@aws/ml-container-creator 0.15.0 → 1.0.0

package/config/parameter-schema-v2.json

@@ -893,7 +893,7 @@
                 ]
             },
             "widget": null,
-            "prompt": null,
+            "prompt": "external",
             "deprecated": false,
             "since": "0.3.0"
         },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aws/ml-container-creator",
-  "version": "0.15.0",
+  "version": "1.0.0",
   "description": "Build and deploy custom ML containers on AWS SageMaker with minimal configuration.",
   "main": "src/index.js",
   "bin": {

package/servers/endpoint-picker/index.js

@@ -78,6 +78,7 @@ function getGpusForInstance(instanceType) {
 let _SageMakerClient = null;
 let _ListEndpointsCommand = null;
 let _DescribeEndpointCommand = null;
+let _DescribeEndpointConfigCommand = null;
 let _ListInferenceComponentsCommand = null;
 let _fromIni = null;
 
@@ -90,6 +91,7 @@ async function _ensureSdkLoaded() {
     _SageMakerClient = sdk.SageMakerClient;
     _ListEndpointsCommand = sdk.ListEndpointsCommand;
     _DescribeEndpointCommand = sdk.DescribeEndpointCommand;
+    _DescribeEndpointConfigCommand = sdk.DescribeEndpointConfigCommand;
     _ListInferenceComponentsCommand = sdk.ListInferenceComponentsCommand;
     try {
         const credentialProviders = await import('@aws-sdk/credential-providers');
@@ -197,9 +199,24 @@ async function fetchEndpoints(client, { limit = 10, showFull = false } = {}) {
             const primaryVariant = variants[0] || {};
 
             const variantName = primaryVariant.VariantName || 'AllTraffic';
-            const instanceType = primaryVariant.CurrentInstanceCount !== null && primaryVariant.CurrentInstanceCount !== undefined
-                ? (primaryVariant.InstanceType || detail.ProductionVariants?.[0]?.InstanceType || 'unknown')
-                : (primaryVariant.InstanceType || 'unknown');
+            let instanceType = primaryVariant.InstanceType || null;
+
+            // For IC-based endpoints, InstanceType may not be in the variant runtime response.
+            // Fall back to DescribeEndpointConfig which always has it.
+            if (!instanceType && detail.EndpointConfigName) {
+                try {
+                    const ecCmd = new _DescribeEndpointConfigCommand({ EndpointConfigName: detail.EndpointConfigName });
+                    const ecDetail = await client.send(ecCmd);
+                    const ecVariant = (ecDetail.ProductionVariants || [])[0];
+                    if (ecVariant?.InstanceType) {
+                        instanceType = ecVariant.InstanceType;
+                    }
+                } catch (ecErr) {
+                    log(`Warning: could not describe endpoint config for "${endpointName}": ${ecErr.message}`);
+                }
+            }
+            instanceType = instanceType || 'unknown';
+
             const instanceCount = primaryVariant.CurrentInstanceCount ?? primaryVariant.DesiredInstanceCount ?? 1;
             const hasInstancePools = !!(primaryVariant.InstancePools && primaryVariant.InstancePools.length > 0);
 
@@ -387,17 +404,9 @@ server.tool(
         limit: z.number().int().positive().default(10).describe('Maximum number of endpoints to return'),
         context: z.record(z.string(), z.any()).optional().describe('Current configuration context (awsRegion, awsProfile, deploymentTarget)')
     },
-    async ({ parameters, limit, context }) => {
-        // Only respond if parameters includes endpointName AND context.deploymentTarget is realtime-inference
-        if (!parameters.includes('endpointName')) {
-            return {
-                content: [{
-                    type: 'text',
-                    text: JSON.stringify({ values: {}, choices: {} })
-                }]
-            };
-        }
-
+    async ({ parameters: _parameters, limit, context }) => {
+        // Only respond if context.deploymentTarget is realtime-inference
+        // Note: parameters may be empty when called on-demand via queryMcpServer()
         if (context?.deploymentTarget && context.deploymentTarget !== 'realtime-inference') {
             return {
                 content: [{

package/src/lib/generated/cli-options.js

@@ -1,6 +1,6 @@
 // AUTO-GENERATED by scripts/codegen-cli.js — DO NOT EDIT
 // Source: config/parameter-schema-v2.json
-// Generated: 2026-06-22T13:49:00.815Z
+// Generated: 2026-06-23T20:55:23.381Z
 
 /**
  * CLI option definitions derived from parameter-schema-v2.json.

package/src/lib/generated/parameter-matrix.js

@@ -1,6 +1,6 @@
 // AUTO-GENERATED by scripts/codegen-parameter-matrix.js — DO NOT EDIT
 // Source: config/parameter-schema-v2.json
-// Generated: 2026-06-22T13:49:00.924Z
+// Generated: 2026-06-23T20:55:23.482Z
 
 /**
  * Parameter matrix defining how each parameter is loaded from various sources.
@@ -225,7 +225,7 @@ export const parameterMatrix = {
         'configFile': true,
         'packageJson': false,
         'mcp': true,
-        'promptable': false,
+        'promptable': true,
         'required': false,
         'default': null,
         'valueSpace': 'unbounded'

package/src/lib/generated/validation-rules.js

@@ -1,6 +1,6 @@
 // AUTO-GENERATED by scripts/codegen-validator.js — DO NOT EDIT
 // Source: config/parameter-schema-v2.json
-// Generated: 2026-06-22T13:49:00.849Z
+// Generated: 2026-06-23T20:55:23.412Z
 
 /**
  * Validation rules derived from parameter-schema-v2.json.

package/src/lib/mcp-client.js

@@ -143,9 +143,23 @@ class McpClient {
         // Build context from bounded parameters that have defaults
         const context = this._buildContext();
 
+        // Auto-discover tool name if using the default (get_ml_config)
+        // Each server registers its own tool name (e.g. get_base_images, get_inference_endpoints)
+        let toolName = this.toolName;
+        if (toolName === DEFAULT_TOOL_NAME) {
+            try {
+                const toolList = await this._client.listTools();
+                if (toolList && toolList.tools && toolList.tools.length > 0) {
+                    toolName = toolList.tools[0].name;
+                }
+            } catch (_listErr) {
+                // Fall through to use default tool name
+            }
+        }
+
         // Call the configured tool
         const result = await this._client.callTool({
-            name: this.toolName,
+            name: toolName,
             arguments: {
                 parameters: unboundedParams,
                 limit: this.limit,

package/src/lib/mcp-query-runner.js

@@ -371,9 +371,13 @@ export default class McpQueryRunner {
         console.log('   🔍 Querying endpoint-picker...');
 
         try {
+            // Pass awsProfile from bootstrap config for credential resolution
+            const awsProfile = this.runner.configManager?.config?.awsProfile
+                || this.runner.options?.profile || process.env.AWS_PROFILE || null;
             const result = await cm.queryMcpServer('endpoint-picker', {
                 awsRegion: infraAnswers.awsRegion,
-                deploymentTarget: 'realtime-inference'
+                deploymentTarget: 'realtime-inference',
+                ...(awsProfile ? { awsProfile } : {})
             });
 
             if (result && result.choices?.endpointName?.length > 0) {

package/src/lib/prompt-runner.js

@@ -224,25 +224,39 @@ export default class PromptRunner {
         // Requirements: 3.3, 4.3, 4.4 — endpoint-picker MCP query
         let existingEndpointAnswers = {};
         if (regionAndTargetAnswers.deploymentTarget === 'realtime-inference') {
-            // Query endpoint-picker MCP server for available endpoints
-            const resolvedRegion = regionAndTargetAnswers.customAwsRegion || regionAndTargetAnswers.awsRegion;
-            await this.mcpQueryRunner._queryMcpForEndpoints({ ...regionAndTargetAnswers, awsRegion: resolvedRegion }, explicitConfig);
-
-            const endpointPreviousAnswers = {
-                ...regionAndTargetAnswers,
-                ...(this._mcpEndpointChoices ? { _mcpEndpointChoices: this._mcpEndpointChoices } : {})
-            };
-            existingEndpointAnswers = await this._runPhase(
-                infraExistingEndpointPrompts,
-                endpointPreviousAnswers,
+            // First ask if user wants to attach to existing endpoint (no MCP call yet)
+            const attachAnswer = await this._runPhase(
+                [infraExistingEndpointPrompts[0]],
+                { ...regionAndTargetAnswers },
                 explicitConfig,
                 existingConfig
             );
 
-            // Resolve custom endpoint name
-            if (existingEndpointAnswers.customExistingEndpointName) {
-                existingEndpointAnswers.existingEndpointName = existingEndpointAnswers.customExistingEndpointName;
-                delete existingEndpointAnswers.customExistingEndpointName;
+            if (attachAnswer.useExistingEndpoint === 'yes') {
+                // Only now query endpoint-picker MCP server
+                const resolvedRegion = regionAndTargetAnswers.customAwsRegion || regionAndTargetAnswers.awsRegion;
+                await this.mcpQueryRunner._queryMcpForEndpoints({ ...regionAndTargetAnswers, awsRegion: resolvedRegion }, explicitConfig);
+
+                const endpointPreviousAnswers = {
+                    ...regionAndTargetAnswers,
+                    ...attachAnswer,
+                    ...(this._mcpEndpointChoices ? { _mcpEndpointChoices: this._mcpEndpointChoices } : {})
+                };
+                existingEndpointAnswers = await this._runPhase(
+                    infraExistingEndpointPrompts.slice(1),
+                    endpointPreviousAnswers,
+                    explicitConfig,
+                    existingConfig
+                );
+                existingEndpointAnswers.useExistingEndpoint = 'yes';
+
+                // Resolve custom endpoint name
+                if (existingEndpointAnswers.customExistingEndpointName) {
+                    existingEndpointAnswers.existingEndpointName = existingEndpointAnswers.customExistingEndpointName;
+                    delete existingEndpointAnswers.customExistingEndpointName;
+                }
+            } else {
+                existingEndpointAnswers = attachAnswer;
             }
         }
 
@@ -376,11 +390,12 @@ export default class PromptRunner {
             const sizerRecs = this._instanceSizerMetadata.recommendations || [];
             const finalInstanceType = instanceAnswers.customInstanceType || instanceAnswers.instanceType;
             const matchingRec = sizerRecs.find(r => r.instanceType === finalInstanceType);
-            const tpRec = matchingRec || sizerRecs[0];
-            if (tpRec && tpRec.tensorParallelism > 1) {
-                this._autoTensorParallelism = tpRec.tensorParallelism;
-                this._autoGpuCount = tpRec.gpuCount;
-                console.log(`   ✓ Auto-set tensor parallelism: TP=${tpRec.tensorParallelism} (${tpRec.gpuCount} GPUs)`);
+            // Only use sizer TP recommendation if user selected a recommended instance
+            // Custom instances resolve TP from the instance catalog in template-variable-resolver
+            if (matchingRec && matchingRec.tensorParallelism > 1) {
+                this._autoTensorParallelism = matchingRec.tensorParallelism;
+                this._autoGpuCount = matchingRec.gpuCount;
+                console.log(`   ✓ Auto-set tensor parallelism: TP=${matchingRec.tensorParallelism} (${matchingRec.gpuCount} GPUs)`);
             }
 
             // Display capacity type confirmation for selected instance

package/templates/do/.benchmark_writer.py

@@ -1401,8 +1401,18 @@ def _load_config_file(config_path):
                     'BASE_IMAGE_VERSION': 'base_image_version',
                     'BENCHMARK_CONCURRENCY': 'benchmark_concurrency',
                 }
+                # Also capture IC_ENV_* serving config vars
+                ic_env_map = {
+                    'IC_ENV_VLLM_MAX_MODEL_LEN': 'max_model_len',
+                    'IC_ENV_VLLM_QUANTIZATION': 'quantization',
+                    'IC_ENV_VLLM_GPU_MEMORY_UTILIZATION': 'gpu_memory_utilization',
+                    'IC_ENV_VLLM_KV_CACHE_DTYPE': 'kv_cache_dtype',
+                    'IC_ENV_VLLM_TENSOR_PARALLEL_SIZE': 'tensor_parallel_degree',
+                }
                 if key in shell_map:
                     context[shell_map[key]] = value
+                elif key in ic_env_map:
+                    context[ic_env_map[key]] = value
 
     except Exception:
         pass
@@ -1419,6 +1429,30 @@ def _load_config_file(config_path):
         parts = context['model_name'].rstrip('/').split('/')
         context['model_name'] = parts[-1] if parts else context['model_name']
 
+    # Also scan IC config files (do/ic/*.conf) for IC_ENV_* serving params
+    # These override do/config values for serving-specific settings
+    try:
+        import glob
+        config_dir = os.path.dirname(os.path.abspath(config_path))
+        ic_dir = os.path.join(config_dir, 'ic')
+        ic_env_map = {
+            'IC_ENV_VLLM_MAX_MODEL_LEN': 'max_model_len',
+            'IC_ENV_VLLM_QUANTIZATION': 'quantization',
+            'IC_ENV_VLLM_GPU_MEMORY_UTILIZATION': 'gpu_memory_utilization',
+            'IC_ENV_VLLM_KV_CACHE_DTYPE': 'kv_cache_dtype',
+            'IC_ENV_VLLM_TENSOR_PARALLEL_SIZE': 'tensor_parallel_degree',
+        }
+        for conf_file in sorted(glob.glob(os.path.join(ic_dir, '*.conf'))):
+            with open(conf_file, 'r') as f:
+                for line in f:
+                    match = re.match(r'^export\s+([A-Z_][A-Z0-9_]*)=["\']?([^"\']*)["\']?\s*$', line.strip())
+                    if match:
+                        key, value = match.group(1), match.group(2)
+                        if key in ic_env_map and value:
+                            context[ic_env_map[key]] = value
+    except Exception:
+        pass  # IC config scanning is best-effort
+
     return context
 
 

package/templates/do/.register_helper.py

@@ -89,6 +89,8 @@ def _truncate_metadata(props):
     result = {}
     for key, value in props.items():
         str_val = str(value) if value is not None else ""
+        if not str_val:
+            continue  # SageMaker requires min length 1 for metadata values — skip empty
         if len(str_val) > MAX_METADATA_VALUE_LEN:
             _warn(f"Metadata '{key}' truncated ({len(str_val)} → {MAX_METADATA_VALUE_LEN} chars)")
             str_val = str_val[: MAX_METADATA_VALUE_LEN - 1] + "…"
@@ -264,33 +266,41 @@ def cmd_register_model(args):
     container_image = args.container_image or ""
     model_data_url = args.model_data_url or ""
 
-    inference_spec = {
-        "Containers": [
-            {
-                "Image": container_image,
-            }
-        ],
-        "SupportedContentTypes": ["application/json"],
-        "SupportedResponseMIMETypes": ["application/json"],
-    }
-    # Only include ModelDataUrl if provided
-    if model_data_url:
-        inference_spec["Containers"][0]["ModelDataUrl"] = model_data_url
-
     # Step 4: Create Model Package version (AC-1.2, AC-1.7)
     description = f"{args.deployment_config or 'model'} on {args.instance_type or 'unknown'}"
 
     print(f"Registering model version in {project_name}...", file=sys.stderr)
     try:
-        pkg = ModelPackage.create(
-            model_package_group_name=project_name,
-            model_package_description=description,
-            inference_specification=inference_spec,
-            customer_metadata_properties=metadata,
-            model_approval_status="Approved",
-        )
+        # Use boto3 directly — sagemaker-core v2.14 has a KeyError bug in ModelPackage.create()
+        # where it tries to read response["ModelPackageName"] but the API returns "ModelPackageArn".
+        import boto3
+        sm_client = boto3.client("sagemaker", region_name=region)
+
+        create_params = {
+            "ModelPackageGroupName": project_name,
+            "ModelPackageDescription": description,
+            "ModelApprovalStatus": "Approved",
+        }
+        if container_image:
+            create_params["InferenceSpecification"] = {
+                "Containers": [{"Image": container_image}],
+                "SupportedContentTypes": ["application/json"],
+                "SupportedResponseMIMETypes": ["application/json"],
+            }
+            if model_data_url:
+                create_params["InferenceSpecification"]["Containers"][0]["ModelDataUrl"] = model_data_url
+        if model_data_url:
+            if "InferenceSpecification" not in create_params:
+                # Store model data URL in metadata if no container image
+                if not metadata:
+                    metadata = {}
+                metadata["modelDataUrl"] = model_data_url[:1024]
+        if metadata:
+            create_params["CustomerMetadataProperties"] = metadata
+
+        response = sm_client.create_model_package(**create_params)
+        model_package_arn = response["ModelPackageArn"]
 
-        model_package_arn = pkg.model_package_arn
         # Extract version number from ARN (format: .../project-name/version)
         version = _extract_version_from_arn(model_package_arn)
 
@@ -407,33 +417,39 @@ def cmd_register_adapter(args):
     container_image = args.container_image or ""
     model_data_url = args.model_data_url or ""
 
-    inference_spec = {
-        "Containers": [
-            {
-                "Image": container_image,
-            }
-        ],
-        "SupportedContentTypes": ["application/json"],
-        "SupportedResponseMIMETypes": ["application/json"],
-    }
-    if model_data_url:
-        inference_spec["Containers"][0]["ModelDataUrl"] = model_data_url
-
     # Step 4: Create adapter Model Package version (AC-2.1)
     technique = args.tune_technique or "unknown"
     description = f"adapter ({technique}) on {args.instance_type or 'unknown'}, parent: {parent_version_arn}"
 
     print(f"Registering adapter version in {project_name}...", file=sys.stderr)
     try:
-        pkg = ModelPackage.create(
-            model_package_group_name=project_name,
-            model_package_description=description,
-            inference_specification=inference_spec,
-            customer_metadata_properties=metadata,
-            model_approval_status="Approved",
-        )
+        # Use boto3 directly — sagemaker-core v2.14 has a KeyError bug in ModelPackage.create()
+        import boto3
+        sm_client = boto3.client("sagemaker", region_name=region)
+
+        create_params = {
+            "ModelPackageGroupName": project_name,
+            "ModelPackageDescription": description,
+            "ModelApprovalStatus": "Approved",
+        }
+        if container_image:
+            create_params["InferenceSpecification"] = {
+                "Containers": [{"Image": container_image}],
+                "SupportedContentTypes": ["application/json"],
+                "SupportedResponseMIMETypes": ["application/json"],
+            }
+            if model_data_url:
+                create_params["InferenceSpecification"]["Containers"][0]["ModelDataUrl"] = model_data_url
+        elif model_data_url:
+            if not metadata:
+                metadata = {}
+            metadata["modelDataUrl"] = model_data_url[:1024]
+        if metadata:
+            create_params["CustomerMetadataProperties"] = metadata
+
+        response = sm_client.create_model_package(**create_params)
+        model_package_arn = response["ModelPackageArn"]
 
-        model_package_arn = pkg.model_package_arn
         version = _extract_version_from_arn(model_package_arn)
 
         print(f"Registered adapter version {version}: {model_package_arn}", file=sys.stderr)
@@ -1133,6 +1149,12 @@ def main():
         parser.print_help()
         sys.exit(1)
 
+    # Set region before any sagemaker-core import (creates boto3 clients at import time)
+    region = getattr(args, 'region', None) or os.environ.get('AWS_DEFAULT_REGION') or os.environ.get('AWS_REGION')
+    if region:
+        os.environ['AWS_DEFAULT_REGION'] = region
+        os.environ.setdefault('AWS_REGION', region)
+
     if args.command == "create-mpg":
         cmd_create_mpg(args)
     elif args.command == "register-model":

package/templates/do/benchmark

@@ -117,10 +117,12 @@ if [ "${ARG_STATUS}" = true ]; then
                     tar_file=""
                     tar_file=$(find "${LOCAL_RESULTS_DIR}" -name "output.tar.gz" -type f 2>/dev/null | head -1)
                     if [ -n "${tar_file}" ]; then
-                        # Detect whether tar has a leading directory prefix
-                        _tar_first=""
-                        _tar_first=$(tar -tzf "${tar_file}" 2>/dev/null | head -1)
-                        if echo "${_tar_first}" | grep -qE '^[^/]+/$'; then
+                        # Detect whether ALL entries share a common leading directory prefix
+                        _tar_prefix_count=""
+                        _tar_prefix_count=$(tar -tzf "${tar_file}" 2>/dev/null | sed 's|/.*||' | sort -u | wc -l | tr -d ' ')
+                        _tar_first_dir=""
+                        _tar_first_dir=$(tar -tzf "${tar_file}" 2>/dev/null | head -1)
+                        if [ "${_tar_prefix_count}" = "1" ] && echo "${_tar_first_dir}" | grep -qE '^[^/]+/$'; then
                             tar -xzf "${tar_file}" --strip-components=1 -C "${LOCAL_RESULTS_DIR}/output/" 2>/dev/null || true
                         else
                             tar -xzf "${tar_file}" -C "${LOCAL_RESULTS_DIR}/output/" 2>/dev/null || true
@@ -1097,11 +1099,14 @@ if [ "${JOB_STATUS}" = "Completed" ]; then
             # Extract any tar.gz archives (benchmark service packages results as output.tar.gz)
             for ARCHIVE in $(find "${LOCAL_RESULTS_DIR}" -name "*.tar.gz" -type f 2>/dev/null); do
                 ARCHIVE_DIR=$(dirname "${ARCHIVE}")
-                # Detect whether tar has a leading directory prefix to strip.
-                # Some AIPerf versions wrap in output/, others are flat.
-                _TAR_FIRST=$(tar -tzf "${ARCHIVE}" 2>/dev/null | head -1)
-                if echo "${_TAR_FIRST}" | grep -qE '^[^/]+/$'; then
-                    # Leading directory (e.g., "output/") — strip it
+                # Detect whether ALL entries share a common leading directory prefix.
+                # Only strip if every entry starts with the same dir (e.g., "output/file1", "output/file2").
+                # A flat archive with mixed top-level files/dirs (e.g., "plots/", "profile_export.jsonl")
+                # must NOT be stripped.
+                _TAR_PREFIX=$(tar -tzf "${ARCHIVE}" 2>/dev/null | sed 's|/.*||' | sort -u | wc -l | tr -d ' ')
+                _TAR_FIRST_DIR=$(tar -tzf "${ARCHIVE}" 2>/dev/null | head -1)
+                if [ "${_TAR_PREFIX}" = "1" ] && echo "${_TAR_FIRST_DIR}" | grep -qE '^[^/]+/$'; then
+                    # Single common leading directory (e.g., all under "output/") — strip it
                     tar -xzf "${ARCHIVE}" --strip-components=1 -C "${ARCHIVE_DIR}" 2>/dev/null || true
                 else
                     # Flat archive — extract as-is

package/templates/do/register

@@ -1137,9 +1137,14 @@ ml-container-creator "${CMD_ARGS[@]}"
 # ============================================================
 
 # Container image URI for the deployed model
-CONTAINER_IMAGE_URI="${ECR_REPOSITORY_NAME}:${PROJECT_NAME}-latest"
-if [ -n "${BASE_IMAGE:-}" ]; then
-    CONTAINER_IMAGE_URI="${BASE_IMAGE}"
+# Build full ECR URI from profile (accountId + region + repoName + tag)
+_ACCOUNT_ID="${_PROFILE_accountId:-}"
+_REGION="${AWS_DEFAULT_REGION:-${_PROFILE_awsRegion:-us-east-1}}"
+if [ -n "${_ACCOUNT_ID}" ] && [ -n "${ECR_REPOSITORY_NAME}" ]; then
+    CONTAINER_IMAGE_URI="${_ACCOUNT_ID}.dkr.ecr.${_REGION}.amazonaws.com/${ECR_REPOSITORY_NAME}:${PROJECT_NAME}-latest"
+else
+    # No ECR info available — MPG will be registered without InferenceSpecification
+    CONTAINER_IMAGE_URI=""
 fi
 
 # Model data S3 URI (from do/config if set)

package/templates/do/tune

@@ -829,7 +829,25 @@ _validate_dataset() {
     if [ -z "${dataset}" ]; then
         if [ -n "${ARG_DATASET_NAME}" ]; then
             # Name-based resolution happens below via resolve-dataset
-            :
+            echo "🔍 Resolving dataset '${ARG_DATASET_NAME}' from registry..."
+            local resolve_result
+            resolve_result=$(python3 "${SCRIPT_DIR}/.register_helper.py" resolve-dataset \
+                --name "${ARG_DATASET_NAME}" 2>/dev/null) || resolve_result=""
+
+            if [ -n "${resolve_result}" ]; then
+                local resolved_uri
+                resolved_uri=$(echo "${resolve_result}" | grep -E '^\{' | tail -1 | python3 -c "import sys,json; print(json.load(sys.stdin).get('s3_uri',''))" 2>/dev/null) || resolved_uri=""
+                if [ -n "${resolved_uri}" ]; then
+                    echo "   Resolved to: ${resolved_uri}"
+                    dataset="${resolved_uri}"
+                    RESOLVED_DATASET_S3_URI="${resolved_uri}"
+                    return 0
+                fi
+            fi
+            echo "❌ Dataset '${ARG_DATASET_NAME}' not found in registry"
+            echo "   Run ./do/tune --list-datasets to see available datasets."
+            echo "   Register: ./do/register dataset <name> --s3-uri <uri> --technique <sft|dpo>"
+            exit 1
         else
             echo "❌ --dataset is required"
             echo "   Provide an S3 URI (s3://bucket/path.jsonl), HF reference (hf://org/name), or registered name"