 'Failed to initialize session');
       socket.close(1011, 'Internal server error');
-      await workspace.shutdown();
-      this.workspaces.delete(authContext.sessionId);
+      if (workspace) {
+        await workspace.shutdown();
+        this.workspaces.delete(authContext.sessionId);
+      }
       if (harness) {
         await harness.cleanup();
       }
@@ -373,6 +395,7 @@ export class WebSocketHandler {
       'get_bars',
       'subscribe_bars',
       'unsubscribe_bars',
+      'evaluate_indicator',
     ];
     return datafeedTypes.includes(payload.type);
   }
@@ -383,7 +406,8 @@ export class WebSocketHandler {
   private async handleDatafeedMessage(
     socket: WebSocket,
     payload: any,
-    logger: any
+    logger: any,
+    authContext?: any
   ): Promise<void> {
     logger.info({ type: payload.type, payload }, 'handleDatafeedMessage called');
     const ohlcService = this.config.ohlcService;
@@ -491,6 +515,7 @@ export class WebSocketHandler {
             payload.to_time,
             payload.countback
           );
+          logger.info({ requestId, barCount: history.bars?.length ?? 0, noData: history.noData, socketState: socket.readyState }, 'Sending get_bars_response');
           socket.send(
             jsonStringifySafe({
               type: 'get_bars_response',
@@ -498,6 +523,7 @@ export class WebSocketHandler {
               history,
             })
           );
+          logger.info({ requestId }, 'get_bars_response sent');
           break;
         }
 
@@ -515,6 +541,69 @@ export class WebSocketHandler {
           );
           break;
 
+        case 'evaluate_indicator': {
+          // Direct MCP call — bypasses the agent/LLM for performance
+          const harness = this.harnesses.get(authContext.sessionId);
+          if (!harness) {
+            socket.send(JSON.stringify({
+              type: 'evaluate_indicator_result',
+              request_id: requestId,
+              error: 'Session not initialized',
+            }));
+            break;
+          }
+          try {
+            const mcpResult = await harness.callMcpTool('evaluate_indicator', {
+              symbol: payload.symbol,
+              from_time: payload.from_time,
+              to_time: payload.to_time,
+              period_seconds: payload.period_seconds,
+              pandas_ta_name: payload.pandas_ta_name,
+              parameters: payload.parameters ?? {},
+            }) as any;
+            // MCP returns { content: [{type: 'text', text: '...json...'}] }
+            // When the tool raises an exception, the MCP framework sets isError: true
+            // and puts the raw exception text in content[0].text (not JSON-wrapped).
+            const rawText = mcpResult?.content?.[0]?.text ?? mcpResult?.[0]?.text;
+            if (mcpResult?.isError || rawText == null) {
+              const errMsg = rawText ?? 'evaluate_indicator returned no content';
+              logger.error({ pandas_ta_name: payload.pandas_ta_name, rawText }, 'evaluate_indicator sandbox error');
+              socket.send(JSON.stringify({
+                type: 'evaluate_indicator_result',
+                request_id: requestId,
+                error: errMsg,
+              }));
+              break;
+            }
+            let data: any;
+            try {
+              data = JSON.parse(rawText);
+            } catch {
+              // Sandbox returned non-JSON (e.g. bare exception text)
+              logger.error({ pandas_ta_name: payload.pandas_ta_name, rawText }, 'evaluate_indicator returned non-JSON');
+              socket.send(JSON.stringify({
+                type: 'evaluate_indicator_result',
+                request_id: requestId,
+                error: rawText,
+              }));
+              break;
+            }
+            socket.send(JSON.stringify({
+              type: 'evaluate_indicator_result',
+              request_id: requestId,
+              ...data,
+            }));
+          } catch (err: any) {
+            logger.error({ err: err?.message, pandas_ta_name: payload.pandas_ta_name }, 'evaluate_indicator handler error');
+            socket.send(JSON.stringify({
+              type: 'evaluate_indicator_result',
+              request_id: requestId,
+              error: err?.message ?? String(err),
+            }));
+          }
+          break;
+        }
+
         default:
           logger.warn({ type: payload.type }, 'Unknown datafeed message type');
       }

gateway/src/clients/duckdb-client.ts

@@ -504,7 +504,11 @@ export class DuckDBClient {
   }
 
   /**
-   * Find missing OHLC data ranges
+   * Find missing OHLC data ranges by checking for absent timestamps.
+   *
+   * Any timestamp slot in [start_time, min(end_time, now)) that has no row in
+   * Iceberg is treated as missing and collected into contiguous ranges that the
+   * caller should request from the relay/ingestor.
    */
   async findMissingOHLCRanges(
     ticker: string,
@@ -517,32 +521,51 @@ export class DuckDBClient {
     try {
       const data = await this.queryOHLC(ticker, period_seconds, start_time, end_time);
 
-      if (data.length === 0) {
-        // All data is missing
-        return [[start_time, end_time]];
-      }
-
-      // Check if we have continuous data
-      // For now, simple check: if we have any data, assume complete
-      // TODO: Implement proper gap detection by checking for missing periods
       const periodNanos = BigInt(period_seconds) * 1_000_000_000n;
-      // end_time is exclusive, so expected count = (end - start) / period (no +1)
-      const expectedBars = Number((end_time - start_time) / periodNanos);
 
-      if (data.length < expectedBars * 0.95) { // Allow 5% tolerance
-        this.logger.debug({
-          ticker,
-          expected: expectedBars,
-          actual: data.length,
-        }, 'Incomplete OHLC data detected');
-        return [[start_time, end_time]]; // Request full range
+      // Cap at current time — future slots are not "missing", they don't exist yet.
+      const nowNanos = BigInt(Date.now()) * 1_000_000n;
+      const effectiveEnd = end_time < nowNanos ? end_time : nowNanos;
+
+      // Build a set of timestamps we already have (all rows are non-null now).
+      const present = new Set(data.map((row: any) => row.timestamp));
+
+      // Collect every expected slot that is absent.
+      const missing: bigint[] = [];
+      for (let t = start_time; t < effectiveEnd; t += periodNanos) {
+        if (!present.has(t)) {
+          missing.push(t);
+        }
       }
 
-      // Data appears complete
-      return [];
+      if (missing.length === 0) {
+        return [];
+      }
+
+      // Coalesce adjacent missing slots into contiguous [rangeStart, rangeEnd) intervals.
+      const ranges: Array<[bigint, bigint]> = [];
+      let rangeStart = missing[0];
+      let prev = missing[0];
+      for (let i = 1; i < missing.length; i++) {
+        if (missing[i] !== prev + periodNanos) {
+          ranges.push([rangeStart, prev + periodNanos]);
+          rangeStart = missing[i];
+        }
+        prev = missing[i];
+      }
+      ranges.push([rangeStart, prev + periodNanos]);
+
+      this.logger.debug({
+        ticker,
+        period_seconds,
+        missingSlots: missing.length,
+        ranges: ranges.length,
+      }, 'OHLC gap detection complete');
+
+      return ranges;
     } catch (error: any) {
       this.logger.error({ error: error.message }, 'Failed to find missing OHLC ranges');
-      // Return full range on error (safe default)
+      // Return full range on error (safe default — triggers a backfill)
       return [[start_time, end_time]];
     }
   }

gateway/src/harness/README.md

@@ -90,7 +90,7 @@ subagents/
 ```yaml
 tools:
   platform: ['symbol_lookup']  # Platform tools
-  mcp: ['category_*']          # MCP tool patterns
+  mcp: ['python_*']             # MCP tool patterns
 ```
 
 **Example:**

gateway/src/harness/agent-harness.ts

@@ -12,10 +12,14 @@ import type { ModelMiddleware } from '../llm/middleware.js';
 import type { WorkspaceManager } from '../workspace/workspace-manager.js';
 import type { ChannelAdapter, PathTriggerContext } from '../workspace/index.js';
 import type { ResearchSubagent } from './subagents/research/index.js';
+import type { IndicatorSubagent } from './subagents/indicator/index.js';
+import type { WebExploreSubagent } from './subagents/web-explore/index.js';
 import type { DynamicStructuredTool } from '@langchain/core/tools';
 import { getToolRegistry } from '../tools/tool-registry.js';
 import type { MCPToolInfo } from '../tools/mcp/mcp-tool-wrapper.js';
 import { createResearchAgentTool } from '../tools/platform/research-agent.tool.js';
+import { createIndicatorAgentTool } from '../tools/platform/indicator-agent.tool.js';
+import { createWebExploreAgentTool } from '../tools/platform/web-explore-agent.tool.js';
 import { createUserContext } from './memory/session-context.js';
 import { readFile } from 'fs/promises';
 import { join, dirname } from 'path';
@@ -52,6 +56,8 @@ export interface AgentHarnessConfig extends HarnessSessionConfig {
   conversationStore?: ConversationStore;
   historyLimit: number;
   researchSubagent?: ResearchSubagent;
+  indicatorSubagent?: IndicatorSubagent;
+  webExploreSubagent?: WebExploreSubagent;
 }
 
 /**
@@ -79,12 +85,17 @@ export class AgentHarness {
   private availableMCPTools: MCPToolInfo[] = [];
   private researchImageCapture: Array<{ data: string; mimeType: string }> = [];
   private conversationStore?: ConversationStore;
+  private indicatorSubagent?: IndicatorSubagent;
+  private webExploreSubagent?: WebExploreSubagent;
+  private abortController: AbortController | null = null;
 
   constructor(config: AgentHarnessConfig) {
     this.config = config;
     this.workspaceManager = config.workspaceManager;
     this.channelAdapter = config.channelAdapter;
     this.researchSubagent = config.researchSubagent;
+    this.indicatorSubagent = config.indicatorSubagent;
+    this.webExploreSubagent = config.webExploreSubagent;
 
     this.modelFactory = new LLMProviderFactory(config.providerConfig, config.logger);
     this.modelRouter = new ModelRouter(this.modelFactory, config.logger);
@@ -117,6 +128,10 @@ export class AgentHarness {
     this.channelAdapter = adapter;
   }
 
+  interrupt(): void {
+    this.abortController?.abort();
+  }
+
   /**
    * Initialize harness and connect to user's MCP server
    */
@@ -132,9 +147,15 @@ export class AgentHarness {
       // Discover available MCP tools from user's server
       await this.discoverMCPTools();
 
+      // Initialize web explore subagent first — research and indicator subagents inject it as a tool
+      await this.initializeWebExploreSubagent();
+
       // Initialize research subagent if not provided
       await this.initializeResearchSubagent();
 
+      // Initialize indicator subagent if not provided
+      await this.initializeIndicatorSubagent();
+
       this.config.logger.info('Agent harness initialized');
     } catch (error) {
       this.config.logger.error({ error }, 'Failed to initialize agent harness');
@@ -214,6 +235,24 @@ export class AgentHarness {
         (img) => this.researchImageCapture.push(img)
       );
 
+      // Inject web_explore tool if the web-explore subagent is ready
+      if (this.webExploreSubagent) {
+        const webExploreContext = {
+          userContext: createUserContext({
+            userId: this.config.userId,
+            sessionId: this.config.sessionId,
+            license: this.config.license,
+            channelType: this.config.channelType ?? ChannelType.WEBSOCKET,
+            channelUserId: this.config.channelUserId ?? this.config.userId,
+          }),
+        };
+        researchTools.push(createWebExploreAgentTool({
+          webExploreSubagent: this.webExploreSubagent,
+          context: webExploreContext,
+          logger: this.config.logger,
+        }));
+      }
+
       // Path resolution: use the compiled output path
       const researchSubagentPath = join(__dirname, 'subagents', 'research');
       this.config.logger.debug({ researchSubagentPath }, 'Using research subagent path');
@@ -243,6 +282,143 @@ export class AgentHarness {
     }
   }
 
+  /**
+   * Initialize indicator subagent
+   */
+  private async initializeIndicatorSubagent(): Promise<void> {
+    if (this.indicatorSubagent) {
+      this.config.logger.debug('Indicator subagent already provided');
+      return;
+    }
+
+    this.config.logger.debug('Creating indicator subagent for session');
+
+    try {
+      const { createIndicatorSubagent } = await import('./subagents/indicator/index.js');
+
+      const { model } = await this.modelRouter.route(
+        'indicator management',
+        this.config.license,
+        RoutingStrategy.COMPLEXITY,
+        this.config.userId
+      );
+
+      const toolRegistry = getToolRegistry();
+      const indicatorTools = await toolRegistry.getToolsForAgent(
+        'indicator',
+        this.mcpClient,
+        this.availableMCPTools,
+        this.workspaceManager,
+        undefined, // no image callback
+        (storeName, newState) => {
+          // After a workspace_patch succeeds in the container, update the gateway's
+          // WorkspaceManager so it pushes a WebSocket patch to the web client.
+          this.workspaceManager?.setState(storeName, newState).catch((err) =>
+            this.config.logger.error({ err, storeName }, 'Failed to sync workspace after indicator mutation')
+          );
+        }
+      );
+
+      // Inject web_explore tool if the web-explore subagent is ready
+      if (this.webExploreSubagent) {
+        const webExploreContext = {
+          userContext: createUserContext({
+            userId: this.config.userId,
+            sessionId: this.config.sessionId,
+            license: this.config.license,
+            channelType: this.config.channelType ?? ChannelType.WEBSOCKET,
+            channelUserId: this.config.channelUserId ?? this.config.userId,
+          }),
+        };
+        indicatorTools.push(createWebExploreAgentTool({
+          webExploreSubagent: this.webExploreSubagent,
+          context: webExploreContext,
+          logger: this.config.logger,
+        }));
+      }
+
+      const indicatorSubagentPath = join(__dirname, 'subagents', 'indicator');
+      this.config.logger.debug({ indicatorSubagentPath }, 'Using indicator subagent path');
+
+      this.indicatorSubagent = await createIndicatorSubagent(
+        model,
+        this.config.logger,
+        indicatorSubagentPath,
+        this.mcpClient,
+        indicatorTools
+      );
+
+      this.config.logger.info(
+        {
+          toolCount: indicatorTools.length,
+          toolNames: indicatorTools.map(t => t.name),
+        },
+        'Indicator subagent created successfully'
+      );
+    } catch (error) {
+      this.config.logger.error(
+        { error, errorMessage: (error as Error).message, stack: (error as Error).stack },
+        'Failed to create indicator subagent'
+      );
+      // Don't throw — indicator subagent is optional
+    }
+  }
+
+  /**
+   * Initialize web explore subagent
+   */
+  private async initializeWebExploreSubagent(): Promise<void> {
+    if (this.webExploreSubagent) {
+      this.config.logger.debug('Web explore subagent already provided');
+      return;
+    }
+
+    this.config.logger.debug('Creating web explore subagent for session');
+
+    try {
+      const { createWebExploreSubagent } = await import('./subagents/web-explore/index.js');
+
+      const { model } = await this.modelRouter.route(
+        'web research and summarization',
+        this.config.license,
+        RoutingStrategy.COMPLEXITY,
+        this.config.userId
+      );
+
+      const toolRegistry = getToolRegistry();
+      const webExploreTools = await toolRegistry.getToolsForAgent(
+        'web-explore',
+        undefined, // no MCP client needed
+        undefined,
+        undefined
+      );
+
+      const webExploreSubagentPath = join(__dirname, 'subagents', 'web-explore');
+      this.config.logger.debug({ webExploreSubagentPath }, 'Using web explore subagent path');
+
+      this.webExploreSubagent = await createWebExploreSubagent(
+        model,
+        this.config.logger,
+        webExploreSubagentPath,
+        webExploreTools
+      );
+
+      this.config.logger.info(
+        {
+          toolCount: webExploreTools.length,
+          toolNames: webExploreTools.map(t => t.name),
+        },
+        'Web explore subagent created successfully'
+      );
+    } catch (error) {
+      this.config.logger.error(
+        { error, errorMessage: (error as Error).message, stack: (error as Error).stack },
+        'Failed to create web explore subagent'
+      );
+      // Don't throw — web explore subagent is optional
+    }
+  }
+
   /**
    * Execute model with tool calling loop
    * Handles multi-turn tool calls until the model produces a final text response
@@ -251,7 +427,8 @@ export class AgentHarness {
     model: any,
     messages: BaseMessage[],
     tools: DynamicStructuredTool[],
-    maxIterations: number = 2
+    maxIterations: number = 2,
+    signal?: AbortSignal
   ): Promise<string> {
     this.config.logger.info(
       { toolCount: tools.length, maxIterations },
@@ -262,6 +439,7 @@ export class AgentHarness {
     let iterations = 0;
 
     while (iterations < maxIterations) {
+      if (signal?.aborted) break;
       iterations++;
       this.config.logger.info(
         {
@@ -275,7 +453,7 @@ export class AgentHarness {
       this.config.logger.debug('Streaming model response...');
       let response: any = null;
       try {
-        const stream = await model.stream(messagesCopy);
+        const stream = await model.stream(messagesCopy, { signal });
         for await (const chunk of stream) {
           if (typeof chunk.content === 'string' && chunk.content.length > 0) {
             this.channelAdapter?.sendChunk(chunk.content);
@@ -415,6 +593,29 @@ export class AgentHarness {
     return 'I apologize, but I encountered an issue processing your request. Please try rephrasing your question.';
   }
 
+  /**
+   * Call a tool on the user's MCP server directly (bypasses the agent/LLM).
+   * Used by channel handlers for direct data requests (e.g. evaluate_indicator).
+   */
+  async callMcpTool(name: string, args: Record<string, unknown>): Promise<unknown> {
+    return this.mcpClient.callTool(name, args);
+  }
+
+  /**
+   * Expose MCP client so channel handlers can wire ContainerSync after harness init.
+   */
+  getMcpClient(): MCPClientConnector {
+    return this.mcpClient;
+  }
+
+  /**
+   * Set workspace manager after construction (used when ContainerSync requires MCP to be connected first).
+   */
+  setWorkspaceManager(workspace: WorkspaceManager): void {
+    this.workspaceManager = workspace;
+    this.registerWorkspaceTriggers();
+  }
+
   /**
    * Handle incoming message from user
    */
@@ -480,18 +681,19 @@ export class AgentHarness {
         this.workspaceManager  // Pass session workspace manager
       );
 
+      // Build shared subagent context
+      const subagentContext = {
+        userContext: createUserContext({
+          userId: this.config.userId,
+          sessionId: this.config.sessionId,
+          license: this.config.license,
+          channelType: this.config.channelType ?? ChannelType.WEBSOCKET,
+          channelUserId: this.config.channelUserId ?? this.config.userId,
+        }),
+      };
+
       // Add research subagent as a tool if available
       if (this.researchSubagent) {
-        const subagentContext = {
-          userContext: createUserContext({
-            userId: this.config.userId,
-            sessionId: this.config.sessionId,
-            license: this.config.license,
-            channelType: this.config.channelType ?? ChannelType.WEBSOCKET,
-            channelUserId: this.config.channelUserId ?? this.config.userId,
-          }),
-        };
-
         tools.push(createResearchAgentTool({
           researchSubagent: this.researchSubagent,
           context: subagentContext,
@@ -499,6 +701,24 @@ export class AgentHarness {
         }));
       }
 
+      // Add indicator subagent as a tool if available
+      if (this.indicatorSubagent) {
+        tools.push(createIndicatorAgentTool({
+          indicatorSubagent: this.indicatorSubagent,
+          context: subagentContext,
+          logger: this.config.logger,
+        }));
+      }
+
+      // Add web explore subagent as a tool if available
+      if (this.webExploreSubagent) {
+        tools.push(createWebExploreAgentTool({
+          webExploreSubagent: this.webExploreSubagent,
+          context: subagentContext,
+          logger: this.config.logger,
+        }));
+      }
+
       this.config.logger.info(
         {
           toolCount: tools.length,
@@ -524,7 +744,9 @@ export class AgentHarness {
 
       // 8. Call LLM with tool calling loop
       this.config.logger.info('Invoking LLM with tool support');
-      const assistantMessage = await this.executeWithToolCalling(modelWithTools, processedMessages, tools, 10);
+      this.abortController = new AbortController();
+      const assistantMessage = await this.executeWithToolCalling(modelWithTools, processedMessages, tools, 10, this.abortController.signal);
+      this.abortController = null;
 
       this.config.logger.info(
         { responseLength: assistantMessage.length },
@@ -587,13 +809,17 @@ export class AgentHarness {
   private getToolLabel(toolName: string): string {
     const labels: Record<string, string> = {
       research: 'Researching...',
+      indicator: 'Adjusting indicators...',
       get_chart_data: 'Fetching chart data...',
       symbol_lookup: 'Searching symbol...',
-      category_list: 'Seeing what we have...',
-      category_edit: 'Coding...',
-      category_write: 'Coding...',
-      category_read: 'Inspecting...',
+      python_list: 'Seeing what we have...',
+      python_edit: 'Coding...',
+      python_write: 'Coding...',
+      python_read: 'Inspecting...',
       execute_research: 'Running script...',
+      backtest_strategy: 'Running backtest...',
+      list_active_strategies: 'Checking active strategies...',
+      web_explore: 'Searching the web...',
     };
     return labels[toolName] ?? `Running ${toolName}...`;
   }

gateway/src/harness/mcp-client.ts

@@ -1,5 +1,5 @@
 import { Client } from '@modelcontextprotocol/sdk/client/index.js';
-import { SSEClientTransport } from '@modelcontextprotocol/sdk/client/sse.js';
+import { StreamableHTTPClientTransport } from '@modelcontextprotocol/sdk/client/streamableHttp.js';
 
 import type { FastifyBaseLogger } from 'fastify';
 
@@ -12,11 +12,12 @@ export interface MCPClientConfig {
 
 /**
  * MCP client connector for user's container
- * Manages connection to user-specific MCP server via SSE transport
+ * Manages connection to user-specific MCP server via Streamable HTTP transport
  */
 export class MCPClientConnector {
   private client: Client | null = null;
   private connected = false;
+  private reconnectPromise: Promise<void> | null = null;
   private config: MCPClientConfig;
 
   constructor(config: MCPClientConfig) {
@@ -24,17 +25,42 @@ export class MCPClientConnector {
   }
 
   /**
-   * Connect to user's MCP server via SSE transport
+   * Connect to user's MCP server via Streamable HTTP transport.
+   * Safe to call when already connecting (concurrent callers wait for the same attempt).
    */
   async connect(): Promise<void> {
     if (this.connected) {
       return;
     }
 
+    // If a reconnect is already in progress, wait for it rather than racing
+    if (this.reconnectPromise) {
+      return this.reconnectPromise;
+    }
+
+    this.reconnectPromise = this._doConnect();
+    try {
+      await this.reconnectPromise;
+    } finally {
+      this.reconnectPromise = null;
+    }
+  }
+
+  private async _doConnect(): Promise<void> {
+    // Close stale client if this is a reconnect attempt
+    if (this.client) {
+      try {
+        await this.client.close();
+      } catch {
+        // Ignore errors closing a stale/broken client
+      }
+      this.client = null;
+    }
+
     try {
       this.config.logger.info(
         { userId: this.config.userId, url: this.config.mcpServerUrl },
-        'Connecting to user MCP server via SSE'
+        'Connecting to user MCP server'
       );
 
       this.client = new Client(
@@ -49,15 +75,32 @@ export class MCPClientConnector {
         }
       );
 
-      // Create SSE transport for HTTP connection to user container
-      const transport = new SSEClientTransport(
-        new URL(`${this.config.mcpServerUrl}/sse`)
+      // Streamable HTTP: single /mcp endpoint, session tracked via mcp-session-id header
+      const transport = new StreamableHTTPClientTransport(
+        new URL(`${this.config.mcpServerUrl}/mcp`)
       );
 
       await this.client.connect(transport);
 
+      // Hook client.onerror to detect transport failures (e.g. sandbox restart returning
+      // 404 "session not found").  When fired, mark disconnected so the next callTool /
+      // listTools call triggers a full reconnect + initialize handshake.
+      const connectedClient = this.client;
+      const origOnError = this.client.onerror;
+      this.client.onerror = (error) => {
+        origOnError?.(error);
+        // Only act on the currently-active client (ignore stale closures after reconnect)
+        if (this.client === connectedClient && this.connected) {
+          this.config.logger.warn(
+            { error },
+            'MCP transport error — marking disconnected for lazy reconnect'
+          );
+          this.connected = false;
+        }
+      };
+
       this.connected = true;
-      this.config.logger.info('Connected to user MCP server via SSE');
+      this.config.logger.info('Connected to user MCP server');
     } catch (error) {
       this.config.logger.error(
         { error, userId: this.config.userId },
@@ -67,18 +110,31 @@ export class MCPClientConnector {
     }
   }
 
+  /**
+   * Ensure the client is connected, reconnecting if necessary.
+   * Used as a preamble for every public method so a sandbox restart is
+   * recovered transparently on the next tool call.
+   */
+  private async ensureConnected(): Promise<void> {
+    if (!this.client || !this.connected) {
+      this.config.logger.info(
+        { userId: this.config.userId },
+        'MCP not connected, attempting reconnect'
+      );
+      await this.connect();
+    }
+  }
+
   /**
    * Call a tool on the user's MCP server
    */
   async callTool(name: string, args: Record<string, unknown>): Promise<unknown> {
-    if (!this.client || !this.connected) {
-      throw new Error('MCP client not connected');
-    }
+    await this.ensureConnected();
 
     try {
       this.config.logger.debug({ tool: name, args }, 'Calling MCP tool');
 
-      const result = await this.client.callTool({ name, arguments: args });
+      const result = await this.client!.callTool({ name, arguments: args });
       return result;
     } catch (error) {
       this.config.logger.error({ error, tool: name }, 'MCP tool call failed');
@@ -91,13 +147,11 @@ export class MCPClientConnector {
    * Returns all available tools from the MCP server
    */
   async listTools(): Promise<Array<{ name: string; description?: string; inputSchema?: any }>> {
-    if (!this.client || !this.connected) {
-      throw new Error('MCP client not connected');
-    }
+    await this.ensureConnected();
 
     try {
       this.config.logger.debug('Requesting tool list from MCP server');
-      const response = await this.client.listTools();
+      const response = await this.client!.listTools();
 
       this.config.logger.debug(
         {
@@ -146,12 +200,10 @@ export class MCPClientConnector {
    * Returns all available resources from the MCP server
    */
   async listResources(): Promise<Array<{ uri: string; name: string; description?: string; mimeType?: string }>> {
-    if (!this.client || !this.connected) {
-      throw new Error('MCP client not connected');
-    }
+    await this.ensureConnected();
 
     try {
-      const response = await this.client.listResources();
+      const response = await this.client!.listResources();
 
       // Return all resources - agent-to-resource binding is handled by the tool registry
       const resources = response.resources.map((resource: any) => ({
@@ -177,14 +229,12 @@ export class MCPClientConnector {
    * Read a resource from user's MCP server
    */
   async readResource(uri: string): Promise<{ uri: string; mimeType?: string; text?: string; blob?: string }> {
-    if (!this.client || !this.connected) {
-      throw new Error('MCP client not connected');
-    }
+    await this.ensureConnected();
 
     try {
       this.config.logger.debug({ uri }, 'Reading MCP resource');
 
-      const response = await this.client.readResource({ uri });
+      const response = await this.client!.readResource({ uri });
 
       // Extract the first content item (MCP returns array of contents)
       const content = response.contents[0];
@@ -206,15 +256,19 @@ export class MCPClientConnector {
    * Disconnect from MCP server
    */
   async disconnect(): Promise<void> {
-    if (this.client && this.connected) {
+    if (this.client) {
       try {
         await this.client.close();
-        this.connected = false;
-        this.config.logger.info('Disconnected from user MCP server');
+        if (this.connected) {
+          this.config.logger.info('Disconnected from user MCP server');
+        }
       } catch (error) {
         this.config.logger.error({ error }, 'Error disconnecting from MCP server');
       }
     }
+    this.connected = false;
+    this.client = null;
+    this.reconnectPromise = null;
   }
 
   isConnected(): boolean {

gateway/src/harness/prompts/system-prompt.md

@@ -3,6 +3,8 @@
 You are a helpful AI assistant for Dexorder, an AI-first trading platform.
 You help users research markets, develop indicators and strategies, and analyze trading data.
 
+Your text responses should be markdown, using emojiis, color, and formatting to create a visually appealing response.
+
 **User License:** {{licenseType}}
 
 **Available Features:**
@@ -10,21 +12,71 @@ You help users research markets, develop indicators and strategies, and analyze
 
 ---
 
+# Platform Capabilities
+
+Dexorder trading platform provides OHLC data at a 1-minute resolution and supports strategies that read one or more OHLC feeds at a 1-minute resolution or coarser. It also offers a wide range of built-in indicators and allows users to create custom indicators for advanced analysis.
+
+Dexorder does not support tick-by-tick trading or high-frequency strategies.
+Dexorder does not support long-running computations like paramater optimizations or training machine learning models.
+Dexorder does not support portfolio optimization or trading strategies that require a large number of symbols.
+
+If the user asks for a capability not provided by Dexorder, decline and offer alternatives.
+
 # Important Instructions
 
+## Investment Advice
+**NEVER** recommend any specific ticker, trade, or strategy. You may suggest mechanical adjustments or improvements to strategies, but you must never recommend that the user adopt a specific trade or position.
+
 ## Task Delegation
-- For ANY research questions, deep analysis, statistical analysis, charting requests, plotting, ML tasks, or market data queries that require computation, you MUST use the 'research' tool
-- The research tool creates and runs Python scripts that generate charts and perform analysis
-- Use 'research' for anything involving: plotting, statistics, calculations, correlations, patterns, volume analysis, technical indicators, or any non-trivial data processing
+- For ANY research questions, deep analysis, statistical analysis, charting requests, or market data queries that require computation, you MUST use the 'research' tool
+- For ANYTHING related to indicators on the chart — reading, adding, removing, modifying, or creating custom indicators — you MUST use the 'indicator' tool
+- For ANY backtesting request — running a strategy against historical data — you MUST use the 'backtest_strategy' tool directly; NEVER use the research tool for backtesting
 - NEVER write Python code directly in your responses to the user
-- NEVER show code to the user - delegate to the research tool instead
-- NEVER attempt to do analysis yourself - let the research subagent handle it
+- NEVER show code to the user — delegate to the research or indicator tool instead
+- NEVER attempt to do analysis yourself — let the subagents handle it
 
 ## Available Tools
-You have access to the following tools:
+
+### indicator
+**Use this tool for all indicator-related requests.**
+
+The indicator subagent manages the chart's indicators: it reads the current indicator set, adds or removes indicators, modifies parameters, and can create custom indicator scripts.
+
+**ALWAYS use indicator for:**
+- "What indicators do I have on the chart?" → read and describe current indicators
+- "Show RSI" / "Add Bollinger Bands" → add indicators to chart
+- "Change MACD fast period to 8" → modify indicator parameters
+- "Remove all moving averages" → remove indicators
+- "Create a custom volume-weighted RSI" → write custom indicator
+- Any question about what an indicator means or how it's configured
+- Recommending indicators for a given strategy
+
+**Custom indicators vs. ad-hoc research scripts:**
+When a user asks for a calculation (e.g. "volume-weighted RSI", "adaptive ATR", "sector relative strength"), prefer creating a **custom indicator** via this tool over writing a one-off pandas/Python script in the research tool. Custom indicators are better because:
+1. **Reusable** — saved permanently and can be applied to any symbol at any time
+2. **First-class UI** — appear in the chart's Indicator picker alongside built-in indicators
+3. **Live chart display** — their values are plotted directly on the chart as the user browses
+4. **Watchlist & trigger support** — can be used to filter symbols (watchlists) and fire alerts/triggers (coming soon)
+
+Use the research tool for exploratory or one-off analysis. Use the indicator tool whenever the user wants to *track* or *reuse* a computed value.
+
+**NEVER modify workspace indicators yourself** — always delegate to the indicator tool.
+
+### web_explore
+**Use this tool to search the web or academic databases.**
+
+The web-explore subagent searches the web (or arXiv for academic topics), fetches relevant pages, and returns a markdown summary with cited sources.
+
+**ALWAYS use web_explore for:**
+- Questions about current events, news, or real-time information
+- Documentation, tutorials, or how-to guides
+- Academic papers, research findings, or scientific topics
+- Any topic that requires up-to-date external sources
+
+**NOT for market data or computation** — use the research tool for analysis, and get_chart_data for OHLC values.
 
 ### research
-**This is your PRIMARY tool for any analysis, computation, charting, or plotting tasks.**
+**This is your PRIMARY tool for data analysis, computation, and charting.**
 
 Creates and runs Python research scripts via a specialized research subagent.
 The subagent autonomously writes code, executes it, handles errors, and generates charts.
@@ -32,7 +84,6 @@ The subagent autonomously writes code, executes it, handles errors, and generate
 **ALWAYS use research for:**
 - Any plotting, charting, or visualization requests
 - Price action analysis and correlations
-- Technical indicators and overlays
 - Statistical analysis of market data
 - Volume analysis and patterns
 - Machine learning or predictive modeling
@@ -41,16 +92,11 @@ The subagent autonomously writes code, executes it, handles errors, and generate
 - Custom calculations or transformations
 - Deep analysis requiring Python libraries (pandas, numpy, scipy, matplotlib, etc.)
 
+**NOT for indicator management** — use the indicator tool for that.
+
 **NEVER attempt to do analysis yourself in the chat.**
 Let the research subagent write and execute the Python code.
 
-**Examples of when to use research:**
-- "Plot BTC with volume overlay" → use research
-- "Calculate correlation between ETH and BTC" → use research
-- "Show me RSI divergences" → use research
-- "Analyze Monday price patterns" → use research
-- "Does volume predict price movement?" → use research
-
 Parameters:
 - instruction: Natural language description of the analysis to perform (be specific!)
 - name: A unique name for the research script (e.g., "BTC Weekly Analysis")
@@ -59,10 +105,37 @@ Example usage:
 - User: "Does Friday price action correlate with Monday?"
 - You: Call research tool with instruction="Analyze correlation between Friday and Monday price action during NY trading hours (9:30-4:00 ET)", name="Friday-Monday Correlation"
 
-### category_list
-List existing research scripts (category="research").
+### backtest_strategy
+**ALWAYS use this tool — and ONLY this tool — for any backtesting request.**
+
+Runs a saved trading strategy against historical OHLC data using the Nautilus Trader backtesting engine.
+Returns structured performance metrics and an equity curve. Any charts generated are automatically sent to the user.
+
+**ALWAYS use backtest_strategy for:**
+- "Backtest my RSI strategy over the last year"
+- "How did this strategy perform on BTC?"
+- "Run a backtest from January to June"
+- Any request to test or evaluate a strategy on historical data
+
+**NEVER use research for backtesting** — the research tool cannot run strategies through the backtesting engine.
+
+After the tool returns, summarize the results clearly: total return, Sharpe ratio, max drawdown, win rate, and trade count. Present the equity curve description in plain language.
+
+Parameters:
+- strategy_name: Display name of the saved strategy (use python_list with category="strategy" to check existing strategies)
+- feeds: Array of `{symbol, period_seconds}` feed objects (e.g. `[{"symbol": "BTC/USDT.BINANCE", "period_seconds": 3600}]`)
+- from_time / to_time: Date strings ("2024-01-01", "90 days ago", "now") or Unix timestamps
+- initial_capital: Starting balance in quote currency (default 10,000)
+
+### list_active_strategies
+Lists all currently active (live or paper) strategies and their status.
+Use this when the user asks what strategies are running.
+
+### python_list
+List existing scripts in a category ("strategy", "indicator", or "research").
 Use this before calling the research tool to check whether a relevant script already exists.
 If one does, pass its exact name to the research tool so the subagent updates it rather than creating a new one.
+Also use before calling backtest_strategy to confirm the strategy name.
 
 ### symbol-lookup
 Look up trading symbols and get metadata.
@@ -102,3 +175,4 @@ You also have access to workspace persistence tools via MCP:
 - **workspace_patch(store_name, patch)**: Apply JSON patch to a workspace store
 
 These are useful for persisting user preferences, analysis results, and custom data across sessions.
+For the `indicators` store specifically, always use the indicator tool rather than calling workspace tools directly.

gateway/src/harness/subagents/base-subagent.ts

@@ -44,12 +44,9 @@ export interface SubagentContext {
  *
  * Structure:
  * subagents/
- *   code-reviewer/
+ *   research/
  *     config.yaml
  *     system-prompt.md
- *     memory/
- *       review-guidelines.md
- *       common-patterns.md
  *     index.ts
  */
 export abstract class BaseSubagent {

gateway/src/harness/subagents/code-reviewer/config.yaml

@@ -1,26 +0,0 @@
-# Code Reviewer Subagent Configuration
-
-name: code-reviewer
-description: Reviews trading strategy code for bugs, performance issues, and best practices
-
-# Model configuration (optional override)
-model: claude-sonnet-4-6
-temperature: 0.3
-maxTokens: 4096
-
-# Memory files to load from memory/ directory
-memoryFiles:
-  - review-guidelines.md
-  - common-patterns.md
-  - best-practices.md
-
-# System prompt file
-systemPromptFile: system-prompt.md
-
-# Capabilities this subagent provides
-capabilities:
-  - static_analysis
-  - performance_review
-  - security_audit
-  - code_quality
-  - best_practices

gateway/src/harness/subagents/code-reviewer/index.ts

@@ -1,93 +0,0 @@
-import { BaseSubagent, type SubagentConfig, type SubagentContext } from '../base-subagent.js';
-import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
-import type { FastifyBaseLogger } from 'fastify';
-
-/**
- * Code Reviewer Subagent
- *
- * Specialized agent for reviewing trading strategy code.
- * Reviews for:
- * - Logic errors and bugs
- * - Performance issues
- * - Security vulnerabilities
- * - Trading best practices
- * - Code quality
- *
- * Loads knowledge from multi-file memory:
- * - review-guidelines.md: What to check for
- * - common-patterns.md: Good and bad examples
- * - best-practices.md: Industry standards
- */
-export class CodeReviewerSubagent extends BaseSubagent {
-  constructor(config: SubagentConfig, model: BaseChatModel, logger: FastifyBaseLogger, mcpClient?: any, tools?: any[]) {
-    super(config, model, logger, mcpClient, tools);
-  }
-
-  /**
-   * Review code and provide structured feedback
-   */
-  async execute(context: SubagentContext, code: string): Promise<string> {
-    this.logger.info(
-      {
-        subagent: this.getName(),
-        userId: context.userContext.userId,
-        codeLength: code.length,
-      },
-      'Reviewing code'
-    );
-
-    const messages = this.buildMessages(context, `Review the following trading strategy code:\n\n\`\`\`typescript\n${code}\n\`\`\``);
-
-    const response = await this.model.invoke(messages);
-
-    return response.content as string;
-  }
-
-  /**
-   * Stream code review
-   */
-  async *stream(context: SubagentContext, code: string): AsyncGenerator<string> {
-    this.logger.info(
-      {
-        subagent: this.getName(),
-        userId: context.userContext.userId,
-        codeLength: code.length,
-      },
-      'Streaming code review'
-    );
-
-    const messages = this.buildMessages(context, `Review the following trading strategy code:\n\n\`\`\`typescript\n${code}\n\`\`\``);
-
-    const stream = await this.model.stream(messages);
-
-    for await (const chunk of stream) {
-      yield chunk.content as string;
-    }
-  }
-}
-
-/**
- * Factory function to create and initialize CodeReviewerSubagent
- */
-export async function createCodeReviewerSubagent(
-  model: BaseChatModel,
-  logger: FastifyBaseLogger,
-  basePath: string,
-  mcpClient?: any,
-  tools?: any[]
-): Promise<CodeReviewerSubagent> {
-  const { readFile } = await import('fs/promises');
-  const { join } = await import('path');
-  const yaml = await import('js-yaml');
-
-  // Load config
-  const configPath = join(basePath, 'config.yaml');
-  const configContent = await readFile(configPath, 'utf-8');
-  const config = yaml.load(configContent) as SubagentConfig;
-
-  // Create and initialize subagent
-  const subagent = new CodeReviewerSubagent(config, model, logger, mcpClient, tools);
-  await subagent.initialize(basePath);
-
-  return subagent;
-}

gateway/src/harness/subagents/code-reviewer/memory/best-practices.md

@@ -1,227 +0,0 @@
-# Trading Strategy Best Practices
-
-## Code Organization
-
-### Separation of Concerns
-```typescript
-// Good: Clear separation
-class Strategy {
-  async analyze(data: MarketData): Promise<Signal> { }
-}
-
-class RiskManager {
-  validateSignal(signal: Signal): boolean { }
-}
-
-class ExecutionEngine {
-  async execute(signal: Signal): Promise<Order> { }
-}
-
-// Bad: Everything in one function
-async function trade() {
-  // Analysis, risk, execution all mixed
-}
-```
-
-### Configuration Management
-```typescript
-// Good: External configuration
-interface StrategyConfig {
-  stopLossPercent: number;
-  takeProfitPercent: number;
-  maxPositionSize: number;
-  riskPerTrade: number;
-}
-
-const config = loadConfig('strategy.yaml');
-
-// Bad: Hardcoded values scattered throughout
-const stopLoss = price * 0.95; // What if you want to change this?
-```
-
-## Testing Considerations
-
-### Testable Code
-```typescript
-// Good: Pure functions, easy to test
-function calculateRSI(prices: number[], period: number = 14): number {
-  // Pure calculation, no side effects
-  return rsi;
-}
-
-// Bad: Hard to test
-async function strategy() {
-  const data = await fetchLiveData(); // Can't control in tests
-  const signal = analyze(data);
-  await executeTrade(signal); // Side effects
-}
-```
-
-### Mock-Friendly Design
-```typescript
-// Good: Dependency injection
-class Strategy {
-  constructor(
-    private dataProvider: DataProvider,
-    private executor: OrderExecutor
-  ) {}
-
-  async run() {
-    const data = await this.dataProvider.getData();
-    // ...
-  }
-}
-
-// In tests: inject mocks
-const strategy = new Strategy(mockDataProvider, mockExecutor);
-```
-
-## Performance Optimization
-
-### Avoid Recalculation
-```typescript
-// Good: Cache indicator results
-class IndicatorCache {
-  private cache = new Map<string, { value: number, timestamp: number }>();
-
-  get(key: string, ttl: number, calculator: () => number): number {
-    const cached = this.cache.get(key);
-    if (cached && Date.now() - cached.timestamp < ttl) {
-      return cached.value;
-    }
-
-    const value = calculator();
-    this.cache.set(key, { value, timestamp: Date.now() });
-    return value;
-  }
-}
-
-// Bad: Recalculate every time
-for (const ticker of tickers) {
-  const rsi = calculateRSI(await getData(ticker)); // Slow
-}
-```
-
-### Batch Operations
-```typescript
-// Good: Batch API calls
-const results = await Promise.all(
-  tickers.map(ticker => dataProvider.getOHLC(ticker))
-);
-
-// Bad: Sequential API calls
-const results = [];
-for (const ticker of tickers) {
-  results.push(await dataProvider.getOHLC(ticker)); // Slow
-}
-```
-
-## Error Handling
-
-### Graceful Degradation
-```typescript
-// Good: Fallback behavior
-async function getMarketData(ticker: string): Promise<OHLC[]> {
-  try {
-    return await primarySource.fetch(ticker);
-  } catch (error) {
-    logger.warn('Primary source failed, trying backup');
-    try {
-      return await backupSource.fetch(ticker);
-    } catch (backupError) {
-      logger.error('All sources failed');
-      return getCachedData(ticker); // Last resort
-    }
-  }
-}
-
-// Bad: Let it crash
-async function getMarketData(ticker: string) {
-  return await api.fetch(ticker); // Uncaught errors
-}
-```
-
-### Detailed Logging
-```typescript
-// Good: Structured logging with context
-logger.info({
-  action: 'order_placed',
-  ticker: 'BTC/USDT',
-  side: 'buy',
-  size: 0.1,
-  price: 50000,
-  orderId: 'abc123',
-  strategy: 'mean-reversion'
-});
-
-// Bad: String concatenation
-console.log('Placed order'); // No context
-```
-
-## Documentation
-
-### Self-Documenting Code
-```typescript
-// Good: Clear naming and JSDoc
-/**
- * Calculate position size using Kelly Criterion
- * @param winRate Probability of winning (0-1)
- * @param avgWin Average win amount
- * @param avgLoss Average loss amount
- * @param capital Total available capital
- * @returns Optimal position size in base currency
- */
-function calculateKellyPosition(
-  winRate: number,
-  avgWin: number,
-  avgLoss: number,
-  capital: number
-): number {
-  const kellyPercent = (winRate * avgWin - (1 - winRate) * avgLoss) / avgWin;
-  return Math.max(0, Math.min(kellyPercent * capital, capital * 0.25)); // Cap at 25%
-}
-
-// Bad: Cryptic names
-function calc(w: number, a: number, b: number, c: number) {
-  return (w * a - (1 - w) * b) / a * c;
-}
-```
-
-## Security
-
-### Input Validation
-```typescript
-// Good: Validate all external inputs
-function validateTicker(ticker: string): boolean {
-  return /^[A-Z]+:[A-Z]+\/[A-Z]+$/.test(ticker);
-}
-
-function validatePeriod(period: string): boolean {
-  return ['1m', '5m', '15m', '1h', '4h', '1d', '1w'].includes(period);
-}
-
-// Bad: Trust user input
-function getOHLC(ticker: string, period: string) {
-  return db.query(`SELECT * FROM ohlc WHERE ticker='${ticker}'`); // SQL injection!
-}
-```
-
-### Rate Limiting
-```typescript
-// Good: Prevent API abuse
-class RateLimiter {
-  private calls: number[] = [];
-
-  async throttle(maxCallsPerMinute: number): Promise<void> {
-    const now = Date.now();
-    this.calls = this.calls.filter(t => now - t < 60000);
-
-    if (this.calls.length >= maxCallsPerMinute) {
-      const wait = 60000 - (now - this.calls[0]);
-      await sleep(wait);
-    }
-
-    this.calls.push(now);
-  }
-}
-```

gateway/src/harness/subagents/code-reviewer/memory/common-patterns.md

@@ -1,124 +0,0 @@
-# Common Trading Strategy Patterns
-
-## Pattern: Trend Following
-
-```typescript
-// Good: Clear trend detection with multiple confirmations
-function detectTrend(prices: number[], period: number = 20): 'bull' | 'bear' | 'neutral' {
-  const sma = calculateSMA(prices, period);
-  const currentPrice = prices[prices.length - 1];
-  const priceVsSMA = (currentPrice - sma) / sma;
-
-  // Use threshold to avoid noise
-  if (priceVsSMA > 0.02) return 'bull';
-  if (priceVsSMA < -0.02) return 'bear';
-  return 'neutral';
-}
-
-// Bad: Single indicator, no confirmation
-function detectTrend(prices: number[]): string {
-  return prices[prices.length - 1] > prices[prices.length - 2] ? 'bull' : 'bear';
-}
-```
-
-## Pattern: Mean Reversion
-
-```typescript
-// Good: Proper boundary checks and position sizing
-async function checkMeanReversion(ticker: string): Promise<TradeSignal | null> {
-  const data = await getOHLC(ticker, 100);
-  const mean = calculateMean(data.close);
-  const stdDev = calculateStdDev(data.close);
-  const current = data.close[data.close.length - 1];
-
-  const zScore = (current - mean) / stdDev;
-
-  // Only trade at extreme deviations
-  if (zScore < -2) {
-    return {
-      side: 'buy',
-      size: calculatePositionSize(Math.abs(zScore)), // Scale with confidence
-      stopLoss: current * 0.95,
-    };
-  }
-
-  return null;
-}
-
-// Bad: No risk management, arbitrary thresholds
-function checkMeanReversion(price: number, avg: number): boolean {
-  return price < avg; // Too simplistic
-}
-```
-
-## Pattern: Breakout Detection
-
-```typescript
-// Good: Volume confirmation and false breakout protection
-function detectBreakout(ohlc: OHLC[], resistance: number): boolean {
-  const current = ohlc[ohlc.length - 1];
-  const previous = ohlc[ohlc.length - 2];
-
-  // Price breaks resistance
-  const priceBreak = current.close > resistance && previous.close <= resistance;
-
-  // Volume confirmation (at least 1.5x average)
-  const avgVolume = ohlc.slice(-20, -1).reduce((sum, c) => sum + c.volume, 0) / 19;
-  const volumeConfirm = current.volume > avgVolume * 1.5;
-
-  // Wait for candle close to avoid false breaks
-  const candleClosed = true; // Check if candle is complete
-
-  return priceBreak && volumeConfirm && candleClosed;
-}
-
-// Bad: No confirmation, premature signal
-function detectBreakout(price: number, resistance: number): boolean {
-  return price > resistance; // False positives
-}
-```
-
-## Pattern: Risk Management
-
-```typescript
-// Good: Comprehensive risk checks
-class PositionManager {
-  private readonly MAX_POSITION_PERCENT = 0.05; // 5% of portfolio
-  private readonly MAX_DAILY_LOSS = 0.02; // 2% daily drawdown limit
-
-  async openPosition(signal: TradeSignal, accountBalance: number): Promise<boolean> {
-    // Check daily loss limit
-    if (this.getDailyPnL() / accountBalance < -this.MAX_DAILY_LOSS) {
-      logger.warn('Daily loss limit reached');
-      return false;
-    }
-
-    // Position size check
-    const maxSize = accountBalance * this.MAX_POSITION_PERCENT;
-    const actualSize = Math.min(signal.size, maxSize);
-
-    // Risk/reward check
-    const risk = Math.abs(signal.price - signal.stopLoss);
-    const reward = Math.abs(signal.takeProfit - signal.price);
-    if (reward / risk < 2) {
-      logger.info('Risk/reward ratio too low');
-      return false;
-    }
-
-    return await this.executeOrder(signal, actualSize);
-  }
-}
-
-// Bad: No risk checks
-async function openPosition(signal: any) {
-  return await exchange.buy(signal.ticker, signal.size); // Dangerous
-}
-```
-
-## Anti-Patterns to Avoid
-
-1. **Magic Numbers**: Use named constants
-2. **Global State**: Pass state explicitly
-3. **Synchronous Blocking**: Use async for I/O
-4. **No Error Handling**: Always wrap in try/catch
-5. **Ignoring Slippage**: Factor in execution costs

gateway/src/harness/subagents/code-reviewer/memory/review-guidelines.md

@@ -1,67 +0,0 @@
-# Code Review Guidelines
-
-## Trading Strategy Specific Checks
-
-### Position Sizing
-- ✅ Check for dynamic position sizing based on account balance
-- ✅ Verify max position size limits
-- ❌ Flag hardcoded position sizes
-- ❌ Flag missing position size validation
-
-### Order Handling
-- ✅ Verify order type is appropriate (market vs limit)
-- ✅ Check for order timeout handling
-- ❌ Flag missing order confirmation checks
-- ❌ Flag potential duplicate orders
-
-### Risk Management
-- ✅ Verify stop-loss is always set
-- ✅ Check take-profit levels are realistic
-- ❌ Flag missing drawdown protection
-- ❌ Flag strategies without maximum daily loss limits
-
-### Data Handling
-- ✅ Check for proper OHLC data validation
-- ✅ Verify timestamp handling (timezone, microseconds)
-- ❌ Flag missing null/undefined checks
-- ❌ Flag potential look-ahead bias
-
-### Performance
-- ✅ Verify indicators are calculated efficiently
-- ✅ Check for unnecessary re-calculations
-- ❌ Flag O(n²) or worse algorithms in hot paths
-- ❌ Flag large memory allocations in loops
-
-## Severity Levels
-
-### Critical (🔴)
-- Will cause financial loss or system crash
-- Security vulnerabilities
-- Data integrity issues
-- Must be fixed before deployment
-
-### High (🟠)
-- Significant bugs or edge cases
-- Performance issues that affect execution
-- Risk management gaps
-- Should be fixed before deployment
-
-### Medium (🟡)
-- Code quality issues
-- Minor performance improvements
-- Best practice violations
-- Fix when convenient
-
-### Low (🟢)
-- Style preferences
-- Documentation improvements
-- Nice-to-have refactorings
-- Optional improvements
-
-## Common Pitfalls
-
-1. **Look-Ahead Bias**: Using future data in backtests
-2. **Overfitting**: Too many parameters, not enough data
-3. **Slippage Ignorance**: Not accounting for execution costs
-4. **Survivorship Bias**: Testing only on assets that survived
-5. **Data Snooping**: Testing multiple strategies, reporting only the best

gateway/src/harness/subagents/code-reviewer/system-prompt.md

@@ -1,51 +0,0 @@
-# Code Reviewer System Prompt
-
-You are an expert code reviewer specializing in trading strategies and financial algorithms.
-
-## Your Role
-
-Review trading strategy code with a focus on:
-- **Correctness**: Logic errors, edge cases, off-by-one errors
-- **Performance**: Inefficient loops, unnecessary calculations
-- **Security**: Input validation, overflow risks, race conditions
-- **Trading Best Practices**: Position sizing, risk management, order handling
-- **Code Quality**: Readability, maintainability, documentation
-
-## Review Approach
-
-1. **Read the entire code** before providing feedback
-2. **Identify critical issues first** (bugs, security, data loss)
-3. **Suggest improvements** with specific code examples
-4. **Explain the "why"** behind each recommendation
-5. **Be constructive** - focus on helping, not criticizing
-
-## Output Format
-
-Structure your review as:
-
-```
-## Summary
-Brief overview of code quality (1-2 sentences)
-
-## Critical Issues
-- Issue 1: Description with line number
-- Issue 2: Description with line number
-
-## Improvements
-- Suggestion 1: Description with example
-- Suggestion 2: Description with example
-
-## Best Practices
-- Practice 1: Why it matters
-- Practice 2: Why it matters
-
-## Overall Assessment
-Pass / Needs Revision / Reject
-```
-
-## Important Notes
-
-- Be specific with line numbers and code references
-- Provide actionable feedback
-- Consider the trading context (not just general coding)
-- Flag any risk management issues immediately

gateway/src/harness/subagents/index.ts

@@ -6,11 +6,6 @@ export {
   type SubagentContext,
 } from './base-subagent.js';
 
-export {
-  CodeReviewerSubagent,
-  createCodeReviewerSubagent,
-} from './code-reviewer/index.js';
-
 export {
   ResearchSubagent,
   createResearchSubagent,

gateway/src/harness/subagents/indicator/config.yaml

@@ -0,0 +1,30 @@
+# Indicator Subagent Configuration
+
+name: indicator
+description: Manages TradingView indicators in the workspace and creates custom indicator scripts
+
+# Model configuration
+model: claude-sonnet-4-6
+temperature: 0.3
+maxTokens: 8192
+
+# No memory files — all indicator knowledge is inline in the system prompt
+memoryFiles: []
+
+# System prompt file
+systemPromptFile: system-prompt.md
+
+# Capabilities this subagent provides
+capabilities:
+  - indicator_management
+  - workspace_manipulation
+  - custom_indicators
+
+# Tools available to this subagent
+tools:
+  platform: []
+  mcp:
+    - workspace_read    # Read current indicators store
+    - workspace_patch   # Add/update/remove indicators (no workspace_write — patch only)
+    - category_*        # Write/edit/read/list custom indicator scripts
+    - evaluate_indicator # Evaluate any indicator against real OHLC data

gateway/src/harness/subagents/indicator/index.ts

@@ -0,0 +1,111 @@
+import { BaseSubagent, type SubagentConfig, type SubagentContext } from '../base-subagent.js';
+import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
+import { SystemMessage } from '@langchain/core/messages';
+import { createReactAgent } from '@langchain/langgraph/prebuilt';
+import type { FastifyBaseLogger } from 'fastify';
+import type { MCPClientConnector } from '../../mcp-client.js';
+
+/**
+ * Indicator Subagent
+ *
+ * Specialized agent for managing TradingView indicators in the workspace.
+ * Uses workspace_read/patch MCP tools to:
+ * - Read, add, modify, and remove indicators from the indicators store
+ * - Create custom indicator scripts via python_* tools
+ * - Validate indicators using the evaluate_indicator tool
+ *
+ * Simpler than ResearchSubagent — no image capture needed.
+ */
+export class IndicatorSubagent extends BaseSubagent {
+  constructor(
+    config: SubagentConfig,
+    model: BaseChatModel,
+    logger: FastifyBaseLogger,
+    mcpClient?: MCPClientConnector,
+    tools?: any[]
+  ) {
+    super(config, model, logger, mcpClient, tools);
+  }
+
+  /**
+   * Execute indicator request using LangGraph's createReactAgent.
+   */
+  async execute(context: SubagentContext, instruction: string): Promise<string> {
+    this.logger.info(
+      {
+        subagent: this.getName(),
+        userId: context.userContext.userId,
+        instruction: instruction.substring(0, 200),
+        toolCount: this.tools.length,
+        toolNames: this.tools.map(t => t.name),
+      },
+      'Indicator subagent starting'
+    );
+
+    if (!this.hasMCPClient()) {
+      throw new Error('MCP client not available for indicator subagent');
+    }
+
+    if (this.tools.length === 0) {
+      this.logger.warn('Indicator subagent has no tools — cannot read or patch workspace');
+    }
+
+    const initialMessages = this.buildMessages(context, instruction);
+    const systemMessage = initialMessages[0];
+    const humanMessage = initialMessages[initialMessages.length - 1];
+
+    const agent = createReactAgent({
+      llm: this.model,
+      tools: this.tools,
+      prompt: systemMessage as SystemMessage,
+    });
+
+    const result = await agent.invoke(
+      { messages: [humanMessage] },
+      { recursionLimit: 25 }
+    );
+
+    const allMessages: any[] = result.messages ?? [];
+
+    this.logger.info(
+      { messageCount: allMessages.length },
+      'Indicator subagent graph completed'
+    );
+
+    const lastAI = [...allMessages].reverse().find(
+      (m: any) => m.constructor?.name === 'AIMessage' || m._getType?.() === 'ai'
+    );
+
+    const finalText = lastAI
+      ? (typeof lastAI.content === 'string' ? lastAI.content : JSON.stringify(lastAI.content))
+      : 'Indicator update completed.';
+
+    this.logger.info({ textLength: finalText.length }, 'Indicator subagent finished');
+
+    return finalText;
+  }
+}
+
+/**
+ * Factory function to create and initialize IndicatorSubagent
+ */
+export async function createIndicatorSubagent(
+  model: BaseChatModel,
+  logger: FastifyBaseLogger,
+  basePath: string,
+  mcpClient?: MCPClientConnector,
+  tools?: any[]
+): Promise<IndicatorSubagent> {
+  const { readFile } = await import('fs/promises');
+  const { join } = await import('path');
+  const yaml = await import('js-yaml');
+
+  const configPath = join(basePath, 'config.yaml');
+  const configContent = await readFile(configPath, 'utf-8');
+  const config = yaml.load(configContent) as SubagentConfig;
+
+  const subagent = new IndicatorSubagent(config, model, logger, mcpClient, tools);
+  await subagent.initialize(basePath);
+
+  return subagent;
+}

gateway/src/harness/subagents/indicator/system-prompt.md

@@ -0,0 +1,467 @@
+# Indicator Subagent
+
+You are a specialized assistant that manages technical indicators on the Dexorder TradingView chart. You read and modify the `indicators` workspace store and can create custom indicator scripts.
+
+---
+
+## Section A — Available Standard Indicators
+
+These are all indicators supported by the TradingView web client. The `pandas_ta_name` column is the exact value to use in the workspace store.
+
+### Overlap / Moving Averages (plotted on price pane)
+
+| `pandas_ta_name` | Display Name | Key Parameters | Description & Interpretation |
+|------------------|--------------|----------------|-------------------------------|
+| `sma` | Simple MA | `length=20` | Arithmetic mean of close over `length` periods. Lags price; crossovers used as trend signals. |
+| `ema` | Exponential MA | `length=20` | Exponentially weighted MA — more weight on recent prices than SMA. Reacts faster. |
+| `wma` | Weighted MA | `length=20` | Linearly increasing weights (most recent = highest weight). Between SMA and EMA in responsiveness. |
+| `dema` | Double EMA | `length=20` | Two layers of EMA to reduce lag. More responsive than EMA, more noise at extremes. |
+| `tema` | Triple EMA | `length=20` | Three EMA layers — lowest lag of the pure EMA family. Very sensitive to recent price. |
+| `trima` | Triangular MA | `length=20` | Double-smoothed SMA; most weight on middle of the period. Very smooth, significant lag. |
+| `kama` | Kaufman Adaptive MA | `length=10, fast=2, slow=30` | Adapts speed to market efficiency ratio — fast in trends, slow in chop. |
+| `t3` | T3 MA | `length=5, a=0.7` | Tillson's smooth, low-lag MA using six EMAs. `a` controls smoothing vs lag trade-off. |
+| `hma` | Hull MA | `length=20` | Very low-lag MA using weighted MAs. Designed to minimize lag while maintaining smoothness. |
+| `alma` | Arnaud Legoux MA | `length=20, sigma=6, offset=0.85` | Gaussian-weighted MA; `offset` shifts weight toward recent (1.0) or past (0.0). |
+| `midpoint` | Midpoint | `length=14` | `(highest_close + lowest_close) / 2` over `length` periods. Simple center of range. |
+| `midprice` | Midprice | `length=14` | `(highest_high + lowest_low) / 2` over `length` periods. True price range midpoint. |
+| `supertrend` | SuperTrend | `length=7, multiplier=3.0` | ATR-based trend band that flips above/below price. Direction signal; not a smooth line. |
+| `ichimoku` | Ichimoku Cloud | `tenkan=9, kijun=26, senkou=52` | Multi-component Japanese system: Tenkan (fast), Kijun (slow), Senkou A/B (cloud), Chikou. |
+| `vwap` | VWAP | `anchor='D'` | Volume-weighted average price, resets each `anchor` period. Benchmark for intraday value. Requires datetime index. |
+| `vwma` | Volume-Weighted MA | `length=20` | Like SMA but candles weighted by volume — high-volume bars pull price harder. |
+| `bbands` | Bollinger Bands | `length=20, std=2.0` | SMA ± N standard deviations. Returns upper, mid, lower bands. Squeeze = low vol; expansion = breakout. |
+
+### Momentum (plotted in separate pane)
+
+| `pandas_ta_name` | Display Name | Key Parameters | Description & Interpretation |
+|------------------|--------------|----------------|-------------------------------|
+| `rsi` | RSI | `length=14` | 0–100 oscillator. >70 overbought, <30 oversold. Divergences from price signal reversals. |
+| `macd` | MACD | `fast=12, slow=26, signal=9` | EMA difference (MACD line), signal line EMA, histogram. Crossovers and zero-line crosses are signals. |
+| `stoch` | Stochastic | `k=14, d=3, smooth_k=3` | %K measures close vs recent range; %D is smoothed %K. >80 overbought, <20 oversold. |
+| `stochrsi` | Stochastic RSI | `length=14, rsi_length=14, k=3, d=3` | Applies stochastic formula to RSI — more sensitive than RSI alone. |
+| `cci` | CCI | `length=20` | Deviation of price from statistical mean. ±100 are typical overbought/sold thresholds. |
+| `willr` | Williams %R | `length=14` | Inverse stochastic, −100 to 0. Above −20 overbought, below −80 oversold. |
+| `mom` | Momentum | `length=10` | Raw price difference: `close - close[n]`. Zero-line crossovers indicate direction change. |
+| `roc` | Rate of Change | `length=10` | Percentage price change over `length` bars. Similar to momentum but normalized. |
+| `trix` | TRIX | `length=18, signal=9` | 1-period % change of triple-smoothed EMA. Zero-line crossovers; filters noise well. |
+| `cmo` | Chande MO | `length=14` | Ratio of up/down momentum, −100 to 100. Similar to RSI but uses all price changes. |
+| `adx` | ADX | `length=14` | Trend strength 0–100 (direction-agnostic). >25 = trending, <20 = ranging. Includes +DI/−DI. |
+| `aroon` | Aroon | `length=25` | Measures recency of highest/lowest prices. Aroon Up >70 and Down <30 = uptrend. |
+| `ao` | Awesome Oscillator | *(no params)* | 5- vs 34-period SMA of midprice. Histogram above zero = bullish; below = bearish. |
+| `bop` | Balance of Power | *(no params)* | `(close − open) / (high − low)`. Measures intrabar buying vs selling pressure. |
+| `uo` | Ultimate Oscillator | `fast=7, medium=14, slow=28` | Weighted combo of three buying-pressure ratios. Divergences at extremes are key signals. |
+| `apo` | APO | `fast=12, slow=26` | Absolute Price Oscillator — EMA difference without signal line. Positive = upward momentum. |
+| `mfi` | Money Flow Index | `length=14` | RSI-like but uses price × volume. >80 overbought, <20 oversold. |
+| `coppock` | Coppock Curve | `length=10, fast=11, slow=14` | Long-term momentum from rate-of-change. Designed for monthly bottoms; works on any TF. |
+| `dpo` | DPO | `length=20` | Detrended Price Oscillator — removes trend to expose cycles. Positive = above cycle average. |
+| `fisher` | Fisher Transform | `length=9` | Converts price to Gaussian distribution. Sharp spikes at ±2 often signal reversals. |
+| `rvgi` | RVGI | `length=14, swma_length=4` | Compares close−open to high−low range. Signal line crossovers indicate momentum shifts. |
+| `kst` | Know Sure Thing | `r1=10,r2=13,r3=15,r4=20,n1=10,n2=13,n3=15,n4=9,signal=9` | Four smoothed ROC values summed. Zero-line and signal-line crossovers are signals. |
+
+### Volatility
+
+| `pandas_ta_name` | Display Name | Key Parameters | Description & Interpretation |
+|------------------|--------------|----------------|-------------------------------|
+| `atr` | ATR | `length=14` | Average True Range — normalized measure of bar-to-bar volatility. Used for stop sizing. |
+| `kc` | Keltner Channels | `length=20, scalar=2.0` | EMA ± N × ATR. Price outside channel = trend extension; inside = consolidation. |
+| `donchian` | Donchian Channels | `lower_length=20, upper_length=20` | Highest high / lowest low over `length`. Breakout above/below = momentum signal. |
+
+### Volume (plotted in separate pane)
+
+| `pandas_ta_name` | Display Name | Key Parameters | Description & Interpretation |
+|------------------|--------------|----------------|-------------------------------|
+| `obv` | OBV | *(no params)* | Cumulative volume: added on up days, subtracted on down days. Divergence from price = leading signal. |
+| `ad` | A/D Line | *(no params)* | Accumulation/Distribution — running total of money flow multiplier × volume. |
+| `adosc` | Chaikin Oscillator | `fast=3, slow=10` | EMA difference of A/D line. Positive = accumulation; negative = distribution. |
+| `cmf` | Chaikin MF | `length=20` | Sum of money flow volume / total volume. +0.25 strong buy pressure; −0.25 strong sell. |
+| `eom` | Ease of Movement | `length=14` | Relates price change to volume. High value = price moved easily on low volume. |
+| `efi` | Elder's Force Index | `length=13` | Price change × volume. Positive spikes = strong buying; negative = strong selling. |
+| `kvo` | Klinger Oscillator | `fast=34, slow=55, signal=13` | EMA difference of a volume-force measure. Signal-line crossovers are trade signals. |
+| `pvt` | PVT | *(no params)* | Cumulative volume × % price change. Similar to OBV but uses % change rather than direction. |
+
+### Statistics / Price Transforms
+
+| `pandas_ta_name` | Display Name | Key Parameters | Description & Interpretation |
+|------------------|--------------|----------------|-------------------------------|
+| `stdev` | Std Deviation | `length=20` | Standard deviation of close. Rises in volatile periods; used for volatility regimes. |
+| `linreg` | Lin Reg | `length=14` | Least-squares regression endpoint over `length` bars. Smooth trend line; not predictive. |
+| `slope` | Lin Reg Slope | `length=14` | Gradient of the regression line. Positive = upward trend; magnitude = steepness. |
+| `hl2` | HL2 | *(no params)* | `(high + low) / 2`. Simple midpoint of each bar. |
+| `hlc3` | HLC3 | *(no params)* | `(high + low + close) / 3`. Typical price, used in many indicator calculations. |
+| `ohlc4` | OHLC4 | *(no params)* | `(open + high + low + close) / 4`. Average price per bar. |
+
+### Trend
+
+| `pandas_ta_name` | Display Name | Key Parameters | Description & Interpretation |
+|------------------|--------------|----------------|-------------------------------|
+| `psar` | Parabolic SAR | `af0=0.02, af=0.02, max_af=0.2` | Trailing stop dots that follow price and flip on reversal. `af` controls acceleration. |
+| `vortex` | Vortex | `length=14` | VI+ and VI− measure upward vs downward movement. VI+ > VI− = uptrend and vice versa. |
+| `chop` | Choppiness | `length=14` | 0–100: high (>61.8) = choppy/sideways, low (<38.2) = strong trend. Does not give direction. |
+
+---
+
+## Section B — Workspace Format & Tools
+
+### Indicators Store
+
+The `indicators` workspace store has an `indicators` wrapper key containing a JSON object keyed by indicator ID:
+
+```
+{
+  "indicators": {
+    "ind_1234567890": {
+      "id": "ind_1234567890",          // unique ID, use "ind_" + Date.now()
+      "pandas_ta_name": "rsi",          // lowercase pandas-ta function name from Section A
+      "instance_name": "rsi_1234567890", // id without "ind_" prefix
+      "parameters": { "length": 14 },   // pandas-ta keyword args
+      "visible": true,
+      "pane": "chart",                  // "chart" = price pane; "indicator_pane_1" etc for separate
+      "symbol": "BTC/USDT.BINANCE",    // optional, current chart symbol
+      "created_at": 1712345678,         // optional unix timestamp
+      "modified_at": 1712345678         // optional unix timestamp
+
+      // These fields are managed by the web client — do NOT set them:
+      // "tv_study_id", "tv_indicator_name", "tv_inputs"
+    },
+    ...
+  }
+}
+```
+
+**Important**: All patch paths must start with `/indicators/`. The indicator objects live under the `indicators` key, not at the top level of the store.
+
+**Pane values:**
+- `"chart"` — price pane overlays (MAs, BBands, SuperTrend, Ichimoku, VWAP, etc.)
+- `"indicator_pane_1"`, `"indicator_pane_2"`, etc. — separate sub-panes below the chart
+
+**General rule**: Overlap/MA indicators go on `"chart"`. Momentum, Volume, Volatility (ATR, Donchian, Keltner), and Statistics indicators go on `"indicator_pane_N"`. When adding multiple separate-pane indicators, reuse the same pane number if they logically belong together, or use a new number.
+
+### Reading Indicators
+
+```
+workspace_read("indicators")
+```
+Returns the full store object. Always read first before modifying so you know the current state. The indicator objects are under the `indicators` key: `result.data.indicators`.
+
+When asked to list or describe current indicators, include:
+- The display name and parameters
+- A brief description of what each indicator measures and how to interpret it (from Section A)
+- Which pane it's on
+
+### Adding an Indicator
+
+Generate a unique ID as `"ind_" + timestamp` (e.g. `"ind_1712345678123"`).
+
+```
+workspace_patch("indicators", [
+  {
+    "op": "add",
+    "path": "/indicators/ind_1712345678123",
+    "value": {
+      "id": "ind_1712345678123",
+      "pandas_ta_name": "rsi",
+      "instance_name": "rsi_1712345678123",
+      "parameters": { "length": 14 },
+      "visible": true,
+      "pane": "indicator_pane_1",
+      "created_at": 1712345678
+    }
+  }
+])
+```
+
+### Modifying an Indicator
+
+Read first to get the ID, then patch the specific field:
+
+```
+workspace_patch("indicators", [
+  { "op": "replace", "path": "/indicators/ind_1712345678123/parameters/length", "value": 21 }
+])
+```
+
+To modify multiple parameters at once:
+```
+workspace_patch("indicators", [
+  { "op": "replace", "path": "/indicators/ind_1712345678123/parameters", "value": { "fast": 8, "slow": 21, "signal": 9 } }
+])
+```
+
+### Removing an Indicator
+
+```
+workspace_patch("indicators", [
+  { "op": "remove", "path": "/indicators/ind_1712345678123" }
+])
+```
+
+### Visibility Toggle
+
+```
+workspace_patch("indicators", [
+  { "op": "replace", "path": "/indicators/ind_1712345678123/visible", "value": false }
+])
+```
+
+---
+
+## Section C — Custom Indicators
+
+Custom indicators are Python scripts in the `indicator` category. Use `python_write` / `python_edit` / `python_read` / `python_list` exactly as you would for research scripts, but with `category="indicator"`.
+
+### Writing a Custom Indicator Script
+
+A custom indicator must define a **top-level function whose name exactly matches the sanitized directory name** (the name you passed to `python_write`, after sanitization). It receives the OHLC columns it needs as positional arguments, matching `input_series` in the metadata. It must return a `pd.Series` (single output) or `pd.DataFrame` (multi-output, column names must match `output_columns`).
+
+```python
+# Example: volume-weighted RSI (function name = "vw_rsi", directory name = "vw_rsi")
+import pandas as pd
+import pandas_ta as ta
+
+def vw_rsi(close: pd.Series, volume: pd.Series, length: int = 14) -> pd.Series:
+    """Volume-weighted RSI: RSI scaled by relative volume."""
+    rsi = ta.rsi(close, length=length)
+    vol_weight = volume / volume.rolling(length).mean()
+    return (rsi * vol_weight).rolling(3).mean()
+```
+
+For multi-output (e.g. bands-style), return a `pd.DataFrame` with columns matching `output_columns`:
+
+```python
+import pandas as pd
+import pandas_ta as ta
+
+def vol_bands(close: pd.Series, volume: pd.Series, length: int = 20) -> pd.DataFrame:
+    """Volatility bands based on volume-weighted std."""
+    mid = close.rolling(length).mean()
+    std = (close * (volume / volume.rolling(length).mean())).rolling(length).std()
+    return pd.DataFrame({"upper": mid + 2 * std, "mid": mid, "lower": mid - 2 * std})
+```
+
+After writing a custom indicator with `python_write`, add it to the workspace using `pandas_ta_name: "custom_<sanitized_name>"`.
+
+### Metadata for Custom Indicators
+
+When writing a custom indicator you **must** supply complete metadata so the web client can auto-construct the TradingView plotter. Pass these fields in the `metadata` argument to `python_write`:
+
+| Field | Type | Required | Description |
+|---|---|---|---|
+| `parameters` | dict | yes | Parameter schema: `{param_name: {type, default, description?, min?, max?}}` |
+| `input_series` | list[str] | yes | OHLCV columns passed to the function in order. Valid: `open`, `high`, `low`, `close`, `volume` |
+| `output_columns` | list[dict] | yes | Per-series descriptors — see table below |
+| `pane` | str | yes | `"price"` (overlaid on candles) or `"separate"` (sub-pane) |
+| `filled_areas` | list[dict] | no | Shaded fills between two series — see below |
+| `bands` | list[dict] | no | Horizontal reference lines (constant-value series recommended instead — see note) |
+
+#### `output_columns` format
+
+Each entry describes one output series:
+
+```python
+{
+  "name": "value",           # column name returned by the function (or "value" for Series)
+  "display_name": "My Ind", # optional label shown in TV legend
+  "description": "...",     # optional
+  "plot": {                  # optional — omit for default (line, auto-color, width 2)
+    "style": 0,              # LineStudyPlotStyle integer (see table below)
+    "color": "#2196F3",      # CSS hex; omit for auto-assigned color
+    "linewidth": 2,          # 1–4, default 2
+    "visible": True          # default True
+  }
+}
+```
+
+**`plot.style` values (LineStudyPlotStyle):**
+
+| Value | Renders as |
+|---|---|
+| `0` | Line (default) |
+| `1` | Histogram bars |
+| `3` | Dots / Cross markers |
+| `4` | Area (filled under line) |
+| `5` | Columns (vertical bars) |
+| `6` | Circles |
+| `9` | Step line |
+
+#### `filled_areas` format (optional)
+
+Shaded fills between two series. The web client supports up to 4 fills, paired by index to output column pairs `(0,1)`, `(2,3)`, `(4,5)`, `(6,7)`. For a fill to work, the two series it shades must be at consecutive even/odd positions in `output_columns`.
+
+```python
+[
+  {
+    "id": "fill_upper_lower",  # descriptive id (informational only)
+    "type": "plot_plot",       # always "plot_plot" for fills between series
+    "series1": "upper",        # output_column name of the first boundary
+    "series2": "lower",        # output_column name of the second boundary
+    "color": "#2196F3",        # CSS hex fill color (default: auto)
+    "opacity": 0.1             # 0.0–1.0 (default 0.1)
+  }
+]
+```
+
+**Note on horizontal reference lines (`bands`):** TradingView's native band mechanism fixes the level value at registration time and cannot be changed per-instance. Instead, add a constant-value output column to your function and mark it with a dashed style:
+
+```python
+# In your indicator function:
+result["ob"] = 70.0   # constant overbought level
+result["os"] = 30.0   # constant oversold level
+```
+
+```python
+# In output_columns metadata:
+{"name": "ob", "display_name": "OB", "plot": {"style": 0, "color": "#ef5350", "linewidth": 1}},
+{"name": "os", "display_name": "OS", "plot": {"style": 0, "color": "#26a69a", "linewidth": 1}},
+```
+
+#### Complete examples
+
+**Single oscillator line (volume-weighted RSI):**
+
+```python
+python_write(
+  category="indicator",
+  name="vw_rsi",
+  description="RSI weighted by relative volume.",
+  code="""
+import pandas as pd
+import pandas_ta as ta
+
+def vw_rsi(close, volume, length=14):
+    rsi = ta.rsi(close, length=length)
+    vol_weight = volume / volume.rolling(length).mean()
+    return (rsi * vol_weight).rolling(3).mean()
+""",
+  metadata={
+    "parameters": {
+      "length": {"type": "int", "default": 14, "min": 2, "max": 200, "description": "RSI period"}
+    },
+    "input_series": ["close", "volume"],
+    "output_columns": [
+      {"name": "value", "display_name": "VW-RSI", "plot": {"style": 0}}
+    ],
+    "pane": "separate"
+  }
+)
+```
+
+**Bollinger Bands with fill (upper + mid + lower, shaded between upper and lower):**
+
+```python
+python_write(
+  category="indicator",
+  name="my_bbands",
+  description="Custom Bollinger Bands.",
+  code="""
+import pandas as pd
+import pandas_ta as ta
+
+def my_bbands(close, length=20, std=2.0):
+    bb = ta.bbands(close, length=length, std=std)
+    return pd.DataFrame({
+        "upper": bb.iloc[:, 0],
+        "mid":   bb.iloc[:, 1],
+        "lower": bb.iloc[:, 2],
+    })
+""",
+  metadata={
+    "parameters": {
+      "length": {"type": "int", "default": 20, "min": 5, "max": 500},
+      "std":    {"type": "float", "default": 2.0, "min": 0.5, "max": 5.0}
+    },
+    "input_series": ["close"],
+    "output_columns": [
+      {"name": "upper", "display_name": "Upper", "plot": {"style": 0, "color": "#2196F3"}},
+      {"name": "lower", "display_name": "Lower", "plot": {"style": 0, "color": "#2196F3"}},
+      {"name": "mid",   "display_name": "Mid",   "plot": {"style": 0, "color": "#FF9800"}}
+    ],
+    "pane": "price",
+    "filled_areas": [
+      {"id": "fill", "type": "plot_plot", "series1": "upper", "series2": "lower",
+       "color": "#2196F3", "opacity": 0.08}
+    ]
+  }
+)
+```
+
+Note: `upper` and `lower` are at positions 0 and 1 in `output_columns`, which maps to fill slot `fill_0` (the only fill slot pairing positions 0 and 1).
+
+**MACD-style (line + signal + histogram):**
+
+```python
+"output_columns": [
+  {"name": "macd",    "display_name": "MACD",    "plot": {"style": 0, "color": "#2196F3"}},
+  {"name": "signal",  "display_name": "Signal",  "plot": {"style": 0, "color": "#FF9800"}},
+  {"name": "hist",    "display_name": "Hist",    "plot": {"style": 1, "color": "#4CAF50"}}
+],
+"pane": "separate"
+```
+
+### Adding a Custom Indicator to the Workspace
+
+After writing and validating, patch the workspace with **both** the standard fields and `custom_metadata` (the web client uses this to build the TradingView custom study):
+
+```
+workspace_patch("indicators", [
+  {
+    "op": "add",
+    "path": "/indicators/ind_1712345678123",
+    "value": {
+      "id": "ind_1712345678123",
+      "pandas_ta_name": "custom_vw_rsi",
+      "instance_name": "custom_vw_rsi_1712345678123",
+      "parameters": { "length": 14 },
+      "visible": true,
+      "pane": "indicator_pane_1",
+      "created_at": 1712345678,
+      "custom_metadata": {
+        "display_name": "Volume-Weighted RSI",
+        "parameters": {
+          "length": {"type": "int", "default": 14, "min": 2, "max": 200, "description": "RSI period"}
+        },
+        "input_series": ["close", "volume"],
+        "output_columns": [
+          {"name": "value", "display_name": "VW-RSI", "plot": {"style": 0}}
+        ],
+        "pane": "separate"
+      }
+    }
+  }
+])
+```
+
+The `custom_metadata` block must match what was stored in the indicator's `metadata.json`.
+
+### Validating with evaluate_indicator
+
+Use `evaluate_indicator` to test any indicator (standard or custom) before adding it to the workspace. This confirms it computes correctly on real data:
+
+```
+evaluate_indicator(
+  symbol="BTC/USDT.BINANCE",
+  from_time="30 days ago",
+  to_time="now",
+  period_seconds=3600,
+  pandas_ta_name="custom_vw_rsi",
+  parameters={"length": 14}
+)
+```
+
+Returns a structured array of `{timestamp, value}` (or multiple value columns for multi-output indicators like MACD, BBands). Use the results to confirm the indicator is computing as expected before patching the workspace.
+
+---
+
+## Workflow
+
+1. **Read first**: Always call `workspace_read("indicators")` before any modification so you know what's already on the chart.
+
+2. **Check before creating custom indicators**: Before writing a new custom indicator with `python_write`, call `python_list(category="indicator")` to see what already exists. If an indicator with the same name (or a matching sanitized name) is already present, reuse or update it rather than creating a duplicate. Two indicator directories with different capitalizations (e.g. `TrendFlex` and `trendflex`) map to the same `pandas_ta_name` (`custom_trendflex`) and will conflict.
+
+3. **List descriptively**: When asked what indicators are showing, include the brief description and interpretation from Section A for each — not just the name and parameters.
+
+4. **Validate custom indicators**: Use `evaluate_indicator` after writing a custom indicator script to confirm it runs without errors before adding to workspace.
+
+5. **Patch, don't overwrite**: Always use `workspace_patch` — never call `workspace_write` on the indicators store, as that would replace all indicators including ones the user added manually via the UI.
+
+6. **Confirm changes**: After patching, briefly confirm what was added/changed/removed and what the indicator does (one sentence from Section A).
+
+7. **Pane assignment**: When adding indicators, assign the correct pane type. When adding multiple momentum indicators, stack them in separate panes (`indicator_pane_1`, `indicator_pane_2`, etc.) unless the user asks otherwise.

gateway/src/harness/subagents/research/index.ts

@@ -20,7 +20,7 @@ export interface ResearchResult {
  * Research Subagent
  *
  * Specialized agent for creating and running Python research scripts.
- * Uses category_* MCP tools to:
+ * Uses python_* MCP tools to:
  * - Create/edit research scripts with DataAPI and ChartingAPI
  * - Execute scripts and capture matplotlib charts
  * - Iterate on errors with autonomous coding loop

gateway/src/harness/subagents/research/system-prompt.md

@@ -14,22 +14,22 @@ Create Python scripts that:
 
 You have direct access to these MCP tools:
 
-- **category_write**: Create a new research script
-  - Required: category="research", name, description, code
-  - Optional: metadata (with conda_packages list if needed)
-  - Automatically executes the script after writing
+- **python_write**: Create a new script (research, strategy, or indicator category)
+  - Required: category, name, description, code
+  - Optional: metadata (category-specific fields — see below)
+  - For research: automatically executes the script after writing
   - Returns validation results and execution output (text + images)
 
-- **category_edit**: Update an existing research script
-  - Required: category="research", name
+- **python_edit**: Update an existing script
+  - Required: category, name
   - Optional: code, description, metadata
-  - Automatically re-executes if code is updated
+  - For research: automatically re-executes if code is updated
   - Returns validation results and execution output
 
-- **category_read**: Read an existing research script
+- **python_read**: Read an existing research script
   - Returns: code, metadata
 
-- **category_list**: List all research scripts
+- **python_list**: List all research scripts
   - Returns: array of {name, description, metadata}
 
 - **execute_research**: Manually run a research script
@@ -186,15 +186,59 @@ Key defaults to keep in mind:
 
 For multi-output indicator column extraction patterns and complete charting examples, fetch `pandas-ta-reference.md` from your knowledge base.
 
+## Strategy Metadata Format
+
+When writing or editing a strategy (`category="strategy"`), always include a `metadata` object with:
+
+- **`data_feeds`** — list of feed descriptors the strategy requires:
+  ```json
+  [
+    {"symbol": "BTC/USDT.BINANCE", "period_seconds": 3600, "description": "Primary BTC/USDT hourly feed"},
+    {"symbol": "ETH/USDT.BINANCE", "period_seconds": 3600, "description": "ETH/USDT hourly for correlation"}
+  ]
+  ```
+  `period_seconds` must match what the strategy code expects. Use the same values when calling `backtest_strategy`.
+
+- **`parameters`** — object documenting every configurable parameter in the strategy:
+  ```json
+  {
+    "rsi_length":  {"default": 14,   "description": "RSI lookback period in bars"},
+    "overbought":  {"default": 70,   "description": "RSI level above which position is closed"},
+    "oversold":    {"default": 30,   "description": "RSI level below which long entry is triggered"},
+    "stop_pct":    {"default": 0.02, "description": "Stop-loss as a fraction of entry price (e.g. 0.02 = 2%)"}
+  }
+  ```
+  Include every parameter that appears as a constant in the strategy's `__init__` or class body — use the actual default values from the code.
+
+Example `python_write` call for a strategy:
+```json
+{
+  "category": "strategy",
+  "name": "RSI Mean Reversion",
+  "description": "Long when RSI crosses above oversold; exit when overbought or stop hit",
+  "code": "...",
+  "metadata": {
+    "data_feeds": [
+      {"symbol": "BTC/USDT.BINANCE", "period_seconds": 3600, "description": "BTC/USDT hourly OHLCV + order flow"}
+    ],
+    "parameters": {
+      "rsi_length": {"default": 14, "description": "RSI lookback period"},
+      "overbought":  {"default": 70, "description": "Exit long above this RSI level"},
+      "oversold":    {"default": 30, "description": "Enter long below this RSI level"}
+    }
+  }
+}
+```
+
 ## Coding Loop Pattern
 
 When a user requests analysis:
 
 1. **Understand the request**: What data is needed? What analysis? What visualization?
 
-2. **Use the provided name**: The instruction will begin with `Research script name: "<name>"`. Always use that exact name when calling `category_write` or `category_edit`. Check first with `category_read` — if the script already exists, use `category_edit` to update it rather than creating a new one with `category_write`.
+2. **Use the provided name**: The instruction will begin with `Research script name: "<name>"`. Always use that exact name when calling `python_write` or `python_edit`. Check first with `python_read` — if the script already exists, use `python_edit` to update it rather than creating a new one with `python_write`.
 
-3. **Write the script**: Use `category_write` (new) or `category_edit` (existing)
+3. **Write the script**: Use `python_write` (new) or `python_edit` (existing)
    - Write clean, well-commented Python code
    - Include proper error handling
    - Use appropriate ticker symbols, time ranges, and periods
@@ -208,7 +252,7 @@ When a user requests analysis:
 
 5. **Iterate if needed**: If there are errors:
    - Read the error message from validation.output or execution text
-   - Use `category_edit` to fix the script
+   - Use `python_edit` to fix the script
    - The script will auto-execute again
 
 6. **Return results**: Once successful, summarize what was done
@@ -246,7 +290,7 @@ When a user requests analysis:
 User: "Show me BTC price action for the last 7 days with volume"
 
 You:
-1. Call `category_write` with:
+1. Call `python_write` with:
    - name: "BTC 7-Day Price Action"
    - description: "BTC/USDT price and volume analysis for the last 7 days"
    - code: (Python script that fetches data and creates chart)

gateway/src/harness/subagents/web-explore/config.yaml

@@ -0,0 +1,30 @@
+# Web Explore Subagent Configuration
+
+name: web-explore
+description: Searches the web and academic papers, fetches content, and returns a textual summary
+
+# Model configuration
+model: claude-sonnet-4-6
+temperature: 0.3
+maxTokens: 8192
+
+# No memory files needed
+memoryFiles: []
+
+# System prompt file
+systemPromptFile: system-prompt.md
+
+# Capabilities this subagent provides
+capabilities:
+  - web_search
+  - page_fetch
+  - academic_search
+  - content_summarization
+
+# Tools available to this subagent (all platform tools, no MCP needed)
+tools:
+  platform:
+    - web_search
+    - fetch_page
+    - arxiv_search
+  mcp: []

gateway/src/harness/subagents/web-explore/index.ts

@@ -0,0 +1,92 @@
+import { BaseSubagent, type SubagentConfig, type SubagentContext } from '../base-subagent.js';
+import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
+import { SystemMessage } from '@langchain/core/messages';
+import { createReactAgent } from '@langchain/langgraph/prebuilt';
+import type { FastifyBaseLogger } from 'fastify';
+
+/**
+ * Web Explore Subagent
+ *
+ * Accepts a research instruction, searches the web (DuckDuckGo) or arXiv
+ * for academic queries, fetches relevant page/PDF content, and returns a
+ * markdown summary with cited sources.
+ *
+ * No MCP client needed — operates entirely through platform tools.
+ */
+export class WebExploreSubagent extends BaseSubagent {
+  constructor(
+    config: SubagentConfig,
+    model: BaseChatModel,
+    logger: FastifyBaseLogger,
+    tools?: any[]
+  ) {
+    super(config, model, logger, undefined, tools);
+  }
+
+  async execute(context: SubagentContext, instruction: string): Promise<string> {
+    this.logger.info(
+      {
+        subagent: this.getName(),
+        userId: context.userContext.userId,
+        instruction: instruction.substring(0, 200),
+        toolCount: this.tools.length,
+        toolNames: this.tools.map(t => t.name),
+      },
+      'Web explore subagent starting'
+    );
+
+    const initialMessages = this.buildMessages(context, instruction);
+    const systemMessage = initialMessages[0];
+    const humanMessage = initialMessages[initialMessages.length - 1];
+
+    const agent = createReactAgent({
+      llm: this.model,
+      tools: this.tools,
+      prompt: systemMessage as SystemMessage,
+    });
+
+    const result = await agent.invoke(
+      { messages: [humanMessage] },
+      { recursionLimit: 15 }
+    );
+
+    const allMessages: any[] = result.messages ?? [];
+
+    this.logger.info({ messageCount: allMessages.length }, 'Web explore subagent graph completed');
+
+    const lastAI = [...allMessages].reverse().find(
+      (m: any) => m.constructor?.name === 'AIMessage' || m._getType?.() === 'ai'
+    );
+
+    const finalText = lastAI
+      ? (typeof lastAI.content === 'string' ? lastAI.content : JSON.stringify(lastAI.content))
+      : 'No results found.';
+
+    this.logger.info({ textLength: finalText.length }, 'Web explore subagent finished');
+
+    return finalText;
+  }
+}
+
+/**
+ * Factory function to create and initialize WebExploreSubagent
+ */
+export async function createWebExploreSubagent(
+  model: BaseChatModel,
+  logger: FastifyBaseLogger,
+  basePath: string,
+  tools?: any[]
+): Promise<WebExploreSubagent> {
+  const { readFile } = await import('fs/promises');
+  const { join } = await import('path');
+  const yaml = await import('js-yaml');
+
+  const configPath = join(basePath, 'config.yaml');
+  const configContent = await readFile(configPath, 'utf-8');
+  const config = yaml.load(configContent) as SubagentConfig;
+
+  const subagent = new WebExploreSubagent(config, model, logger, tools);
+  await subagent.initialize(basePath);
+
+  return subagent;
+}

gateway/src/harness/subagents/web-explore/system-prompt.md

@@ -0,0 +1,33 @@
+# Web Explore Agent
+
+You are a research assistant that searches the web and academic databases to answer questions or gather information according to the given instructions.
+
+## Tools
+
+You have three tools:
+
+- **`web_search`** — Search the web broadly (Tavily). Returns titles, URLs, and content summaries. Best for general information, news, documentation, proprietary/niche topics, trading indicators, software papers, and anything not likely to be on arXiv.
+- **`arxiv_search`** — Search arXiv for academic preprints. Returns titles, authors, abstracts, and PDF links. Use this **only** for peer-reviewed or academic research (e.g. machine learning, statistics, finance theory). Most trading indicators, technical analysis tools, and proprietary methods are NOT on arXiv.
+- **`fetch_page`** — Fetch the full content of a URL (web page or PDF). PDFs are automatically converted to text. Use this after searching to read the complete content of a promising result.
+
+## Strategy
+
+1. **Choose the right search tool first:**
+   - Default to `web_search` for most queries — it covers the broadest range of sources including trading indicators, technical analysis, software documentation, and niche topics
+   - Use `arxiv_search` only when the instruction is explicitly academic in nature (e.g. "find papers on", "peer-reviewed research on", "academic study of")
+   - If `arxiv_search` returns nothing clearly relevant after 1–2 queries → switch to `web_search` immediately
+
+2. **Search, then fetch:** After getting results, call `fetch_page` on the 2–3 most promising URLs to get full content.
+
+3. **Don't loop on the same query:** If a search returns results but nothing useful, change your approach — try different keywords or a different tool. Never repeat the same search query.
+
+4. **Synthesize:** Write a clear, well-structured markdown summary that directly addresses the instruction. Cite sources with inline links.
+
+## Output format
+
+Return a markdown response with:
+- A direct answer or summary addressing the instruction
+- Key findings or takeaways
+- Sources cited inline (e.g. `[Title](url)`)
+
+Keep the response focused and concise — avoid padding or restating the question.

gateway/src/harness/workflows/index.ts

@@ -9,11 +9,6 @@ export {
   type WorkflowEdgeCondition,
 } from './base-workflow.js';
 
-export {
-  StrategyValidationWorkflow,
-  createStrategyValidationWorkflow,
-} from './strategy-validation/graph.js';
-
 export {
   TradingRequestWorkflow,
   createTradingRequestWorkflow,

gateway/src/harness/workflows/strategy-validation/config.yaml

@@ -1,19 +0,0 @@
-# Strategy Validation Workflow Configuration
-
-name: strategy-validation
-description: Validates trading strategies with code review, backtest, and risk assessment
-
-# Workflow settings
-timeout: 300000  # 5 minutes
-maxRetries: 3
-requiresApproval: true
-approvalNodes:
-  - human_approval
-
-# Validation loop settings
-maxValidationRetries: 3  # Max times to retry fixing errors
-minBacktestScore: 0.5    # Minimum Sharpe ratio to pass
-
-# Model override (optional)
-model: claude-sonnet-4-6
-temperature: 0.3

gateway/src/harness/workflows/strategy-validation/graph.ts

@@ -1,138 +0,0 @@
-import { StateGraph } from '@langchain/langgraph';
-import { BaseWorkflow, type WorkflowConfig } from '../base-workflow.js';
-import { StrategyValidationState, type StrategyValidationStateType } from './state.js';
-import {
-  createCodeReviewNode,
-  createFixCodeNode,
-  createBacktestNode,
-  createRiskAssessmentNode,
-  createHumanApprovalNode,
-  createRecommendationNode,
-} from './nodes.js';
-import type { FastifyBaseLogger } from 'fastify';
-import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
-import type { CodeReviewerSubagent } from '../../subagents/code-reviewer/index.js';
-
-/**
- * Strategy Validation Workflow
- *
- * Multi-step workflow with validation loop:
- * 1. Code Review (using CodeReviewerSubagent)
- * 2. If issues found → Fix Code → Loop back to Code Review
- * 3. Backtest (using user's MCP server)
- * 4. If backtest fails → Fix Code → Loop back to Code Review
- * 5. Risk Assessment
- * 6. Human Approval (pause for user input)
- * 7. Final Recommendation
- *
- * Features:
- * - Validation loop with max retries
- * - Human-in-the-loop approval gate
- * - Multi-file memory from CodeReviewerSubagent
- * - Comprehensive state tracking
- */
-export class StrategyValidationWorkflow extends BaseWorkflow<StrategyValidationStateType> {
-  constructor(
-    config: WorkflowConfig,
-    private model: BaseChatModel,
-    private codeReviewer: CodeReviewerSubagent,
-    private mcpBacktestFn: (code: string, ticker: string, timeframe: string) => Promise<Record<string, unknown>>,
-    logger: FastifyBaseLogger
-  ) {
-    super(config, logger);
-  }
-
-  buildGraph(): any {
-    const graph = new StateGraph(StrategyValidationState);
-
-    // Create nodes
-    const codeReviewNode = createCodeReviewNode(this.codeReviewer, this.logger);
-    const fixCodeNode = createFixCodeNode(this.model, this.logger);
-    const backtestNode = createBacktestNode(this.mcpBacktestFn, this.logger);
-    const riskAssessmentNode = createRiskAssessmentNode(this.model, this.logger);
-    const humanApprovalNode = createHumanApprovalNode(this.logger);
-    const recommendationNode = createRecommendationNode(this.model, this.logger);
-
-    // Add nodes to graph
-    graph
-      .addNode('code_review', codeReviewNode)
-      .addNode('fix_code', fixCodeNode)
-      .addNode('backtest', backtestNode)
-      .addNode('risk_assessment', riskAssessmentNode)
-      .addNode('human_approval', humanApprovalNode)
-      .addNode('recommendation', recommendationNode);
-
-    // Define edges
-    (graph as any).addEdge('__start__', 'code_review');
-
-    // Conditional: After code review, fix if needed or proceed to backtest
-    (graph as any).addConditionalEdges('code_review', (state: any) => {
-      if (state.needsFixing && state.validationRetryCount < 3) {
-        return 'fix_code';
-      }
-      if (state.needsFixing && state.validationRetryCount >= 3) {
-        return 'recommendation'; // Give up, generate rejection
-      }
-      return 'backtest';
-    });
-
-    // After fixing code, loop back to code review
-    (graph as any).addEdge('fix_code', 'code_review');
-
-    // Conditional: After backtest, fix if failed or proceed to risk assessment
-    (graph as any).addConditionalEdges('backtest', (state: any) => {
-      if (!state.backtestPassed && state.validationRetryCount < 3) {
-        return 'fix_code';
-      }
-      if (!state.backtestPassed && state.validationRetryCount >= 3) {
-        return 'recommendation'; // Give up
-      }
-      return 'risk_assessment';
-    });
-
-    // After risk assessment, go to human approval
-    (graph as any).addEdge('risk_assessment', 'human_approval');
-
-    // Conditional: After human approval, proceed to recommendation or reject
-    (graph as any).addConditionalEdges('human_approval', (state: any) => {
-      return state.humanApproved ? 'recommendation' : '__end__';
-    });
-
-    // Final recommendation is terminal
-    (graph as any).addEdge('recommendation', '__end__');
-
-    return graph;
-  }
-}
-
-/**
- * Factory function to create and compile workflow
- */
-export async function createStrategyValidationWorkflow(
-  model: BaseChatModel,
-  codeReviewer: CodeReviewerSubagent,
-  mcpBacktestFn: (code: string, ticker: string, timeframe: string) => Promise<Record<string, unknown>>,
-  logger: FastifyBaseLogger,
-  configPath: string
-): Promise<StrategyValidationWorkflow> {
-  const { readFile } = await import('fs/promises');
-  const yaml = await import('js-yaml');
-
-  // Load config
-  const configContent = await readFile(configPath, 'utf-8');
-  const config = yaml.load(configContent) as WorkflowConfig;
-
-  // Create workflow
-  const workflow = new StrategyValidationWorkflow(
-    config,
-    model,
-    codeReviewer,
-    mcpBacktestFn,
-    logger
-  );
-
-  // Compile graph
-  workflow.compile();
-
-  return workflow;
-}

gateway/src/harness/workflows/strategy-validation/nodes.ts

@@ -1,233 +0,0 @@
-import type { StrategyValidationStateType } from './state.js';
-import type { FastifyBaseLogger } from 'fastify';
-import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
-import type { CodeReviewerSubagent } from '../../subagents/code-reviewer/index.js';
-import { HumanMessage, SystemMessage } from '@langchain/core/messages';
-
-/**
- * Node: Code Review
- * Reviews strategy code using CodeReviewerSubagent
- */
-export function createCodeReviewNode(
-  codeReviewer: CodeReviewerSubagent,
-  logger: FastifyBaseLogger
-) {
-  return async (state: StrategyValidationStateType): Promise<Partial<StrategyValidationStateType>> => {
-    logger.info('Strategy validation: Code review');
-
-    const review = await codeReviewer.execute(
-      { userContext: state.userContext },
-      state.strategyCode
-    );
-
-    // Simple issue detection (in production, parse structured output)
-    const hasIssues = review.toLowerCase().includes('critical') ||
-                     review.toLowerCase().includes('reject');
-
-    return {
-      codeReview: review,
-      codeIssues: hasIssues ? ['Issues detected in code review'] : [],
-      needsFixing: hasIssues,
-    };
-  };
-}
-
-/**
- * Node: Fix Code Issues
- * Uses LLM to fix issues identified in code review
- */
-export function createFixCodeNode(
-  model: BaseChatModel,
-  logger: FastifyBaseLogger
-) {
-  return async (state: StrategyValidationStateType): Promise<Partial<StrategyValidationStateType>> => {
-    logger.info('Strategy validation: Fixing code issues');
-
-    const systemPrompt = `You are a trading strategy developer.
-Fix the issues identified in the code review while maintaining the strategy's logic.
-Return only the corrected code without explanation.`;
-
-    const userPrompt = `Original code:
-\`\`\`typescript
-${state.strategyCode}
-\`\`\`
-
-Code review feedback:
-${state.codeReview}
-
-Provide the corrected code:`;
-
-    const response = await model.invoke([
-      new SystemMessage(systemPrompt),
-      new HumanMessage(userPrompt),
-    ]);
-
-    const fixedCode = (response.content as string)
-      .replace(/```typescript\n?/g, '')
-      .replace(/```\n?/g, '')
-      .trim();
-
-    return {
-      strategyCode: fixedCode,
-      validationRetryCount: state.validationRetryCount + 1,
-    };
-  };
-}
-
-/**
- * Node: Backtest Strategy
- * Runs backtest using user's MCP server
- */
-export function createBacktestNode(
-  mcpBacktestFn: (code: string, ticker: string, timeframe: string) => Promise<Record<string, unknown>>,
-  logger: FastifyBaseLogger
-) {
-  return async (state: StrategyValidationStateType): Promise<Partial<StrategyValidationStateType>> => {
-    logger.info('Strategy validation: Running backtest');
-
-    try {
-      const results = await mcpBacktestFn(
-        state.strategyCode,
-        state.ticker,
-        state.timeframe
-      );
-
-      // Check if backtest passed (simplified)
-      const sharpeRatio = (results.sharpeRatio as number) || 0;
-      const passed = sharpeRatio > 0.5;
-
-      return {
-        backtestResults: results,
-        backtestPassed: passed,
-        needsFixing: !passed,
-      };
-    } catch (error) {
-      logger.error({ error }, 'Backtest failed');
-      return {
-        backtestResults: { error: (error as Error).message },
-        backtestPassed: false,
-        needsFixing: true,
-      };
-    }
-  };
-}
-
-/**
- * Node: Risk Assessment
- * Analyzes backtest results for risk
- */
-export function createRiskAssessmentNode(
-  model: BaseChatModel,
-  logger: FastifyBaseLogger
-) {
-  return async (state: StrategyValidationStateType): Promise<Partial<StrategyValidationStateType>> => {
-    logger.info('Strategy validation: Risk assessment');
-
-    const systemPrompt = `You are a risk management expert.
-Analyze the strategy and backtest results to assess risk level.
-Provide: risk level (low/medium/high) and detailed assessment.`;
-
-    const userPrompt = `Strategy code:
-\`\`\`typescript
-${state.strategyCode}
-\`\`\`
-
-Backtest results:
-${JSON.stringify(state.backtestResults, null, 2)}
-
-Provide risk assessment in format:
-RISK_LEVEL: [low/medium/high]
-ASSESSMENT: [detailed explanation]`;
-
-    const response = await model.invoke([
-      new SystemMessage(systemPrompt),
-      new HumanMessage(userPrompt),
-    ]);
-
-    const assessment = response.content as string;
-
-    // Parse risk level (simplified)
-    let riskLevel: 'low' | 'medium' | 'high' = 'medium';
-    if (assessment.includes('RISK_LEVEL: low')) riskLevel = 'low';
-    if (assessment.includes('RISK_LEVEL: high')) riskLevel = 'high';
-
-    return {
-      riskAssessment: assessment,
-      riskLevel,
-    };
-  };
-}
-
-/**
- * Node: Human Approval
- * Pauses workflow for human review
- */
-export function createHumanApprovalNode(logger: FastifyBaseLogger) {
-  return async (state: StrategyValidationStateType): Promise<Partial<StrategyValidationStateType>> => {
-    logger.info('Strategy validation: Awaiting human approval');
-
-    // In real implementation, this would:
-    // 1. Send approval request to user's channel
-    // 2. Store workflow state with interrupt
-    // 3. Wait for user response
-    // 4. Resume with approval decision
-
-    // For now, auto-approve if risk is low/medium and backtest passed
-    const autoApprove = state.backtestPassed &&
-                       (state.riskLevel === 'low' || state.riskLevel === 'medium');
-
-    return {
-      humanApproved: autoApprove,
-      approvalComment: autoApprove ? 'Auto-approved: passed validation' : 'Needs manual review',
-    };
-  };
-}
-
-/**
- * Node: Final Recommendation
- * Generates final recommendation based on all steps
- */
-export function createRecommendationNode(
-  model: BaseChatModel,
-  logger: FastifyBaseLogger
-) {
-  return async (state: StrategyValidationStateType): Promise<Partial<StrategyValidationStateType>> => {
-    logger.info('Strategy validation: Generating recommendation');
-
-    const systemPrompt = `You are the final decision maker for strategy deployment.
-Based on all validation steps, provide a clear recommendation: approve, reject, or revise.`;
-
-    const userPrompt = `Strategy validation summary:
-
-Code Review: ${state.codeIssues.length === 0 ? 'Passed' : 'Issues found'}
-Backtest: ${state.backtestPassed ? 'Passed' : 'Failed'}
-Risk Level: ${state.riskLevel}
-Human Approved: ${state.humanApproved}
-
-Backtest Results:
-${JSON.stringify(state.backtestResults, null, 2)}
-
-Risk Assessment:
-${state.riskAssessment}
-
-Provide final recommendation (approve/reject/revise) and reasoning:`;
-
-    const response = await model.invoke([
-      new SystemMessage(systemPrompt),
-      new HumanMessage(userPrompt),
-    ]);
-
-    const recommendation = response.content as string;
-
-    // Parse recommendation (simplified)
-    let decision: 'approve' | 'reject' | 'revise' = 'revise';
-    if (recommendation.toLowerCase().includes('approve')) decision = 'approve';
-    if (recommendation.toLowerCase().includes('reject')) decision = 'reject';
-
-    return {
-      recommendation: decision,
-      recommendationReason: recommendation,
-      output: recommendation,
-    };
-  };
-}

gateway/src/harness/workflows/strategy-validation/state.ts

@@ -1,78 +0,0 @@
-import { Annotation } from '@langchain/langgraph';
-import { BaseWorkflowState } from '../base-workflow.js';
-
-/**
- * Strategy validation workflow state
- *
- * Extends base workflow state with strategy-specific fields
- */
-export const StrategyValidationState = Annotation.Root({
-  ...BaseWorkflowState.spec,
-
-  // Input
-  strategyCode: Annotation<string>(),
-  ticker: Annotation<string>(),
-  timeframe: Annotation<string>(),
-
-  // Code review step
-  codeReview: Annotation<string | null>({
-    value: (left, right) => right ?? left,
-    default: () => null,
-  }),
-  codeIssues: Annotation<string[]>({
-    value: (left, right) => right ?? left,
-    default: () => [],
-  }),
-
-  // Backtest step
-  backtestResults: Annotation<Record<string, unknown> | null>({
-    value: (left, right) => right ?? left,
-    default: () => null,
-  }),
-  backtestPassed: Annotation<boolean>({
-    value: (left, right) => right ?? left,
-    default: () => false,
-  }),
-
-  // Risk assessment step
-  riskAssessment: Annotation<string | null>({
-    value: (left, right) => right ?? left,
-    default: () => null,
-  }),
-  riskLevel: Annotation<'low' | 'medium' | 'high' | null>({
-    value: (left, right) => right ?? left,
-    default: () => null,
-  }),
-
-  // Human approval step
-  humanApproved: Annotation<boolean>({
-    value: (left, right) => right ?? left,
-    default: () => false,
-  }),
-  approvalComment: Annotation<string | null>({
-    value: (left, right) => right ?? left,
-    default: () => null,
-  }),
-
-  // Validation loop control
-  validationRetryCount: Annotation<number>({
-    value: (left, right) => right ?? left,
-    default: () => 0,
-  }),
-  needsFixing: Annotation<boolean>({
-    value: (left, right) => right ?? left,
-    default: () => false,
-  }),
-
-  // Final output
-  recommendation: Annotation<'approve' | 'reject' | 'revise' | null>({
-    value: (left, right) => right ?? left,
-    default: () => null,
-  }),
-  recommendationReason: Annotation<string | null>({
-    value: (left, right) => right ?? left,
-    default: () => null,
-  }),
-});
-
-export type StrategyValidationStateType = typeof StrategyValidationState.State;

gateway/src/llm/provider.ts

@@ -1,5 +1,5 @@
 import type { BaseChatModel } from '@langchain/core/language_models/chat_models';
-import { ChatAnthropic } from '@langchain/anthropic';
+import { ChatOpenAI } from '@langchain/openai';
 import type { FastifyBaseLogger } from 'fastify';
 import { type ModelMiddleware, NoopMiddleware, AnthropicCachingMiddleware } from './middleware.js';
 
@@ -10,7 +10,7 @@ export { NoopMiddleware, AnthropicCachingMiddleware };
  * Supported LLM providers
  */
 export enum LLMProvider {
-  ANTHROPIC = 'anthropic',
+  DEEP_INFRA = 'deepinfra',
 }
 
 /**
@@ -47,11 +47,13 @@ export interface LicenseModelsConfig {
  * Provider configuration with API keys
  */
 export interface ProviderConfig {
-  anthropicApiKey?: string;
+  deepinfraApiKey?: string;
   defaultModel?: ModelConfig;
   licenseModels?: LicenseModelsConfig;
 }
 
+const DEEP_INFRA_BASE_URL = 'https://api.deepinfra.com/v1/openai';
+
 /**
  * LLM Provider factory
  * Creates model instances with unified interface across providers
@@ -75,8 +77,8 @@ export class LLMProviderFactory {
     );
 
     switch (modelConfig.provider) {
-      case LLMProvider.ANTHROPIC:
-        return this.createAnthropicModel(modelConfig);
+      case LLMProvider.DEEP_INFRA:
+        return this.createDeepInfraModel(modelConfig);
 
       default:
         throw new Error(`Unsupported provider: ${modelConfig.provider}`);
@@ -84,22 +86,24 @@ export class LLMProviderFactory {
   }
 
   /**
-   * Create Anthropic Claude model
+   * Create Deep Infra model via OpenAI-compatible API
    */
-  private createAnthropicModel(config: ModelConfig): { model: ChatAnthropic; middleware: AnthropicCachingMiddleware } {
-    if (!this.config.anthropicApiKey) {
-      throw new Error('Anthropic API key not configured');
+  private createDeepInfraModel(config: ModelConfig): { model: ChatOpenAI; middleware: NoopMiddleware } {
+    if (!this.config.deepinfraApiKey) {
+      throw new Error('Deep Infra API key not configured');
     }
 
-    const model = new ChatAnthropic({
+    const model = new ChatOpenAI({
       model: config.model,
       temperature: config.temperature ?? 0.7,
       maxTokens: config.maxTokens ?? 4096,
-      anthropicApiKey: this.config.anthropicApiKey,
-      clientOptions: { defaultHeaders: { 'anthropic-beta': 'prompt-caching-2024-07-31' } },
+      apiKey: this.config.deepinfraApiKey,
+      configuration: {
+        baseURL: DEEP_INFRA_BASE_URL,
+      },
     });
 
-    return { model, middleware: new AnthropicCachingMiddleware() };
+    return { model, middleware: new NoopMiddleware() };
   }
 
   /**
@@ -110,13 +114,13 @@ export class LLMProviderFactory {
       return this.config.defaultModel;
     }
 
-    if (!this.config.anthropicApiKey) {
-      throw new Error('Anthropic API key not configured');
+    if (!this.config.deepinfraApiKey) {
+      throw new Error('Deep Infra API key not configured');
     }
 
     return {
-      provider: LLMProvider.ANTHROPIC,
-      model: 'claude-sonnet-4-6',
+      provider: LLMProvider.DEEP_INFRA,
+      model: 'zai-org/GLM-5',
     };
   }
 
@@ -132,16 +136,12 @@ export class LLMProviderFactory {
  * Predefined model configurations
  */
 export const MODELS = {
-  CLAUDE_SONNET: {
-    provider: LLMProvider.ANTHROPIC,
-    model: 'claude-sonnet-4-6',
+  GLM_5: {
+    provider: LLMProvider.DEEP_INFRA,
+    model: 'zai-org/GLM-5',
   },
-  CLAUDE_HAIKU: {
-    provider: LLMProvider.ANTHROPIC,
-    model: 'claude-haiku-4-5-20251001',
-  },
-  CLAUDE_OPUS: {
-    provider: LLMProvider.ANTHROPIC,
-    model: 'claude-opus-4-6',
+  QWEN_235B: {
+    provider: LLMProvider.DEEP_INFRA,
+    model: 'Qwen/Qwen3-235B-A22B-Instruct-2507',
   },
 } as const satisfies Record<string, ModelConfig>;

gateway/src/llm/router.ts

@@ -113,17 +113,17 @@ export class ModelRouter {
     // Fallback to hardcoded defaults
     if (license.licenseType === 'enterprise') {
       return isComplex
-        ? { provider: LLMProvider.ANTHROPIC, model: 'claude-opus-4-6' }
-        : { provider: LLMProvider.ANTHROPIC, model: 'claude-sonnet-4-6' };
+        ? { provider: LLMProvider.DEEP_INFRA, model: 'Qwen/Qwen3-235B-A22B-Instruct-2507' }
+        : { provider: LLMProvider.DEEP_INFRA, model: 'zai-org/GLM-5' };
     }
 
     if (license.licenseType === 'pro') {
       return isComplex
-        ? { provider: LLMProvider.ANTHROPIC, model: 'claude-sonnet-4-6' }
-        : { provider: LLMProvider.ANTHROPIC, model: 'claude-haiku-4-5-20251001' };
+        ? { provider: LLMProvider.DEEP_INFRA, model: 'zai-org/GLM-5' }
+        : { provider: LLMProvider.DEEP_INFRA, model: 'zai-org/GLM-5' };
     }
 
-    return { provider: LLMProvider.ANTHROPIC, model: 'claude-haiku-4-5-20251001' };
+    return { provider: LLMProvider.DEEP_INFRA, model: 'zai-org/GLM-5' };
   }
 
   /**
@@ -141,13 +141,13 @@ export class ModelRouter {
     // Fallback to hardcoded defaults
     switch (license.licenseType) {
       case 'enterprise':
-        return { provider: LLMProvider.ANTHROPIC, model: 'claude-sonnet-4-6' };
+        return { provider: LLMProvider.DEEP_INFRA, model: 'zai-org/GLM-5' };
 
       case 'pro':
-        return { provider: LLMProvider.ANTHROPIC, model: 'claude-sonnet-4-6' };
+        return { provider: LLMProvider.DEEP_INFRA, model: 'zai-org/GLM-5' };
 
       case 'free':
-        return { provider: LLMProvider.ANTHROPIC, model: 'claude-haiku-4-5-20251001' };
+        return { provider: LLMProvider.DEEP_INFRA, model: 'zai-org/GLM-5' };
 
       default:
         return this.defaultModel;
@@ -166,8 +166,8 @@ export class ModelRouter {
       }
     }
 
-    // Fallback: use Haiku for cost efficiency
-    return { provider: LLMProvider.ANTHROPIC, model: 'claude-haiku-4-5-20251001' };
+    // Fallback: use GLM-5
+    return { provider: LLMProvider.DEEP_INFRA, model: 'zai-org/GLM-5' };
   }
 
   /**
@@ -195,12 +195,12 @@ export class ModelRouter {
 
     // Fallback to hardcoded defaults
     if (license.licenseType === 'free') {
-      const allowedModels = ['claude-haiku-4-5-20251001'];
+      const allowedModels = ['zai-org/GLM-5'];
       return allowedModels.includes(model.model);
     }
 
     if (license.licenseType === 'pro') {
-      const blockedModels = ['claude-opus-4-6'];
+      const blockedModels = ['Qwen/Qwen3-235B-A22B-Instruct-2507'];
       return !blockedModels.includes(model.model);
     }
 

gateway/src/main.ts

@@ -90,31 +90,28 @@ function loadConfig() {
 
     // LLM provider API keys and model configuration
     providerConfig: {
-      anthropicApiKey: secretsData.llm_providers?.anthropic_api_key || process.env.ANTHROPIC_API_KEY,
-      openaiApiKey: secretsData.llm_providers?.openai_api_key || process.env.OPENAI_API_KEY,
-      googleApiKey: secretsData.llm_providers?.google_api_key || process.env.GOOGLE_API_KEY,
-      openrouterApiKey: secretsData.llm_providers?.openrouter_api_key || process.env.OPENROUTER_API_KEY,
+      deepinfraApiKey: secretsData.llm_providers?.deepinfra_api_key || process.env.DEEPINFRA_API_KEY,
       defaultModel: {
-        provider: configData.defaults?.model_provider || 'anthropic',
-        model: configData.defaults?.model || 'claude-sonnet-4-6',
+        provider: configData.defaults?.model_provider || 'deepinfra',
+        model: configData.defaults?.model || 'zai-org/GLM-5',
       },
       licenseModels: {
         free: {
-          default: configData.license_models?.free?.default || 'claude-haiku-4-5-20251001',
-          cost_optimized: configData.license_models?.free?.cost_optimized || 'claude-haiku-4-5-20251001',
-          complex: configData.license_models?.free?.complex || 'claude-haiku-4-5-20251001',
-          allowed_models: configData.license_models?.free?.allowed_models || ['claude-haiku-4-5-20251001'],
+          default: configData.license_models?.free?.default || 'zai-org/GLM-5',
+          cost_optimized: configData.license_models?.free?.cost_optimized || 'zai-org/GLM-5',
+          complex: configData.license_models?.free?.complex || 'zai-org/GLM-5',
+          allowed_models: configData.license_models?.free?.allowed_models || ['zai-org/GLM-5'],
         },
         pro: {
-          default: configData.license_models?.pro?.default || 'claude-sonnet-4-6',
-          cost_optimized: configData.license_models?.pro?.cost_optimized || 'claude-haiku-4-5-20251001',
-          complex: configData.license_models?.pro?.complex || 'claude-sonnet-4-6',
-          blocked_models: configData.license_models?.pro?.blocked_models || ['claude-opus-4-6'],
+          default: configData.license_models?.pro?.default || 'zai-org/GLM-5',
+          cost_optimized: configData.license_models?.pro?.cost_optimized || 'zai-org/GLM-5',
+          complex: configData.license_models?.pro?.complex || 'zai-org/GLM-5',
+          blocked_models: configData.license_models?.pro?.blocked_models || ['Qwen/Qwen3-235B-A22B-Instruct-2507'],
         },
         enterprise: {
-          default: configData.license_models?.enterprise?.default || 'claude-sonnet-4-6',
-          cost_optimized: configData.license_models?.enterprise?.cost_optimized || 'claude-haiku-4-5-20251001',
-          complex: configData.license_models?.enterprise?.complex || 'claude-opus-4-6',
+          default: configData.license_models?.enterprise?.default || 'zai-org/GLM-5',
+          cost_optimized: configData.license_models?.enterprise?.cost_optimized || 'zai-org/GLM-5',
+          complex: configData.license_models?.enterprise?.complex || 'Qwen/Qwen3-235B-A22B-Instruct-2507',
         },
       },
     },
@@ -181,6 +178,9 @@ function loadConfig() {
       storageClass: configData.kubernetes?.storage_class || process.env.SANDBOX_STORAGE_CLASS || '',
       imagePullPolicy: configData.kubernetes?.image_pull_policy || process.env.IMAGE_PULL_POLICY || 'Always',
     },
+
+    // Search API keys
+    tavilyApiKey: secretsData.search?.tavily_api_key || process.env.TAVILY_API_KEY,
   };
 }
 
@@ -200,10 +200,9 @@ const app = Fastify({
   },
 });
 
-// Validate at least one LLM provider is configured
-const hasAnyProvider = Object.values(config.providerConfig).some(key => !!key);
-if (!hasAnyProvider) {
-  app.log.error('At least one LLM provider API key is required (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, or OPENROUTER_API_KEY)');
+// Validate LLM provider is configured
+if (!config.providerConfig.deepinfraApiKey) {
+  app.log.error('DEEPINFRA_API_KEY is required');
   process.exit(1);
 }
 
@@ -407,6 +406,8 @@ app.log.debug('Initializing auth routes...');
 const authRoutes = new AuthRoutes({
   authService,
   betterAuth,
+  containerManager,
+  userService,
 });
 
 // Register routes
@@ -581,6 +582,7 @@ try {
       ohlcService: () => ohlcService,
       symbolIndexService: () => symbolIndexService,
       workspaceManager: undefined, // Will be set per-session
+      tavilyApiKey: config.tavilyApiKey,
     });
 
     // Register agent tool configurations
@@ -588,20 +590,27 @@ try {
     toolRegistry.registerAgentTools({
       agentName: 'main',
       platformTools: ['symbol_lookup', 'get_chart_data'],
-      mcpTools: ['category_list'], // category_list lets the main agent see existing research scripts
+      mcpTools: ['python_list', 'backtest_strategy', 'list_active_strategies'],
     });
 
     // Research subagent: only MCP tools for script creation/execution
     toolRegistry.registerAgentTools({
       agentName: 'research',
       platformTools: [], // No platform tools (works at script level)
-      mcpTools: ['category_*', 'execute_research'],
+      mcpTools: ['python_*', 'execute_research'],
     });
 
-    // Code reviewer subagent: no tools by default
+    // Indicator subagent: workspace patch + category tools + evaluate_indicator
     toolRegistry.registerAgentTools({
-      agentName: 'code-reviewer',
+      agentName: 'indicator',
       platformTools: [],
+      mcpTools: ['workspace_read', 'workspace_patch', 'python_*', 'evaluate_indicator'],
+    });
+
+    // Web explore subagent: platform search/fetch tools only (no MCP needed)
+    toolRegistry.registerAgentTools({
+      agentName: 'web-explore',
+      platformTools: ['web_search', 'fetch_page', 'arxiv_search'],
       mcpTools: [],
     });
 

gateway/src/routes/auth-routes.ts

@@ -1,10 +1,14 @@
 import type { FastifyInstance, FastifyRequest, FastifyReply } from 'fastify';
 import type { AuthService } from '../auth/auth-service.js';
 import type { BetterAuthInstance } from '../auth/better-auth-config.js';
+import type { ContainerManager } from '../k8s/container-manager.js';
+import type { UserService } from '../db/user-service.js';
 
 export interface AuthRoutesConfig {
   authService: AuthService;
   betterAuth: BetterAuthInstance;
+  containerManager: ContainerManager;
+  userService: UserService;
 }
 
 /**
@@ -74,6 +78,14 @@ export class AuthRoutes {
         // Ensure user has a license
         await this.config.authService.ensureUserLicense(result.userId, email);
 
+        // Warm up the sandbox container so it's likely ready by first login
+        this.config.userService.getUserLicense(result.userId).then((license) => {
+          if (license) {
+            this.config.containerManager.ensureContainerRunning(result.userId, license.license, false)
+              .catch((err) => app.log.warn({ err, userId: result.userId }, 'Container warmup on registration failed'));
+          }
+        }).catch((err) => app.log.warn({ err, userId: result.userId }, 'Failed to fetch license for container warmup'));
+
         // Auto sign in after registration
         const signInResult = await this.config.authService.signIn(email, password);
 

gateway/src/services/ohlc-service.ts

@@ -104,34 +104,41 @@ export class OHLCService {
       end_time
     );
 
+    this.logger.info({ ticker, period_seconds, dataCount: data.length, missingRangeCount: missingRanges.length, missingRanges }, 'OHLC cache check result');
+
     if (missingRanges.length === 0 && data.length > 0) {
       // All data exists in Iceberg
-      this.logger.debug({ ticker, period_seconds, cached: true }, 'OHLC data found in cache');
+      this.logger.info({ ticker, period_seconds, cached: true }, 'OHLC data found in cache, returning immediately');
       return this.formatHistoryResult(data, start_time, end_time, period_seconds, countback);
     }
 
-    // Step 3: Request missing data via relay
-    this.logger.debug({ ticker, period_seconds, missingRanges: missingRanges.length }, 'Requesting missing OHLC data');
+    // Step 3: Request each missing range from the relay individually so we
+    // only fetch what's actually absent, not the whole requested window.
+    this.logger.info({ ticker, period_seconds, missingRanges: missingRanges.length, dataCount: data.length }, 'Requesting missing OHLC data from relay');
 
     try {
-      const notification = await this.relayClient.requestHistoricalOHLC(
-        ticker,
-        period_seconds,
-        start_time,
-        end_time
-        // countback is NOT passed as a limit — the ingestor must fetch the full range.
-        // Countback is applied below after we have the complete dataset.
-      );
-
-      this.logger.info({
-        ticker,
-        period_seconds,
-        row_count: notification.row_count,
-        status: notification.status,
-      }, 'Historical data request completed');
+      for (const [rangeStart, rangeEnd] of missingRanges) {
+        const notification = await this.relayClient.requestHistoricalOHLC(
+          ticker,
+          period_seconds,
+          rangeStart,
+          rangeEnd
+          // countback is NOT passed as a limit — the ingestor must fetch the full range.
+          // Countback is applied below after we have the complete dataset.
+        );
+        this.logger.info({
+          ticker,
+          period_seconds,
+          rangeStart: rangeStart.toString(),
+          rangeEnd: rangeEnd.toString(),
+          row_count: notification.row_count,
+          status: notification.status,
+        }, 'Relay range request completed');
+      }
 
       // Step 4: Query Iceberg again for complete dataset
       data = await this.icebergClient.queryOHLC(ticker, period_seconds, start_time, end_time);
+      this.logger.info({ ticker, period_seconds, dataCount: data.length }, 'Final Iceberg query complete, returning result');
 
       return this.formatHistoryResult(data, start_time, end_time, period_seconds, countback);
 

gateway/src/tools/mcp/mcp-tool-wrapper.ts

@@ -27,7 +27,8 @@ export function createMCPToolWrapper(
   toolInfo: MCPToolInfo,
   mcpClient: MCPClientConnector,
   logger: FastifyBaseLogger,
-  onImage?: (image: { data: string; mimeType: string }) => void
+  onImage?: (image: { data: string; mimeType: string }) => void,
+  onWorkspaceMutation?: (storeName: string, newState: unknown) => void
 ): DynamicStructuredTool {
   // Convert MCP input schema to Zod schema
   const zodSchema = mcpInputSchemaToZod(toolInfo.inputSchema);
@@ -42,6 +43,28 @@ export function createMCPToolWrapper(
 
         logger.info({ tool: toolInfo.name }, 'MCP tool call completed');
 
+        // Fire workspace mutation callback when workspace_patch or workspace_write succeeds.
+        // The sandbox returns {"success": true, "data": <newState>} as a text content item.
+        if (
+          onWorkspaceMutation &&
+          (toolInfo.name === 'workspace_patch' || toolInfo.name === 'workspace_write')
+        ) {
+          const content = (result as any)?.content;
+          if (Array.isArray(content)) {
+            for (const item of content) {
+              if (item.type === 'text' && item.text) {
+                try {
+                  const parsed = JSON.parse(item.text);
+                  if (parsed?.success && parsed?.data !== undefined) {
+                    onWorkspaceMutation((input as any).store_name as string, parsed.data);
+                  }
+                } catch { /* ignore parse errors */ }
+                break; // only need first text item
+              }
+            }
+          }
+        }
+
         // Handle different MCP result formats
         if (typeof result === 'string') {
           return result;
@@ -180,7 +203,10 @@ export function createMCPToolWrappers(
   toolInfos: MCPToolInfo[],
   mcpClient: MCPClientConnector,
   logger: FastifyBaseLogger,
-  onImage?: (image: { data: string; mimeType: string }) => void
+  onImage?: (image: { data: string; mimeType: string }) => void,
+  onWorkspaceMutation?: (storeName: string, newState: unknown) => void
 ): DynamicStructuredTool[] {
-  return toolInfos.map(toolInfo => createMCPToolWrapper(toolInfo, mcpClient, logger, onImage));
+  return toolInfos.map(toolInfo =>
+    createMCPToolWrapper(toolInfo, mcpClient, logger, onImage, onWorkspaceMutation)
+  );
 }

gateway/src/tools/platform/arxiv-search.tool.ts

@@ -0,0 +1,65 @@
+import { DynamicStructuredTool } from '@langchain/core/tools';
+import { z } from 'zod';
+import type { FastifyBaseLogger } from 'fastify';
+
+/**
+ * ArXiv Search Tool
+ *
+ * Searches arXiv for academic papers using the LangChain ArxivRetriever.
+ * Free, no API key required.
+ */
+
+export interface ArxivSearchToolConfig {
+  logger: FastifyBaseLogger;
+}
+
+export function createArxivSearchTool(config: ArxivSearchToolConfig): DynamicStructuredTool {
+  const { logger } = config;
+
+  return new DynamicStructuredTool({
+    name: 'arxiv_search',
+    description: 'Search arXiv for academic papers. Returns titles, authors, abstracts, and PDF links. Use this for scientific or technical research queries instead of web_search.',
+    schema: z.object({
+      query: z.string().describe('The research query'),
+      max_results: z.number().optional().default(5).describe('Maximum number of papers to return (default: 5)'),
+    }),
+    func: async ({ query, max_results }) => {
+      logger.debug({ query, max_results }, 'Executing arxiv_search tool');
+
+      try {
+        const { ArxivRetriever } = await import('@langchain/community/retrievers/arxiv');
+
+        const retriever = new ArxivRetriever({
+          getFullDocuments: false,
+          maxSearchResults: max_results,
+        });
+
+        const docs = await retriever.invoke(query);
+
+        const results = docs.map(doc => {
+          const meta = doc.metadata as Record<string, any>;
+          // Derive PDF URL from abstract URL: /abs/ID -> /pdf/ID
+          const pdfUrl = typeof meta.url === 'string'
+            ? meta.url.replace('/abs/', '/pdf/')
+            : undefined;
+
+          return {
+            title: meta.title,
+            authors: Array.isArray(meta.authors) ? meta.authors : [],
+            abstract: doc.pageContent,
+            published: meta.published,
+            url: meta.url,
+            pdf_url: pdfUrl,
+          };
+        });
+
+        logger.info({ query, resultCount: results.length }, 'arXiv search completed');
+
+        return JSON.stringify({ query, results });
+      } catch (error) {
+        logger.error({ error, query }, 'arxiv_search tool failed');
+        return JSON.stringify({ error: error instanceof Error ? error.message : String(error) });
+      }
+    },
+  });
+}

gateway/src/tools/platform/fetch-page.tool.ts

@@ -0,0 +1,80 @@
+import { DynamicStructuredTool } from '@langchain/core/tools';
+import { z } from 'zod';
+import type { FastifyBaseLogger } from 'fastify';
+
+const MAX_CONTENT_LENGTH = 50_000;
+
+/**
+ * Fetch Page Tool
+ *
+ * Fetches a URL and returns its content as text/markdown.
+ * - PDFs are converted to text using pdf-parse
+ * - HTML pages are scraped with cheerio
+ * - Output is truncated to 50k characters
+ */
+
+export interface FetchPageToolConfig {
+  logger: FastifyBaseLogger;
+}
+
+export function createFetchPageTool(config: FetchPageToolConfig): DynamicStructuredTool {
+  const { logger } = config;
+
+  return new DynamicStructuredTool({
+    name: 'fetch_page',
+    description: 'Fetch a web page or PDF and return its text content. PDFs are automatically converted to markdown. Use this after web_search or arxiv_search to read the full content of a result.',
+    schema: z.object({
+      url: z.string().url().describe('The URL to fetch'),
+    }),
+    func: async ({ url }) => {
+      logger.debug({ url }, 'Executing fetch_page tool');
+
+      try {
+        const response = await fetch(url, {
+          headers: { 'User-Agent': 'Mozilla/5.0 (compatible; research-agent/1.0)' },
+          signal: AbortSignal.timeout(30_000),
+        });
+
+        if (!response.ok) {
+          return JSON.stringify({ error: `HTTP ${response.status}: ${response.statusText}`, url });
+        }
+
+        const contentType = response.headers.get('content-type') ?? '';
+        const isPdf = contentType.includes('pdf') || url.toLowerCase().endsWith('.pdf');
+
+        let content: string;
+
+        if (isPdf) {
+          const buffer = Buffer.from(await response.arrayBuffer());
+          const { PDFParse } = await import('pdf-parse');
+          const arrayBuffer = buffer.buffer.slice(buffer.byteOffset, buffer.byteOffset + buffer.byteLength);
+          const parser = new PDFParse({ data: arrayBuffer });
+          const result = await parser.getText();
+          content = result.text;
+          logger.debug({ url, chars: content.length, pages: result.pages.length }, 'PDF text extracted');
+        } else {
+          const html = await response.text();
+          const { load } = await import('cheerio');
+          const $ = load(html);
+
+          // Remove non-content elements
+          $('script, style, nav, footer, header, aside, [role="navigation"]').remove();
+
+          // Prefer article/main content
+          const main = $('article, main, [role="main"]').first();
+          content = (main.length ? main : $('body')).text().replace(/\s{3,}/g, '\n\n').trim();
+
+          logger.debug({ url, chars: content.length }, 'HTML page scraped');
+        }
+
+        const truncated = content.length > MAX_CONTENT_LENGTH;
+        const output = truncated ? content.slice(0, MAX_CONTENT_LENGTH) + '\n\n[content truncated]' : content;
+
+        return JSON.stringify({ url, content: output, truncated });
+      } catch (error) {
+        logger.error({ error, url }, 'fetch_page tool failed');
+        return JSON.stringify({ error: error instanceof Error ? error.message : String(error), url });
+      }
+    },
+  });
+}

gateway/src/tools/platform/indicator-agent.tool.ts

@@ -0,0 +1,53 @@
+import { DynamicStructuredTool } from '@langchain/core/tools';
+import { z } from 'zod';
+import type { FastifyBaseLogger } from 'fastify';
+import type { IndicatorSubagent } from '../../harness/subagents/indicator/index.js';
+import type { SubagentContext } from '../../harness/subagents/base-subagent.js';
+
+export interface IndicatorAgentToolConfig {
+  indicatorSubagent: IndicatorSubagent;
+  context: SubagentContext;
+  logger: FastifyBaseLogger;
+}
+
+/**
+ * Creates a LangChain tool that delegates to the indicator subagent.
+ * Mirrors the pattern of research-agent.tool.ts.
+ */
+export function createIndicatorAgentTool(config: IndicatorAgentToolConfig): DynamicStructuredTool {
+  const { indicatorSubagent, context, logger } = config;
+
+  return new DynamicStructuredTool({
+    name: 'indicator',
+    description: `Delegate to the indicator subagent for all indicator-related tasks on the chart.
+
+Use this tool for:
+- Reading which indicators are currently on the chart and explaining what they show
+- Adding indicators to the chart ("show RSI", "add Bollinger Bands with std=1.5")
+- Modifying indicator parameters ("change MACD fast to 8", "set RSI length to 21")
+- Removing indicators ("remove all moving averages", "clear the volume indicators")
+- Toggling indicator visibility
+- Creating custom indicators using Python scripts
+- Recommending indicators for a given strategy or analysis goal
+
+ALWAYS use this tool for any request about the chart's indicators.
+NEVER modify the indicators workspace store directly.`,
+    schema: z.object({
+      instruction: z.string().describe(
+        'The indicator task to perform. Be specific about which indicators, parameters, ' +
+        'and what changes are needed. Include relevant context like the current symbol ' +
+        'if the user mentioned it.'
+      ),
+    }),
+    func: async ({ instruction }: { instruction: string }): Promise<string> => {
+      logger.info({ instruction: instruction.substring(0, 100) }, 'Delegating to indicator subagent');
+
+      try {
+        return await indicatorSubagent.execute(context, instruction);
+      } catch (error) {
+        logger.error({ error, errorMessage: (error as Error)?.message }, 'Indicator subagent failed');
+        throw error;
+      }
+    },
+  });
+}

gateway/src/tools/platform/web-explore-agent.tool.ts

@@ -0,0 +1,49 @@
+import { DynamicStructuredTool } from '@langchain/core/tools';
+import { z } from 'zod';
+import type { FastifyBaseLogger } from 'fastify';
+import type { WebExploreSubagent } from '../../harness/subagents/web-explore/index.js';
+import type { SubagentContext } from '../../harness/subagents/base-subagent.js';
+
+export interface WebExploreAgentToolConfig {
+  webExploreSubagent: WebExploreSubagent;
+  context: SubagentContext;
+  logger: FastifyBaseLogger;
+}
+
+/**
+ * Creates a LangChain tool that delegates to the web-explore subagent.
+ * The subagent decides whether to use web search or arXiv based on the instruction.
+ */
+export function createWebExploreAgentTool(config: WebExploreAgentToolConfig): DynamicStructuredTool {
+  const { webExploreSubagent, context, logger } = config;
+
+  return new DynamicStructuredTool({
+    name: 'web_explore',
+    description: `Search the web or academic databases and return a summarized answer.
+
+Use this tool when the user asks about:
+- Current events, news, or real-time information
+- Documentation, tutorials, or how-to guides
+- Academic papers, research findings, or scientific topics
+- Any topic that benefits from external sources
+
+The subagent will search the web (or arXiv for academic queries), fetch relevant content, and return a markdown summary with cited sources.`,
+    schema: z.object({
+      instruction: z.string().describe(
+        'What to search for and summarize. Be specific — include the topic, what aspects matter, ' +
+        'and any context that helps narrow the search (e.g. "recent papers on momentum factor in equities" ' +
+        'or "how to configure rate limiting in Fastify").'
+      ),
+    }),
+    func: async ({ instruction }: { instruction: string }): Promise<string> => {
+      logger.info({ instruction: instruction.substring(0, 100) }, 'Delegating to web-explore subagent');
+
+      try {
+        return await webExploreSubagent.execute(context, instruction);
+      } catch (error) {
+        logger.error({ error, errorMessage: (error as Error)?.message }, 'Web explore subagent failed');
+        throw error;
+      }
+    },
+  });
+}

gateway/src/tools/platform/web-search.tool.ts

@@ -0,0 +1,65 @@
+import { DynamicStructuredTool } from '@langchain/core/tools';
+import { z } from 'zod';
+import type { FastifyBaseLogger } from 'fastify';
+
+/**
+ * Web Search Tool
+ *
+ * Calls the Tavily REST API directly. The config interface is intentionally
+ * minimal so the underlying provider can be swapped without touching callers.
+ */
+
+export interface WebSearchToolConfig {
+  apiKey: string;
+  logger: FastifyBaseLogger;
+}
+
+export function createWebSearchTool(config: WebSearchToolConfig): DynamicStructuredTool {
+  const { apiKey, logger } = config;
+
+  return new DynamicStructuredTool({
+    name: 'web_search',
+    description: 'Search the web. Returns titles, URLs, and content summaries. Use this for general web searches. For academic/scientific papers, prefer arxiv_search instead.',
+    schema: z.object({
+      query: z.string().describe('The search query'),
+      max_results: z.number().optional().default(8).describe('Maximum number of results to return (default: 8)'),
+    }),
+    func: async ({ query, max_results }) => {
+      logger.debug({ query, max_results }, 'Executing web_search tool');
+
+      try {
+        const response = await fetch('https://api.tavily.com/search', {
+          method: 'POST',
+          headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify({
+            api_key: apiKey,
+            query,
+            max_results,
+            search_depth: 'basic',
+          }),
+          signal: AbortSignal.timeout(30_000),
+        });
+
+        if (!response.ok) {
+          const text = await response.text();
+          throw new Error(`Tavily API error ${response.status}: ${text}`);
+        }
+
+        const data = await response.json() as { results?: Array<{ title: string; url: string; content: string }> };
+
+        const items = (data.results ?? []).map(r => ({
+          title: r.title,
+          url: r.url,
+          snippet: r.content,
+        }));
+
+        logger.info({ query, resultCount: items.length }, 'Web search completed');
+
+        return JSON.stringify({ query, results: items });
+      } catch (error) {
+        logger.error({ error, query, errorMessage: error instanceof Error ? error.message : String(error) }, 'web_search tool failed');
+        return JSON.stringify({ error: error instanceof Error ? error.message : String(error) });
+      }
+    },
+  });
+}

gateway/src/tools/tool-registry.ts

@@ -6,6 +6,9 @@ import type { SymbolIndexService } from '../services/symbol-index-service.js';
 import type { WorkspaceManager } from '../workspace/workspace-manager.js';
 import { createSymbolLookupTool } from './platform/symbol-lookup.tool.js';
 import { createGetChartDataTool } from './platform/get-chart-data.tool.js';
+import { createWebSearchTool } from './platform/web-search.tool.js';
+import { createFetchPageTool } from './platform/fetch-page.tool.js';
+import { createArxivSearchTool } from './platform/arxiv-search.tool.js';
 import { createMCPToolWrappers, type MCPToolInfo } from './mcp/mcp-tool-wrapper.js';
 
 /**
@@ -13,13 +16,13 @@ import { createMCPToolWrappers, type MCPToolInfo } from './mcp/mcp-tool-wrapper.
  * Specifies which tools are available to which agent
  */
 export interface AgentToolConfig {
-  /** Agent name (e.g., 'main', 'research', 'code-reviewer') */
+  /** Agent name (e.g., 'main', 'research', 'web-explore') */
   agentName: string;
 
   /** Platform tool names to include */
   platformTools: string[];
 
-  /** MCP tool patterns/names to include (supports wildcards like 'category_*') */
+  /** MCP tool patterns/names to include (supports wildcards like 'python_*') */
   mcpTools: string[];
 }
 
@@ -31,6 +34,7 @@ export interface PlatformServices {
   ohlcService?: OHLCService | (() => OHLCService | undefined);
   symbolIndexService?: SymbolIndexService | (() => SymbolIndexService | undefined);
   workspaceManager?: WorkspaceManager | (() => WorkspaceManager | undefined);
+  tavilyApiKey?: string;
 }
 
 /**
@@ -81,7 +85,8 @@ export class ToolRegistry {
     mcpClient?: MCPClientConnector,
     availableMCPTools?: MCPToolInfo[],
     workspaceManager?: WorkspaceManager,
-    onImage?: (image: { data: string; mimeType: string }) => void
+    onImage?: (image: { data: string; mimeType: string }) => void,
+    onWorkspaceMutation?: (storeName: string, newState: unknown) => void
   ): Promise<DynamicStructuredTool[]> {
     const config = this.agentToolConfigs.get(agentName);
 
@@ -105,7 +110,7 @@ export class ToolRegistry {
     // Add MCP tools (if MCP client and tools are available)
     if (mcpClient && availableMCPTools && availableMCPTools.length > 0) {
       const filteredMCPTools = this.filterMCPTools(availableMCPTools, config.mcpTools);
-      const mcpToolInstances = createMCPToolWrappers(filteredMCPTools, mcpClient, this.logger, onImage);
+      const mcpToolInstances = createMCPToolWrappers(filteredMCPTools, mcpClient, this.logger, onImage, onWorkspaceMutation);
       tools.push(...mcpToolInstances);
 
       this.logger.debug(
@@ -180,6 +185,25 @@ export class ToolRegistry {
         break;
       }
 
+      case 'web_search': {
+        if (this.platformServices.tavilyApiKey) {
+          tool = createWebSearchTool({ apiKey: this.platformServices.tavilyApiKey, logger: this.logger });
+        } else {
+          this.logger.warn('TAVILY_API_KEY not configured — web_search tool unavailable');
+        }
+        break;
+      }
+
+      case 'fetch_page': {
+        tool = createFetchPageTool({ logger: this.logger });
+        break;
+      }
+
+      case 'arxiv_search': {
+        tool = createArxivSearchTool({ logger: this.logger });
+        break;
+      }
+
       default:
         this.logger.warn({ tool: toolName }, 'Unknown platform tool');
         return null;
@@ -202,7 +226,7 @@ export class ToolRegistry {
 
   /**
    * Filter MCP tools based on patterns/names
-   * Supports wildcards like 'category_*' or exact names like 'execute_research'
+   * Supports wildcards like 'python_*' or exact names like 'execute_research'
    */
   private filterMCPTools(availableTools: MCPToolInfo[], patterns: string[]): MCPToolInfo[] {
     if (patterns.length === 0) {
@@ -221,7 +245,7 @@ export class ToolRegistry {
 
   /**
    * Check if a tool name matches a pattern
-   * Supports wildcards: 'category_*' matches 'category_write', 'category_read', etc.
+   * Supports wildcards: 'python_*' matches 'python_write', 'python_read', etc.
    */
   private matchesPattern(toolName: string, pattern: string): boolean {
     if (pattern === toolName) {

gateway/src/types/ohlc.ts

@@ -11,11 +11,11 @@
  * TradingView bar format (used by web frontend)
  */
 export interface TradingViewBar {
-  time: number;           // Unix timestamp in SECONDS
-  open: number | null;    // null for gap bars (no trades that period)
-  high: number | null;
-  low: number | null;
-  close: number | null;
+  time: number;   // Unix timestamp in SECONDS
+  open: number;   // always non-null — ingestor forward-fills interior gaps
+  high: number;
+  low: number;
+  close: number;
   volume?: number | null;
   // Optional extra columns from ohlc.proto
   buy_vol?: number;
@@ -31,13 +31,13 @@ export interface TradingViewBar {
  * Backend OHLC format (from Iceberg)
  */
 export interface BackendOHLC {
-  timestamp: bigint;          // Unix timestamp in NANOSECONDS — kept as bigint to preserve precision
-  ticker: string;             // Nautilus format: "BTC/USDT.BINANCE"
+  timestamp: bigint;  // Unix timestamp in NANOSECONDS — kept as bigint to preserve precision
+  ticker: string;     // Nautilus format: "BTC/USDT.BINANCE"
   period_seconds: number;
-  open: number | null;        // null for gap bars (no trades that period)
-  high: number | null;
-  low: number | null;
-  close: number | null;
+  open: number;       // always non-null — ingestor forward-fills interior gaps
+  high: number;
+  low: number;
+  close: number;
   volume: number | null;
 }
 

gateway/src/types/user.ts

@@ -96,7 +96,7 @@ export const LICENSE_TIER_TEMPLATES: Record<LicenseTier, License> = {
       memoryRequest: '512Mi', memoryLimit: '2Gi',
       cpuRequest: '250m',     cpuLimit: '2000m',
       storage: '10Gi',        tmpSizeLimit: '256Mi',
-      enableIdleShutdown: true, idleTimeoutMinutes: 60,
+      enableIdleShutdown: false, idleTimeoutMinutes: 0,
     },
   },
   enterprise: {

gateway/src/workspace/container-sync.ts

@@ -55,6 +55,20 @@ export class ContainerSync {
     this.logger = logger.child({ component: 'ContainerSync' });
   }
 
+  /**
+   * Parse a raw MCP callTool response into the tool's return value.
+   * MCP tool results are wrapped as: { content: [{ type: 'text', text: '<json>' }] }
+   */
+  private parseMcpResult(raw: unknown): unknown {
+    const r = raw as any;
+    const text = r?.content?.[0]?.text ?? r?.[0]?.text;
+    if (typeof text === 'string') {
+      return JSON.parse(text);
+    }
+    // Already unwrapped (shouldn't happen in practice)
+    return raw;
+  }
+
   /**
    * Load a workspace store from the container.
    * Returns the stored state or indicates the store doesn't exist.
@@ -68,7 +82,7 @@ export class ContainerSync {
     try {
       this.logger.debug({ store: storeName }, 'Loading store from container');
 
-      const result = (await this.mcpClient.callTool('workspace_read', {
+      const result = this.parseMcpResult(await this.mcpClient.callTool('workspace_read', {
         store_name: storeName,
       })) as { exists: boolean; data?: unknown; error?: string };
 
@@ -104,7 +118,7 @@ export class ContainerSync {
     try {
       this.logger.debug({ store: storeName }, 'Saving store to container');
 
-      const result = (await this.mcpClient.callTool('workspace_write', {
+      const result = this.parseMcpResult(await this.mcpClient.callTool('workspace_write', {
         store_name: storeName,
         data: state,
       })) as { success: boolean; error?: string };
@@ -136,7 +150,7 @@ export class ContainerSync {
     try {
       this.logger.debug({ store: storeName, patchOps: patch.length }, 'Patching store in container');
 
-      const result = (await this.mcpClient.callTool('workspace_patch', {
+      const result = this.parseMcpResult(await this.mcpClient.callTool('workspace_patch', {
         store_name: storeName,
         patch,
       })) as { success: boolean; data?: unknown; error?: string };

gateway/src/workspace/sync-registry.ts

@@ -59,12 +59,12 @@ class SyncEntry {
 
   /**
    * Set state directly (used for loading from container).
-   * Resets sequence to 0.
+   * Sets sequence to 1 so clients at seq 0 (empty state) receive a full snapshot.
    */
   setState(newState: unknown): void {
     this.state = deepClone(newState);
     this.lastSnapshot = deepClone(newState);
-    this.seq = 0;
+    this.seq = 1;
     this.history = [];
   }
 

gateway/src/workspace/types.ts

@@ -272,12 +272,84 @@ export interface Shape {
  */
 export type ShapesStore = Record<string, Shape>;
 
+/**
+ * Parameter schema entry for a custom indicator.
+ */
+export interface CustomIndicatorParam {
+  type: 'int' | 'float' | 'bool' | 'string';
+  default: any;
+  description?: string;
+  min?: number;
+  max?: number;
+}
+
+/**
+ * Per-series plot configuration for a custom indicator output column.
+ * style maps to LineStudyPlotStyle: 0=Line, 1=Histogram, 3=Dots/Cross,
+ * 4=Area, 5=Columns, 6=Circles, 9=StepLine.
+ */
+export interface PlotConfig {
+  style: number;
+  color?: string;
+  linewidth?: number;
+  visible?: boolean;
+}
+
+/**
+ * Shaded region between two plots ("plot_plot") or two bands ("hline_hline").
+ */
+export interface FilledAreaConfig {
+  id: string;
+  type: 'plot_plot' | 'hline_hline';
+  series1: string;
+  series2: string;
+  color?: string;
+  opacity?: number;
+}
+
+/**
+ * Horizontal reference line (e.g. RSI overbought/oversold level).
+ * linestyle: 0=solid, 1=dotted, 2=dashed.
+ */
+export interface BandConfig {
+  id: string;
+  value: number;
+  color?: string;
+  linewidth?: number;
+  linestyle?: number;
+  visible?: boolean;
+}
+
+/**
+ * Output column descriptor for a custom indicator.
+ */
+export interface CustomIndicatorColumn {
+  name: string;
+  display_name?: string;
+  description?: string;
+  plot?: PlotConfig;
+}
+
+/**
+ * Metadata needed to auto-construct a TradingView custom study.
+ * Populated by the indicator subagent when adding a custom_ indicator.
+ */
+export interface CustomIndicatorMetadata {
+  display_name: string;
+  parameters: Record<string, CustomIndicatorParam>;
+  input_series: string[];
+  output_columns: CustomIndicatorColumn[];
+  pane: 'price' | 'separate';
+  filled_areas?: FilledAreaConfig[];
+  bands?: BandConfig[];
+}
+
 /**
  * Indicator instance on TradingView chart.
  */
 export interface IndicatorInstance {
   id: string;
-  talib_name: string;
+  pandas_ta_name: string;
   instance_name: string;
   parameters: Record<string, any>;
   tv_study_id?: string;
@@ -289,6 +361,8 @@ export interface IndicatorInstance {
   created_at?: number;
   modified_at?: number;
   original_id?: string;
+  /** Populated for custom_ indicators; drives TV custom study auto-construction. */
+  custom_metadata?: CustomIndicatorMetadata;
 }
 
 /**

ingestor/src/ccxt-fetcher.js

@@ -45,6 +45,7 @@ export class CCXTFetcher {
     const exchange = this.getExchange(exchangeName);
 
     // Load market info from CCXT
+    this.logger.info({ exchangeName, symbol }, 'Loading markets for metadata');
     await exchange.loadMarkets();
     const market = exchange.market(symbol);
 
@@ -108,8 +109,9 @@ export class CCXTFetcher {
     // Map period seconds to CCXT timeframe
     const timeframe = this.secondsToTimeframe(periodSeconds);
 
+    const marketsLoaded = exchange.markets != null && Object.keys(exchange.markets).length > 0;
     this.logger.info(
-      { ticker, timeframe, startMs, endMs, limit },
+      { ticker, timeframe, startMs, endMs, limit, marketsLoaded },
       'Fetching historical OHLC'
     );
 
@@ -120,44 +122,76 @@ export class CCXTFetcher {
     // The caller's limit/countback is irrelevant to how much we need to fetch from the exchange.
     const PAGE_SIZE = 1000;
 
+    const FETCH_RETRIES = 3;
+    const FETCH_RETRY_DELAY_MS = 5000;
+
     while (since < endMs) {
-      try {
-        const candles = await exchange.fetchOHLCV(
-          symbol,
-          timeframe,
-          since,
-          PAGE_SIZE
-        );
-
-        if (candles.length === 0) {
+      let candles;
+      let lastError;
+      for (let attempt = 1; attempt <= FETCH_RETRIES; attempt++) {
+        try {
+          candles = await exchange.fetchOHLCV(symbol, timeframe, since, PAGE_SIZE);
+          lastError = null;
           break;
+        } catch (error) {
+          lastError = error;
+          const isRetryable = error.constructor?.name === 'NetworkError' ||
+            error.constructor?.name === 'RequestTimeout' ||
+            error.constructor?.name === 'ExchangeNotAvailable';
+          this.logger.warn(
+            {
+              errorType: error.constructor?.name,
+              error: error.message,
+              errorUrl: error.url,
+              ticker,
+              since,
+              attempt,
+              retryable: isRetryable
+            },
+            'OHLC fetch attempt failed'
+          );
+          if (!isRetryable || attempt === FETCH_RETRIES) break;
+          await exchange.sleep(FETCH_RETRY_DELAY_MS * attempt);
         }
-
-        // Filter candles within the requested time range
-        const filteredCandles = candles.filter(c => {
-          const timestamp = c[0];
-          return timestamp >= startMs && timestamp < endMs;  // endMs is exclusive
-        });
-
-        fetchedCandles.push(...filteredCandles);
-
-        // Advance to next batch start
-        const lastTimestamp = candles[candles.length - 1][0];
-        since = lastTimestamp + (periodSeconds * 1000);
-
-        if (since >= endMs) {
-          break;
-        }
-
-        // Apply rate limiting
-        await exchange.sleep(exchange.rateLimit);
-      } catch (error) {
+      }
+      if (lastError) {
         this.logger.error(
-          { error: error.message, ticker, since },
+          {
+            errorType: lastError.constructor?.name,
+            error: lastError.message,
+            errorUrl: lastError.url,
+            ticker,
+            since,
+            marketsLoaded: exchange.markets != null && Object.keys(exchange.markets).length > 0,
+            stack: lastError.stack
+          },
           'Error fetching OHLC'
         );
-        throw error;
+        throw lastError;
       }
+
+      if (candles.length === 0) {
+        break;
+      }
+
+      // Filter candles within the requested time range
+      const filteredCandles = candles.filter(c => {
+        const timestamp = c[0];
+        return timestamp >= startMs && timestamp < endMs;  // endMs is exclusive
+      });
+
+      fetchedCandles.push(...filteredCandles);
+
+      // Advance to next batch start
+      const lastTimestamp = candles[candles.length - 1][0];
+      since = lastTimestamp + (periodSeconds * 1000);
+
+      if (since >= endMs) {
+        break;
+      }
+
+      // Apply rate limiting
+      await exchange.sleep(exchange.rateLimit);
     }
 
     // Get metadata for proper denomination
@@ -173,32 +207,44 @@ export class CCXTFetcher {
 
     const periodMs = periodSeconds * 1000;
 
-    // Only create null gap bars for interior gaps — periods where real data exists
-    // on BOTH sides (i.e., between the first and last real bar). Do NOT append
-    // null bars before the first real bar or after the last real bar: those edge
-    // positions may be in-progress candles or simply outside the exchange's history,
-    // and we have no positive signal that a gap exists there.
+    // Forward-fill interior gaps — periods between the first and last real bar
+    // where the exchange returned no candle. Edge gaps (before firstRealTs or
+    // after lastRealTs) are left absent; they'll be caught by gap detection and
+    // trigger a targeted backfill request.
     const realTimestamps = [...fetchedByTs.keys()].sort((a, b) => a - b);
     const firstRealTs = realTimestamps[0];
     const lastRealTs  = realTimestamps[realTimestamps.length - 1];
 
     const allCandles = [];
     let gapCount = 0;
+    let prevClose = null;
 
     for (let ts = firstRealTs; ts <= lastRealTs; ts += periodMs) {
       if (fetchedByTs.has(ts)) {
-        allCandles.push(this.convertToOHLC(fetchedByTs.get(ts), ticker, periodSeconds, metadata));
-      } else {
-        // Interior gap — confirmed by real bars on both sides
+        const bar = this.convertToOHLC(fetchedByTs.get(ts), ticker, periodSeconds, metadata);
+        prevClose = bar.close;
+        allCandles.push(bar);
+      } else if (prevClose !== null) {
+        // Interior gap — forward-fill with previous close, zero volume
         gapCount++;
-        allCandles.push(this.createGapBar(ts, ticker, periodSeconds, metadata));
+        allCandles.push({
+          ticker,
+          timestamp:  (ts * 1_000_000).toString(),
+          open:  prevClose,
+          high:  prevClose,
+          low:   prevClose,
+          close: prevClose,
+          volume: '0',
+          open_time:  (ts * 1_000_000).toString(),
+          close_time: ((ts + periodSeconds * 1000) * 1_000_000).toString()
+        });
       }
     }
 
     if (gapCount > 0) {
       this.logger.info(
         { ticker, gapCount, total: allCandles.length },
-        'Filled interior gap bars for missing periods in source data'
+        'Forward-filled interior gap bars with previous close price'
       );
     }
 
@@ -264,24 +310,6 @@ export class CCXTFetcher {
     };
   }
 
-  /**
-   * Create a gap bar for a period with no trade data.
-   * All OHLC/volume fields are null — the timestamp slot is reserved but unpopulated.
-   */
-  createGapBar(timestampMs, ticker, periodSeconds, metadata) {
-    return {
-      ticker,
-      timestamp: (timestampMs * 1_000_000).toString(),   // Convert ms to nanoseconds
-      open: null,
-      high: null,
-      low: null,
-      close: null,
-      volume: null,
-      open_time: (timestampMs * 1_000_000).toString(),
-      close_time: ((timestampMs + periodSeconds * 1000) * 1_000_000).toString()
-    };
-  }
-
   /**
    * Convert CCXT trade to our Tick format
    * Uses precision fields from market metadata for proper integer representation

ingestor/src/index.js

@@ -285,7 +285,14 @@ class IngestorWorker {
 
     } catch (error) {
       this.logger.error(
-        { error: error.message, request_id, ticker },
+        {
+          errorType: error.constructor?.name,
+          error: error.message,
+          errorUrl: error.url,
+          request_id,
+          ticker,
+          stack: error.stack
+        },
         'Failed to process historical request'
       );
 

ingestor/src/kafka-producer.js

@@ -181,16 +181,14 @@ export class KafkaProducer {
         errorMessage: metadata.error_message
       },
       rows: ohlcData.map(candle => {
-        // null open/high/low/close signals a gap bar (no trades that period).
-        // Omit fields from the protobuf message when null so hasOpen() etc. return false.
         const row = {
           timestamp: candle.timestamp,
-          ticker: candle.ticker,
+          ticker:    candle.ticker,
+          open:      candle.open,
+          high:      candle.high,
+          low:       candle.low,
+          close:     candle.close,
         };
-        if (candle.open  != null) row.open   = candle.open;
-        if (candle.high  != null) row.high   = candle.high;
-        if (candle.low   != null) row.low    = candle.low;
-        if (candle.close != null) row.close  = candle.close;
         if (candle.volume != null) row.volume = candle.volume;
         return row;
       })

protobuf/ohlc.proto

@@ -8,12 +8,12 @@ message OHLC {
   // Timestamp in nanoseconds since epoch
   uint64 timestamp = 1;
 
-  // Prices are stored as doubles (Nautilus-aligned, no denominator needed).
-  // Optional to support null bars for periods with no trades.
-  optional int64 open = 2;
-  optional int64 high = 3;
-  optional int64 low = 4;
-  optional int64 close = 5;
+  // Prices are stored as integers (Nautilus-aligned with precision denominator).
+  // Always non-null — ingestor forward-fills interior gaps with the previous close.
+  int64 open = 2;
+  int64 high = 3;
+  int64 low = 4;
+  int64 close = 5;
   optional int64 volume = 6;
   optional int64 buy_vol = 7;
   optional int64 sell_vol = 8;

sandbox/Dockerfile

@@ -40,6 +40,7 @@ WORKDIR /app
 # Install runtime dependencies only
 RUN apt-get update && apt-get install -y --no-install-recommends \
     libzmq5 \
+    git \
     && rm -rf /var/lib/apt/lists/*
 
 # Create non-root user
@@ -76,6 +77,7 @@ USER dexorder
 # Environment variables (can be overridden in k8s)
 ENV PYTHONUNBUFFERED=1 \
     MPLCONFIGDIR=/tmp \
+    NUMBA_CACHE_DIR=/tmp/numba_cache \
     LOG_LEVEL=INFO \
     CONFIG_PATH=/app/config/config.yaml \
     SECRETS_PATH=/app/config/secrets.yaml \

sandbox/dexorder/nautilus/__init__.py

@@ -0,0 +1,34 @@
+"""
+dexorder.nautilus — Nautilus Trader integration for strategy backtesting.
+
+Quants import PandasStrategy to write strategies:
+
+    from dexorder.nautilus import PandasStrategy
+    import pandas as pd
+    import pandas_ta as ta
+
+    class MyStrategy(PandasStrategy):
+        def evaluate(self, dfs):
+            df = dfs.get("BTC/USDT.BINANCE:3600")
+            if df is None or len(df) < 14:
+                return
+            rsi = ta.rsi(df["close"], length=14)
+            if rsi.iloc[-1] < 30:
+                self.buy(0.01)
+            elif rsi.iloc[-1] > 70:
+                self.sell(0.01)
+
+SecretsVault provides the interface for user-owned exchange API keys
+(stub until the user-local vault is implemented):
+
+    from dexorder.nautilus import SecretsVault
+"""
+
+from dexorder.nautilus.pandas_strategy import PandasStrategy, PandasStrategyConfig
+from dexorder.secrets_vault import SecretsVault
+
+__all__ = [
+    "PandasStrategy",
+    "PandasStrategyConfig",
+    "SecretsVault",
+]

sandbox/dexorder/nautilus/backtest_runner.py

@@ -0,0 +1,358 @@
+"""
+Backtest runner — sets up Nautilus BacktestEngine and runs a PandasStrategy.
+
+Entry points
+------------
+run_backtest()             — called from execute_strategy MCP tool (via thread executor)
+_load_strategy_class()     — exec() the user's implementation.py, find PandasStrategy subclass
+_setup_custom_indicators() — register user indicators with pandas-ta via ta.import_dir()
+_compute_metrics()         — extract P&L metrics from completed BacktestEngine
+"""
+
+from __future__ import annotations
+
+import inspect
+import logging
+from pathlib import Path
+from typing import Any
+
+import pandas as pd
+
+log = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Custom indicator setup
+# ---------------------------------------------------------------------------
+
+def _setup_custom_indicators(data_dir: Path) -> None:
+    """Register user's custom indicators with pandas-ta (delegates to python_tools)."""
+    from dexorder.tools.python_tools import setup_custom_indicators
+    setup_custom_indicators(data_dir)
+
+
+# ---------------------------------------------------------------------------
+# Strategy class loading
+# ---------------------------------------------------------------------------
+
+def _load_strategy_class(impl_path: Path) -> type:
+    """
+    Execute implementation.py and return the unique PandasStrategy subclass.
+
+    The exec namespace is seeded with:
+        PandasStrategy — base class (so the subclass check works)
+        pd             — pandas
+        ta             — pandas_ta (if available)
+
+    Raises:
+        ValueError: if zero or multiple PandasStrategy subclasses are defined
+        SyntaxError / Exception: if the file fails to parse or execute
+    """
+    from dexorder.nautilus.pandas_strategy import PandasStrategy
+
+    namespace: dict[str, Any] = {
+        "__builtins__": __builtins__,
+        "PandasStrategy": PandasStrategy,
+        "pd": pd,
+    }
+
+    try:
+        import pandas_ta as ta
+        namespace["ta"] = ta
+    except ImportError:
+        pass
+
+    code = impl_path.read_text()
+    exec(compile(code, str(impl_path), "exec"), namespace)  # noqa: S102
+
+    subclasses = [
+        obj for obj in namespace.values()
+        if (
+            inspect.isclass(obj)
+            and issubclass(obj, PandasStrategy)
+            and obj is not PandasStrategy
+        )
+    ]
+
+    if len(subclasses) == 0:
+        raise ValueError(
+            f"No PandasStrategy subclass found in {impl_path}. "
+            "The strategy file must define exactly one class that inherits from PandasStrategy."
+        )
+    if len(subclasses) > 1:
+        names = [c.__name__ for c in subclasses]
+        raise ValueError(
+            f"Multiple PandasStrategy subclasses found in {impl_path}: {names}. "
+            "Define exactly one concrete strategy class per file."
+        )
+
+    return subclasses[0]
+
+
+# ---------------------------------------------------------------------------
+# Metrics extraction
+# ---------------------------------------------------------------------------
+
+def _compute_metrics(
+    engine,
+    venue_strs: list[str],
+    initial_capital: float,
+    all_bars: list,
+) -> dict[str, Any]:
+    """
+    Extract performance metrics from a completed BacktestEngine.
+
+    Returns dict with:
+        total_return   float  — fractional (0.15 = +15%)
+        sharpe_ratio   float  — annualized; 0.0 if no trades or constant equity
+        max_drawdown   float  — max peak-to-trough as fraction (0.10 = 10% drawdown)
+        win_rate       float  — fraction of trades with positive realized PnL
+        trade_count    int
+        equity_curve   list[{timestamp: int_unix_s, equity: float}]
+    """
+    # Reconstruct equity curve from fills
+    equity_points: list[dict] = []
+    if all_bars:
+        equity_points.append({
+            "timestamp": all_bars[0].ts_event // 1_000_000_000,
+            "equity": initial_capital,
+        })
+
+    running_equity = initial_capital
+    trade_count = 0
+    winning_trades = 0
+
+    try:
+        fills_df = engine.trader.generate_order_fills_report()
+    except Exception as exc:
+        log.debug("generate_order_fills_report() failed: %s", exc)
+        fills_df = None
+
+    if fills_df is not None and len(fills_df) > 0:
+        # Sort by event time
+        if "ts_event" in fills_df.columns:
+            fills_df = fills_df.sort_values("ts_event")
+
+        for _, fill in fills_df.iterrows():
+            rpnl = fill.get("realized_pnl") if hasattr(fill, "get") else None
+            if rpnl is None:
+                continue
+
+            # Nautilus Money objects: str form is "15.32 USDT"
+            rpnl_float: float | None = None
+            try:
+                if hasattr(rpnl, "as_decimal"):
+                    rpnl_float = float(rpnl.as_decimal())
+                elif rpnl is not None:
+                    rpnl_str = str(rpnl).strip()
+                    if rpnl_str and rpnl_str.lower() not in ("none", "nan"):
+                        rpnl_float = float(rpnl_str.split()[0])
+            except (ValueError, TypeError, IndexError):
+                pass
+
+            if rpnl_float is not None and rpnl_float != 0.0:
+                ts_s: int | None = None
+                raw_ts = fill.get("ts_event") if hasattr(fill, "get") else None
+                if raw_ts is not None:
+                    try:
+                        ts_s = int(raw_ts) // 1_000_000_000
+                    except (TypeError, ValueError):
+                        pass
+
+                running_equity += rpnl_float
+                trade_count += 1
+                if rpnl_float > 0:
+                    winning_trades += 1
+
+                if ts_s is not None:
+                    equity_points.append({"timestamp": ts_s, "equity": running_equity})
+
+    if all_bars:
+        equity_points.append({
+            "timestamp": all_bars[-1].ts_event // 1_000_000_000,
+            "equity": running_equity,
+        })
+
+    # Try to get actual final balance from the account (more accurate than fill reconstruction)
+    try:
+        from nautilus_trader.model.identifiers import Venue
+        for venue_str in venue_strs:
+            account = engine.cache.account_for_venue(Venue(venue_str))
+            if account is None:
+                continue
+            # Sum all balances (quote currency is what we started with)
+            for bal in account.balances().values():
+                total = getattr(bal, "total", None)
+                if total is not None:
+                    final_val = float(str(total).split()[0]) if not hasattr(total, "as_decimal") else float(total.as_decimal())
+                    # Use the account balance as the definitive final equity
+                    running_equity = final_val
+                    if equity_points:
+                        equity_points[-1]["equity"] = running_equity
+                    break
+    except Exception as exc:
+        log.debug("Account balance extraction failed: %s", exc)
+
+    # Core metrics
+    total_return = (running_equity - initial_capital) / initial_capital if initial_capital else 0.0
+    win_rate = winning_trades / trade_count if trade_count > 0 else 0.0
+
+    # Sharpe ratio (annualized) from equity curve returns
+    sharpe = 0.0
+    if len(equity_points) > 2 and all_bars and len(all_bars) > 1:
+        equity_series = pd.Series([p["equity"] for p in equity_points])
+        returns = equity_series.pct_change().dropna()
+        if len(returns) > 1 and returns.std() > 0:
+            bar_duration_ns = (all_bars[-1].ts_event - all_bars[0].ts_event) / max(len(all_bars) - 1, 1)
+            if bar_duration_ns > 0:
+                bars_per_year = (365 * 24 * 3600 * 1e9) / bar_duration_ns
+                sharpe = float((returns.mean() / returns.std()) * (bars_per_year ** 0.5))
+
+    # Max drawdown
+    max_drawdown = 0.0
+    if len(equity_points) > 1:
+        equity_arr = pd.Series([p["equity"] for p in equity_points])
+        rolling_max = equity_arr.cummax()
+        drawdowns = (equity_arr - rolling_max) / rolling_max.replace(0, float("nan"))
+        max_drawdown = float(abs(drawdowns.min())) if len(drawdowns) > 0 else 0.0
+
+    return {
+        "total_return":  round(total_return, 6),
+        "sharpe_ratio":  round(sharpe, 4),
+        "max_drawdown":  round(max_drawdown, 6),
+        "win_rate":      round(win_rate, 4),
+        "trade_count":   trade_count,
+        "equity_curve":  equity_points,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Main entry point
+# ---------------------------------------------------------------------------
+
+def run_backtest(
+    strategy_class: type,
+    feeds: list[tuple[str, int]],
+    ohlc_dfs: dict[str, pd.DataFrame],
+    initial_capital: float = 10_000.0,
+    paper: bool = True,
+) -> dict[str, Any]:
+    """
+    Configure and run a BacktestEngine synchronously.
+
+    Designed to be called from asyncio via loop.run_in_executor() since
+    BacktestEngine does not support async.
+
+    Args:
+        strategy_class: Concrete PandasStrategy subclass to instantiate
+        feeds: List of (ticker, period_seconds) pairs, e.g. [("BTC/USDT.BINANCE", 3600)]
+        ohlc_dfs: Dict of feed_key → full OHLC+ DataFrame (with buy_vol, sell_vol, etc.)
+        initial_capital: Starting account balance in quote currency
+        paper: Always True for historical backtest (flag reserved for forward testing)
+
+    Returns:
+        Dict of performance metrics (see _compute_metrics)
+    """
+    from nautilus_trader.backtest.engine import BacktestEngine, BacktestEngineConfig
+    from nautilus_trader.backtest.models import FillModel
+    from nautilus_trader.config import LoggingConfig
+    from nautilus_trader.model.enums import OmsType, AccountType
+    from nautilus_trader.model.identifiers import Venue
+    from nautilus_trader.model.objects import Money
+
+    from dexorder.nautilus.pandas_strategy import PandasStrategyConfig, make_feed_key
+    from dexorder.nautilus.data_adapter import (
+        make_instrument_from_metadata,
+        make_bar_type,
+        df_to_bars,
+        extras_lookup,
+    )
+
+    # --- Engine config ---
+    engine_config = BacktestEngineConfig(
+        trader_id="DEXORDER-BACKTEST-001",
+        logging=LoggingConfig(log_level="ERROR"),
+    )
+    engine = BacktestEngine(config=engine_config)
+
+    # --- Per-venue setup (unique venues from feeds) ---
+    venues_seen: set[str] = set()
+    all_bars: list = []
+    feed_keys: list[str] = []
+
+    instruments: dict[str, Any] = {}
+    price_precisions: dict[str, int] = {}
+    size_precisions: dict[str, int] = {}
+
+    for ticker, period_seconds in feeds:
+        feed_key = make_feed_key(ticker, period_seconds)
+        feed_keys.append(feed_key)
+
+        from dexorder.symbol_metadata_client import parse_ticker
+        exchange_id, _ = parse_ticker(ticker)
+
+        if exchange_id not in venues_seen:
+            venues_seen.add(exchange_id)
+            # Determine quote currency from ticker (e.g. USDT from BTC/USDT)
+            _, market_id = parse_ticker(ticker)
+            quote_str = market_id.split("/")[1] if "/" in market_id else "USDT"
+            from nautilus_trader.model.currencies import Currency
+            quote_currency = Currency.from_str(quote_str)
+            engine.add_venue(
+                venue=Venue(exchange_id),
+                oms_type=OmsType.NETTING,
+                account_type=AccountType.CASH,
+                base_currency=None,
+                starting_balances=[Money(initial_capital, quote_currency)],
+                fill_model=FillModel(),
+            )
+
+        # Instrument and bars
+        instrument, pp, sp = make_instrument_from_metadata(ticker)
+        instruments[feed_key] = instrument
+        price_precisions[feed_key] = pp
+        size_precisions[feed_key] = sp
+
+        engine.add_instrument(instrument)
+
+        df = ohlc_dfs.get(feed_key)
+        if df is not None and not df.empty:
+            bar_type = make_bar_type(ticker, period_seconds)
+            bars = df_to_bars(df, bar_type, pp, sp)
+            engine.add_data(bars)
+            all_bars.extend(bars)
+        else:
+            log.warning("No OHLC data for feed %s — strategy will receive no bars", feed_key)
+
+    if not all_bars:
+        return {
+            "total_return": 0.0, "sharpe_ratio": 0.0, "max_drawdown": 0.0,
+            "win_rate": 0.0, "trade_count": 0, "equity_curve": [],
+        }
+
+    # Sort combined bars by timestamp for metrics computation
+    all_bars.sort(key=lambda b: b.ts_event)
+
+    # --- Instantiate and configure strategy ---
+    strategy_config = PandasStrategyConfig(
+        strategy_id=f"{strategy_class.__name__}-001",
+        feed_keys=tuple(feed_keys),
+        initial_capital=initial_capital,
+    )
+    strategy = strategy_class(config=strategy_config)
+
+    # Inject OHLC+ extras before run
+    for feed_key, df in ohlc_dfs.items():
+        if df is not None and not df.empty:
+            strategy._inject_extras(feed_key, extras_lookup(df))
+
+    engine.add_strategy(strategy)
+
+    # --- Run ---
+    engine.run()
+
+    # --- Extract metrics ---
+    metrics = _compute_metrics(engine, list(venues_seen), initial_capital, all_bars)
+    engine.dispose()
+
+    return metrics

sandbox/dexorder/nautilus/data_adapter.py

@@ -0,0 +1,235 @@
+"""
+Data adapter — converts our OHLC DataFrames to Nautilus objects.
+
+Functions
+---------
+make_instrument   — CurrencyPair from ticker string
+make_bar_type     — BarType from ticker + period_seconds
+df_to_bars        — OHLC DataFrame → list[Bar]
+extras_lookup     — extract OHLC+ extras dict from DataFrame
+make_instrument_from_metadata — instrument with best-effort precision
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Optional
+
+import pandas as pd
+
+from nautilus_trader.model.currencies import Currency
+from nautilus_trader.model.data import Bar, BarType, BarSpecification
+from nautilus_trader.model.enums import BarAggregation, PriceType, AggregationSource
+from nautilus_trader.model.identifiers import InstrumentId, Symbol, Venue
+from nautilus_trader.model.instruments import CurrencyPair
+from nautilus_trader.model.objects import Price, Quantity
+
+from dexorder.symbol_metadata_client import parse_ticker
+from dexorder.nautilus.pandas_strategy import (
+    bar_type_from_feed_key,
+    _PERIOD_TO_AGGREGATION,
+)
+
+log = logging.getLogger(__name__)
+
+# Columns in our OHLC+ DataFrames that are extras (not part of Nautilus Bar)
+_EXTRA_COLS = ("buy_vol", "sell_vol", "open_interest")
+
+
+def make_bar_type(ticker: str, period_seconds: int) -> BarType:
+    """
+    Construct a Nautilus BarType from our ticker and period_seconds.
+
+    Period mapping:
+        period_seconds < 60    → SECOND, step = period_seconds
+        period_seconds < 3600  → MINUTE, step = period_seconds // 60
+        period_seconds < 86400 → HOUR,   step = period_seconds // 3600
+        else                   → DAY,    step = period_seconds // 86400
+
+    Price type = MID (standard for crypto OHLC).
+    Source = EXTERNAL (we supply pre-aggregated data, not Nautilus aggregation).
+    """
+    exchange_id, market_id = parse_ticker(ticker)
+    instrument_id = InstrumentId(Symbol(market_id), Venue(exchange_id))
+
+    for threshold, agg, divisor in _PERIOD_TO_AGGREGATION:
+        if period_seconds < threshold:
+            step = max(1, period_seconds // divisor)
+            break
+    else:
+        agg = BarAggregation.DAY
+        step = max(1, period_seconds // 86400)
+
+    spec = BarSpecification(step=step, aggregation=agg, price_type=PriceType.MID)
+    return BarType(instrument_id=instrument_id, bar_spec=spec,
+                   aggregation_source=AggregationSource.EXTERNAL)
+
+
+def make_instrument(
+    ticker: str,
+    price_precision: int = 8,
+    size_precision: int = 8,
+    tick_size: Optional[float] = None,
+    lot_size: Optional[float] = None,
+    maker_fee: float = 0.001,
+    taker_fee: float = 0.001,
+    margin_init: float = 0.0,
+    margin_maint: float = 0.0,
+) -> CurrencyPair:
+    """
+    Create a minimal CurrencyPair instrument from a Nautilus-format ticker.
+
+    Args:
+        ticker: e.g. "BTC/USDT.BINANCE"
+        price_precision: decimal places for price (default 8)
+        size_precision: decimal places for quantity (default 8)
+        tick_size: minimum price increment (defaults to 10^-price_precision)
+        lot_size: minimum order size (defaults to 10^-size_precision)
+        maker_fee, taker_fee: fee rates as fractions (0.001 = 0.1%)
+        margin_init, margin_maint: margin ratios (0.0 = spot/no margin)
+    """
+    exchange_id, market_id = parse_ticker(ticker)
+    base_str, quote_str = market_id.split("/")
+
+    instrument_id = InstrumentId(Symbol(market_id), Venue(exchange_id))
+
+    if tick_size is None:
+        tick_size = 10.0 ** (-price_precision)
+    if lot_size is None:
+        lot_size = 10.0 ** (-size_precision)
+
+    ts_now = 0  # static instrument — timestamp not relevant for backtesting
+
+    return CurrencyPair(
+        instrument_id=instrument_id,
+        raw_symbol=Symbol(market_id),
+        base_currency=Currency.from_str(base_str),
+        quote_currency=Currency.from_str(quote_str),
+        price_precision=price_precision,
+        size_precision=size_precision,
+        price_increment=Price(tick_size, price_precision),
+        size_increment=Quantity(lot_size, size_precision),
+        lot_size=Quantity(lot_size, size_precision),
+        max_quantity=None,
+        min_quantity=Quantity(lot_size, size_precision),
+        max_notional=None,
+        min_notional=None,
+        max_price=None,
+        min_price=None,
+        margin_init=margin_init,
+        margin_maint=margin_maint,
+        maker_fee=maker_fee,
+        taker_fee=taker_fee,
+        ts_event=ts_now,
+        ts_init=ts_now,
+    )
+
+
+def make_instrument_from_metadata(ticker: str) -> tuple[CurrencyPair, int, int]:
+    """
+    Create a CurrencyPair using SymbolMetadata when available.
+
+    Returns:
+        (instrument, price_precision, size_precision)
+
+    Falls back to (instrument with 8/8 defaults) if metadata is unavailable.
+    """
+    try:
+        from dexorder.symbol_metadata_client import SymbolMetadataClient
+        from dexorder.api import get_api
+        # DataAPIImpl stores the catalog URI as an attribute on the OHLCClient
+        api = get_api()
+        ohlc_client = getattr(api.data, '_ohlc_client', None) or getattr(api.data, 'ohlc_client', None)
+        iceberg_client = getattr(ohlc_client, 'iceberg', None) if ohlc_client else None
+        catalog_uri = getattr(iceberg_client, 'catalog_uri', None) if iceberg_client else None
+
+        if catalog_uri:
+            meta_client = SymbolMetadataClient(catalog_uri=catalog_uri)
+            meta = meta_client.get_metadata(ticker)
+            pp = meta.price_precision or 8
+            sp = meta.size_precision or 8
+            instrument = make_instrument(
+                ticker,
+                price_precision=pp,
+                size_precision=sp,
+                tick_size=meta.tick_size,
+                lot_size=meta.lot_size,
+                maker_fee=meta.maker_fee or 0.001,
+                taker_fee=meta.taker_fee or 0.001,
+                margin_init=meta.margin_init or 0.0,
+                margin_maint=meta.margin_maint or 0.0,
+            )
+            return instrument, pp, sp
+    except Exception:
+        log.debug("make_instrument_from_metadata: metadata unavailable for %s, using defaults", ticker)
+
+    instrument = make_instrument(ticker)
+    return instrument, 8, 8
+
+
+def df_to_bars(
+    df: pd.DataFrame,
+    bar_type: BarType,
+    price_precision: int = 8,
+    size_precision: int = 8,
+) -> list[Bar]:
+    """
+    Convert an OHLC DataFrame to a list of Nautilus Bar objects.
+
+    Args:
+        df: DataFrame with columns [timestamp (ns), open, high, low, close].
+            volume column is optional; defaults to 0.0 if absent.
+        bar_type: BarType to tag each bar with.
+        price_precision: decimal precision for Price construction.
+        size_precision: decimal precision for Quantity construction.
+
+    Returns:
+        list[Bar] sorted ascending by timestamp.
+    """
+    has_volume = "volume" in df.columns
+    bars = []
+
+    for row in df.itertuples(index=False):
+        ts_ns = int(row.timestamp)
+        volume = float(row.volume) if has_volume else 0.0
+
+        bar = Bar(
+            bar_type=bar_type,
+            open=Price(float(row.open),  price_precision),
+            high=Price(float(row.high),  price_precision),
+            low=Price(float(row.low),   price_precision),
+            close=Price(float(row.close), price_precision),
+            volume=Quantity(volume, size_precision),
+            ts_event=ts_ns,
+            ts_init=ts_ns,
+        )
+        bars.append(bar)
+
+    return bars
+
+
+def extras_lookup(df: pd.DataFrame) -> dict[int, dict]:
+    """
+    Build a {ts_event_ns → {buy_vol, sell_vol, open_interest}} mapping.
+
+    Values are None for columns absent from the DataFrame.
+
+    Used by PandasStrategy._inject_extras() to enrich each bar with
+    OHLC+ fields that Nautilus Bar does not carry natively.
+    """
+    result: dict[int, dict] = {}
+
+    present = {col: col in df.columns for col in _EXTRA_COLS}
+
+    for row in df.itertuples(index=False):
+        ts_ns = int(row.timestamp)
+        entry: dict = {}
+        for col in _EXTRA_COLS:
+            if present[col]:
+                val = getattr(row, col)
+                entry[col] = None if (val is None or (isinstance(val, float) and pd.isna(val))) else float(val)
+            else:
+                entry[col] = None
+        result[ts_ns] = entry
+
+    return result

sandbox/dexorder/nautilus/pandas_strategy.py

@@ -0,0 +1,315 @@
+"""
+PandasStrategy — Nautilus Strategy base class with a DataFrame-oriented API.
+
+Quants subclass PandasStrategy and implement evaluate(dfs) — the same function
+they'd write in a research notebook. No Nautilus objects appear in quant code.
+
+Features:
+- Multiple data feeds: subscribe to N (ticker, period_seconds) pairs
+- evaluate(dfs) receives a dict[feed_key, DataFrame] where feed_key is "TICKER:period_seconds"
+- Every feed's DataFrame includes OHLC + volume, buy_vol, sell_vol, open_interest
+- Timer hook (on_timer) reserved as extension point — TBD
+
+Feed key format: "BTC/USDT.BINANCE:3600"
+"""
+
+from __future__ import annotations
+
+import logging
+from abc import abstractmethod
+from typing import TYPE_CHECKING
+
+import pandas as pd
+
+from nautilus_trader.config import StrategyConfig
+from nautilus_trader.model.data import Bar, BarType, BarSpecification
+from nautilus_trader.model.enums import BarAggregation, PriceType, AggregationSource, OrderSide, TimeInForce
+from nautilus_trader.model.identifiers import InstrumentId, Symbol, Venue
+from nautilus_trader.model.objects import Quantity
+from nautilus_trader.trading.strategy import Strategy
+
+from dexorder.symbol_metadata_client import parse_ticker
+
+log = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Feed key helpers
+# ---------------------------------------------------------------------------
+
+_PERIOD_TO_AGGREGATION: list[tuple[int, BarAggregation, int]] = [
+    # (threshold_exclusive, aggregation, divisor)
+    # period_seconds < 60 → SECOND, step = period_seconds
+    (60,    BarAggregation.SECOND, 1),
+    # 60 <= period_seconds < 3600 → MINUTE
+    (3600,  BarAggregation.MINUTE, 60),
+    # 3600 <= period_seconds < 86400 → HOUR
+    (86400, BarAggregation.HOUR, 3600),
+]
+
+_AGG_TO_SECONDS: dict[BarAggregation, int] = {
+    BarAggregation.SECOND: 1,
+    BarAggregation.MINUTE: 60,
+    BarAggregation.HOUR:   3600,
+    BarAggregation.DAY:    86400,
+}
+
+
+def make_feed_key(ticker: str, period_seconds: int) -> str:
+    """Return canonical feed key, e.g. 'BTC/USDT.BINANCE:3600'."""
+    return f"{ticker}:{period_seconds}"
+
+
+def parse_feed_key(feed_key: str) -> tuple[str, int]:
+    """Split 'BTC/USDT.BINANCE:3600' → ('BTC/USDT.BINANCE', 3600)."""
+    ticker, period_str = feed_key.rsplit(":", 1)
+    return ticker, int(period_str)
+
+
+def bar_type_from_feed_key(feed_key: str) -> BarType:
+    """Build a Nautilus BarType from a feed key string."""
+    ticker, period_seconds = parse_feed_key(feed_key)
+    exchange_id, market_id = parse_ticker(ticker)
+    instrument_id = InstrumentId(Symbol(market_id), Venue(exchange_id))
+
+    for threshold, agg, divisor in _PERIOD_TO_AGGREGATION:
+        if period_seconds < threshold:
+            step = period_seconds // divisor
+            break
+    else:
+        agg = BarAggregation.DAY
+        step = period_seconds // 86400
+
+    spec = BarSpecification(step=step, aggregation=agg, price_type=PriceType.MID)
+    return BarType(instrument_id=instrument_id, bar_spec=spec,
+                   aggregation_source=AggregationSource.EXTERNAL)
+
+
+def feed_key_from_bar_type(bar_type: BarType) -> str:
+    """Reconstruct the feed key from a BarType."""
+    iid = bar_type.instrument_id
+    ticker = f"{iid.symbol}.{iid.venue}"
+    multiplier = _AGG_TO_SECONDS.get(bar_type.spec.aggregation, 1)
+    period_seconds = bar_type.spec.step * multiplier
+    return f"{ticker}:{period_seconds}"
+
+
+# ---------------------------------------------------------------------------
+# Config
+# ---------------------------------------------------------------------------
+
+class PandasStrategyConfig(StrategyConfig, frozen=True):
+    """
+    Configuration for PandasStrategy.
+
+    feed_keys: tuple of feed key strings, e.g. ("BTC/USDT.BINANCE:3600",)
+               Set by the backtest/activate runner — not by the quant's code.
+    initial_capital: informational; actual account balance is set in BacktestEngine.
+    """
+    feed_keys: tuple[str, ...] = ()
+    initial_capital: float = 10_000.0
+
+
+# ---------------------------------------------------------------------------
+# Base class
+# ---------------------------------------------------------------------------
+
+class PandasStrategy(Strategy):
+    """
+    Base class for quant strategies.
+
+    Quants implement evaluate(dfs) — the same function they'd write in a research
+    notebook. All bar accumulation, OHLC+ field injection, and DataFrame management
+    is handled internally.
+
+    Example
+    -------
+    ::
+
+        from dexorder.nautilus import PandasStrategy
+        import pandas as pd
+        import pandas_ta as ta
+
+        class MyStrategy(PandasStrategy):
+            def evaluate(self, dfs):
+                df = dfs.get("BTC/USDT.BINANCE:3600")
+                if df is None or len(df) < 14:
+                    return
+                rsi = ta.rsi(df["close"], length=14)
+                if rsi.iloc[-1] < 30:
+                    self.buy(0.01)
+                elif rsi.iloc[-1] > 70:
+                    self.sell(0.01)
+    """
+
+    def __init__(self, config: PandasStrategyConfig) -> None:
+        super().__init__(config)
+        # Per-feed row accumulator
+        self._rows: dict[str, list[dict]] = {}
+        # Per-feed DataFrame (updated after each bar)
+        self._dfs: dict[str, pd.DataFrame] = {}
+        # Per-feed extras lookup: {ts_event_ns: {buy_vol, sell_vol, open_interest}}
+        self._extras: dict[str, dict[int, dict]] = {}
+        # Resolved BarType objects (populated in on_start)
+        self._bar_types: dict[str, BarType] = {}
+
+    # ------------------------------------------------------------------
+    # Nautilus lifecycle
+    # ------------------------------------------------------------------
+
+    def on_start(self) -> None:
+        for feed_key in self.config.feed_keys:
+            bar_type = bar_type_from_feed_key(feed_key)
+            self._bar_types[feed_key] = bar_type
+            self.subscribe_bars(bar_type)
+
+    def on_bar(self, bar: Bar) -> None:
+        feed_key = feed_key_from_bar_type(bar.bar_type)
+        ts_ns = bar.ts_event
+
+        # Merge OHLC+ extras (buy_vol, sell_vol, open_interest) by timestamp
+        extras = self._extras.get(feed_key, {}).get(ts_ns, {})
+
+        row = {
+            "timestamp":     ts_ns,
+            "open":          float(bar.open),
+            "high":          float(bar.high),
+            "low":           float(bar.low),
+            "close":         float(bar.close),
+            "volume":        float(bar.volume),
+            "buy_vol":       extras.get("buy_vol"),
+            "sell_vol":      extras.get("sell_vol"),
+            "open_interest": extras.get("open_interest"),
+        }
+
+        if feed_key not in self._rows:
+            self._rows[feed_key] = []
+        self._rows[feed_key].append(row)
+        self._dfs[feed_key] = pd.DataFrame(self._rows[feed_key])
+
+        self.evaluate(self._dfs)
+
+    def on_stop(self) -> None:
+        pass
+
+    # ------------------------------------------------------------------
+    # Quant API — override in subclass
+    # ------------------------------------------------------------------
+
+    @abstractmethod
+    def evaluate(self, dfs: dict[str, pd.DataFrame]) -> None:
+        """
+        Implement your strategy logic here.
+
+        Called after every new bar on any subscribed feed.
+
+        Args:
+            dfs: Dict mapping feed_key → DataFrame.
+                 Feed key format: "TICKER:period_seconds", e.g. "BTC/USDT.BINANCE:3600".
+                 DataFrame columns: timestamp (ns), open, high, low, close, volume,
+                                    buy_vol, sell_vol, open_interest.
+                 All rows up to and including the latest bar are included.
+                 A feed's DataFrame is absent (key missing) until its first bar arrives.
+
+        Trading methods available:
+            self.buy(quantity, feed_key=None)     — market buy
+            self.sell(quantity, feed_key=None)    — market sell
+            self.flatten(feed_key=None)           — close all positions for feed
+        """
+
+    def on_timer(self, timer_name: str) -> None:
+        """
+        Called on timer ticks (TBD — timer wiring not yet implemented).
+
+        Override to handle time-based evaluation independent of bar arrival.
+        Default implementation calls evaluate() with current DataFrames.
+        """
+        self.evaluate(self._dfs)
+
+    # ------------------------------------------------------------------
+    # Order helpers
+    # ------------------------------------------------------------------
+
+    def _resolve_feed_key(self, feed_key: str | None) -> str | None:
+        """Return feed_key if given, else the first configured feed key."""
+        if feed_key is not None:
+            return feed_key
+        keys = self.config.feed_keys
+        return keys[0] if keys else None
+
+    def _instrument_id_for_feed(self, feed_key: str) -> InstrumentId | None:
+        ticker, _ = parse_feed_key(feed_key)
+        try:
+            exchange_id, market_id = parse_ticker(ticker)
+            return InstrumentId(Symbol(market_id), Venue(exchange_id))
+        except ValueError:
+            log.error("Cannot parse ticker from feed key: %s", feed_key)
+            return None
+
+    def buy(self, quantity: float, feed_key: str | None = None) -> None:
+        """Submit a market buy order. Defaults to the first configured feed."""
+        fk = self._resolve_feed_key(feed_key)
+        if not fk:
+            log.error("buy(): no feed key available")
+            return
+        instrument_id = self._instrument_id_for_feed(fk)
+        if instrument_id is None:
+            return
+        instrument = self.cache.instrument(instrument_id)
+        if instrument is None:
+            log.error("buy(): instrument not found for %s", instrument_id)
+            return
+        order = self.order_factory.market(
+            instrument_id=instrument_id,
+            order_side=OrderSide.BUY,
+            quantity=Quantity(quantity, instrument.size_precision),
+            time_in_force=TimeInForce.GTC,
+        )
+        self.submit_order(order)
+
+    def sell(self, quantity: float, feed_key: str | None = None) -> None:
+        """Submit a market sell order. Defaults to the first configured feed."""
+        fk = self._resolve_feed_key(feed_key)
+        if not fk:
+            log.error("sell(): no feed key available")
+            return
+        instrument_id = self._instrument_id_for_feed(fk)
+        if instrument_id is None:
+            return
+        instrument = self.cache.instrument(instrument_id)
+        if instrument is None:
+            log.error("sell(): instrument not found for %s", instrument_id)
+            return
+        order = self.order_factory.market(
+            instrument_id=instrument_id,
+            order_side=OrderSide.SELL,
+            quantity=Quantity(quantity, instrument.size_precision),
+            time_in_force=TimeInForce.GTC,
+        )
+        self.submit_order(order)
+
+    def flatten(self, feed_key: str | None = None) -> None:
+        """Close all open positions for the specified feed (defaults to first feed)."""
+        fk = self._resolve_feed_key(feed_key)
+        if not fk:
+            return
+        instrument_id = self._instrument_id_for_feed(fk)
+        if instrument_id is None:
+            return
+        positions = self.cache.positions_open(instrument_id=instrument_id)
+        for pos in positions:
+            self.close_position(pos)
+
+    # ------------------------------------------------------------------
+    # Runner API — called by backtest_runner, not by quant code
+    # ------------------------------------------------------------------
+
+    def _inject_extras(self, feed_key: str, extras: dict[int, dict]) -> None:
+        """
+        Pre-load OHLC+ extras for a feed before the backtest runs.
+
+        Args:
+            feed_key: e.g. "BTC/USDT.BINANCE:3600"
+            extras: {ts_event_ns: {"buy_vol": float|None, "sell_vol": float|None,
+                                   "open_interest": float|None}}
+        """
+        self._extras[feed_key] = extras

sandbox/dexorder/ohlc_client.py

@@ -141,39 +141,6 @@ class OHLCClient:
         # Step 5: Query Iceberg again for complete dataset
         df = self.iceberg.query_ohlc(ticker, period_seconds, start_time, end_time)
 
-        return self._forward_fill_gaps(df, period_seconds)
-
-    def _forward_fill_gaps(self, df: pd.DataFrame, period_seconds: int) -> pd.DataFrame:
-        """
-        Forward-fill interior missing bars by carrying the last known close into
-        open, high, low, and close of any gap bar.
-
-        Only interior gaps (rows already present with null OHLC from the ingestor,
-        or timestamp slots missing between real bars) are filled. Edge gaps (before
-        the first real bar or after the last real bar) are left as-is.
-        """
-        if df.empty:
-            return df
-
-        df = df.sort_index()
-
-        # Identify rows that are gap bars (null close)
-        is_gap = df['close'].isna()
-
-        if not is_gap.any():
-            return df
-
-        # Forward-fill close across gap rows, then copy into open/high/low
-        df['close'] = df['close'].ffill()
-        price_cols = ['open', 'high', 'low']
-        for col in price_cols:
-            if col in df.columns:
-                df[col] = df[col].where(~is_gap, df['close'])
-
-        # Zero out volume for filled gap rows
-        if 'volume' in df.columns:
-            df['volume'] = df['volume'].where(~is_gap, 0.0)
-
         return df
 
     async def __aenter__(self):

sandbox/dexorder/secrets_vault.py

@@ -0,0 +1,71 @@
+"""
+User Secrets Vault
+
+Stores user-owned API keys for live exchange execution. Secured with the user's
+password — Dexorder cannot read these secrets. This is entirely separate from
+secrets.yaml, which holds Dexorder infrastructure credentials (Iceberg, MinIO, etc.).
+
+Currently a stub — raises NotImplementedError on all calls. Will be backed by
+a user-local encrypted store in a future iteration.
+"""
+
+
+class SecretsVault:
+    """
+    Interface for the user secrets vault.
+
+    The vault is secured with the user's own password; the Dexorder platform
+    cannot decrypt or access its contents. This distinguishes it from the
+    system-level secrets.yaml, which stores infrastructure credentials managed
+    by Dexorder operators.
+
+    Typical keys stored here:
+        "BINANCE_API_KEY", "BINANCE_API_SECRET"
+        "COINBASE_API_KEY", "COINBASE_API_SECRET"
+        etc.
+    """
+
+    def get_secret(self, key: str) -> str:
+        """
+        Retrieve a user secret by key.
+
+        Args:
+            key: Identifier for the secret, e.g. "BINANCE_API_KEY"
+
+        Returns:
+            The secret value as a string.
+
+        Raises:
+            NotImplementedError: Always — vault not yet implemented.
+        """
+        raise NotImplementedError(
+            "User secrets vault is not yet implemented. "
+            "Live execution API key management is a future feature."
+        )
+
+    def set_secret(self, key: str, value: str) -> None:
+        """
+        Store a secret in the vault.
+
+        Args:
+            key: Identifier for the secret
+            value: Secret value to store (stored encrypted with user's password)
+
+        Raises:
+            NotImplementedError: Always — vault not yet implemented.
+        """
+        raise NotImplementedError(
+            "User secrets vault is not yet implemented. "
+            "Live execution API key management is a future feature."
+        )
+
+    def delete_secret(self, key: str) -> None:
+        """
+        Remove a secret from the vault.
+
+        Raises:
+            NotImplementedError: Always — vault not yet implemented.
+        """
+        raise NotImplementedError(
+            "User secrets vault is not yet implemented."
+        )

sandbox/dexorder/tools/activate_strategy.py

@@ -0,0 +1,173 @@
+"""
+activate_strategy / deactivate_strategy — start and stop live or paper trading.
+
+paper=True  (default): forward paper trading — strategy runs on live data with
+            simulated fills. No API keys required.
+
+paper=False: live trading — real order execution via user's exchange API keys,
+             retrieved from the user secrets vault. Currently raises
+             NotImplementedError until the vault is implemented.
+
+Full live-data feed streaming for forward testing is TBD (requires a live bar
+source). This module establishes the interface and stubs the runtime loop.
+"""
+
+import json
+import logging
+from typing import Any
+
+log = logging.getLogger(__name__)
+
+# Registry of active strategies: {strategy_name → runtime state dict}
+# In a future implementation this will hold live strategy runners.
+_active_strategies: dict[str, dict] = {}
+
+
+async def activate_strategy(
+    strategy_name: str,
+    feeds: list[dict],
+    allocation: float,
+    paper: bool = True,
+) -> list:
+    """
+    Activate a strategy for live or paper forward trading.
+
+    Args:
+        strategy_name: Display name as saved via python_write("strategy", ...)
+        feeds: List of feed dicts, e.g. [{"symbol": "BTC/USDT.BINANCE", "period_seconds": 3600}]
+        allocation: Capital allocated in quote currency (e.g. 5000.0 USDT)
+        paper: True = paper/simulated fills (default); False = live execution
+
+    Returns:
+        list[TextContent] with JSON:
+        {"status": "activated", "strategy_name": str, "paper": bool, "allocation": float}
+
+        On error:
+        {"error": str}
+    """
+    from mcp.types import TextContent
+
+    def _err(msg: str) -> list:
+        log.error("activate_strategy '%s': %s", strategy_name, msg)
+        return [TextContent(type="text", text=json.dumps({"error": msg}))]
+
+    if strategy_name in _active_strategies:
+        return _err(
+            f"Strategy '{strategy_name}' is already active. "
+            "Call deactivate_strategy first."
+        )
+
+    if not paper:
+        # Live execution requires the user secrets vault for API keys.
+        # The vault is not yet implemented.
+        try:
+            from dexorder.secrets_vault import SecretsVault
+            _vault = SecretsVault()
+            _vault.get_secret("__probe__")  # will raise NotImplementedError
+        except NotImplementedError:
+            return _err(
+                "Live trading (paper=False) requires the user secrets vault, "
+                "which is not yet implemented. Use paper=True for paper forward testing."
+            )
+
+    # Validate feeds
+    if not feeds:
+        return _err("feeds list is empty")
+
+    parsed_feeds: list[tuple[str, int]] = []
+    for f in feeds:
+        sym = f.get("symbol", "")
+        ps = f.get("period_seconds", 3600)
+        if not sym:
+            return _err(f"Feed entry missing 'symbol': {f}")
+        parsed_feeds.append((sym, int(ps)))
+
+    # TODO: Full implementation — start a live/paper trading loop:
+    # 1. Load strategy class from category files
+    # 2. Set up custom indicators via _setup_custom_indicators()
+    # 3. Subscribe to live bar stream for each feed
+    # 4. Initialize paper account (Nautilus SimulatedExchange) or live account
+    # 5. Run strategy event loop (on_bar → evaluate → submit orders)
+    # This requires a live data feed adapter (TBD).
+
+    log.info(
+        "activate_strategy: registering '%s' (paper=%s, allocation=%.2f) — "
+        "live feed loop is TBD",
+        strategy_name, paper, allocation,
+    )
+
+    _active_strategies[strategy_name] = {
+        "strategy_name": strategy_name,
+        "feeds": [{"symbol": t, "period_seconds": p} for t, p in parsed_feeds],
+        "allocation": allocation,
+        "paper": paper,
+        "status": "registered",
+        "pnl": 0.0,
+    }
+
+    payload = {
+        "status":          "activated",
+        "strategy_name":   strategy_name,
+        "paper":           paper,
+        "allocation":      allocation,
+        "feeds":           [{"symbol": t, "period_seconds": p} for t, p in parsed_feeds],
+        "note":            (
+            "Strategy registered. Live data feed streaming is not yet implemented — "
+            "forward trading will begin when the live feed adapter is available."
+        ),
+    }
+    return [TextContent(type="text", text=json.dumps(payload))]
+
+
+async def deactivate_strategy(strategy_name: str) -> list:
+    """
+    Deactivate a running strategy and return its final P&L summary.
+
+    Args:
+        strategy_name: Display name of the active strategy
+
+    Returns:
+        list[TextContent] with JSON:
+        {"status": "deactivated", "strategy_name": str, "final_pnl": float}
+
+        On error:
+        {"error": str}
+    """
+    from mcp.types import TextContent
+
+    def _err(msg: str) -> list:
+        log.error("deactivate_strategy '%s': %s", strategy_name, msg)
+        return [TextContent(type="text", text=json.dumps({"error": msg}))]
+
+    if strategy_name not in _active_strategies:
+        return _err(f"Strategy '{strategy_name}' is not active")
+
+    state = _active_strategies.pop(strategy_name)
+
+    # TODO: Stop the live feed loop and collect final P&L from the running engine.
+    final_pnl = state.get("pnl", 0.0)
+
+    log.info("deactivate_strategy: stopped '%s', final_pnl=%.4f", strategy_name, final_pnl)
+
+    payload = {
+        "status":        "deactivated",
+        "strategy_name": strategy_name,
+        "final_pnl":     final_pnl,
+    }
+    return [TextContent(type="text", text=json.dumps(payload))]
+
+
+async def list_active_strategies() -> list:
+    """
+    Return a list of currently active strategies and their status.
+
+    Returns:
+        list[TextContent] with JSON:
+        {"active_strategies": [{strategy_name, paper, allocation, feeds, pnl}, ...]}
+    """
+    from mcp.types import TextContent
+
+    payload = {
+        "active_strategies": list(_active_strategies.values()),
+    }
+    return [TextContent(type="text", text=json.dumps(payload))]

sandbox/dexorder/tools/backtest_strategy.py

@@ -0,0 +1,163 @@
+"""
+backtest_strategy — run a PandasStrategy against historical OHLC data.
+
+Called directly from the MCP server's async handle_tool_call.
+
+Returns a JSON payload with backtest metrics and equity curve, following the
+same pattern as evaluate_indicator.py.
+"""
+
+import json
+import logging
+from pathlib import Path
+from typing import Any
+
+log = logging.getLogger(__name__)
+
+# All OHLC+ columns to request from the DataAPI
+_OHLC_EXTRA_COLUMNS = ["volume", "buy_vol", "sell_vol", "open_interest"]
+
+
+async def backtest_strategy(
+    strategy_name: str,
+    feeds: list[dict],
+    from_time: Any,
+    to_time: Any,
+    initial_capital: float = 10_000.0,
+    paper: bool = True,
+) -> list:
+    """
+    Load a saved strategy, fetch OHLC+ data for each feed, and run a backtest.
+
+    Args:
+        strategy_name: Display name as saved via python_write("strategy", ...)
+        feeds: List of feed dicts, e.g. [{"symbol": "BTC/USDT.BINANCE", "period_seconds": 3600}]
+        from_time: Backtest start (Unix timestamp or date string)
+        to_time:   Backtest end   (Unix timestamp or date string)
+        initial_capital: Starting balance in quote currency (default 10,000)
+        paper: Always True for historical backtest (flag reserved for forward testing)
+
+    Returns:
+        list[TextContent] with JSON payload:
+        {
+            "strategy_name": str,
+            "feeds": [...],
+            "initial_capital": float,
+            "paper": bool,
+            "total_candles": int,
+            "total_return": float,       # fractional (0.15 = +15%)
+            "sharpe_ratio": float,
+            "max_drawdown": float,       # fractional (0.10 = 10% drawdown)
+            "win_rate": float,
+            "trade_count": int,
+            "equity_curve": [{"timestamp": int, "equity": float}, ...]
+        }
+
+        On error:
+        {"error": str}
+    """
+    from mcp.types import TextContent
+
+    def _err(msg: str) -> list:
+        log.error("backtest_strategy '%s': %s", strategy_name, msg)
+        return [TextContent(type="text", text=json.dumps({"error": msg}))]
+
+    # --- 1. Validate feeds input ---
+    if not feeds:
+        return _err("feeds list is empty — provide at least one {symbol, period_seconds} entry")
+
+    parsed_feeds: list[tuple[str, int]] = []
+    for f in feeds:
+        sym = f.get("symbol", "")
+        ps = f.get("period_seconds", 3600)
+        if not sym:
+            return _err(f"Feed entry missing 'symbol': {f}")
+        parsed_feeds.append((sym, int(ps)))
+
+    # --- 2. Resolve strategy implementation file ---
+    try:
+        from dexorder.tools.python_tools import get_category_manager, sanitize_name
+        category_manager = get_category_manager()
+        safe_name = sanitize_name(strategy_name)
+        impl_path = category_manager.src_dir / "strategy" / safe_name / "implementation.py"
+        if not impl_path.exists():
+            return _err(f"Strategy '{strategy_name}' not found (looked at {impl_path})")
+    except Exception as exc:
+        return _err(f"Failed to locate strategy: {exc}")
+
+    # --- 3. Register custom indicators with pandas-ta ---
+    try:
+        from dexorder.nautilus.backtest_runner import _setup_custom_indicators
+        _setup_custom_indicators(category_manager.src_dir)
+    except Exception as exc:
+        log.warning("backtest_strategy: custom indicator setup failed: %s", exc)
+
+    # --- 4. Load strategy class ---
+    try:
+        from dexorder.nautilus.backtest_runner import _load_strategy_class
+        strategy_class = _load_strategy_class(impl_path)
+    except Exception as exc:
+        log.exception("backtest_strategy: strategy load failed")
+        return _err(f"Strategy load failed: {exc}")
+
+    # --- 5. Fetch OHLC+ data for each feed ---
+    try:
+        from dexorder.api import get_api
+        api = get_api()
+    except Exception as exc:
+        return _err(f"API not available: {exc}")
+
+    ohlc_dfs: dict[str, Any] = {}
+    total_candles = 0
+
+    for ticker, period_seconds in parsed_feeds:
+        from dexorder.nautilus.pandas_strategy import make_feed_key
+        feed_key = make_feed_key(ticker, period_seconds)
+        try:
+            df = await api.data.historical_ohlc(
+                ticker=ticker,
+                period_seconds=period_seconds,
+                start_time=from_time,
+                end_time=to_time,
+                extra_columns=_OHLC_EXTRA_COLUMNS,
+            )
+        except Exception as exc:
+            log.exception("backtest_strategy: OHLC fetch failed for %s", feed_key)
+            return _err(f"OHLC fetch failed for {feed_key}: {exc}")
+
+        if df.empty:
+            return _err(f"No OHLC data for {feed_key} in the requested range")
+
+        ohlc_dfs[feed_key] = df
+        total_candles += len(df)
+
+    # --- 6. Run backtest in thread executor (BacktestEngine is synchronous) ---
+    try:
+        import asyncio
+        from dexorder.nautilus.backtest_runner import run_backtest
+
+        loop = asyncio.get_event_loop()
+        metrics = await loop.run_in_executor(
+            None,
+            lambda: run_backtest(
+                strategy_class=strategy_class,
+                feeds=parsed_feeds,
+                ohlc_dfs=ohlc_dfs,
+                initial_capital=initial_capital,
+                paper=paper,
+            ),
+        )
+    except Exception as exc:
+        log.exception("backtest_strategy: backtest run failed")
+        return _err(f"Backtest failed: {exc}")
+
+    # --- 7. Return results ---
+    payload = {
+        "strategy_name":  strategy_name,
+        "feeds": [{"symbol": t, "period_seconds": p} for t, p in parsed_feeds],
+        "initial_capital": initial_capital,
+        "paper":           paper,
+        "total_candles":   total_candles,
+        **metrics,
+    }
+    return [TextContent(type="text", text=json.dumps(payload))]

sandbox/dexorder/tools/evaluate_indicator.py

@@ -0,0 +1,243 @@
+"""
+evaluate_indicator — runs a pandas-ta (or custom) indicator against real OHLC data.
+
+Called directly from the MCP server's async handle_tool_call, so it can await
+the DataAPI without subprocess overhead.
+
+Returns a JSON object with a `values` array of {timestamp, ...} records, where
+timestamp is a Unix second integer and value fields hold floats (or null for NaN).
+"""
+
+import json
+import logging
+from pathlib import Path
+from typing import Any
+
+import pandas as pd
+
+log = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Input routing — which series each pandas-ta function expects
+# ---------------------------------------------------------------------------
+
+# Maps pandas_ta_name → tuple of column names from the OHLCV dataframe
+# Columns available: open, high, low, close, volume
+# "volume" is fetched via extra_columns=["volume"]
+
+_INPUTS: dict[str, tuple[str, ...]] = {
+    # Close only
+    "sma": ("close",),
+    "ema": ("close",),
+    "wma": ("close",),
+    "dema": ("close",),
+    "tema": ("close",),
+    "trima": ("close",),
+    "kama": ("close",),
+    "t3": ("close",),
+    "hma": ("close",),
+    "alma": ("close",),
+    "midpoint": ("close",),
+    "rsi": ("close",),
+    "macd": ("close",),
+    "mom": ("close",),
+    "roc": ("close",),
+    "trix": ("close",),
+    "cmo": ("close",),
+    "ao": ("high", "low"),   # ao uses midprice (high, low)
+    "apo": ("close",),
+    "coppock": ("close",),
+    "dpo": ("close",),
+    "fisher": ("high", "low"),
+    "rvgi": ("open", "high", "low", "close"),
+    "kst": ("close",),
+    "stdev": ("close",),
+    "linreg": ("close",),
+    "slope": ("close",),
+    "vwma": ("close", "volume"),
+    "obv": ("close", "volume"),
+    "pvt": ("close", "volume"),
+    "efi": ("close", "volume"),
+    # High + Low
+    "hl2": ("high", "low"),
+    "midprice": ("high", "low"),
+    # High + Low + Close
+    "hlc3": ("high", "low", "close"),
+    "atr": ("high", "low", "close"),
+    "kc": ("high", "low", "close"),
+    "donchian": ("high", "low", "close"),
+    "stoch": ("high", "low", "close"),
+    "stochrsi": ("high", "low", "close"),
+    "cci": ("high", "low", "close"),
+    "willr": ("high", "low", "close"),
+    "adx": ("high", "low", "close"),
+    "aroon": ("high", "low", "close"),
+    "uo": ("high", "low", "close"),
+    "psar": ("high", "low", "close"),
+    "vortex": ("high", "low", "close"),
+    "chop": ("high", "low", "close"),
+    "supertrend": ("high", "low", "close"),
+    "ichimoku": ("high", "low", "close"),
+    # Open + High + Low + Close
+    "ohlc4": ("open", "high", "low", "close"),
+    "bop": ("open", "high", "low", "close"),
+    # High + Low + Close + Volume
+    "mfi": ("high", "low", "close", "volume"),
+    "ad": ("high", "low", "close", "volume"),
+    "adosc": ("high", "low", "close", "volume"),
+    "cmf": ("high", "low", "close", "volume"),
+    "eom": ("high", "low", "close", "volume"),
+    "kvo": ("high", "low", "close", "volume"),
+    # VWAP needs datetime index — handled specially
+    "vwap": ("high", "low", "close", "volume"),
+}
+
+_NEEDS_VOLUME = {name for name, cols in _INPUTS.items() if "volume" in cols}
+
+
+async def evaluate_indicator(
+    symbol: str,
+    from_time: Any,
+    to_time: Any,
+    period_seconds: int,
+    pandas_ta_name: str,
+    parameters: dict,
+) -> list:
+    """
+    Fetch OHLC data and evaluate a pandas-ta indicator.
+
+    Returns a list containing a single MCP TextContent with JSON:
+    {
+        "symbol": ...,
+        "period_seconds": ...,
+        "pandas_ta_name": ...,
+        "parameters": {...},
+        "candle_count": N,
+        "columns": ["timestamp", "value"] or ["timestamp", "col1", "col2", ...],
+        "values": [{"timestamp": <unix_s>, "value": <float|null>}, ...]
+    }
+    """
+    from mcp.types import TextContent
+
+    try:
+        import pandas_ta as ta
+    except ImportError:
+        return [TextContent(type="text", text=json.dumps({"error": "pandas_ta not installed"}))]
+
+    name_lower = pandas_ta_name.lower()
+
+    # For custom indicators, register them with pandas-ta first, then resolve
+    # input columns from their stored metadata.
+    if name_lower.startswith("custom_"):
+        import os
+        from dexorder.tools.python_tools import setup_custom_indicators, get_category_manager
+        setup_custom_indicators(Path(os.environ.get("DATA_DIR", "data")))
+
+        fn = getattr(ta, name_lower, None)
+        if fn is None:
+            return [TextContent(type="text", text=json.dumps({
+                "error": (
+                    f"Custom indicator '{pandas_ta_name}' not found after registering "
+                    "custom indicators. Make sure the indicator was created with "
+                    "python_write(category='indicator', name='...') and that its "
+                    "implementation.py defines a function matching the sanitized name."
+                )
+            }))]
+
+        # Get input_series from the indicator's metadata
+        indicator_name = pandas_ta_name[len("custom_"):]
+        mgr = get_category_manager()
+        read_result = mgr.read("indicator", indicator_name)
+        if read_result.get("exists") and read_result.get("metadata"):
+            raw_series = read_result["metadata"].get("input_series") or ["close"]
+            input_cols = tuple(raw_series)
+        else:
+            input_cols = ("close",)
+    else:
+        # Look up the pandas-ta function for built-in indicators
+        fn = getattr(ta, name_lower, None)
+        if fn is None:
+            return [TextContent(type="text", text=json.dumps({
+                "error": f"Unknown indicator '{pandas_ta_name}'. Check pandas_ta_name against the supported list."
+            }))]
+
+        # Determine required columns
+        input_cols = _INPUTS.get(name_lower, ("close",))
+    needs_volume = "volume" in input_cols
+
+    # Fetch OHLC
+    try:
+        from dexorder.api import get_api
+        api = get_api()
+        df = await api.data.historical_ohlc(
+            ticker=symbol,
+            period_seconds=period_seconds,
+            start_time=from_time,
+            end_time=to_time,
+            extra_columns=["volume"] if needs_volume else [],
+        )
+    except Exception as exc:
+        log.exception("evaluate_indicator: OHLC fetch failed")
+        return [TextContent(type="text", text=json.dumps({"error": f"OHLC fetch failed: {exc}"}))]
+
+    if df.empty:
+        return [TextContent(type="text", text=json.dumps({
+            "error": f"No OHLC data for {symbol} in the requested range"
+        }))]
+
+    # VWAP already requires a DatetimeIndex — the OHLC df index is already a
+    # DatetimeIndex, so no extra work needed here.
+
+    # Build positional args
+    args = []
+    for col in input_cols:
+        if col not in df.columns:
+            return [TextContent(type="text", text=json.dumps({
+                "error": f"Column '{col}' not in fetched dataframe (columns: {list(df.columns)})"
+            }))]
+        args.append(df[col])
+
+    # Compute
+    try:
+        result = fn(*args, **parameters)
+    except Exception as exc:
+        log.exception("evaluate_indicator: computation failed")
+        return [TextContent(type="text", text=json.dumps({
+            "error": f"Indicator computation failed: {exc}"
+        }))]
+
+    # Convert DatetimeIndex → Unix seconds
+    timestamps = (df.index.astype("int64") // 1_000_000_000).tolist()
+
+    # Serialize output
+    if isinstance(result, pd.DataFrame):
+        columns = ["timestamp"] + list(result.columns)
+        values = []
+        for i, ts in enumerate(timestamps):
+            row: dict[str, Any] = {"timestamp": int(ts)}
+            for col in result.columns:
+                v = result.iloc[i][col]
+                row[col] = None if (isinstance(v, float) and pd.isna(v)) else float(v)
+            values.append(row)
+    elif isinstance(result, pd.Series):
+        columns = ["timestamp", "value"]
+        values = [
+            {"timestamp": int(ts), "value": None if pd.isna(v) else float(v)}
+            for ts, v in zip(timestamps, result.tolist())
+        ]
+    else:
+        return [TextContent(type="text", text=json.dumps({
+            "error": f"Unexpected indicator output type: {type(result).__name__}"
+        }))]
+
+    payload = {
+        "symbol": symbol,
+        "period_seconds": period_seconds,
+        "pandas_ta_name": pandas_ta_name,
+        "parameters": parameters,
+        "candle_count": len(df),
+        "columns": columns,
+        "values": values,
+    }
+
+    return [TextContent(type="text", text=json.dumps(payload))]

sandbox/dexorder/tools/indicator_harness.py

@@ -0,0 +1,182 @@
+#!/usr/bin/env python3
+"""
+Indicator harness — tests a custom indicator against synthetic OHLC data.
+
+Runs in a subprocess so the indicator code is isolated from the MCP server process.
+
+Usage: python indicator_harness.py <impl_path> <metadata_path>
+
+Outputs JSON to stdout:
+{
+    "success": bool,
+    "output": str,   # human-readable summary of the indicator output
+    "error": str     # error message / traceback if failed (null on success)
+}
+"""
+import importlib.util
+import json
+import os
+import sys
+import traceback
+import types
+from pathlib import Path
+
+# Ensure dexorder package is importable (same as research_harness.py)
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent.parent))
+
+
+# ---------------------------------------------------------------------------
+# Synthetic OHLCV data — 200 deterministic bars, no network required
+# ---------------------------------------------------------------------------
+
+def make_synthetic_ohlcv(n: int = 200):
+    import numpy as np
+    import pandas as pd
+
+    rng = np.random.default_rng(42)
+
+    # Realistic BTC-style price random walk
+    returns = rng.normal(0, 0.015, n)
+    closes = 40_000.0 * np.cumprod(1.0 + returns)
+
+    opens = np.empty(n)
+    opens[0] = closes[0]
+    opens[1:] = closes[:-1]  # open = previous close
+
+    noise = np.abs(rng.normal(0, 0.005, n))
+    highs = np.maximum(opens, closes) * (1.0 + noise)
+    lows  = np.minimum(opens, closes) * (1.0 - noise)
+    volumes = rng.uniform(1e6, 1e8, n)
+
+    return pd.DataFrame({
+        "open":   opens,
+        "high":   highs,
+        "low":    lows,
+        "close":  closes,
+        "volume": volumes,
+    })
+
+
+def summarize(result, n: int) -> str:
+    import pandas as pd
+
+    if isinstance(result, pd.Series):
+        nan_count = int(result.isna().sum())
+        valid = result.dropna()
+        sample = [round(float(v), 4) for v in valid.tail(5).values] if len(valid) else []
+        return (
+            f"Series({n} bars), NaN: {nan_count}/{n}, "
+            f"last 5 valid values: {sample}"
+        )
+    elif isinstance(result, pd.DataFrame):
+        cols = list(result.columns)
+        nan_counts = {c: int(result[c].isna().sum()) for c in cols}
+        sample = {}
+        for col in cols:
+            valid = result[col].dropna()
+            if len(valid):
+                sample[col] = [round(float(v), 4) for v in valid.tail(3).values]
+        return (
+            f"DataFrame({n} bars × {len(cols)} cols {cols}), "
+            f"NaN counts: {nan_counts}, last 3 valid per col: {sample}"
+        )
+    else:
+        return f"Unexpected return type: {type(result).__name__}"
+
+
+def main():
+    if len(sys.argv) < 3:
+        print(json.dumps({"success": False, "error": "Usage: indicator_harness.py <impl_path> <metadata_path>"}))
+        sys.exit(1)
+
+    impl_path = sys.argv[1]
+    metadata_path = sys.argv[2]
+
+    # --- Load metadata ---
+    input_series = ["close"]
+    parameters: dict = {}
+    try:
+        with open(metadata_path) as f:
+            meta = json.load(f)
+        input_series = meta.get("input_series") or ["close"]
+        param_schema = meta.get("parameters") or {}
+        for pname, pinfo in param_schema.items():
+            if isinstance(pinfo, dict) and "default" in pinfo:
+                parameters[pname] = pinfo["default"]
+            elif not isinstance(pinfo, dict):
+                # bare value (legacy)
+                parameters[pname] = pinfo
+    except Exception as e:
+        print(json.dumps({"success": False, "error": f"Failed to read metadata: {e}"}))
+        sys.exit(0)
+
+    # --- Generate synthetic data ---
+    try:
+        import numpy  # noqa: F401 — verify numpy available
+        import pandas as pd
+    except ImportError as e:
+        print(json.dumps({"success": False, "error": f"Missing required package: {e}"}))
+        sys.exit(0)
+
+    df = make_synthetic_ohlcv(n=200)
+    n = len(df)
+
+    # --- Load implementation ---
+    try:
+        spec = importlib.util.spec_from_file_location("_indicator_impl", impl_path)
+        module = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(module)  # type: ignore[union-attr]
+    except Exception:
+        tb = traceback.format_exc()
+        print(json.dumps({"success": False, "error": f"Import failed:\n{tb}"}))
+        sys.exit(0)
+
+    # --- Find the indicator function ---
+    # Prefer a function whose name matches the sanitized directory name,
+    # fall back to the first public function in the module.
+    fn_name = os.path.basename(os.path.dirname(impl_path)).lower()
+    fn = getattr(module, fn_name, None)
+    if fn is None:
+        candidates = [
+            v for k, v in vars(module).items()
+            if isinstance(v, types.FunctionType) and not k.startswith("_")
+        ]
+        fn = candidates[0] if candidates else None
+
+    if fn is None:
+        print(json.dumps({"success": False, "error": "No callable function found in implementation.py"}))
+        sys.exit(0)
+
+    # --- Build positional args from input_series ---
+    args = []
+    for col in input_series:
+        if col not in df.columns:
+            print(json.dumps({"success": False, "error": f"input_series '{col}' not in synthetic df columns {list(df.columns)}"}))
+            sys.exit(0)
+        args.append(df[col])
+
+    # --- Execute ---
+    try:
+        result = fn(*args, **parameters)
+    except Exception:
+        tb = traceback.format_exc()
+        print(json.dumps({"success": False, "error": f"Execution failed:\n{tb}"}))
+        sys.exit(0)
+
+    # --- Validate output type ---
+    if not isinstance(result, (pd.Series, pd.DataFrame)):
+        print(json.dumps({
+            "success": False,
+            "error": (
+                f"Indicator must return pd.Series or pd.DataFrame, "
+                f"got {type(result).__name__}. "
+                "Wrap the output if using pandas-ta internally."
+            ),
+        }))
+        sys.exit(0)
+
+    print(json.dumps({"success": True, "output": summarize(result, n)}))
+
+
+if __name__ == "__main__":
+    main()

sandbox/dexorder/tools/python_tools.py

@@ -30,8 +30,9 @@ from typing import Any, Optional
 
 log = logging.getLogger(__name__)
 
-# Path to the research harness script (written to disk, not inline)
+# Path to the harness scripts (written to disk, not inline)
 _RESEARCH_HARNESS = Path(__file__).parent / "research_harness.py"
+_INDICATOR_HARNESS = Path(__file__).parent / "indicator_harness.py"
 
 # Import conda manager for package installation
 try:
@@ -62,12 +63,15 @@ class BaseMetadata:
 @dataclass
 class StrategyMetadata(BaseMetadata):
     """Metadata for trading strategies."""
-    data_feeds: list[str] = None  # Required data feeds (e.g., ["BTC/USD", "ETH/USD"])
+    data_feeds: list[dict] = None  # Required data feeds: [{"symbol": "BTC/USDT.BINANCE", "period_seconds": 3600, "description": "..."}]
+    parameters: dict = None  # Strategy parameters: {"param_name": {"default": value, "description": "..."}}
     conda_packages: list[str] = None  # Additional conda packages required
 
     def __post_init__(self):
         if self.data_feeds is None:
             self.data_feeds = []
+        if self.parameters is None:
+            self.parameters = {}
         if self.conda_packages is None:
             self.conda_packages = []
 
@@ -75,12 +79,78 @@ class StrategyMetadata(BaseMetadata):
 @dataclass
 class IndicatorMetadata(BaseMetadata):
     """Metadata for technical indicators."""
-    default_length: int = 14  # Default period/length parameter
     conda_packages: list[str] = None  # Additional conda packages required
 
+    # Fields for TradingView custom study auto-construction:
+    parameters: dict = None
+    # Parameter schema: {param_name: {type: "int"|"float"|"bool"|"string",
+    #   default: value, description?: str, min?: num, max?: num}}
+    # Example: {"length": {"type": "int", "default": 14, "min": 1, "max": 500}}
+
+    input_series: list = None
+    # OHLCV columns the indicator function receives as positional args.
+    # Valid values: "open", "high", "low", "close", "volume"
+    # Example: ["close"] or ["high", "low", "close", "volume"]
+
+    output_columns: list = None
+    # Output series produced by the function. Each entry:
+    #   {
+    #     name: str,           # column name (or "value" for plain Series)
+    #     display_name?: str,  # label shown in TV legend
+    #     description?: str,
+    #     plot?: {
+    #       style: int,        # LineStudyPlotStyle: 0=Line, 1=Histogram, 3=Dots/Cross,
+    #                          #   4=Area, 5=Columns, 6=Circles, 9=StepLine
+    #       color?: str,       # CSS hex e.g. "#2196F3" (auto-assigned if omitted)
+    #       linewidth?: int,   # 1–4 (default 2)
+    #       visible?: bool     # default true
+    #     }
+    #   }
+    # Example (single line):   [{"name": "value", "display_name": "My Indicator"}]
+    # Example (multi-line):    [{"name": "upper", "plot": {"style": 0}}, {"name": "lower", "plot": {"style": 0}}]
+    # Example (histogram):     [{"name": "value", "plot": {"style": 1}}]
+    # Example (MACD-style):    [{"name": "macd", "plot": {"style": 0}}, {"name": "signal", "plot": {"style": 0}}, {"name": "hist", "plot": {"style": 1}}]
+
+    pane: str = "separate"
+    # Where to render: "price" (overlaid on candles) or "separate" (sub-pane)
+
+    filled_areas: list = None
+    # Optional shaded regions between two plots or two bands. Each entry:
+    #   {
+    #     id: str,           # unique id e.g. "fill_upper_lower"
+    #     type: str,         # "plot_plot" (between two series) or "hline_hline" (between two bands)
+    #     series1: str,      # output_column name (for plot_plot) or band id (for hline_hline)
+    #     series2: str,
+    #     color?: str,       # CSS hex fill color (default semi-transparent blue)
+    #     opacity?: float    # 0.0–1.0 (default 0.1)
+    #   }
+    # Example (Bollinger fill): [{"id": "fill", "type": "plot_plot", "series1": "upper", "series2": "lower", "color": "#2196F3", "opacity": 0.1}]
+
+    bands: list = None
+    # Optional horizontal reference lines (e.g. RSI overbought/oversold). Each entry:
+    #   {
+    #     id: str,          # unique id e.g. "ob"
+    #     value: float,     # fixed y-level
+    #     color?: str,      # CSS hex (default "#787B86")
+    #     linewidth?: int,  # default 1
+    #     linestyle?: int,  # 0=solid, 1=dotted, 2=dashed (default 2)
+    #     visible?: bool    # default true
+    #   }
+    # Example (RSI levels): [{"id": "ob", "value": 70}, {"id": "os", "value": 30}]
+
     def __post_init__(self):
         if self.conda_packages is None:
             self.conda_packages = []
+        if self.input_series is None:
+            self.input_series = ["close"]
+        if self.output_columns is None:
+            self.output_columns = [{"name": "value"}]
+        if self.parameters is None:
+            self.parameters = {}
+        if self.filled_areas is None:
+            self.filled_areas = []
+        if self.bands is None:
+            self.bands = []
 
 
 @dataclass
@@ -141,21 +211,212 @@ def get_category_path(data_dir: Path, category: Category, name: str) -> Path:
     return data_dir / category.value / safe_name
 
 
+# =============================================================================
+# Git Manager
+# =============================================================================
+
+class GitManager:
+    """
+    Thin wrapper around git subprocess calls for category revision tracking.
+    All operations are non-fatal: errors are logged as warnings.
+    """
+
+    def __init__(self, repo_dir: Path):
+        self.repo_dir = repo_dir
+
+    def _run(self, *args, check: bool = True) -> subprocess.CompletedProcess:
+        return subprocess.run(
+            ["git"] + list(args),
+            cwd=self.repo_dir,
+            capture_output=True,
+            text=True,
+            check=check,
+        )
+
+    def ensure_init(self):
+        """Init git repo if not exists; initial commit if files already present."""
+        if (self.repo_dir / ".git").exists():
+            return
+        self._run("init", "-b", "main")
+        self._run("config", "user.email", "sandbox@dexorder.ai")
+        self._run("config", "user.name", "Dexorder Sandbox")
+        # Commit any pre-existing files (migrated from old layout)
+        status = self._run("status", "--porcelain")
+        if status.stdout.strip():
+            self._run("add", "-A")
+            self._run("commit", "-m", "init: migrate existing category files")
+        log.info(f"Git repo initialized at {self.repo_dir}")
+
+    def commit(self, message: str) -> Optional[str]:
+        """Stage all changes and commit. Returns short hash or None if nothing to commit / on error."""
+        try:
+            self._run("add", "-A")
+            status = self._run("status", "--porcelain")
+            if not status.stdout.strip():
+                return None  # nothing changed
+            self._run("commit", "-m", message)
+            result = self._run("rev-parse", "--short", "HEAD")
+            return result.stdout.strip()
+        except Exception as e:
+            log.warning(f"Git commit failed (non-fatal): {e}")
+            return None
+
+    def log(self, path: Optional[Path] = None, n: int = 20) -> list[dict]:
+        """Return recent commits, optionally filtered to a path."""
+        cmd = ["log", f"-{n}", "--pretty=format:%H|%h|%s|%ai"]
+        if path:
+            cmd += ["--", str(path.relative_to(self.repo_dir))]
+        result = self._run(*cmd, check=False)
+        entries = []
+        for line in result.stdout.strip().splitlines():
+            if line:
+                parts = line.split("|", 3)
+                if len(parts) == 4:
+                    entries.append({
+                        "hash": parts[0],
+                        "short_hash": parts[1],
+                        "message": parts[2],
+                        "date": parts[3],
+                    })
+        return entries
+
+    def restore(self, revision: str, path: Optional[Path] = None) -> Optional[str]:
+        """Restore path (or entire tree) to revision state. Returns new commit hash."""
+        try:
+            rel = str(path.relative_to(self.repo_dir)) if path else "."
+            self._run("checkout", revision, "--", rel)
+            return self.commit(f"revert: restore to {revision[:8]}")
+        except subprocess.CalledProcessError as e:
+            raise RuntimeError(e.stderr.strip()) from e
+
+
+# =============================================================================
+# Custom Indicator Setup
+# =============================================================================
+
+def setup_custom_indicators(data_dir: Path) -> None:
+    """
+    Register user's custom indicators with pandas-ta.
+
+    Loads each indicator's implementation.py directly via importlib and binds
+    the function as ``ta.custom_{sanitized_name}`` so that evaluate_indicator
+    can call it as ``getattr(ta, "custom_trendflex", None)``.
+
+    The binding is idempotent — indicators already registered are skipped.
+
+    Note: pandas-ta's ta.import_dir() requires a category-based directory
+    structure (e.g. tmpdir/momentum/trendflex.py) plus a companion
+    ``{name}_method`` function.  Our indicators don't follow that convention,
+    so we bind directly instead.
+    """
+    try:
+        import pandas_ta as ta
+    except ImportError:
+        log.warning("pandas-ta not available — custom indicators will not be registered")
+        return
+
+    src_dir = data_dir / "src"
+    indicator_root = src_dir / "indicator"
+    if not indicator_root.exists():
+        return
+
+    import importlib.util
+    import types
+
+    # Track which sanitized names we've seen to handle duplicate directories
+    # (e.g. "TrendFlex" and "trendflex" both sanitise to "custom_trendflex").
+    seen: set[str] = set()
+    registered = 0
+
+    # Sort so that exact-lowercase names (e.g. "trendflex") come before mixed-case
+    # variants (e.g. "TrendFlex") — when duplicates exist the lowercase one wins.
+    for ind_dir in sorted(indicator_root.iterdir(), key=lambda p: (p.name != p.name.lower(), p.name.lower())):
+        if not ind_dir.is_dir():
+            continue
+        impl = ind_dir / "implementation.py"
+        if not impl.exists():
+            continue
+
+        sanitized = ind_dir.name.lower().replace("-", "_").replace(" ", "_")
+        ta_name = f"custom_{sanitized}"
+
+        if ta_name in seen:
+            log.warning(
+                "Duplicate custom indicator name '%s' from directory '%s' — skipping",
+                ta_name, ind_dir.name,
+            )
+            continue
+        seen.add(ta_name)
+
+        # Skip if already bound (e.g. called multiple times in a process)
+        if getattr(ta, ta_name, None) is not None:
+            continue
+
+        try:
+            spec = importlib.util.spec_from_file_location(ta_name, impl)
+            module = importlib.util.module_from_spec(spec)
+            spec.loader.exec_module(module)  # type: ignore[union-attr]
+
+            # Find the callable: prefer the function whose name matches the
+            # sanitized directory name, fall back to any top-level function.
+            fn = getattr(module, sanitized, None)
+            if fn is None:
+                candidates = [
+                    v for k, v in vars(module).items()
+                    if isinstance(v, types.FunctionType) and not k.startswith("_")
+                ]
+                fn = candidates[0] if candidates else None
+
+            if fn is None:
+                log.warning("No callable found in %s — skipping", impl)
+                continue
+
+            setattr(ta, ta_name, fn)
+            registered += 1
+            log.debug("Registered custom indicator '%s' from %s", ta_name, impl)
+
+        except Exception:
+            log.warning("Could not register indicator '%s':", ind_dir.name, exc_info=True)
+
+    if registered > 0:
+        log.info("Registered %d custom indicator(s) with pandas-ta", registered)
+
+
 # =============================================================================
 # Category File Manager
 # =============================================================================
 
 class CategoryFileManager:
     """
-    Manages category-based file operations with validation.
+    Manages category-based file operations with validation and git revision tracking.
+    Category files live under {data_dir}/src/ which is the git repo root.
+    Workspace and other ephemeral data remain under {data_dir}/ but outside the repo.
     """
 
     def __init__(self, data_dir: Path):
         self.data_dir = data_dir
 
-        # Ensure category directories exist
-        for category in Category:
-            (data_dir / category.value).mkdir(parents=True, exist_ok=True)
+        src = self.src_dir
+        src.mkdir(parents=True, exist_ok=True)
+
+        # Migrate: move existing top-level category dirs into src/ (one-time)
+        for cat in Category:
+            old = data_dir / cat.value
+            new = src / cat.value
+            if old.exists() and not new.exists():
+                old.rename(new)
+                log.info(f"Migrated {old} → {new}")
+            else:
+                new.mkdir(exist_ok=True)
+
+        # Init git repo in src/
+        self.git = GitManager(src)
+        self.git.ensure_init()
+
+    @property
+    def src_dir(self) -> Path:
+        """Root of the versioned category code (git repo root)."""
+        return self.data_dir / "src"
 
     def write(
         self,
@@ -191,7 +452,7 @@ class CategoryFileManager:
             }
 
         # Get item directory
-        item_dir = get_category_path(self.data_dir, cat, name)
+        item_dir = get_category_path(self.src_dir, cat, name)
         item_dir.mkdir(parents=True, exist_ok=True)
 
         # Write implementation
@@ -228,11 +489,19 @@ class CategoryFileManager:
             "validation": validation,
         }
 
-        # Auto-execute research scripts after successful write
-        if cat == Category.RESEARCH and validation["success"]:
-            log.info(f"Auto-executing research script: {name}")
-            execution_result = self.execute_research(name)
-            result["execution"] = execution_result
+        # Auto-execute after successful write to give the agent immediate runtime feedback
+        if validation["success"]:
+            if cat == Category.RESEARCH:
+                log.info(f"Auto-executing research script: {name}")
+                result["execution"] = self.execute_research(name)
+            elif cat == Category.INDICATOR:
+                log.info(f"Auto-executing indicator test: {name}")
+                result["execution"] = self._execute_indicator(item_dir)
+
+        # Commit to git
+        commit_hash = self.git.commit(f"create({category}): {name}")
+        if commit_hash:
+            result["revision"] = commit_hash
 
         return result
 
@@ -241,6 +510,7 @@ class CategoryFileManager:
         category: str,
         name: str,
         code: Optional[str] = None,
+        patches: Optional[list[dict]] = None,
         description: Optional[str] = None,
         metadata: Optional[dict] = None
     ) -> dict[str, Any]:
@@ -250,7 +520,8 @@ class CategoryFileManager:
         Args:
             category: Category name
             name: Display name for the item
-            code: Python implementation code (optional, omit to keep existing)
+            code: Full Python implementation code to replace existing (optional)
+            patches: List of {old_string, new_string} replacements (optional, preferred for small changes)
             description: Updated description (optional, omit to keep existing)
             metadata: Additional metadata updates (optional)
 
@@ -261,12 +532,15 @@ class CategoryFileManager:
             - validation: dict - results from test harness (if code updated)
             - error: str (if any)
         """
+        if code is not None and patches is not None:
+            return {"success": False, "error": "Provide either 'code' or 'patches', not both"}
+
         try:
             cat = Category(category)
         except ValueError:
             return {"success": False, "error": f"Invalid category '{category}'"}
 
-        item_dir = get_category_path(self.data_dir, cat, name)
+        item_dir = get_category_path(self.src_dir, cat, name)
 
         if not item_dir.exists():
             return {"success": False, "error": f"Item '{name}' does not exist in category '{category}'"}
@@ -282,8 +556,28 @@ class CategoryFileManager:
         except Exception as e:
             return {"success": False, "error": f"Failed to read existing metadata: {e}"}
 
-        # Update code if provided
-        if code is not None:
+        # Apply string-replacement patches if provided
+        if patches is not None:
+            if not impl_path.exists():
+                return {"success": False, "error": "Cannot patch: implementation file does not exist"}
+            try:
+                current_code = impl_path.read_text()
+                for i, patch in enumerate(patches):
+                    old = patch.get("old_string", "")
+                    new = patch.get("new_string", "")
+                    if old not in current_code:
+                        return {"success": False, "error": f"Patch {i}: old_string not found in file"}
+                    if current_code.count(old) > 1:
+                        return {"success": False, "error": f"Patch {i}: old_string is not unique — add more surrounding context"}
+                    current_code = current_code.replace(old, new, 1)
+                impl_path.write_text(current_code)
+                log.info(f"Applied {len(patches)} patch(es) to {impl_path}")
+                code = current_code  # trigger validation below
+            except Exception as e:
+                return {"success": False, "error": f"Failed to apply patches: {e}"}
+
+        # Update code if provided (full replace)
+        if code is not None and patches is None:
             try:
                 impl_path.write_text(code)
                 log.info(f"Updated {cat.value} implementation: {impl_path}")
@@ -321,11 +615,21 @@ class CategoryFileManager:
             result["validation"] = validation
             result["success"] = validation["success"]
 
-        # Auto-execute research scripts after successful edit (if code was updated)
-        if cat == Category.RESEARCH and code is not None and result["success"]:
-            log.info(f"Auto-executing research script after edit: {name}")
-            execution_result = self.execute_research(name)
-            result["execution"] = execution_result
+        # Auto-execute after successful edit to give the agent immediate runtime feedback
+        if code is not None and result["success"]:
+            if cat == Category.RESEARCH:
+                log.info(f"Auto-executing research script after edit: {name}")
+                result["execution"] = self.execute_research(name)
+            elif cat == Category.INDICATOR:
+                log.info(f"Auto-executing indicator test after edit: {name}")
+                result["execution"] = self._execute_indicator(item_dir)
+
+        # Commit to git if code changed
+        if code is not None and result["success"]:
+            action = "patch" if patches is not None else "edit"
+            commit_hash = self.git.commit(f"{action}({category}): {name}")
+            if commit_hash:
+                result["revision"] = commit_hash
 
         return result
 
@@ -349,7 +653,7 @@ class CategoryFileManager:
         except ValueError:
             return {"exists": False, "error": f"Invalid category '{category}'"}
 
-        item_dir = get_category_path(self.data_dir, cat, name)
+        item_dir = get_category_path(self.src_dir, cat, name)
 
         if not item_dir.exists():
             return {"exists": False}
@@ -385,7 +689,7 @@ class CategoryFileManager:
         except ValueError:
             return {"error": f"Invalid category '{category}'"}
 
-        cat_dir = self.data_dir / cat.value
+        cat_dir = self.src_dir / cat.value
         items = []
 
         for item_dir in cat_dir.iterdir():
@@ -487,33 +791,58 @@ class CategoryFileManager:
 
     def _validate_indicator(self, impl_path: Path) -> dict[str, Any]:
         """
-        Validate an indicator implementation.
+        Validate an indicator by running it against synthetic OHLC data.
 
-        Runs basic syntax check and imports.
+        Uses indicator_harness.py in a subprocess so the indicator code is
+        isolated from the MCP server process. Catches import errors, runtime
+        errors, and wrong return types — not just syntax.
         """
+        meta_path = impl_path.parent / "metadata.json"
+        return self._execute_indicator(impl_path.parent, timeout=30)
+
+    def _execute_indicator(self, item_dir: Path, timeout: int = 30) -> dict[str, Any]:
+        """
+        Run an indicator against synthetic OHLC data via indicator_harness.py.
+
+        Returns:
+            dict with success, output (human-readable summary), error
+        """
+        impl_path = item_dir / "implementation.py"
+        meta_path = item_dir / "metadata.json"
+
+        if not impl_path.exists():
+            return {"success": False, "error": "implementation.py not found"}
+        if not meta_path.exists():
+            return {"success": False, "error": "metadata.json not found"}
+
         try:
             result = subprocess.run(
-                [sys.executable, "-m", "py_compile", str(impl_path)],
+                [sys.executable, str(_INDICATOR_HARNESS), str(impl_path), str(meta_path)],
                 capture_output=True,
                 text=True,
-                timeout=10,
+                timeout=timeout,
+                cwd=str(item_dir),
             )
-
-            if result.returncode == 0:
-                return {
-                    "success": True,
-                    "output": "Indicator syntax valid",
-                }
-            else:
-                return {
-                    "success": False,
-                    "output": result.stderr,
-                    "error": "Syntax error in indicator",
-                }
         except subprocess.TimeoutExpired:
-            return {"success": False, "error": "Validation timeout"}
+            return {"success": False, "error": f"Indicator test timed out after {timeout}s"}
         except Exception as e:
-            return {"success": False, "error": f"Validation failed: {e}"}
+            return {"success": False, "error": f"Harness launch failed: {e}"}
+
+        if result.returncode != 0:
+            return {
+                "success": False,
+                "error": f"Harness process failed:\n{result.stderr}",
+            }
+
+        try:
+            data = json.loads(result.stdout)
+        except json.JSONDecodeError:
+            return {
+                "success": False,
+                "error": f"Harness produced invalid JSON:\n{result.stdout[:500]}",
+            }
+
+        return data
 
     def _run_research_harness(self, impl_path: Path, item_dir: Path, timeout: int = 30) -> dict[str, Any]:
         """
@@ -594,7 +923,7 @@ class CategoryFileManager:
             - content: list of TextContent and ImageContent objects (MCP format)
             - error: str (if any)
         """
-        item_dir = get_category_path(self.data_dir, Category.RESEARCH, name)
+        item_dir = get_category_path(self.src_dir, Category.RESEARCH, name)
 
         if not item_dir.exists():
             return {"error": f"Research script '{name}' does not exist"}
@@ -654,6 +983,66 @@ class CategoryFileManager:
         return {"content": content}
 
 
+    def git_log(
+        self,
+        category: Optional[str] = None,
+        name: Optional[str] = None,
+        limit: int = 20
+    ) -> dict[str, Any]:
+        """
+        List recent git commits, optionally filtered to a category or item.
+
+        Returns:
+            dict with:
+            - success: bool
+            - commits: list of {hash, short_hash, message, date}
+        """
+        path = None
+        if category:
+            try:
+                cat = Category(category)
+            except ValueError:
+                return {"success": False, "error": f"Invalid category '{category}'"}
+            if name:
+                path = get_category_path(self.src_dir, cat, name)
+            else:
+                path = self.src_dir / cat.value
+        entries = self.git.log(path=path, n=limit)
+        return {"success": True, "commits": entries}
+
+    def git_revert(self, revision: str, category: str, name: str) -> dict[str, Any]:
+        """
+        Restore a category item to a previous git revision (creates a new commit).
+
+        Returns:
+            dict with:
+            - success: bool
+            - revision: str - new commit hash
+            - validation: dict
+            - error: str (if any)
+        """
+        try:
+            cat = Category(category)
+        except ValueError:
+            return {"success": False, "error": f"Invalid category '{category}'"}
+
+        item_dir = get_category_path(self.src_dir, cat, name)
+        if not item_dir.exists():
+            return {"success": False, "error": f"Item '{name}' not found in '{category}'"}
+
+        try:
+            commit_hash = self.git.restore(revision, path=item_dir)
+        except RuntimeError as e:
+            return {"success": False, "error": str(e)}
+
+        validation = self._validate(cat, item_dir)
+        return {
+            "success": validation["success"],
+            "revision": commit_hash,
+            "validation": validation,
+        }
+
+
 # =============================================================================
 # Global Manager Instance
 # =============================================================================

sandbox/dexorder/tools/research_harness.py

@@ -3,7 +3,7 @@
 Research script harness - runs implementation.py in a subprocess with API
 initialization, stdout/stderr capture, and matplotlib figure capture.
 
-This file is written to disk and invoked by category_tools.py rather than
+This file is written to disk and invoked by python_tools.py rather than
 being passed inline via `python -c`, so the harness code is inspectable and
 not regenerated on every call.
 
@@ -77,6 +77,16 @@ try:
 except Exception as e:
     print(f"WARNING: API initialization failed: {e}", file=sys.stderr)
 
+# ---------------------------------------------------------------------------
+# Register custom indicators so research scripts can use df.ta.my_indicator()
+# ---------------------------------------------------------------------------
+try:
+    from dexorder.tools.python_tools import setup_custom_indicators
+    _data_dir = Path(os.environ.get("DATA_DIR", "/app/data"))
+    setup_custom_indicators(_data_dir)
+except Exception as e:
+    print(f"WARNING: Custom indicator registration failed: {e}", file=sys.stderr)
+
 
 def main():
     if len(sys.argv) < 2:

sandbox/dexorder/tools/workspace_tools.py

@@ -43,6 +43,23 @@ class WorkspaceStore:
         # Map of "store_name/json/pointer/path" -> list of callbacks
         self._triggers: dict[str, list[Callable[[Any, Any], None]]] = {}
 
+    def _ensure_intermediate_paths(self, state: dict, patch: list[dict]) -> dict:
+        """Create missing intermediate objects for deep patch paths (mirrors gateway logic)."""
+        import copy
+        state = copy.deepcopy(state)
+        for op in patch:
+            if op.get("op") not in ("add", "replace"):
+                continue
+            parts = [p for p in op.get("path", "").split("/") if p]
+            if len(parts) <= 1:
+                continue
+            current = state
+            for part in parts[:-1]:
+                if not isinstance(current.get(part), dict):
+                    current[part] = {}
+                current = current[part]
+        return state
+
     def _store_path(self, store_name: str) -> Path:
         """Get the filesystem path for a store."""
         # Sanitize store name to prevent directory traversal
@@ -136,6 +153,9 @@ class WorkspaceStore:
                 with open(path, "r") as f:
                     old_state = json.load(f)
 
+            # Create missing intermediate objects for deep paths (mirrors gateway logic)
+            old_state = self._ensure_intermediate_paths(old_state, patch)
+
             # Apply patch
             new_state = jsonpatch.apply_patch(old_state, patch)
 

sandbox/environment.yml

@@ -50,3 +50,4 @@ dependencies:
     - starlette>=0.27.0
     - uvicorn>=0.27.0
     - sse-starlette>=1.6.0
+    - nautilus_trader>=1.200.0

sandbox/main.py

@@ -10,6 +10,7 @@ Brings together:
 """
 
 import asyncio
+import contextlib
 import logging
 import os
 import signal
@@ -20,8 +21,8 @@ from typing import Optional
 import uvicorn
 import yaml
 from mcp.server import Server
-from mcp.server.sse import SseServerTransport
 from mcp.server.stdio import stdio_server
+from mcp.server.streamable_http_manager import StreamableHTTPSessionManager
 from mcp.types import Tool, TextContent, ImageContent
 from starlette.applications import Starlette
 from starlette.requests import Request
@@ -34,8 +35,11 @@ from dexorder.conda_manager import sync_packages, install_packages
 from dexorder.events import EventType, UserEvent, DeliverySpec
 from dexorder.impl.charting_api_impl import ChartingAPIImpl
 from dexorder.impl.data_api_impl import DataAPIImpl
-from dexorder.tools.category_tools import get_category_manager
+from dexorder.tools.python_tools import get_category_manager
 from dexorder.tools.workspace_tools import get_workspace_store
+from dexorder.tools.evaluate_indicator import evaluate_indicator
+from dexorder.tools.backtest_strategy import backtest_strategy
+from dexorder.tools.activate_strategy import activate_strategy, deactivate_strategy, list_active_strategies
 
 # =============================================================================
 # Global Data Directory
@@ -249,7 +253,7 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                 }
             ),
             Tool(
-                name="category_write",
+                name="python_write",
                 description="Write a new strategy, indicator, or research script with validation",
                 inputSchema={
                     "type": "object",
@@ -273,15 +277,27 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                         },
                         "metadata": {
                             "type": "object",
-                            "description": "Optional category-specific metadata (e.g., default_length for indicators, data_feeds for strategies)"
+                            "description": (
+                                "Optional category-specific metadata. "
+                                "For strategy: include 'data_feeds' (list of {symbol, period_seconds, description}) "
+                                "and 'parameters' (object mapping param_name → {default, description}). "
+                                "Example: {\"data_feeds\": [{\"symbol\": \"BTC/USDT.BINANCE\", \"period_seconds\": 3600, \"description\": \"Primary BTC/USDT hourly feed\"}], "
+                                "\"parameters\": {\"rsi_length\": {\"default\": 14, \"description\": \"RSI lookback period\"}, \"threshold\": {\"default\": 70, \"description\": \"Overbought level\"}}}. "
+                                "For indicator: include 'default_length' (int). "
+                                "For any category: 'conda_packages' (list of package names) if extra dependencies are needed."
+                            )
                         }
                     },
                     "required": ["category", "name", "description", "code"]
                 }
             ),
             Tool(
-                name="category_edit",
-                description="Edit an existing category script (updates code, description, or metadata)",
+                name="python_edit",
+                description=(
+                    "Edit an existing category script. "
+                    "Use 'patches' for targeted string replacements (preferred for small changes), "
+                    "or 'code' to replace the full implementation. Do not supply both."
+                ),
                 inputSchema={
                     "type": "object",
                     "properties": {
@@ -296,7 +312,24 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                         },
                         "code": {
                             "type": "string",
-                            "description": "Updated Python code (optional, omit to keep existing)"
+                            "description": "Full replacement Python code. Use only when rewriting the entire implementation; prefer 'patches' for targeted edits."
+                        },
+                        "patches": {
+                            "type": "array",
+                            "description": (
+                                "Targeted code edits as old/new string pairs. Preferred over 'code' for small changes. "
+                                "Each patch: {\"old_string\": \"exact text to find\", \"new_string\": \"replacement text\"}. "
+                                "old_string must be unique in the file (add surrounding context if needed). "
+                                "Patches are applied in order."
+                            ),
+                            "items": {
+                                "type": "object",
+                                "properties": {
+                                    "old_string": {"type": "string"},
+                                    "new_string": {"type": "string"}
+                                },
+                                "required": ["old_string", "new_string"]
+                            }
                         },
                         "description": {
                             "type": "string",
@@ -304,14 +337,20 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                         },
                         "metadata": {
                             "type": "object",
-                            "description": "Updated metadata fields (optional)"
+                            "description": (
+                                "Updated metadata fields (optional). "
+                                "For strategy: 'data_feeds' (list of {symbol, period_seconds, description}) "
+                                "and/or 'parameters' (object mapping param_name → {default, description}). "
+                                "For indicator: 'default_length' (int). "
+                                "For any category: 'conda_packages' (list of package names)."
+                            )
                         }
                     },
                     "required": ["category", "name"]
                 }
             ),
             Tool(
-                name="category_read",
+                name="python_read",
                 description="Read a category script and its metadata",
                 inputSchema={
                     "type": "object",
@@ -330,7 +369,7 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                 }
             ),
             Tool(
-                name="category_list",
+                name="python_list",
                 description="List all items in a category with names and descriptions",
                 inputSchema={
                     "type": "object",
@@ -344,6 +383,53 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                     "required": ["category"]
                 }
             ),
+            Tool(
+                name="python_log",
+                description="Show git commit history for category items. Filter by category and/or name to see history for a specific item.",
+                inputSchema={
+                    "type": "object",
+                    "properties": {
+                        "category": {
+                            "type": "string",
+                            "enum": ["strategy", "indicator", "research"],
+                            "description": "Filter to this category (optional)"
+                        },
+                        "name": {
+                            "type": "string",
+                            "description": "Filter to this item (optional, requires category)"
+                        },
+                        "limit": {
+                            "type": "integer",
+                            "description": "Max commits to return (default 20)",
+                            "default": 20
+                        }
+                    },
+                    "required": []
+                }
+            ),
+            Tool(
+                name="python_revert",
+                description="Restore a category item to a previous git revision. Creates a new commit — non-destructive.",
+                inputSchema={
+                    "type": "object",
+                    "properties": {
+                        "revision": {
+                            "type": "string",
+                            "description": "Git commit hash (full or short) to restore to"
+                        },
+                        "category": {
+                            "type": "string",
+                            "enum": ["strategy", "indicator", "research"],
+                            "description": "Category of the item"
+                        },
+                        "name": {
+                            "type": "string",
+                            "description": "Display name of the item to restore"
+                        }
+                    },
+                    "required": ["revision", "category", "name"]
+                }
+            ),
             Tool(
                 name="conda_sync",
                 description="Sync conda packages: scan all metadata, remove unused packages (excluding base environment)",
@@ -381,13 +467,179 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                     },
                     "required": ["name"]
                 }
-            )
+            ),
+            Tool(
+                name="evaluate_indicator",
+                description=(
+                    "Evaluate a pandas-ta indicator against real OHLC data and return a structured "
+                    "array of timestamped values. Use this to validate that an indicator computes "
+                    "correctly before adding it to the workspace, or to inspect its output values."
+                ),
+                inputSchema={
+                    "type": "object",
+                    "properties": {
+                        "symbol": {
+                            "type": "string",
+                            "description": "Market symbol in 'MARKET.EXCHANGE' format, e.g. 'BTC/USDT.BINANCE'"
+                        },
+                        "from_time": {
+                            "description": "Start of time range. Unix timestamp (int) or date string e.g. '30 days ago', '2024-01-01'"
+                        },
+                        "to_time": {
+                            "description": "End of time range. Unix timestamp (int) or date string e.g. 'now', '2024-03-01'"
+                        },
+                        "period_seconds": {
+                            "type": "integer",
+                            "description": "Candle period in seconds (e.g. 3600 for 1h, 900 for 15m, 86400 for 1d)",
+                            "default": 3600
+                        },
+                        "pandas_ta_name": {
+                            "type": "string",
+                            "description": "Lowercase pandas-ta function name, e.g. 'rsi', 'macd', 'bbands'"
+                        },
+                        "parameters": {
+                            "type": "object",
+                            "description": "pandas-ta keyword arguments, e.g. {\"length\": 14} or {\"fast\": 12, \"slow\": 26, \"signal\": 9}",
+                            "default": {}
+                        }
+                    },
+                    "required": ["symbol", "from_time", "to_time", "pandas_ta_name"]
+                }
+            ),
+            Tool(
+                name="backtest_strategy",
+                description=(
+                    "Run a saved trading strategy against historical OHLC data using Nautilus Trader "
+                    "BacktestEngine. Returns performance metrics (total return, Sharpe ratio, "
+                    "max drawdown, win rate, trade count) and a full equity curve. "
+                    "Supports multiple data feeds and includes order-flow fields (buy_vol, sell_vol, "
+                    "open_interest) in the strategy's DataFrame."
+                ),
+                inputSchema={
+                    "type": "object",
+                    "properties": {
+                        "strategy_name": {
+                            "type": "string",
+                            "description": "Display name of the strategy as saved via python_write"
+                        },
+                        "feeds": {
+                            "type": "array",
+                            "description": "Data feeds to backtest against",
+                            "items": {
+                                "type": "object",
+                                "properties": {
+                                    "symbol": {
+                                        "type": "string",
+                                        "description": "Market symbol in 'MARKET.EXCHANGE' format, e.g. 'BTC/USDT.BINANCE'"
+                                    },
+                                    "period_seconds": {
+                                        "type": "integer",
+                                        "description": "Candle period in seconds (e.g. 3600 for 1h)",
+                                        "default": 3600
+                                    }
+                                },
+                                "required": ["symbol"]
+                            },
+                            "minItems": 1
+                        },
+                        "from_time": {
+                            "description": "Backtest start. Unix timestamp or date string e.g. '2024-01-01', '90 days ago'"
+                        },
+                        "to_time": {
+                            "description": "Backtest end. Unix timestamp or date string e.g. '2025-01-01', 'now'"
+                        },
+                        "initial_capital": {
+                            "type": "number",
+                            "description": "Starting capital in quote currency (e.g. 10000.0 USDT)",
+                            "default": 10000.0
+                        },
+                        "paper": {
+                            "type": "boolean",
+                            "description": "Always true for historical backtest (reserved for forward testing)",
+                            "default": True
+                        }
+                    },
+                    "required": ["strategy_name", "feeds", "from_time", "to_time"]
+                }
+            ),
+            Tool(
+                name="activate_strategy",
+                description=(
+                    "Activate a strategy for paper or live forward trading with a capital allocation. "
+                    "paper=true (default): simulated fills on live data — no API keys required. "
+                    "paper=false: real execution via user secrets vault (not yet implemented). "
+                    "Note: live data streaming is TBD; this registers the strategy for when it becomes available."
+                ),
+                inputSchema={
+                    "type": "object",
+                    "properties": {
+                        "strategy_name": {
+                            "type": "string",
+                            "description": "Display name of the strategy as saved via python_write"
+                        },
+                        "feeds": {
+                            "type": "array",
+                            "description": "Data feeds for the strategy",
+                            "items": {
+                                "type": "object",
+                                "properties": {
+                                    "symbol": {"type": "string"},
+                                    "period_seconds": {"type": "integer", "default": 3600}
+                                },
+                                "required": ["symbol"]
+                            },
+                            "minItems": 1
+                        },
+                        "allocation": {
+                            "type": "number",
+                            "description": "Capital allocated in quote currency (e.g. 5000.0 USDT)"
+                        },
+                        "paper": {
+                            "type": "boolean",
+                            "description": "True = paper/simulated (default); False = live execution",
+                            "default": True
+                        }
+                    },
+                    "required": ["strategy_name", "feeds", "allocation"]
+                }
+            ),
+            Tool(
+                name="deactivate_strategy",
+                description="Stop an active strategy and return its final P&L summary.",
+                inputSchema={
+                    "type": "object",
+                    "properties": {
+                        "strategy_name": {
+                            "type": "string",
+                            "description": "Display name of the active strategy to stop"
+                        }
+                    },
+                    "required": ["strategy_name"]
+                }
+            ),
+            Tool(
+                name="list_active_strategies",
+                description="List all currently active (live or paper) strategies and their status.",
+                inputSchema={
+                    "type": "object",
+                    "properties": {},
+                    "required": []
+                }
+            ),
         ]
 
 
     @server.call_tool()
     async def handle_tool_call(name: str, arguments: dict):
         """Handle tool calls including workspace and category tools"""
+        get_lifecycle_manager().record_activity()
+        try:
+            return await _handle_tool_call_inner(name, arguments)
+        except Exception:
+            logging.exception("Unhandled exception in tool '%s'", name)
+            raise
+
+    async def _handle_tool_call_inner(name: str, arguments: dict):
         if name == "workspace_read":
             return workspace_store.read(arguments.get("store_name", ""))
         elif name == "workspace_write":
@@ -400,7 +652,7 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                 arguments.get("store_name", ""),
                 arguments.get("patch", [])
             )
-        elif name == "category_write":
+        elif name == "python_write":
             result = category_manager.write(
                 category=arguments.get("category", ""),
                 name=arguments.get("name", ""),
@@ -410,6 +662,8 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
             )
             content = []
             meta_parts = [f"success: {result['success']}", f"path: {result['path']}"]
+            if result.get("revision"):
+                meta_parts.append(f"revision: {result['revision']}")
             if result.get("validation") and not result["validation"].get("success"):
                 meta_parts.append(f"validation errors: {result['validation'].get('errors', [])}")
             content.append(TextContent(type="text", text="\n".join(meta_parts)))
@@ -417,20 +671,23 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                 exec_content = result["execution"].get("content", [])
                 content.extend(exec_content)
                 image_count = sum(1 for item in exec_content if item.type == "image")
-                logging.info(f"category_write '{arguments.get('name')}': returning {len(content)} items, {image_count} images")
+                logging.info(f"python_write '{arguments.get('name')}': returning {len(content)} items, {image_count} images")
             else:
-                logging.info(f"category_write '{arguments.get('name')}': no execution result (category={arguments.get('category')})")
+                logging.info(f"python_write '{arguments.get('name')}': no execution result (category={arguments.get('category')})")
             return content
-        elif name == "category_edit":
+        elif name == "python_edit":
             result = category_manager.edit(
                 category=arguments.get("category", ""),
                 name=arguments.get("name", ""),
                 code=arguments.get("code"),
+                patches=arguments.get("patches"),
                 description=arguments.get("description"),
                 metadata=arguments.get("metadata")
             )
             content = []
             meta_parts = [f"success: {result['success']}", f"path: {result['path']}"]
+            if result.get("revision"):
+                meta_parts.append(f"revision: {result['revision']}")
             if result.get("validation") and not result["validation"].get("success"):
                 meta_parts.append(f"validation errors: {result['validation'].get('errors', [])}")
             content.append(TextContent(type="text", text="\n".join(meta_parts)))
@@ -438,19 +695,43 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
                 exec_content = result["execution"].get("content", [])
                 content.extend(exec_content)
                 image_count = sum(1 for item in exec_content if item.type == "image")
-                logging.info(f"category_edit '{arguments.get('name')}': returning {len(content)} items, {image_count} images")
+                logging.info(f"python_edit '{arguments.get('name')}': returning {len(content)} items, {image_count} images")
             else:
-                logging.info(f"category_edit '{arguments.get('name')}': no execution result")
+                logging.info(f"python_edit '{arguments.get('name')}': no execution result")
             return content
-        elif name == "category_read":
+        elif name == "python_read":
             return category_manager.read(
                 category=arguments.get("category", ""),
                 name=arguments.get("name", "")
             )
-        elif name == "category_list":
+        elif name == "python_list":
             return category_manager.list_items(
                 category=arguments.get("category", "")
             )
+        elif name == "python_log":
+            result = category_manager.git_log(
+                category=arguments.get("category"),
+                name=arguments.get("name"),
+                limit=int(arguments.get("limit", 20))
+            )
+            lines = [f"success: {result['success']}"]
+            for c in result.get("commits", []):
+                lines.append(f"{c['short_hash']} {c['date'][:10]}  {c['message']}")
+            return [TextContent(type="text", text="\n".join(lines))]
+        elif name == "python_revert":
+            result = category_manager.git_revert(
+                revision=arguments.get("revision", ""),
+                category=arguments.get("category", ""),
+                name=arguments.get("name", "")
+            )
+            meta_parts = [f"success: {result['success']}"]
+            if result.get("revision"):
+                meta_parts.append(f"revision: {result['revision']}")
+            if result.get("error"):
+                meta_parts.append(f"error: {result['error']}")
+            if result.get("validation") and not result["validation"].get("success"):
+                meta_parts.append(f"validation errors: {result['validation'].get('errors', [])}")
+            return [TextContent(type="text", text="\n".join(meta_parts))]
         elif name == "conda_sync":
             # Get environment.yml path relative to main.py
             env_yml = Path(__file__).parent / "environment.yml"
@@ -469,6 +750,37 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
             image_count = sum(1 for item in content if item.type == "image")
             logging.info(f"execute_research '{arguments.get('name')}': returning {len(content)} items, {image_count} images")
             return content
+        elif name == "evaluate_indicator":
+            return await evaluate_indicator(
+                symbol=arguments.get("symbol", ""),
+                from_time=arguments.get("from_time"),
+                to_time=arguments.get("to_time"),
+                period_seconds=int(arguments.get("period_seconds", 3600)),
+                pandas_ta_name=arguments.get("pandas_ta_name", ""),
+                parameters=arguments.get("parameters") or {},
+            )
+        elif name == "backtest_strategy":
+            return await backtest_strategy(
+                strategy_name=arguments.get("strategy_name", ""),
+                feeds=arguments.get("feeds", []),
+                from_time=arguments.get("from_time"),
+                to_time=arguments.get("to_time"),
+                initial_capital=float(arguments.get("initial_capital", 10_000.0)),
+                paper=bool(arguments.get("paper", True)),
+            )
+        elif name == "activate_strategy":
+            return await activate_strategy(
+                strategy_name=arguments.get("strategy_name", ""),
+                feeds=arguments.get("feeds", []),
+                allocation=float(arguments.get("allocation", 0.0)),
+                paper=bool(arguments.get("paper", True)),
+            )
+        elif name == "deactivate_strategy":
+            return await deactivate_strategy(
+                strategy_name=arguments.get("strategy_name", ""),
+            )
+        elif name == "list_active_strategies":
+            return await list_active_strategies()
         else:
             raise ValueError(f"Unknown tool: {name}")
 
@@ -477,26 +789,18 @@ def create_mcp_server(config: Config, event_publisher: EventPublisher) -> Server
 
 
 # =============================================================================
-# SSE Transport Setup
+# Streamable HTTP Transport Setup
 # =============================================================================
 
-def create_sse_app(mcp_server: Server) -> Starlette:
-    """Create Starlette app with SSE endpoint for MCP"""
+def create_streamable_http_app(mcp_server: Server) -> Starlette:
+    """Create Starlette app with Streamable HTTP endpoint for MCP"""
 
-    # Create SSE transport instance
-    sse = SseServerTransport("/messages/")
+    session_manager = StreamableHTTPSessionManager(app=mcp_server)
 
-    async def handle_sse(request: Request) -> Response:
-        """Handle SSE connections for MCP"""
-        async with sse.connect_sse(
-            request.scope, request.receive, request._send
-        ) as streams:
-            await mcp_server.run(
-                streams[0],
-                streams[1],
-                mcp_server.create_initialization_options()
-            )
-        return Response()
+    @contextlib.asynccontextmanager
+    async def lifespan(app: Starlette):
+        async with session_manager.run():
+            yield
 
     async def handle_health(request: Request) -> Response:
         """Health check endpoint for k8s probes and gateway readiness checks"""
@@ -506,9 +810,9 @@ def create_sse_app(mcp_server: Server) -> Starlette:
         )
 
     app = Starlette(
+        lifespan=lifespan,
         routes=[
-            Route("/sse", handle_sse),
-            Mount("/messages/", app=sse.handle_post_message),
+            Mount("/mcp", app=session_manager.handle_request),
             Route("/health", handle_health),
         ]
     )
@@ -648,9 +952,9 @@ class UserContainer:
                         self.mcp_server.create_initialization_options()
                     )
             elif self.config.mcp_transport == "sse":
-                # Run MCP server via HTTP/SSE (for production)
-                logging.info(f"Starting MCP server with SSE transport on {self.config.mcp_http_host}:{self.config.mcp_http_port}")
-                app = create_sse_app(self.mcp_server)
+                # Run MCP server via Streamable HTTP (for production)
+                logging.info(f"Starting MCP server with Streamable HTTP transport on {self.config.mcp_http_host}:{self.config.mcp_http_port}")
+                app = create_streamable_http_app(self.mcp_server)
                 config = uvicorn.Config(
                     app,
                     host=self.config.mcp_http_host,

web/src/components/ChartView.vue

@@ -4,9 +4,11 @@ import Card from 'primevue/card'
 import { createTradingViewDatafeed } from '../composables/useTradingViewDatafeed'
 import { useTradingViewShapes } from '../composables/useTradingViewShapes'
 import { useTradingViewIndicators } from '../composables/useTradingViewIndicators'
+import { useCustomIndicators, getCustomIndicatorsGetter } from '../composables/useCustomIndicators'
 import { useChartStore } from '../stores/chart'
 import type { IChartingLibraryWidget } from '../types/tradingview'
 import { intervalToSeconds } from '../utils'
+import { wsManager } from '../composables/useWebSocket'
 
 // Convert seconds to TradingView interval string
 function secondsToInterval(seconds: number): string {
@@ -22,12 +24,25 @@ let datafeed: any = null
 let isUpdatingFromChart = false // Flag to prevent circular updates
 let shapeCleanup: (() => void) | null = null // Cleanup function for shape sync
 let indicatorCleanup: (() => void) | null = null // Cleanup function for indicator sync
+let customIndicatorCleanup: (() => void) | null = null // Cleanup for custom TV studies
+let chartInitialized = false // Guard against double-init on reconnect
+
+const maybeInitChart = () => {
+  if (chartInitialized || !chartContainer.value) return
+  chartInitialized = true
+  initChart()
+}
 
 onMounted(() => {
-  if (!chartContainer.value) return
+  // Wait for workspace to be ready (persistent stores loaded from container)
+  // before initializing TradingView, so stores are populated when onChartReady fires.
+  watch(wsManager.sessionStatus, (status) => {
+    if (status === 'ready') maybeInitChart()
+  }, { immediate: true })
+})
 
-  // Wait for TradingView library to load
-  const initChart = () => {
+// Wait for TradingView library to load
+function initChart() {
     if (!window.TradingView) {
       setTimeout(initChart, 100)
       return
@@ -43,16 +58,23 @@ onMounted(() => {
         container: chartContainer.value!,
         library_path: '/charting_library/',
         locale: 'en',
+        // Register the two generic custom study dispatch types.
+        // Must be provided here — TV has no dynamic study registration API.
+        custom_indicators_getter: getCustomIndicatorsGetter(),
         disabled_features: [
           'use_localstorage_for_settings',
           'header_symbol_search',
           'symbol_search_hot_key'
         ],
-        enabled_features: ['study_templates'],
-        // Restrict indicators to only those supported by both TA-Lib and TradingView
+        enabled_features: [],
+        // Restrict indicators to only those supported by both TA-Lib and TradingView.
+        // Custom AI-generated indicators (from custom_indicators_getter) must also be listed here.
         studies_access: {
           type: 'white',
           tools: [
+            // AI custom indicator dispatch studies
+            { name: 'dxo_customstudy_overlay' },
+            { name: 'dxo_customstudy_pane' },
             // Overlap Studies (14)
             { name: 'Moving Average' },
             { name: 'Moving Average Exponential' },
@@ -150,15 +172,13 @@ onMounted(() => {
         if (tvWidget) {
           shapeCleanup = useTradingViewShapes(tvWidget)
           indicatorCleanup = useTradingViewIndicators(tvWidget)
+          customIndicatorCleanup = useCustomIndicators(tvWidget)
         }
       })
     } catch (error) {
       console.error('Failed to initialize TradingView widget:', error)
     }
-  }
-
-  initChart()
-})
+}
 
 function initializeVisibleRange() {
   if (!tvWidget) return
@@ -281,6 +301,12 @@ onBeforeUnmount(() => {
     indicatorCleanup = null
   }
 
+  // Cleanup custom TV studies
+  if (customIndicatorCleanup) {
+    customIndicatorCleanup()
+    customIndicatorCleanup = null
+  }
+
   if (tvWidget) {
     tvWidget.remove()
     tvWidget = null

web/src/components/ChatPanel.vue

@@ -238,14 +238,7 @@ const handleMessage = (data: WebSocketMessage) => {
 
 // Stop agent processing
 const stopAgent = () => {
-  // Send empty message to trigger interrupt without new agent round
-  const wsMessage = {
-    type: 'agent_user_message',
-    session_id: SESSION_ID,
-    content: '',
-    attachments: []
-  }
-  wsManager.send(wsMessage)
+  wsManager.send({ type: 'agent_stop', session_id: SESSION_ID })
   isAgentProcessing.value = false
   removeToolCallBubble()
   lastSentMessageId = null
@@ -586,7 +579,9 @@ onUnmounted(() => {
 }
 
 .workspace-loading {
-  flex: 1;
+  position: fixed;
+  inset: 0;
+  z-index: 9999;
   display: flex;
   flex-direction: column;
   align-items: center;
@@ -639,7 +634,7 @@ onUnmounted(() => {
 .stop-button-container {
   position: absolute;
   bottom: 80px;
-  left: 20px;
+  right: 20px;
   z-index: 1000;
 }
 

web/src/composables/useCustomIndicators.ts

@@ -0,0 +1,551 @@
+/**
+ * TradingView custom study integration for pandas-ta custom indicators.
+ *
+ * Architecture overview
+ * ---------------------
+ * TV's custom study API only allows registering study types via the
+ * `custom_indicators_getter` widget constructor option — there is no
+ * dynamic registration API (createCustomStudy does not exist on the widget
+ * or chart APIs).
+ *
+ * To support custom indicators that arrive at runtime (e.g. from the AI
+ * agent), we pre-register two generic dispatch studies in
+ * `custom_indicators_getter`:
+ *
+ *   dxo_customstudy_overlay  — is_price_study: true  (drawn on price pane)
+ *   dxo_customstudy_pane     — is_price_study: false (separate pane)
+ *
+ * Each has a single text input `_cfg` (a config key) and MAX_PLOTS
+ * line plots. The constructor dispatches to `customStudyRegistry[cfgKey]`
+ * to look up the per-indicator configuration and data.
+ *
+ * These study type names MUST also appear in the `studies_access` whitelist
+ * in ChartView.vue — TV treats unlisted studies as nonexistent.
+ *
+ * Registration flow
+ * -----------------
+ * 1. Widget constructor calls getCustomIndicatorsGetter() which registers
+ *    the two generic study types.
+ * 2. When a custom_ indicator appears in the store, registerCustomStudy():
+ *    a. Stores the config in customStudyRegistry under a unique cfgKey.
+ *    b. Calls chart.createStudy('dxo_customstudy_*', ..., { _cfg: cfgKey }).
+ *    c. Calls study.setStudyTitle(indicator name) for a human-readable header.
+ * 3. TV calls the study's init(ctx, inputs):
+ *    a. Reads symbol/period from ctx; builds the data cache key.
+ *    b. Fires an async evaluateIndicator WebSocket request.
+ * 4. When data arrives the constructor calls the registered refreshCallback
+ *    which calls IStudyApi.setInputValues() with a new config key, causing
+ *    TV to re-run init()+main() with the now-populated cache.
+ *
+ * IMPORTANT: call getCustomIndicatorsGetter() and pass it as the
+ * `custom_indicators_getter` option when creating the TradingView widget.
+ */
+
+import { watch } from 'vue'
+import { useIndicatorStore, type IndicatorInstance, type CustomIndicatorMetadata } from '../stores/indicators'
+import { useChartStore } from '../stores/chart'
+import { wsManager, type MessageHandler } from './useWebSocket'
+import { intervalToSeconds } from '../utils'
+
+// ---------------------------------------------------------------------------
+// WebSocket helper — evaluate_indicator request/response
+// ---------------------------------------------------------------------------
+
+interface EvaluateResult {
+  symbol: string
+  period_seconds: number
+  pandas_ta_name: string
+  parameters: Record<string, any>
+  candle_count: number
+  columns: string[]
+  values: Array<Record<string, any>>
+  error?: string
+}
+
+function evaluateIndicator(
+  symbol: string,
+  fromTime: number,
+  toTime: number,
+  periodSeconds: number,
+  pandasTaName: string,
+  parameters: Record<string, any>,
+  timeoutMs = 30_000
+): Promise<EvaluateResult> {
+  const requestId = `cind_${Date.now()}_${Math.random().toString(36).substring(7)}`
+
+  return new Promise((resolve, reject) => {
+    const timer = window.setTimeout(() => {
+      wsManager.removeHandler(handler)
+      reject(new Error(`evaluate_indicator timeout for ${pandasTaName}`))
+    }, timeoutMs)
+
+    const handler: MessageHandler = (message: any) => {
+      if (message.type !== 'evaluate_indicator_result') return
+      if (message.request_id !== requestId) return
+      clearTimeout(timer)
+      wsManager.removeHandler(handler)
+      if (message.error) reject(new Error(message.error))
+      else resolve(message as EvaluateResult)
+    }
+
+    wsManager.addHandler(handler)
+    wsManager.send({
+      type: 'evaluate_indicator',
+      request_id: requestId,
+      symbol,
+      from_time: fromTime,
+      to_time: toTime,
+      period_seconds: periodSeconds,
+      pandas_ta_name: pandasTaName,
+      parameters,
+    })
+  })
+}
+
+// ---------------------------------------------------------------------------
+// Data cache — keyed by "indicatorId_symbol_periodSeconds_paramsHash"
+// Each entry maps timestamp-ms → row object
+// ---------------------------------------------------------------------------
+
+type DataRow = Record<string, number | null>
+const dataCache = new Map<string, Map<number, DataRow>>()
+
+function cacheKey(indicatorId: string, symbol: string, periodSeconds: number, paramsHash: string): string {
+  return `${indicatorId}_${symbol}_${periodSeconds}_${paramsHash}`
+}
+
+function buildDataCache(result: EvaluateResult): Map<number, DataRow> {
+  const map = new Map<number, DataRow>()
+  for (const point of result.values) {
+    const tsMs = (point.timestamp as number) * 1000  // server sends Unix seconds → ms
+    map.set(tsMs, point as DataRow)
+  }
+  return map
+}
+
+// ---------------------------------------------------------------------------
+// Custom study registry — config map shared between getter and composable
+// ---------------------------------------------------------------------------
+
+interface CustomStudyEntry {
+  indicatorId: string
+  pandasTaName: string
+  parameters: Record<string, any>
+  metadata: CustomIndicatorMetadata
+}
+
+// cfgKey → per-instance config; populated by registerCustomStudy()
+const customStudyRegistry = new Map<string, CustomStudyEntry>()
+
+// indicatorId → callback(newCfgKey); set by registerCustomStudy()
+// Called by the constructor when async data arrives to trigger TV re-run.
+const refreshCallbacks = new Map<string, (newCfgKey: string) => void>()
+
+// TradingView widget reference — set by useCustomIndicators() so the
+// constructor can query the current visible range.
+let _tvWidget: any = null
+
+// ---------------------------------------------------------------------------
+// Generic study design constants
+// ---------------------------------------------------------------------------
+
+const MAX_PLOTS = 8
+
+const MULTI_LINE_COLORS = [
+  '#2196F3', '#FF9800', '#4CAF50', '#E91E63', '#9C27B0',
+  '#00BCD4', '#FF5722', '#795548',
+]
+
+// ---------------------------------------------------------------------------
+// Custom indicators getter
+// Pass the result of this function as the widget option:
+//   custom_indicators_getter: getCustomIndicatorsGetter()
+//
+// The study type names must also be listed in studies_access in ChartView.vue.
+// ---------------------------------------------------------------------------
+
+export function getCustomIndicatorsGetter(): (_PineJS: any) => Promise<any[]> {
+  function makeGenericStudy(name: string, isPriceStudy: boolean): any {
+    const plots = Array.from({ length: MAX_PLOTS }, (_, i) => ({ id: `plot_${i}`, type: 'line' }))
+    const styles: Record<string, any> = {}
+    const defaultStyles: Record<string, any> = {}
+    for (let i = 0; i < MAX_PLOTS; i++) {
+      styles[`plot_${i}`] = { title: `Plot ${i}` }
+      defaultStyles[`plot_${i}`] = {
+        linestyle: 0,
+        linewidth: 1,
+        plottype: 0,
+        color: MULTI_LINE_COLORS[i % MULTI_LINE_COLORS.length],
+        visible: i === 0,
+      }
+    }
+
+    return {
+      name,
+      metainfo: {
+        _metainfoVersion: 51,
+        // Use @tv-custom-1 — @tv-basicstudies-1 is reserved for built-ins
+        // and TV throws "unexpected study id" if a custom indicator uses it.
+        id: `${name}@tv-custom-1`,
+        scriptIdPart: '',
+        name,
+        description: name,
+        shortDescription: name,
+        is_price_study: isPriceStudy,
+        isCustomIndicator: true,
+        format: { type: 'inherit' },
+        // Single text input carries the per-instance config key.
+        inputs: [
+          { id: '_cfg', name: 'Config Key', type: 'text', defval: '' },
+        ],
+        plots,
+        styles,
+        defaults: {
+          inputs: { _cfg: '' },
+          styles: defaultStyles,
+        },
+      },
+      // ES5 constructor — TV instantiates this with `new`
+      constructor: function (this: any) {
+        // Per-instance mutable state stored on the constructor instance
+        let _cfgKey = ''   // current config key (from inputs(0))
+        let _dataKey = ''  // data cache key (built from ctx symbol/period/params)
+        let _fetchGen = 0  // incremented each init(); used to cancel stale fetches
+
+        this.init = function (ctx: any, inputs: (i: number) => any) {
+          const cfgKey = inputs(0) as string
+          _cfgKey = cfgKey
+          _fetchGen++
+          const myGen = _fetchGen
+
+          const entry = customStudyRegistry.get(cfgKey)
+          if (!entry) return
+
+          // Derive symbol and period from the TV context object.
+          // ctx.symbol.ticker — symbol name without exchange prefix
+          // ctx.symbol.period — TV interval string ("15", "1D", etc.)
+          const symbol: string = ctx.symbol.ticker
+          const periodStr: string = ctx.symbol.period
+          const periodSeconds = intervalToSeconds(periodStr)
+          const paramsHash = JSON.stringify(entry.parameters)
+          const dk = cacheKey(entry.indicatorId, symbol, periodSeconds, paramsHash)
+          _dataKey = dk
+
+          if (dataCache.has(dk)) return  // Data already fetched for this symbol/period/params
+
+          // Determine time range: prefer chart's visible range, fall back to 500-bar window
+          let fromTime: number
+          let toTime: number
+          const now = Math.floor(Date.now() / 1000)
+          toTime = now
+          fromTime = now - periodSeconds * 500
+          if (_tvWidget) {
+            try {
+              const range = _tvWidget.activeChart().getVisibleRange()
+              if (range?.from && range?.to) {
+                const dur = Math.floor(range.to) - Math.floor(range.from)
+                fromTime = Math.floor(range.from) - Math.floor(dur * 0.5)
+                toTime = Math.floor(range.to)
+              }
+            } catch { /* chart not yet ready */ }
+          }
+
+          // Capture mutable vars before async gap
+          const capturedDk = dk
+          const capturedCfgKey = cfgKey
+
+          evaluateIndicator(symbol, fromTime, toTime, periodSeconds, entry.pandasTaName, entry.parameters)
+            .then((result) => {
+              if (myGen !== _fetchGen) return  // Superseded by a newer init() call
+              dataCache.set(capturedDk, buildDataCache(result))
+              // Create a sibling config key pointing to the same entry.
+              // Calling setInputValues() with this new key causes TV to
+              // re-invoke init()+main() with the now-populated cache.
+              const refreshKey = `${capturedCfgKey}__r`
+              customStudyRegistry.set(refreshKey, entry)
+              const cb = refreshCallbacks.get(entry.indicatorId)
+              if (cb) cb(refreshKey)
+            })
+            .catch((err) => {
+              console.error('[CustomIndicators] Failed to fetch data for', entry.pandasTaName, err)
+            })
+        }
+
+        this.main = function (ctx: any, _inputs: (i: number) => any) {
+          // ctx.symbol.bartime() returns the bar timestamp in milliseconds (documented)
+          const ts: number = ctx.symbol.bartime()
+          if (!_cfgKey || !_dataKey) return new Array(MAX_PLOTS).fill(NaN)
+          const entry = customStudyRegistry.get(_cfgKey)
+          if (!entry) return new Array(MAX_PLOTS).fill(NaN)
+          const cache = dataCache.get(_dataKey)
+          if (!cache) return new Array(MAX_PLOTS).fill(NaN)
+          const row = cache.get(ts)
+          return Array.from({ length: MAX_PLOTS }, (_, i) => {
+            const col = entry.metadata.output_columns[i]
+            return col && row ? (row[col.name] as number) ?? NaN : NaN
+          })
+        }
+      },
+    }
+  }
+
+  return (_PineJS: any): Promise<any[]> => {
+    return Promise.resolve([
+      makeGenericStudy('dxo_customstudy_overlay', true),
+      makeGenericStudy('dxo_customstudy_pane', false),
+    ])
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Main composable
+// ---------------------------------------------------------------------------
+
+export function useCustomIndicators(tvWidget: any) {
+  _tvWidget = tvWidget
+
+  const indicatorStore = useIndicatorStore()
+  const chartStore = useChartStore()
+
+  // Maps indicator id → { cfgKey, tvStudyId, symbol }
+  const registered = new Map<string, { cfgKey: string; tvStudyId: string | null; symbol: string }>()
+  // Monotonic version counter per indicator for unique config keys
+  const cfgVersions = new Map<string, number>()
+  // Last-seen parameter hash per indicator id for change detection.
+  // Needed because Pinia $patch mutates in place (oldValue === newValue).
+  const lastParams = new Map<string, string>()
+
+  let isChartReady = false
+
+  function nextCfgKey(indicatorId: string): string {
+    const v = (cfgVersions.get(indicatorId) || 0) + 1
+    cfgVersions.set(indicatorId, v)
+    return `cfg_${indicatorId.replace(/[^a-zA-Z0-9]/g, '_')}_v${v}`
+  }
+
+  // Apply per-indicator visual overrides after createStudy() returns.
+  // Uses per-column plot config (style, color, linewidth, visible) from metadata.
+  function applyStudyOverrides(studyId: string, meta: CustomIndicatorMetadata) {
+    try {
+      const study = tvWidget.activeChart().getStudyById(studyId)
+      if (!study) return
+      const cols = meta.output_columns
+      const overrides: Record<string, any> = {}
+
+      for (let i = 0; i < MAX_PLOTS; i++) {
+        const col = cols[i]
+        if (col == null) {
+          overrides[`styles.plot_${i}.visible`] = false
+          continue
+        }
+        const p = col.plot
+        overrides[`styles.plot_${i}.visible`] = p?.visible ?? true
+        overrides[`styles.plot_${i}.plottype`] = p?.style ?? 0
+        overrides[`styles.plot_${i}.linewidth`] = p?.linewidth ?? 2
+        overrides[`styles.plot_${i}.linestyle`] = 0
+        overrides[`styles.plot_${i}.color`] = p?.color ?? MULTI_LINE_COLORS[i % MULTI_LINE_COLORS.length]
+      }
+
+      // Note: TV band `value` is fixed at metainfo-declaration time and cannot be changed
+      // via overrides. Indicators that need horizontal reference lines at configurable
+      // values (e.g. RSI at 70/30) should instead include a constant-value output column
+      // rather than relying on meta.bands.
+
+      study.applyOverrides(overrides)
+    } catch (err) {
+      console.warn('[CustomIndicators] Could not apply overrides:', err)
+    }
+  }
+
+  // ------------------------------------------------------------------
+  // Register a custom indicator as a TV study instance
+  // ------------------------------------------------------------------
+  async function registerCustomStudy(indicator: IndicatorInstance) {
+    const meta = indicator.custom_metadata
+    if (!meta) {
+      console.warn('[CustomIndicators] No custom_metadata on indicator:', indicator.id)
+      return
+    }
+
+    const symbol = indicator.symbol || chartStore.symbol
+    const cfgKey = nextCfgKey(indicator.id)
+    const forceOverlay = meta.pane === 'price'
+    const studyTypeName = meta.pane === 'price' ? 'dxo_customstudy_overlay' : 'dxo_customstudy_pane'
+
+    // Store per-instance config in the registry so the constructor can find it
+    customStudyRegistry.set(cfgKey, {
+      indicatorId: indicator.id,
+      pandasTaName: indicator.pandas_ta_name,
+      parameters: indicator.parameters,
+      metadata: meta,
+    })
+
+    // Register the callback invoked by the constructor after async data loads.
+    // We change the study's _cfg input to a sibling key, which causes TV to
+    // re-run init()+main() and pick up the freshly populated cache.
+    refreshCallbacks.set(indicator.id, (newCfgKey: string) => {
+      const entry = registered.get(indicator.id)
+      if (!entry?.tvStudyId) return
+      try {
+        const study = tvWidget.activeChart().getStudyById(entry.tvStudyId)
+        if (study) {
+          registered.set(indicator.id, { ...entry, cfgKey: newCfgKey })
+          study.setInputValues([{ id: '_cfg', value: newCfgKey }])
+        }
+      } catch (err) {
+        console.warn('[CustomIndicators] Could not refresh study after data load:', err)
+      }
+    })
+
+    try {
+      const tvStudyId = (await tvWidget.activeChart().createStudy(
+        studyTypeName, forceOverlay, false,
+        { _cfg: cfgKey }
+      )) as string | null
+
+      registered.set(indicator.id, { cfgKey, tvStudyId: tvStudyId ?? null, symbol })
+      lastParams.set(indicator.id, JSON.stringify(indicator.parameters))
+
+      if (tvStudyId) {
+        // Set human-readable panel title (falls back to pandas_ta_name if no display name)
+        const displayName = meta.display_name || indicator.pandas_ta_name.replace(/^custom_/, '')
+        try {
+          const study = tvWidget.activeChart().getStudyById(tvStudyId)
+          if (study && typeof study.setStudyTitle === 'function') {
+            study.setStudyTitle(displayName)
+          }
+        } catch { /* setStudyTitle not available in this TV build */ }
+
+        applyStudyOverrides(tvStudyId, meta)
+        if (tvStudyId !== indicator.tv_study_id) {
+          indicatorStore.updateIndicator(indicator.id, { tv_study_id: tvStudyId })
+        }
+      }
+
+      console.log('[CustomIndicators] Registered:', indicator.pandas_ta_name, '→', studyTypeName, '(', tvStudyId, ')')
+    } catch (err) {
+      console.error('[CustomIndicators] Failed to create TV custom study:', studyTypeName, err)
+    }
+  }
+
+  // ------------------------------------------------------------------
+  // Remove a custom study from the chart
+  // ------------------------------------------------------------------
+  function removeCustomStudy(indicatorId: string) {
+    const entry = registered.get(indicatorId)
+    if (!entry) return
+    registered.delete(indicatorId)
+    lastParams.delete(indicatorId)
+    refreshCallbacks.delete(indicatorId)
+
+    if (entry.tvStudyId) {
+      try { tvWidget.activeChart().removeStudy(entry.tvStudyId) } catch { /* already gone */ }
+    }
+  }
+
+  // ------------------------------------------------------------------
+  // Re-register when parameters/symbol/period change (forces new data fetch)
+  // ------------------------------------------------------------------
+  async function refreshCustomStudy(indicator: IndicatorInstance) {
+    // Purge stale cache entries so init() fetches fresh data
+    for (const key of Array.from(dataCache.keys())) {
+      if (key.startsWith(`${indicator.id}_`)) {
+        dataCache.delete(key)
+      }
+    }
+    removeCustomStudy(indicator.id)
+    await registerCustomStudy(indicator)
+  }
+
+  // ------------------------------------------------------------------
+  // Store watcher — respond to indicator additions, changes, removals
+  //
+  // NOTE: Pinia $patch mutates in place, so oldValue === newValue for
+  // backend-originated updates. We track state manually via lastParams.
+  // ------------------------------------------------------------------
+  watch(
+    () => indicatorStore.indicators,
+    async (newIndicators) => {
+      if (!isChartReady) return
+
+      for (const [id, indicator] of Object.entries(newIndicators)) {
+        if (!indicator.pandas_ta_name.startsWith('custom_')) continue
+
+        if (!registered.has(id)) {
+          lastParams.set(id, JSON.stringify(indicator.parameters))
+          await registerCustomStudy(indicator)
+        } else {
+          const entry = registered.get(id)!
+          const currParams = JSON.stringify(indicator.parameters)
+          const prevParams = lastParams.get(id)
+          const currSymbol = indicator.symbol || chartStore.symbol
+          if (currParams !== prevParams || currSymbol !== entry.symbol) {
+            lastParams.set(id, currParams)
+            await refreshCustomStudy(indicator)
+          }
+        }
+      }
+
+      // Handle removals
+      for (const id of registered.keys()) {
+        if (!(id in newIndicators)) {
+          lastParams.delete(id)
+          removeCustomStudy(id)
+        }
+      }
+    },
+    { deep: true }
+  )
+
+  // Re-fetch when chart resolution changes
+  watch(
+    () => chartStore.period,
+    () => {
+      if (!isChartReady) return
+      for (const [id, indicator] of Object.entries(indicatorStore.indicators)) {
+        if (!indicator.pandas_ta_name.startsWith('custom_')) continue
+        if (registered.has(id)) {
+          lastParams.set(id, JSON.stringify(indicator.parameters))
+          refreshCustomStudy(indicator)
+        }
+      }
+    }
+  )
+
+  // ------------------------------------------------------------------
+  // Chart ready — apply any indicators already in the store
+  // ------------------------------------------------------------------
+  // useCustomIndicators is always called from within tvWidget.onChartReady in ChartView,
+  // so the chart is already ready.
+  isChartReady = true
+
+  // TV processes custom_indicators_getter asynchronously (Promise microtask), so the
+  // custom study types are not yet available at onChartReady time. Defer the initial
+  // registration of any pending indicators until chart data loads — by that point the
+  // getter Promise has resolved and the study types are registered in TV's internal
+  // study index (and the studies_access whitelist check passes).
+  let initialApplied = false
+  tvWidget.activeChart().onDataLoaded().subscribe(null, () => {
+    if (initialApplied) return
+    initialApplied = true
+    const pending = Object.values(indicatorStore.indicators).filter(
+      (ind) => ind.pandas_ta_name.startsWith('custom_') && !registered.has(ind.id)
+    )
+    for (const indicator of pending) {
+      lastParams.set(indicator.id, JSON.stringify(indicator.parameters))
+      registerCustomStudy(indicator)
+    }
+  })
+
+  // Cleanup
+  return () => {
+    for (const id of [...registered.keys()]) {
+      removeCustomStudy(id)
+    }
+    registered.clear()
+    cfgVersions.clear()
+    lastParams.clear()
+    if (_tvWidget === tvWidget) _tvWidget = null
+    isChartReady = false
+  }
+}

web/src/composables/useTradingViewDatafeed.ts

@@ -259,37 +259,17 @@ export class WebSocketDatafeed implements IBasicDataFeed {
 
           const rawBars: any[] = response.history.bars || []
 
-          // Parse bars, preserving null OHLC for gap bars (no trades that period)
-          const parsedBars: Bar[] = rawBars.map((bar: any) => {
-            if (bar.open === null || bar.close === null) {
-              return { time: bar.time * 1000, open: null, high: null, low: null, close: null }
-            }
-            return {
-              time: bar.time * 1000,
-              open: parseFloat(bar.open) / denoms.tick,
-              high: parseFloat(bar.high) / denoms.tick,
-              low: parseFloat(bar.low) / denoms.tick,
-              close: parseFloat(bar.close) / denoms.tick,
-              volume: parseFloat(bar.volume) / denoms.base
-            }
-          })
+          // All bars have non-null prices — ingestor forward-fills interior gaps.
+          const bars: Bar[] = rawBars.map((bar: any) => ({
+            time:   bar.time * 1000,
+            open:   parseFloat(bar.open)   / denoms.tick,
+            high:   parseFloat(bar.high)   / denoms.tick,
+            low:    parseFloat(bar.low)    / denoms.tick,
+            close:  parseFloat(bar.close)  / denoms.tick,
+            volume: parseFloat(bar.volume) / denoms.base
+          }))
 
-          parsedBars.sort((a, b) => a.time - b.time)
-
-          // Fill any gaps between returned bars with null bars so TradingView
-          // receives a contiguous array of the correct length.
-          const periodMs = intervalToSeconds(resolution) * 1000
-          const bars: Bar[] = []
-          for (let i = 0; i < parsedBars.length; i++) {
-            if (i > 0) {
-              const prev = parsedBars[i - 1].time
-              const curr = parsedBars[i].time
-              for (let t = prev + periodMs; t < curr; t += periodMs) {
-                bars.push({ time: t, open: null, high: null, low: null, close: null })
-              }
-            }
-            bars.push(parsedBars[i])
-          }
+          bars.sort((a, b) => a.time - b.time)
 
           console.log('[TradingView Datafeed] Scaled bar sample:', bars[0])
 

web/src/composables/useTradingViewIndicators.ts

@@ -367,6 +367,22 @@ export function useTradingViewIndicators(tvWidget: IChartingLibraryWidget) {
       }
       isChartReady = true
 
+      // Apply any indicators that arrived before chart was ready (e.g. from workspace sync on page load)
+      const pendingIndicators = Object.values(indicatorStore.indicators).filter(ind => !ind.tv_study_id)
+      if (pendingIndicators.length > 0) {
+        console.log('[Indicators] Chart ready, applying', pendingIndicators.length, 'pending indicators from store')
+        isApplyingTVUpdate = true
+        ;(async () => {
+          try {
+            for (const indicator of pendingIndicators) {
+              await createTVStudy(indicator)
+            }
+          } finally {
+            isApplyingTVUpdate = false
+          }
+        })()
+      }
+
       console.log('[Indicators] Setting up indicator event subscriptions')
       console.log('[Indicators] Chart ready, widget:', tvWidget)
 
@@ -781,6 +797,10 @@ export function useTradingViewIndicators(tvWidget: IChartingLibraryWidget) {
   async function createTVStudy(indicator: IndicatorInstance) {
     if (!isChartReady) return
 
+    // Custom indicators (pandas_ta_name starts with "custom_") are handled by
+    // useCustomIndicators — they use TV createCustomStudy, not createStudy.
+    if (indicator.pandas_ta_name.startsWith('custom_')) return
+
     try {
       const chart = tvWidget.activeChart()
       if (!chart) return

web/src/composables/useWebSocket.ts

@@ -22,7 +22,7 @@ class WebSocketManager {
   private reconnectAttempts = 0
   private maxReconnectAttempts = Infinity // Keep trying indefinitely
   private reconnectDelay = 1000 // Start with 1 second
-  private maxReconnectDelay = 15000 // Max 15 seconds
+  private maxReconnectDelay = 50000 // Max 50 seconds
 
   /**
    * Connect to WebSocket with JWT token for authentication
@@ -70,6 +70,11 @@ class WebSocketManager {
 
       this.ws.onopen = () => {
         console.log('[WebSocket] Connected successfully')
+        // Cancel any pending reconnect timer — we're already connected
+        if (this.reconnectTimeout) {
+          clearTimeout(this.reconnectTimeout)
+          this.reconnectTimeout = null
+        }
         this.isConnected.value = true
         this.isAuthenticated.value = false // Wait for 'connected' message from server
         this.reconnectAttempts = 0 // Reset reconnection counter
@@ -149,8 +154,8 @@ class WebSocketManager {
     } else {
       console.log('[WebSocket] Queuing message (not connected yet):', message.type, '- Queue size:', this.messageQueue.length + 1)
       this.messageQueue.push(message)
-      // Trigger reconnection if not already in progress
-      if (this.token && !this.reconnectTimeout) {
+      // Only schedule reconnect if socket is CLOSED/undefined — not when it's still CONNECTING
+      if (this.token && !this.reconnectTimeout && this.ws?.readyState !== WebSocket.CONNECTING) {
         this.scheduleReconnect()
       }
     }

web/src/stores/indicators.ts

@@ -1,6 +1,63 @@
 import { defineStore } from 'pinia'
 import { ref } from 'vue'
 
+export interface CustomIndicatorParam {
+  type: 'int' | 'float' | 'bool' | 'string'
+  default: any
+  description?: string
+  min?: number
+  max?: number
+}
+
+/**
+ * Per-series plot configuration.
+ * style maps to LineStudyPlotStyle: 0=Line, 1=Histogram, 3=Dots/Cross,
+ * 4=Area, 5=Columns, 6=Circles, 9=StepLine.
+ */
+export interface PlotConfig {
+  style: number
+  color?: string
+  linewidth?: number
+  visible?: boolean
+}
+
+/** Shaded region between two plots ("plot_plot") or two bands ("hline_hline"). */
+export interface FilledAreaConfig {
+  id: string
+  type: 'plot_plot' | 'hline_hline'
+  series1: string
+  series2: string
+  color?: string
+  opacity?: number
+}
+
+/** Horizontal reference line (e.g. RSI ob/os). linestyle: 0=solid, 1=dotted, 2=dashed. */
+export interface BandConfig {
+  id: string
+  value: number
+  color?: string
+  linewidth?: number
+  linestyle?: number
+  visible?: boolean
+}
+
+export interface CustomIndicatorColumn {
+  name: string
+  display_name?: string
+  description?: string
+  plot?: PlotConfig
+}
+
+export interface CustomIndicatorMetadata {
+  display_name: string
+  parameters: Record<string, CustomIndicatorParam>
+  input_series: string[]
+  output_columns: CustomIndicatorColumn[]
+  pane: 'price' | 'separate'
+  filled_areas?: FilledAreaConfig[]
+  bands?: BandConfig[]
+}
+
 export interface IndicatorInstance {
   id: string
   pandas_ta_name: string
@@ -15,6 +72,8 @@ export interface IndicatorInstance {
   created_at?: number
   modified_at?: number
   original_id?: string
+  /** Populated for custom_ indicators; drives TV custom study auto-construction. */
+  custom_metadata?: CustomIndicatorMetadata
 }
 
 export const useIndicatorStore = defineStore('indicators', () => {