@inspectr/mcplab-mcp-server 1.7.2 → 1.8.1

package/dist/runtime.js

@@ -5,10 +5,11 @@ import { basename, dirname, extname, resolve, join, sep, relative } from 'node:p
 import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
 import { StreamableHTTPServerTransport } from '@modelcontextprotocol/sdk/server/streamableHttp.js';
 import { isInitializeRequest } from '@modelcontextprotocol/sdk/types.js';
-import { loadConfig, hashConfig, runAll, selectScenarios, loadOrBuildSearchIndex, indexNeedsRefresh, getResultsIndexPaths, resolveRunArtifactPath, searchDocs, getContext, applyRuntimeServerOverrides, readLibraryAgentsAndServers } from '@inspectr/mcplab-core';
+import { loadConfig, hashConfig, runAll, selectScenarios, loadOrBuildSearchIndex, indexNeedsRefresh, getResultsIndexPaths, resolveRunArtifactPath, searchDocs, getContext, applyRuntimeServerOverrides, readLibraryAgentsAndServers, resolveScenarioLibraryDir, createEvaluationConfigFile, updateEvaluationConfigFile, readEvaluationConfigFile, createTestCaseFile, readTestCaseFile, updateTestCaseFile, buildScenarioEntry } from '@inspectr/mcplab-core';
 import { renderReport } from '@inspectr/mcplab-reporting';
 import { parse as parseYaml, stringify as stringifyYaml } from 'yaml';
 import { z } from 'zod';
+import { loadSkills, readSkillFile, registerSkills } from './skills.js';
 import { buildAggregateRunsReport, buildCompareRunsReport } from './mcp-run-calculations.js';
 const PACKAGE_JSON = JSON.parse(readFileSync(new URL('../package.json', import.meta.url), 'utf8'));
 const SERVER_VERSION = typeof PACKAGE_JSON.version === 'string' && PACKAGE_JSON.version.trim().length > 0
@@ -59,12 +60,7 @@ const AggregateRowSchema = z.object({
         max: z.string()
     })
         .optional(),
-    total_runs: z.number().int().nonnegative(),
-    passed_runs: z.number().int().nonnegative(),
-    failed_runs: z.number().int().nonnegative(),
-    pass_rate: z.number(),
-    avg_tool_calls_per_run: z.number(),
-    avg_tool_latency_ms: z.number().nullable()
+    summary: MetricSummarySchema
 });
 const CompareRowSchema = z.object({
     key: z.string(),
@@ -122,8 +118,28 @@ const LibraryEntrySchema = z.object({
     bundleRoot: z.string(),
     servers: z.array(LibraryServerEntrySchema),
     agents: z.array(LibraryAgentEntrySchema),
+    test_cases: z.array(LibraryScenarioEntrySchema),
     scenarios: z.array(LibraryScenarioEntrySchema)
 });
+const EvaluationConfigListItemSchema = z.object({
+    id: z.string(),
+    name: z.string(),
+    relative_path: z.string(),
+    config_path: z.string(),
+    suite_path: z.string(),
+    updated_at: z.string(),
+    scenario_count: z.number().int().nonnegative(),
+    agent_count: z.number().int().nonnegative(),
+    error: z.string().optional()
+});
+const EvaluationConfigListSchema = z.object({
+    evals_dir: z.string(),
+    total: z.number().int().nonnegative(),
+    matching: z.number().int().nonnegative(),
+    filters: GenericObjectSchema,
+    sort: z.object({ by: z.string(), direction: z.enum(['asc', 'desc']) }),
+    configs: z.array(EvaluationConfigListItemSchema)
+});
 const ServerAuthSchema = z.union([
     z.object({
         type: z.literal('bearer'),
@@ -148,12 +164,148 @@ const ServerEntrySchema = z.object({
     url: z.string(),
     auth: ServerAuthSchema.optional()
 });
+const ResponseAssertionSchema = z
+    .discriminatedUnion('type', [
+    z.object({
+        type: z.literal('regex').describe('Match the final response against a regular expression.'),
+        pattern: z.string().describe('JavaScript regex pattern; matching is case-insensitive.')
+    }),
+    z.object({
+        type: z.literal('contains').describe('Require literal text in the final response.'),
+        value: z.string().describe('Case-insensitive text that must appear.')
+    }),
+    z.object({
+        type: z
+            .literal('not_contains')
+            .describe('Require literal text to be absent from the final response.'),
+        value: z.string().describe('Case-insensitive text that must not appear.')
+    }),
+    z.object({
+        type: z
+            .literal('starts_with')
+            .describe('Require the final response to start with literal text.'),
+        value: z.string().describe('Case-insensitive prefix.')
+    }),
+    z.object({
+        type: z.literal('ends_with').describe('Require the final response to end with literal text.'),
+        value: z.string().describe('Case-insensitive suffix.')
+    }),
+    z.object({
+        type: z
+            .literal('equals')
+            .describe('Require the entire final response to equal literal text.'),
+        value: z.string().describe('Case-insensitive expected response.')
+    }),
+    z.object({
+        type: z.literal('jsonpath').describe('Inspect a value in a JSON final response.'),
+        path: z.string().describe('JSONPath expression.'),
+        equals: z
+            .union([z.string(), z.number(), z.boolean()])
+            .optional()
+            .describe('Optional primitive value that must match.')
+    }),
+    z.object({
+        type: z
+            .literal('jsonpath_exists')
+            .describe('Require a JSONPath to exist in a JSON final response.'),
+        path: z.string().describe('JSONPath expression.')
+    }),
+    z.object({
+        type: z
+            .literal('jsonpath_not_exists')
+            .describe('Require a JSONPath to be absent from a JSON final response.'),
+        path: z.string().describe('JSONPath expression.')
+    })
+])
+    .describe('Assertions evaluated against the final response text or JSON.');
+const ToolInputAssertionSchema = z
+    .discriminatedUnion('type', [
+    z.object({
+        type: z.literal('contains').describe('Require text in serialized tool arguments.'),
+        tool: z.string().describe('Raw MCP tool name.'),
+        value: z.string().describe('Case-insensitive text that must appear in arguments.')
+    }),
+    z.object({
+        type: z.literal('regex').describe('Match serialized tool arguments with a regex.'),
+        tool: z.string().describe('Raw MCP tool name.'),
+        pattern: z.string().describe('JavaScript regex pattern.')
+    }),
+    z.object({
+        type: z.literal('jsonpath').describe('Inspect a value in tool arguments.'),
+        tool: z.string().describe('Raw MCP tool name.'),
+        path: z.string().describe('JSONPath expression over the tool arguments.'),
+        equals: z
+            .union([z.string(), z.number(), z.boolean()])
+            .optional()
+            .describe('Optional primitive value that must match.')
+    })
+])
+    .describe('Assertions evaluated against calls to a named MCP tool.');
+const AgentAssertionSchema = z
+    .object({
+    label: z.string().describe('Stable name for the semantic check.'),
+    prompt: z.string().describe('Question the judge should answer about the scenario result.')
+})
+    .describe('Semantic checks for requirements that are not reliably deterministic.');
+const AgentContextSchema = z
+    .object({
+    include_prompt: z
+        .boolean()
+        .optional()
+        .describe('Include the scenario prompt in judge context.'),
+    include_tool_sequence: z
+        .boolean()
+        .optional()
+        .describe('Include called raw MCP tool names in judge context.'),
+    include_tool_inputs: z
+        .boolean()
+        .optional()
+        .describe('Include raw MCP tool arguments in judge context.')
+})
+    .describe('Shared context sent to every agent assertion.');
+const EvalRulesSchema = z
+    .object({
+    tool_constraints: z
+        .object({
+        required_tools: z
+            .array(z.string())
+            .optional()
+            .describe('Raw MCP tools that must be called.'),
+        forbidden_tools: z
+            .array(z.string())
+            .optional()
+            .describe('Raw MCP tools that must not be called.')
+    })
+        .strict()
+        .describe('Deterministic MCP tool-usage constraints.')
+        .optional(),
+    tool_sequence: z
+        .array(z.string())
+        .optional()
+        .describe('Flat ordered list of raw MCP tool names that must be called.'),
+    tool_input_assertions: z
+        .array(ToolInputAssertionSchema)
+        .optional()
+        .describe('Checks over MCP tool arguments.'),
+    response_assertions: z
+        .array(ResponseAssertionSchema)
+        .optional()
+        .describe('Checks over the final response.'),
+    agent_assertions: z
+        .array(AgentAssertionSchema)
+        .optional()
+        .describe('Semantic checks judged by an evaluation agent.'),
+    agent_context: AgentContextSchema.optional()
+})
+    .strict()
+    .describe('Full MCPLab evaluation rules. Use this object for all supported assertion types.');
 const ScenarioEntrySchema = z.object({
     id: z.string(),
+    name: z.string().optional(),
     agent: z.string().optional(),
-    servers: z.array(z.string()),
+    mcp_servers: z.array(z.object({ ref: z.string() })),
     prompt: z.string(),
-    eval: GenericObjectSchema.optional(),
+    eval: EvalRulesSchema.optional(),
     extract: z
         .array(z.object({
         name: z.string(),
@@ -180,14 +332,7 @@ const ResolvedScenarioSchema = z.object({
     servers: z.array(z.string()),
     agent: z.string().optional(),
     prompt: z.string().optional(),
-    eval: z
-        .object({
-        type: z.string(),
-        assertions: z.array(z.string()).optional(),
-        rubric: GenericObjectSchema.optional()
-    })
-        .passthrough()
-        .optional(),
+    eval: GenericObjectSchema.optional(),
     extract: z
         .array(z
         .object({
@@ -532,6 +677,7 @@ export function createConfiguredServer() {
         websiteUrl: 'https://mcplab.inspectr.dev',
         icons: [{ src: SERVER_ICON_URL, mimeType: 'image/svg+xml' }]
     });
+    registerSkills(server);
     registerTools(server);
     registerPrompts(server);
     return server;
@@ -547,6 +693,79 @@ export function registerTools(server) {
             annotations: inferToolAnnotations(name, resolvedTitle, config.annotations)
         }, cb);
     };
+    registerTool('mcplab_build_app_link', {
+        description: 'Build a safe deep link to a local MCPLab app view.',
+        inputSchema: {
+            view: z.enum([
+                'home',
+                'evaluations',
+                'run_evaluation',
+                'results',
+                'result_detail',
+                'compare',
+                'tool_analysis',
+                'tool_analysis_results',
+                'tool_analysis_result',
+                'library_servers',
+                'library_agents',
+                'library_test_cases',
+                'settings'
+            ]),
+            run_id: z.string().min(1).max(200).optional(),
+            tool_analysis_id: z.string().min(1).max(200).optional(),
+            config_id: z.string().min(1).max(200).optional(),
+            agent: z.string().min(1).max(200).optional(),
+            scenario: z.string().min(1).max(200).optional(),
+            time_filter: z.enum(['last', 'custom']).optional(),
+            time_preset: z.enum(['15min', '30min', '1h', '24h', '7d', '14d', '30d']).optional(),
+            time_start: z.string().datetime().optional(),
+            time_end: z.string().datetime().optional()
+        },
+        outputSchema: { url: z.string().url(), view: z.string(), path: z.string() }
+    }, async ({ view, run_id, tool_analysis_id, config_id, agent, scenario, time_filter, time_preset, time_start, time_end }) => withToolHandling(() => {
+        const paths = {
+            home: '/',
+            evaluations: '/evaluations',
+            run_evaluation: '/run',
+            results: '/results',
+            result_detail: '/results/detail',
+            compare: '/compare',
+            tool_analysis: '/tool-analysis',
+            tool_analysis_results: '/tool-analysis/results',
+            tool_analysis_result: '/tool-analysis/result',
+            library_servers: '/servers',
+            library_agents: '/agents',
+            library_test_cases: '/test-cases',
+            settings: '/settings'
+        };
+        const path = paths[view];
+        if (!path)
+            throw new Error(`Unsupported app view: ${view}`);
+        const url = new URL(path, process.env.MCPLAB_APP_URL ?? 'http://127.0.0.1:8787');
+        if (run_id)
+            url.searchParams.set('run', run_id);
+        if (tool_analysis_id)
+            url.searchParams.set('tool_analysis_id', tool_analysis_id);
+        if (config_id)
+            url.searchParams.set('config', config_id);
+        if (agent)
+            url.searchParams.set('agent', agent);
+        if (scenario)
+            url.searchParams.set('scenario', scenario);
+        if (time_filter)
+            url.searchParams.set('time_filter', time_filter);
+        if (time_preset)
+            url.searchParams.set('time_preset', time_preset);
+        if (time_start)
+            url.searchParams.set('time_start', time_start);
+        if (time_end)
+            url.searchParams.set('time_end', time_end);
+        if (view === 'result_detail' && !run_id)
+            throw new Error('run_id is required for result_detail');
+        if (view === 'tool_analysis_result' && !tool_analysis_id)
+            throw new Error('tool_analysis_id is required for tool_analysis_result');
+        return ok(`Built app link for ${view}`, { url: url.toString(), view, path });
+    }));
     registerTool('mcplab_write_markdown_report', {
         description: 'Write a Markdown (.md or .markdown) file to a path within the current workspace. Returns structured output with ok:true and the resolved path on success. On failure, returns ok:false with an error_code from: PATH_ESCAPE (path traversal attempt), FILE_EXISTS (file already exists and overwrite is false), PERMISSION_DENIED, PARENT_DIR_MISSING (create_dirs is false and parent does not exist), INVALID_EXTENSION (not .md or .markdown), IO_ERROR.',
         outputSchema: WriteMarkdownReportOutputSchema,
@@ -688,7 +907,7 @@ export function registerTools(server) {
         }
     });
     registerTool('mcplab_search_markdown_reports', {
-        description: 'List saved markdown reports under mcplab/reports. Supports filtering by run id substring to find reports linked to a result.',
+        description: 'Search saved Markdown report FILES under mcplab/reports. Use this to locate report documents by filename, path, run-id substring, or report text metadata; it does not search evaluation answers, traces, or tool-analysis records.',
         outputSchema: {
             reports_dir: z.string(),
             run_id_filter: z.string().optional(),
@@ -819,7 +1038,7 @@ export function registerTools(server) {
         outputSchema: LibraryEntrySchema,
         inputSchema: {
             kind: z
-                .enum(['all', 'servers', 'agents', 'scenarios'])
+                .enum(['all', 'servers', 'agents', 'test_cases', 'scenarios'])
                 .optional()
                 .describe('Which library category to list. Defaults to all.'),
             includeContent: z
@@ -838,6 +1057,7 @@ export function registerTools(server) {
                     bundleRoot: data.bundleRoot,
                     servers: selectedKind === 'servers' ? data.servers : [],
                     agents: selectedKind === 'agents' ? data.agents : [],
+                    test_cases: selectedKind === 'test_cases' ? data.test_cases : [],
                     scenarios: selectedKind === 'scenarios' ? data.scenarios : []
                 };
             return ok(`Loaded MCPLab library from ${root}`, structured);
@@ -847,14 +1067,16 @@ export function registerTools(server) {
         description: 'Get a specific reusable server, agent, or scenario definition from a MCPLab library bundle and return both structured data and YAML.',
         outputSchema: {
             bundleRoot: z.string(),
-            kind: z.enum(['servers', 'agents', 'scenarios']),
+            kind: z.enum(['servers', 'agents', 'test_cases', 'scenarios']),
             id: z.string(),
             file: z.string().optional(),
             yaml: z.string(),
             content: GenericObjectSchema
         },
         inputSchema: {
-            kind: z.enum(['servers', 'agents', 'scenarios']).describe('Library category.'),
+            kind: z
+                .enum(['servers', 'agents', 'test_cases', 'scenarios'])
+                .describe('Library category.'),
             id: z.string().describe('Entry id (for scenarios this is scenario.id, not filename).')
         }
     }, async ({ kind, id }) => {
@@ -910,8 +1132,246 @@ export function registerTools(server) {
             });
         });
     });
+    registerTool('mcplab_create_test_case', {
+        description: 'Create one reviewed MCPLab Test Case in mcplab/test-cases/. Put evaluation checks in the eval object for full support of response, tool-input, and agent assertions. Simple shorthand fields are supported for basic cases; do not provide a shorthand field together with its equivalent eval field. This is a scoped YAML write and refuses duplicate IDs.',
+        outputSchema: { id: z.string(), path: z.string(), test_case: ScenarioEntrySchema },
+        inputSchema: {
+            id: z
+                .string()
+                .min(1)
+                .describe('Stable Test Case id; letters, numbers, hyphens, and underscores.'),
+            name: z
+                .string()
+                .optional()
+                .describe('Human-readable Test Case label; defaults to id when blank.'),
+            servers: z.array(z.string()).min(1).describe('MCP server ids available to this Test Case.'),
+            prompt: z
+                .string()
+                .min(1)
+                .max(4000)
+                .describe('Task prompt executed by the evaluation agent.'),
+            required_tools: z
+                .array(z.string())
+                .optional()
+                .describe('Simple shorthand for eval.tool_constraints.required_tools. Use for basic checks only; put new or advanced rules under eval. Do not provide both.'),
+            forbidden_tools: z
+                .array(z.string())
+                .optional()
+                .describe('Simple shorthand for eval.tool_constraints.forbidden_tools. Use for basic checks only; put new or advanced rules under eval. Do not provide both.'),
+            tool_sequence: z
+                .array(z.string())
+                .min(1)
+                .optional()
+                .describe('Simple shorthand for eval.tool_sequence: a flat ordered list of raw MCP tool names. Use eval for new or advanced rules. Do not provide both.'),
+            extract_rules: z
+                .array(z.object({
+                name: z.string().describe('Extracted field name.'),
+                regex: z.string().describe('Regex applied to final_text.')
+            }))
+                .optional()
+                .describe('Extract values from the final response for trend analysis.'),
+            response_regex_patterns: z
+                .array(z.string())
+                .optional()
+                .describe('Simple shorthand for response regex checks. Use eval.response_assertions for new or advanced checks; choose type=contains for literal text. Do not provide both.'),
+            eval: EvalRulesSchema.optional().describe('Full evaluation-rule object. Use this for new authoring or whenever you need more than one rule type; do not combine it with equivalent shorthand fields.')
+        }
+    }, async ({ id, name, servers, prompt, required_tools, forbidden_tools, tool_sequence, extract_rules, response_regex_patterns, eval: evalRules }) => withToolHandling(async () => {
+        const bundleRoot = resolveBundleRoot();
+        const library = readLibrary(bundleRoot, false);
+        const created = createTestCaseFile({
+            librariesDir: bundleRoot,
+            knownServerIds: library.servers.map((item) => item.id),
+            testCase: {
+                id,
+                name,
+                servers,
+                prompt,
+                requiredTools: required_tools,
+                forbiddenTools: forbidden_tools,
+                toolSequence: tool_sequence,
+                extractRules: extract_rules,
+                responseRegexPatterns: response_regex_patterns,
+                eval: evalRules
+            }
+        });
+        return ok(`Created Test Case '${created.id}'`, {
+            id: created.id,
+            path: created.path,
+            test_case: created.testCase
+        });
+    }));
+    registerTool('mcplab_get_test_case', {
+        description: 'Read one existing MCPLab test case, including its full parsed YAML content.',
+        inputSchema: { file_path: z.string().describe('Path relative to mcplab/test-cases/.') },
+        outputSchema: { path: z.string(), test_case: ScenarioEntrySchema }
+    }, async ({ file_path }) => withToolHandling(() => {
+        const result = readTestCaseFile({ librariesDir: resolveBundleRoot(), filePath: file_path });
+        return ok(`Read Test Case '${result.id}'`, {
+            path: result.path,
+            test_case: result.testCase
+        });
+    }));
+    registerTool('mcplab_update_test_case', {
+        description: 'Replace one existing MCPLab test case in place. Put evaluation checks in the eval object for full support of response, tool-input, and agent assertions. Simple shorthand fields are supported for basic cases; do not provide a shorthand field together with its equivalent eval field. Requires confirm=true and never creates a new file.',
+        inputSchema: {
+            file_path: z
+                .string()
+                .describe('Path to an existing test case relative to mcplab/test-cases/.'),
+            confirm: z
+                .literal(true)
+                .describe('Required confirmation that the existing file should be overwritten.'),
+            id: z
+                .string()
+                .min(1)
+                .describe('Stable Test Case id; letters, numbers, hyphens, and underscores.'),
+            name: z
+                .string()
+                .optional()
+                .describe('Human-readable Test Case label; defaults to id when blank.'),
+            servers: z.array(z.string()).min(1).describe('MCP server ids available to this Test Case.'),
+            prompt: z
+                .string()
+                .min(1)
+                .max(4000)
+                .describe('Task prompt executed by the evaluation agent.'),
+            required_tools: z
+                .array(z.string())
+                .optional()
+                .describe('Simple shorthand for eval.tool_constraints.required_tools. Use for basic checks only; put new or advanced rules under eval. Do not provide both.'),
+            forbidden_tools: z
+                .array(z.string())
+                .optional()
+                .describe('Simple shorthand for eval.tool_constraints.forbidden_tools. Use for basic checks only; put new or advanced rules under eval. Do not provide both.'),
+            tool_sequence: z
+                .array(z.string())
+                .min(1)
+                .optional()
+                .describe('Simple shorthand for eval.tool_sequence: a flat ordered list of raw MCP tool names. Use eval for new or advanced rules. Do not provide both.'),
+            extract_rules: z
+                .array(z.object({
+                name: z.string().describe('Extracted field name.'),
+                regex: z.string().describe('Regex applied to final_text.')
+            }))
+                .optional()
+                .describe('Extract values from the final response for trend analysis.'),
+            response_regex_patterns: z
+                .array(z.string())
+                .optional()
+                .describe('Simple shorthand for response regex checks. Use eval.response_assertions for new or advanced checks; choose type=contains for literal text. Do not provide both.'),
+            eval: EvalRulesSchema.optional().describe('Full evaluation-rule object. Use this for new authoring or whenever you need more than one rule type; do not combine it with equivalent shorthand fields.')
+        },
+        outputSchema: { id: z.string(), path: z.string(), test_case: ScenarioEntrySchema }
+    }, async ({ file_path, confirm: _confirm, id, name, servers, prompt, required_tools, forbidden_tools, tool_sequence, extract_rules, response_regex_patterns, eval: evalRules }) => withToolHandling(() => {
+        const library = readLibrary(resolveBundleRoot(), false);
+        const updated = updateTestCaseFile({
+            librariesDir: resolveBundleRoot(),
+            knownServerIds: library.servers.map((item) => item.id),
+            filePath: file_path,
+            testCase: {
+                id,
+                name,
+                servers,
+                prompt,
+                requiredTools: required_tools,
+                forbiddenTools: forbidden_tools,
+                toolSequence: tool_sequence,
+                extractRules: extract_rules,
+                responseRegexPatterns: response_regex_patterns,
+                eval: evalRules
+            }
+        });
+        return ok(`Updated Test Case '${updated.id}'`, {
+            id: updated.id,
+            path: updated.path,
+            test_case: updated.testCase
+        });
+    }));
+    registerTool('mcplab_create_evaluation_config', {
+        description: 'Create one reviewed MCPLab evaluation configuration in mcplab/evals/. This is a scoped YAML write with automatic filename normalization and collision handling. In scenario eval blocks, use the full evaluation rules for response, tool-input, and agent checks: tool_constraints, flat tool_sequence, response_assertions (contains, equals, regex, and JSONPath variants), tool_input_assertions, agent_assertions, and agent_context. Use contains for literal text; use regex only for patterns. Do not combine eval rules with equivalent shorthand fields.',
+        outputSchema: {
+            file_name: z.string(),
+            path: z.string(),
+            relative_path: z.string(),
+            config: GenericObjectSchema
+        },
+        inputSchema: {
+            file_name: z
+                .string()
+                .min(1)
+                .describe('Destination YAML filename under mcplab/evals/. Extension is added if omitted.'),
+            config: GenericObjectSchema.describe('Complete MCPLab evaluation configuration. Scenario eval blocks support tool_constraints, a flat tool_sequence, response_assertions (contains, not_contains, starts_with, ends_with, equals, regex, and JSONPath variants), tool_input_assertions, agent_assertions, and agent_context. Use contains for literal response text; use regex only when pattern matching is required.')
+        }
+    }, async ({ file_name, config }) => withToolHandling(async () => {
+        const created = createEvaluationConfigFile({
+            evalsDir: join(resolveBundleRoot(), 'evals'),
+            fileName: file_name,
+            config: config
+        });
+        return ok(`Created evaluation configuration '${created.fileName}'`, {
+            file_name: created.fileName,
+            path: created.path,
+            relative_path: created.relativePath,
+            config: created.config
+        });
+    }));
+    registerTool('mcplab_get_evaluation_config', {
+        description: 'Read one existing MCPLab evaluation config, including its full parsed YAML content. Returned scenario eval blocks may use full tool constraints, flat tool sequences, response/tool-input assertions, agent assertions, and agent context.',
+        inputSchema: {
+            config_path: z
+                .string()
+                .describe('Path to an existing eval YAML, relative to the MCPLab evals directory. Invalid, missing, or out-of-root paths return an error; the tool never silently coerces them.')
+        },
+        outputSchema: z.object({
+            path: z.string(),
+            relative_path: z.string(),
+            config: GenericObjectSchema
+        })
+    }, async ({ config_path }) => withToolHandling(() => {
+        const result = readEvaluationConfigFile({
+            evalsDir: join(resolveBundleRoot(), 'evals'),
+            filePath: config_path
+        });
+        return ok(`Read evaluation config '${result.relativePath}'`, {
+            path: result.path,
+            relative_path: result.relativePath,
+            config: result.config
+        });
+    }));
+    registerTool('mcplab_update_evaluation_config', {
+        description: 'Replace one existing MCPLab evaluation config in place after normalization. Requires confirm=true; refuses missing or out-of-root paths and never creates a new file. In scenario eval blocks, use the full evaluation rules for response, tool-input, and agent checks: tool_constraints, flat tool_sequence, response_assertions (contains, equals, regex, and JSONPath variants), tool_input_assertions, agent_assertions, and agent_context. Use contains for literal text; use regex only for patterns. Do not combine eval rules with equivalent shorthand fields.',
+        inputSchema: {
+            config_path: z
+                .string()
+                .describe('Path to an existing eval YAML, relative to the MCPLab evals directory. Invalid, missing, or out-of-root paths return an error; the tool never silently coerces them.'),
+            config: GenericObjectSchema.describe('Complete replacement evaluation configuration. Scenario eval blocks support tool_constraints, a flat tool_sequence, response_assertions (contains, not_contains, starts_with, ends_with, equals, regex, and JSONPath variants), tool_input_assertions, agent_assertions, and agent_context. Use contains for literal response text; use regex only when pattern matching is required.'),
+            confirm: z
+                .literal(true)
+                .describe('Required confirmation that the existing file should be overwritten.')
+        },
+        outputSchema: {
+            file_name: z.string(),
+            path: z.string(),
+            relative_path: z.string(),
+            config: GenericObjectSchema,
+            warnings: z.array(z.string())
+        }
+    }, async ({ config_path, config }) => withToolHandling(() => {
+        const updated = updateEvaluationConfigFile({
+            evalsDir: join(resolveBundleRoot(), 'evals'),
+            filePath: config_path,
+            config: config
+        });
+        return ok(`Updated evaluation config '${updated.relativePath}'`, {
+            file_name: updated.fileName,
+            path: updated.path,
+            relative_path: updated.relativePath,
+            config: updated.config,
+            warnings: updated.warnings
+        });
+    }));
     registerTool('mcplab_generate_scenario_entry', {
-        description: 'Generate a MCPLab scenario YAML snippet with prompt, server links, and optional evaluation/extract rules. Optimized for scenario authoring workflows.',
+        description: 'Generate a MCPLab scenario YAML snippet with prompt, server links, and optional evaluation/extract rules. Put checks in the eval object for full assertion support; simple shorthand fields are supported for basic cases and must not be combined with equivalent eval fields.',
         outputSchema: {
             scenario: ScenarioEntrySchema,
             yaml: z.string(),
@@ -928,7 +1388,7 @@ export function registerTools(server) {
             name: z
                 .string()
                 .optional()
-                .describe('Optional human label used only to derive id when id is omitted.'),
+                .describe('Optional human-readable scenario label. Also used to derive id when omitted.'),
             agent: z
                 .string()
                 .optional()
@@ -936,22 +1396,30 @@ export function registerTools(server) {
             servers: z
                 .array(z.string())
                 .min(1)
-                .describe('One or more server ids available to the scenario.'),
+                .describe('One or more server ids; emitted as scenario-owned mcp_servers refs.'),
             prompt: z
                 .string()
                 .min(1)
                 .max(4000)
                 .describe('The task prompt the evaluation agent should execute (1-4000 chars).'),
-            required_tools: z.array(z.string()).optional().describe('Tools that must be called.'),
-            forbidden_tools: z.array(z.string()).optional().describe('Tools that must not be called.'),
-            allowed_tool_sequences: z
-                .array(z.array(z.string()).min(1))
+            required_tools: z
+                .array(z.string())
+                .optional()
+                .describe('Simple shorthand for eval.tool_constraints.required_tools. Use for basic checks only; put new or advanced rules under eval. Do not provide both.'),
+            forbidden_tools: z
+                .array(z.string())
                 .optional()
-                .describe('Allowed tool call sequences (exact order groups).'),
+                .describe('Simple shorthand for eval.tool_constraints.forbidden_tools. Use for basic checks only; put new or advanced rules under eval. Do not provide both.'),
+            tool_sequence: z
+                .array(z.string())
+                .min(1)
+                .optional()
+                .describe('Simple shorthand for eval.tool_sequence: a flat ordered list of raw MCP tool names. Use eval for new or advanced rules. Do not provide both.'),
             response_regex_patterns: z
                 .array(z.string())
                 .optional()
-                .describe('Regex patterns that must match the final response text.'),
+                .describe('Simple shorthand for response regex checks. Use eval.response_assertions for new or advanced checks; choose type=contains for literal text. Do not provide both.'),
+            eval: EvalRulesSchema.optional().describe('Full evaluation-rule object. Use this for new authoring or whenever you need more than one rule type; do not combine it with equivalent shorthand fields.'),
             extract_rules: z
                 .array(z.object({
                 name: z.string().describe('Extracted field name.'),
@@ -981,6 +1449,16 @@ export function registerTools(server) {
             });
         });
     });
+    registerTool('mcplab_list_evaluation_configs', {
+        description: 'List MCPLab evaluation configurations with nested-suite filtering and deterministic sorting.',
+        inputSchema: {
+            suite: z.string().optional(),
+            query: z.string().optional(),
+            sort_by: z.enum(['name', 'scenarios', 'agents', 'updated_at']).optional(),
+            sort_direction: z.enum(['asc', 'desc']).optional()
+        },
+        outputSchema: EvaluationConfigListSchema
+    }, async ({ suite, query, sort_by, sort_direction }) => withToolHandling(() => ok('Listed evaluation configurations', listEvaluationConfigs({ suite, query, sortBy: sort_by, sortDirection: sort_direction }))));
     registerTool('mcplab_validate_config', {
         description: 'Validate and expand a MCPLab config file via mcplab-core loadConfig(), including server/agent/scenario library references.',
         outputSchema: {
@@ -1021,7 +1499,7 @@ export function registerTools(server) {
         });
     });
     registerTool('mcplab_run_eval', {
-        description: 'Run a MCPLab evaluation using mcplab-core runAll() from a config file and return the run directory plus summary metrics.',
+        description: 'Execute a MCPLab evaluation immediately in this MCP server process using mcplab-core runAll(). Use this when the caller needs results now; use mcplab_queue_run to hand the run to the APP queue instead.',
         outputSchema: {
             run_dir: z.string(),
             total_scenarios: z.number().int().nonnegative(),
@@ -1117,6 +1595,133 @@ export function registerTools(server) {
             });
         });
     });
+    registerTool('mcplab_list_runs', {
+        description: 'List MCPLab evaluation runs with optional inclusive ISO time bounds.',
+        inputSchema: {
+            since: z.string().datetime().optional(),
+            until: z.string().datetime().optional(),
+            limit: z.number().int().positive().max(200).optional()
+        },
+        outputSchema: {
+            since: z.string().optional(),
+            until: z.string().optional(),
+            total_matching: z.number(),
+            runs: z.array(z.object({
+                run_id: z.string(),
+                timestamp: z.string().optional(),
+                config_hash: z.string().optional(),
+                summary: z.any().optional()
+            }))
+        }
+    }, async ({ since, until, limit }) => withToolHandling(() => {
+        const sinceMs = since ? new Date(since).getTime() : undefined;
+        const untilMs = until ? new Date(until).getTime() : undefined;
+        if (sinceMs !== undefined && untilMs !== undefined && sinceMs > untilMs)
+            throw new Error('since must be earlier than or equal to until');
+        const matching = listRunsWithFallback(resolveRunsDir(), undefined, true).filter((entry) => {
+            const timestamp = String(entry.metadata && typeof entry.metadata === 'object'
+                ? entry.metadata.timestamp ?? ''
+                : '');
+            const time = new Date(timestamp).getTime();
+            if (sinceMs === undefined && untilMs === undefined)
+                return true;
+            return (!Number.isNaN(time) &&
+                (sinceMs === undefined || time >= sinceMs) &&
+                (untilMs === undefined || time <= untilMs));
+        });
+        const runs = matching.slice(0, limit ?? 50).map((entry) => ({
+            run_id: String(entry.run_id),
+            timestamp: typeof entry.metadata?.timestamp === 'string'
+                ? entry.metadata.timestamp
+                : undefined,
+            config_hash: typeof entry.metadata?.config_hash === 'string'
+                ? entry.metadata.config_hash
+                : undefined,
+            summary: entry.summary
+        }));
+        return ok(`Found ${matching.length} run(s)`, {
+            since,
+            until,
+            total_matching: matching.length,
+            runs
+        });
+    }));
+    registerTool('mcplab_queue_run', {
+        description: 'Add a MCPLab evaluation to the local APP queue for the APP worker to execute later; this tool does not execute the evaluation or return results. Use mcplab_run_eval when results are needed immediately. Returns the queued job id, position, and APP /run URL.',
+        inputSchema: {
+            config_path: z
+                .string()
+                .describe('Path to the eval YAML. Relative paths resolve under the APP evals directory; absolute paths must remain inside that directory. Invalid, missing, or disallowed paths return an error—the tool never silently coerces them.'),
+            scenario_id: z.string().optional().describe('Optional single scenario id to run.'),
+            scenario_ids: z
+                .array(z.string())
+                .min(1)
+                .optional()
+                .describe('Optional scenario ids to run.'),
+            agents: z.array(z.string()).min(1).optional().describe('Optional agent ids to use.'),
+            runs_per_scenario: z
+                .number()
+                .int()
+                .positive()
+                .optional()
+                .describe('Runs per scenario (default 1).'),
+            run_note: z
+                .string()
+                .max(500)
+                .optional()
+                .describe('Optional note attached to the queued run.'),
+            server_override_all: z
+                .array(z.string())
+                .min(1)
+                .optional()
+                .describe('Replace MCP servers for every selected scenario for this queued run.'),
+            scenario_server_overrides: z
+                .record(z.array(z.string()))
+                .optional()
+                .describe('Map scenario ids to replacement MCP server ids for this queued run.')
+        },
+        outputSchema: z.object({
+            jobId: z.string().describe('APP queue job identifier.'),
+            queued: z.boolean().optional().describe('Whether the APP accepted the job into its queue.'),
+            position: z
+                .number()
+                .int()
+                .nonnegative()
+                .optional()
+                .describe('Current position in the APP queue.'),
+            queue_url: z.string().url().describe('URL of the APP run queue page.')
+        })
+    }, async (input) => withToolHandling(async () => {
+        const appBaseUrl = process.env.MCPLAB_APP_URL ?? 'http://127.0.0.1:8787';
+        const response = await fetch(new URL('/api/runs', appBaseUrl), {
+            method: 'POST',
+            headers: { 'content-type': 'application/json', accept: 'application/json' },
+            body: JSON.stringify({
+                configPath: input.config_path,
+                runsPerScenario: input.runs_per_scenario ?? 1,
+                ...(input.scenario_id ? { scenarioId: input.scenario_id } : {}),
+                ...(input.scenario_ids ? { scenarioIds: input.scenario_ids } : {}),
+                ...(input.agents ? { agents: input.agents } : {}),
+                ...(input.run_note ? { runNote: input.run_note } : {}),
+                ...(input.server_override_all ? { serverOverrideAll: input.server_override_all } : {}),
+                ...(input.scenario_server_overrides
+                    ? { scenarioServerOverrides: input.scenario_server_overrides }
+                    : {})
+            })
+        });
+        const payload = (await response.json().catch(() => ({})));
+        if (!response.ok) {
+            throw new Error(typeof payload.error === 'string'
+                ? payload.error
+                : `APP queue request failed with HTTP ${response.status}`);
+        }
+        if (typeof payload.jobId !== 'string')
+            throw new Error('APP queue response did not include a jobId');
+        return ok(`Queued MCPLab run ${payload.jobId}`, {
+            ...payload,
+            queue_url: new URL('/run', appBaseUrl).toString()
+        });
+    }));
     registerTool('mcplab_aggregate_runs', {
         description: 'Aggregate metrics across historical MCPLab runs with compact summary-first output.',
         outputSchema: z.object({
@@ -1262,7 +1867,7 @@ export function registerTools(server) {
         });
     });
     registerTool('mcplab_search_tool_analysis_results', {
-        description: 'Search saved MCP tool analysis reports from mcplab/results/tool-analysis. Use the optional query parameter to filter by report_id, server name, agent name, or summary metadata (case-insensitive substring).',
+        description: 'Search saved MCP TOOL-ANALYSIS records under mcplab/results/tool-analysis. Use this for aggregate tool usage, latency, and tool-analysis summaries; do not use it for final answers, raw execution traces, or Markdown report files.',
         outputSchema: {
             tool_analysis_results_dir: z.string(),
             total: z.number().int().nonnegative(),
@@ -1525,7 +2130,7 @@ export function registerTools(server) {
         });
     });
     registerTool('mcplab_trace_get_conversation', {
-        description: 'Return a structured conversation timeline (messages + tool blocks) for a specific scenario+agent in a scenario_run trace.',
+        description: 'Return one structured conversation timeline (messages and tool blocks) for a specific scenario and agent in a run. If you do not know the scenario_id and agent, call mcplab_trace_list_conversations first.',
         outputSchema: {
             run_id: z.string(),
             scenario_id: z.string(),
@@ -1557,6 +2162,10 @@ export function registerTools(server) {
             const { runId, records, legacyDetected } = readScenarioRunTraceRecordsForRun(run_id);
             const textMax = max_text_chars ?? 4000;
             const record = records.find((r) => r.scenario_id === scenario_id && r.agent === agent);
+            if (!record) {
+                const available = Array.from(new Set(records.map((r) => `${r.scenario_id} / ${r.agent}`)));
+                throw new Error(`No conversation found for ${scenario_id} / ${agent}. Available conversations: ${available.join(', ') || 'none'}`);
+            }
             const timeline = record
                 ? buildConversationTimeline(record, textMax).slice(0, max_items ?? 300)
                 : [];
@@ -1569,8 +2178,35 @@ export function registerTools(server) {
             });
         });
     });
+    registerTool('mcplab_trace_list_conversations', {
+        description: 'List the scenario and agent conversation targets available in one run. Use this before mcplab_trace_get_conversation when you need to discover the required scenario_id and agent values.',
+        outputSchema: {
+            run_id: z.string(),
+            legacy_trace_detected: z.boolean().optional(),
+            conversations: z.array(z.object({
+                scenario_id: z.string(),
+                agent: z.string(),
+                timeline_item_count: z.number().int().nonnegative()
+            }))
+        },
+        inputSchema: {
+            run_id: z.string().describe("Run id directory name or 'LATEST'.")
+        }
+    }, async ({ run_id }) => withToolHandling(() => {
+        const { runId, records, legacyDetected } = readScenarioRunTraceRecordsForRun(run_id);
+        const conversations = records.map((record) => ({
+            scenario_id: record.scenario_id,
+            agent: record.agent,
+            timeline_item_count: buildConversationTimeline(record, 4000).length
+        }));
+        return ok(`Found ${conversations.length} conversation(s) in run ${runId}`, {
+            run_id: runId,
+            legacy_trace_detected: legacyDetected || undefined,
+            conversations
+        });
+    }));
     registerTool('mcplab_trace_search', {
-        description: 'Search scenario_run trace content for a text query and return matching message/block items.',
+        description: 'Search RAW SCENARIO-RUN TRACE events and message/block content for a text query. Use this for execution-level evidence such as tool calls, arguments, ordering, or intermediate messages; use mcplab_results_search for indexed result/answer content instead.',
         outputSchema: {
             run_id: z.string(),
             query: z.string().trim().min(1),
@@ -1821,7 +2457,7 @@ export function registerTools(server) {
         });
     });
     registerTool('mcplab_results_search', {
-        description: 'Search MCPLab run results in compact LLM-first format. Auto-refreshes index when artifacts changed.',
+        description: 'Search INDEXED MCPLab evaluation RESULTS and final-answer content in compact LLM-first format. Use this for finding passed/failed scenarios, answer text, and result metadata; use mcplab_trace_search for raw execution events, mcplab_search_tool_analysis_results for aggregate tool-analysis records, and mcplab_search_markdown_reports for report files.',
         outputSchema: {
             query: z.string(),
             runs_dir: z.string(),
@@ -1872,7 +2508,7 @@ export function registerTools(server) {
         });
     });
     registerTool('mcplab_results_context', {
-        description: 'Fetch focused context for a scenario/run from results, trace, or summary. Returns bounded excerpts only. around is trace-line context and is only valid with source=trace (or when source is omitted).',
+        description: 'Fetch bounded evidence for one run/scenario after a search hit. Use this to inspect a focused result, summary, or trace excerpt; it is not a broad search tool. For broad final-answer/result searches use mcplab_results_search, and for raw trace discovery use mcplab_trace_search.',
         outputSchema: {
             run_id: z.string(),
             scenario_id: z.string(),
@@ -1922,6 +2558,30 @@ export function registerTools(server) {
             });
         });
     });
+    registerTool('mcplab_list_skills', {
+        description: 'List MCPLab skills bundled with this MCP server.',
+        inputSchema: {},
+        outputSchema: {
+            skills: z.array(z.object({ name: z.string(), description: z.string(), files: z.array(z.string()) }))
+        }
+    }, async () => withToolHandling(() => {
+        const skills = loadSkills();
+        return ok(`Found ${skills.length} MCPLab skill(s)`, {
+            skills: skills.map(({ name, description, files }) => ({ name, description, files }))
+        });
+    }));
+    registerTool('mcplab_get_skill', {
+        description: 'Read one file from a bundled MCPLab skill. Defaults to SKILL.md.',
+        inputSchema: { name: z.string(), file: z.string().optional() },
+        outputSchema: { name: z.string(), file: z.string(), content: z.string() }
+    }, async ({ name, file }) => withToolHandling(() => {
+        const targetFile = file ?? 'SKILL.md';
+        return ok(`Read ${name}/${targetFile}`, {
+            name,
+            file: targetFile,
+            content: readSkillFile(name, targetFile)
+        });
+    }));
 }
 export function registerPrompts(server) {
     server.registerPrompt('mcplab-scenario-author', {
@@ -1988,17 +2648,33 @@ export function registerPrompts(server) {
 const DESTRUCTIVE_TOOLS = new Set([
     'mcplab_delete_tool_analysis_result',
     'mcplab_write_markdown_report',
-    'mcplab_run_eval'
+    'mcplab_run_eval',
+    'mcplab_queue_run',
+    'mcplab_create_test_case',
+    'mcplab_update_test_case',
+    'mcplab_create_evaluation_config',
+    'mcplab_update_evaluation_config'
+]);
+const MUTATING_TOOLS = new Set([
+    'mcplab_write_markdown_report',
+    'mcplab_run_eval',
+    'mcplab_queue_run',
+    'mcplab_create_test_case',
+    'mcplab_update_test_case',
+    'mcplab_create_evaluation_config',
+    'mcplab_update_evaluation_config'
 ]);
-const MUTATING_TOOLS = new Set(['mcplab_write_markdown_report', 'mcplab_run_eval']);
-const OPEN_WORLD_TOOLS = new Set(['mcplab_run_eval']);
+const OPEN_WORLD_TOOLS = new Set(['mcplab_run_eval', 'mcplab_queue_run']);
 const PREFERRED_TOOL_TITLES = {
     mcplab_write_markdown_report: 'Write Markdown Report to Disk',
     mcplab_search_markdown_reports: 'Search Markdown Reports',
     mcplab_list_library: 'Search Library Entries',
     mcplab_generate_agent_entry: 'Generate MCPLab agents.yaml Entry',
     mcplab_search_tool_analysis_results: 'Search Tool Analysis Results',
-    mcplab_trace_search: 'Search Trace Events'
+    mcplab_trace_search: 'Search Raw Trace Events',
+    mcplab_results_search: 'Search Indexed Evaluation Results',
+    mcplab_results_context: 'Fetch Focused Result or Trace Context',
+    mcplab_queue_run: 'Queue MCPLab Evaluation in APP'
 };
 function normalizeOptionalNonEmpty(value) {
     if (typeof value !== 'string')
@@ -2089,7 +2765,7 @@ function mergeLibraryServersAndAgents(config, bundleRoot) {
 function readLibrary(bundleRoot, includeContent) {
     const serversPath = join(bundleRoot, 'servers.yaml');
     const agentsPath = join(bundleRoot, 'agents.yaml');
-    const scenariosDir = join(bundleRoot, 'scenarios');
+    const scenariosDir = resolveScenarioLibraryDir(bundleRoot).path;
     const servers = existsSync(serversPath)
         ? parseYaml(readFileSync(serversPath, 'utf8')) ?? {}
         : {};
@@ -2126,6 +2802,7 @@ function readLibrary(bundleRoot, includeContent) {
             id,
             ...(includeContent ? { entry: agents[id] } : {})
         })),
+        test_cases: scenarioEntries,
         scenarios: scenarioEntries
     };
     return out;
@@ -2150,7 +2827,7 @@ function getLibraryItem(bundleRoot, kind, id) {
             content: entry
         };
     }
-    const dir = join(bundleRoot, 'scenarios');
+    const dir = resolveScenarioLibraryDir(bundleRoot).path;
     if (!existsSync(dir)) {
         throw new Error(`Scenario library directory not found: ${dir}`);
     }
@@ -2172,6 +2849,93 @@ function getLibraryItem(bundleRoot, kind, id) {
     }
     throw new Error(`Scenario '${id}' not found in ${dir}`);
 }
+function listEvaluationConfigs(params) {
+    const evalsDir = join(resolveBundleRoot(), 'evals');
+    const files = [];
+    const walk = (dir) => {
+        if (!existsSync(dir))
+            return;
+        for (const entry of readdirSync(dir, { withFileTypes: true })) {
+            const path = join(dir, entry.name);
+            if (entry.isDirectory())
+                walk(path);
+            else if (entry.isFile() && /\.ya?ml$/i.test(entry.name))
+                files.push(path);
+        }
+    };
+    walk(evalsDir);
+    const all = files.map((path) => {
+        const relativePath = relative(evalsDir, path).split(sep).join('/');
+        const suitePath = dirname(relativePath) === '.' ? '' : dirname(relativePath).split(sep).join('/');
+        const fallbackName = basename(path, extname(path));
+        const stat = statSync(path);
+        try {
+            const parsed = parseYaml(readFileSync(path, 'utf8'));
+            return {
+                id: Buffer.from(relativePath, 'utf8').toString('base64url'),
+                name: typeof parsed?.name === 'string' && parsed.name.trim()
+                    ? parsed.name.trim()
+                    : fallbackName,
+                relative_path: relativePath,
+                config_path: path,
+                suite_path: suitePath,
+                updated_at: stat.mtime.toISOString(),
+                scenario_count: Array.isArray(parsed?.scenarios) ? parsed.scenarios.length : 0,
+                agent_count: Array.isArray(parsed?.agents) ? parsed.agents.length : 0
+            };
+        }
+        catch (error) {
+            return {
+                id: Buffer.from(relativePath, 'utf8').toString('base64url'),
+                name: fallbackName,
+                relative_path: relativePath,
+                config_path: path,
+                suite_path: suitePath,
+                updated_at: stat.mtime.toISOString(),
+                scenario_count: 0,
+                agent_count: 0,
+                error: error instanceof Error ? error.message : String(error)
+            };
+        }
+    });
+    const suite = params.suite?.trim().replace(/^suite:/, '');
+    const query = params.query?.trim().toLowerCase();
+    const matching = all.filter((config) => (!suite || config.suite_path === suite || config.suite_path.startsWith(`${suite}/`)) &&
+        (!query ||
+            `${config.name} ${config.relative_path} ${config.error ?? ''}`
+                .toLowerCase()
+                .includes(query)));
+    const sortBy = params.sortBy ?? 'name';
+    const direction = params.sortDirection ?? 'asc';
+    matching.sort((a, b) => {
+        const left = sortBy === 'scenarios'
+            ? a.scenario_count
+            : sortBy === 'agents'
+                ? a.agent_count
+                : sortBy === 'updated_at'
+                    ? Date.parse(a.updated_at)
+                    : a.name;
+        const right = sortBy === 'scenarios'
+            ? b.scenario_count
+            : sortBy === 'agents'
+                ? b.agent_count
+                : sortBy === 'updated_at'
+                    ? Date.parse(b.updated_at)
+                    : b.name;
+        const comparison = typeof left === 'number' && typeof right === 'number'
+            ? left - right
+            : String(left).localeCompare(String(right));
+        return direction === 'asc' ? comparison : -comparison;
+    });
+    return {
+        evals_dir: evalsDir,
+        total: all.length,
+        matching: matching.length,
+        filters: removeUndefined({ suite, query }),
+        sort: { by: sortBy, direction },
+        configs: matching
+    };
+}
 function buildServerEntry(input) {
     const transport = input.transport ?? 'http';
     const authType = input.auth_type ?? 'none';
@@ -2220,46 +2984,7 @@ function buildServerEntry(input) {
     };
 }
 function buildScenario(input) {
-    const id = input.id?.trim() || slugify(input.name?.trim() || input.prompt.slice(0, 40));
-    if (!id) {
-        throw new Error('Unable to derive scenario id. Provide id or name.');
-    }
-    const scenario = removeUndefined({
-        id,
-        agent: input.agent?.trim() || undefined,
-        servers: input.servers,
-        prompt: input.prompt,
-        eval: buildEvalRules(input),
-        extract: input.extract_rules?.map((rule) => ({
-            name: rule.name,
-            from: 'final_text',
-            regex: rule.regex
-        }))
-    });
-    return scenario;
-}
-function buildEvalRules(input) {
-    const toolConstraints = input.required_tools?.length || input.forbidden_tools?.length
-        ? removeUndefined({
-            required_tools: normalizeStringArray(input.required_tools),
-            forbidden_tools: normalizeStringArray(input.forbidden_tools)
-        })
-        : undefined;
-    const toolSequence = input.allowed_tool_sequences && input.allowed_tool_sequences.length > 0
-        ? { allow: input.allowed_tool_sequences }
-        : undefined;
-    const responseAssertions = input.response_regex_patterns && input.response_regex_patterns.length > 0
-        ? input.response_regex_patterns.map((pattern) => ({ type: 'regex', pattern }))
-        : undefined;
-    const evalRules = removeUndefined({
-        tool_constraints: toolConstraints,
-        tool_sequence: toolSequence,
-        response_assertions: responseAssertions
-    });
-    if (Object.keys(evalRules ?? {}).length === 0) {
-        return undefined;
-    }
-    return evalRules;
+    return buildScenarioEntry(input);
 }
 function validateScenarioHeuristics(scenario) {
     const warnings = [];
@@ -2981,19 +3706,6 @@ function searchableText(value) {
     visit(value);
     return tokens.join(' ');
 }
-function normalizeStringArray(values) {
-    if (!values)
-        return undefined;
-    const out = values.map((value) => value.trim()).filter(Boolean);
-    return out.length > 0 ? out : undefined;
-}
-function slugify(input) {
-    return input
-        .toLowerCase()
-        .replace(/[^a-z0-9]+/g, '-')
-        .replace(/^-+|-+$/g, '')
-        .slice(0, 80);
-}
 function indentBlock(text, spaces) {
     const prefix = ' '.repeat(spaces);
     return text
@@ -3033,7 +3745,14 @@ async function handleHttpRequest(req, res, sessions, mcpPath) {
             await runtime.transport.handleRequest(req, res, body);
             return;
         }
-        if (!sessionId && isInitializeRequest(body)) {
+        // Per the MCP Streamable HTTP transport, an unknown session indicates that
+        // the server was restarted or the session expired. The client must receive
+        // 404 so it can discard the stale session and initialize a new connection.
+        if (sessionId) {
+            sendPlain(res, 404, 'Not Found');
+            return;
+        }
+        if (isInitializeRequest(body)) {
             let runtime;
             const transport = new StreamableHTTPServerTransport({
                 sessionIdGenerator: () => randomUUID(),
@@ -3067,7 +3786,11 @@ async function handleHttpRequest(req, res, sessions, mcpPath) {
     }
     if (method === 'GET' || method === 'DELETE') {
         const sessionId = getSessionId(req);
-        if (!sessionId || !sessions.has(sessionId)) {
+        if (sessionId && !sessions.has(sessionId)) {
+            sendPlain(res, 404, 'Not Found');
+            return;
+        }
+        if (!sessionId) {
             sendPlain(res, 400, 'Invalid or missing session ID');
             return;
         }