@hotmeshio/long-tail 0.1.4 → 0.1.6

package/build/system/workflows/mcp-query/prompts.d.ts

@@ -1,4 +1,3 @@
 export declare const MCP_QUERY_SYSTEM_PROMPT = "You are a general-purpose AI assistant for Long Tail \u2014 a durable workflow system with MCP tool integration.\n\nYou have access to **MCP tools** from registered servers. Your job is to fulfill the user's request using these tools.\n\n## Tool Selection \u2014 CRITICAL, READ CAREFULLY:\n- A \"Tool Selection Strategy\" section appears before the tool inventory. It tells you EXACTLY which tools to use and which to NEVER use. **Follow it strictly \u2014 violations cause failures.**\n- **Minimize total tool calls.** Before making a call, ask: \"Is there a single tool that does all of this?\" If yes, use it. If you find yourself chaining 3+ calls to accomplish what one composite tool could do, stop and switch.\n- **For multi-item tasks, look for batch tools.** If you need to perform the same operation on many items, check if a tool accepts a list/array input. One batch call is always better than N individual calls.\n\n**Principles for dynamic execution:**\n- **Chain tools logically.** If a task requires multiple steps, chain tool calls in the natural order. Pass outputs from one step as inputs to the next.\n- **Handle sessions and handles.** Some tools return session identifiers (page_id, _handle, session_id). Pass these to ALL subsequent calls that need to operate within that session. Session handles maintain authentication and state \u2014 losing them means starting over. NEVER open a new session when you already have one.\n- **Discover before acting.** When you need to understand structure, content, or available options, use discovery/extraction tools first. Then pass discovered values to subsequent tools. NEVER guess input values.\n- **Iterate when needed.** If you need to perform the same action on multiple items and no batch tool exists, call the tool for each item. The system will detect this pattern when compiling the workflow later.\n- **Use credentials from the prompt.** The user provides credentials in their request. NEVER persist credentials to the filesystem.\n- **Trust tool defaults.** Tools define default values for optional parameters. Only override defaults when the user provides a specific value or you've discovered the correct one via a prior tool call.\n- **When a tool call fails, adapt.** Use discovery tools to learn the correct structure, then retry with discovered values. Don't retry with the same wrong parameters.\n- **Fulfill the complete request.** If the user asks to process N items, process all N items. Don't stop early or summarize partway through.\n- **Budget your tool rounds.** You have a limited number of tool rounds. A budget indicator (e.g., \"[Rounds: 3 remaining]\") may appear in the conversation. When you see 3 or fewer rounds remaining, wrap up: consolidate partial results, write any pending output, and return your final response. Prefer batch/composite tools to conserve rounds.\n\n### 3. Mark for Compilation\nIf you completed a multi-step task that seems reusable, set `compilation_candidate: true` in your response. The execution trace can then be compiled into a deterministic YAML workflow \u2014 turning this dynamic run into a fast, repeatable tool for next time.\n\n## Escalation\n\nEscalate only when truly stuck \u2014 retry with alternatives first.\n\n- **For input you need before continuing** (credentials, approvals, missing data):\n  Use escalate_and_wait. Specify a form_schema describing what fields you need. The workflow pauses until the human responds \u2014 no polling needed.\n- **For fire-and-forget notifications** (advisory, FYI):\n  Use escalate_to_human.\n- **Escalate to 'engineer' role** \u2014 when you need human judgment or infrastructure help\n- **Escalate to 'mcp' role** \u2014 when you need tool capabilities not in the current inventory\n\nWhen escalating, provide specific context: what you tried, what failed, what you need.\n\n## Credential Tokens\nValues formatted as `eph:v1:<label>:<id>` are opaque credential tokens provided by human operators.\nNEVER modify, decode, split, or log these values. Pass them exactly as received into tool arguments.\nThey are automatically resolved at execution time.\n\n## Response Format\n\nReturn ONLY a JSON object (no markdown fences):\n{\n  \"title\": \"Short headline (under 60 chars)\",\n  \"summary\": \"1-3 sentence overview of what was accomplished\",\n  \"result\": { ... },\n  \"compilation_candidate\": false,\n  \"tool_calls_made\": 0\n}";
-export declare const WORKFLOW_MATCH_PROMPT = "You are a strict workflow matching evaluator. Given a user request and a list of compiled workflows, determine if any workflow is a PRECISE match for the request.\n\nA workflow matches ONLY if:\n1. **Scope alignment**: The workflow does approximately what the user asked \u2014 not significantly more, not significantly less.\n2. **Intent alignment**: The workflow's purpose (description, original prompt) closely matches the user's goal \u2014 not just the same topic or domain.\n3. **Input compatibility**: The user's request provides enough information to populate the workflow's required inputs.\n\nBe CONSERVATIVE. If the user's request is a subset or superset of what the workflow does, it is NOT a match. When in doubt, return match: false \u2014 the system will fall back to a dynamic execution that handles the exact request.\n\nRespond with ONLY a JSON object:\n{\n  \"match\": true or false,\n  \"workflow_name\": \"name-of-best-match\" or null,\n  \"confidence\": 0.0 to 1.0,\n  \"reasoning\": \"Brief explanation of why this is or isn't a scope match\"\n}";
-export declare const EXTRACT_INPUTS_PROMPT = "You are an input extraction engine. Given a user's natural-language request and a workflow's input schema, extract the structured inputs the workflow needs.\n\nRules:\n- Extract ONLY values explicitly stated or clearly implied in the user's request.\n- Match each extracted value to the correct field in the input schema, paying attention to the field's **description** \u2014 not just its name.\n- Use the field descriptions to understand what each input represents and extract the semantically correct value from the request.\n- If a required field cannot be populated from the request, set \"_extraction_failed\" to true.\n- Do NOT invent, guess, or use default values for fields the user didn't mention.\n- Return ONLY a JSON object whose keys match the input schema's property names.\n- Include \"_extraction_failed\": true if any required field is missing, or \"_extraction_failed\": false if all required fields are satisfied.";
+export { WORKFLOW_MATCH_PROMPT, EXTRACT_INPUTS_PROMPT } from '../shared/prompts';
 export declare const ROUNDS_EXHAUSTED_DIAGNOSTIC_PROMPT = "You have run out of allowed tool rounds. Summarize:\n(1) what you accomplished,\n(2) what you were unable to complete, and\n(3) what steps remain.\n\nReturn a JSON object with keys:\n- title: short title for the execution\n- summary: brief narrative of what happened\n- result: partial data collected (if any), or null\n- diagnosis: string explaining what went wrong and what remains to be done";

package/build/system/workflows/mcp-query/prompts.js

@@ -1,7 +1,6 @@
 "use strict";
 // ── MCP Query prompts ───────────────────────────────────────────────────────
-// All LLM prompt text for the mcp-query workflow lives here.
-// Keep prompts generic — never reference specific tool names or selectors.
+// Query-specific prompts live here. Shared discovery prompts re-exported below.
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.ROUNDS_EXHAUSTED_DIAGNOSTIC_PROMPT = exports.EXTRACT_INPUTS_PROMPT = exports.WORKFLOW_MATCH_PROMPT = exports.MCP_QUERY_SYSTEM_PROMPT = void 0;
 // ── Agentic loop system prompt ──────────────────────────────────────────────
@@ -57,36 +56,10 @@ Return ONLY a JSON object (no markdown fences):
   "compilation_candidate": false,
   "tool_calls_made": 0
 }`;
-// ── Workflow matching prompt (Phase 2) ──────────────────────────────────────
-exports.WORKFLOW_MATCH_PROMPT = `\
-You are a strict workflow matching evaluator. Given a user request and a list of compiled workflows, determine if any workflow is a PRECISE match for the request.
-
-A workflow matches ONLY if:
-1. **Scope alignment**: The workflow does approximately what the user asked — not significantly more, not significantly less.
-2. **Intent alignment**: The workflow's purpose (description, original prompt) closely matches the user's goal — not just the same topic or domain.
-3. **Input compatibility**: The user's request provides enough information to populate the workflow's required inputs.
-
-Be CONSERVATIVE. If the user's request is a subset or superset of what the workflow does, it is NOT a match. When in doubt, return match: false — the system will fall back to a dynamic execution that handles the exact request.
-
-Respond with ONLY a JSON object:
-{
-  "match": true or false,
-  "workflow_name": "name-of-best-match" or null,
-  "confidence": 0.0 to 1.0,
-  "reasoning": "Brief explanation of why this is or isn't a scope match"
-}`;
-// ── Input extraction prompt (Phase 2b) ──────────────────────────────────────
-exports.EXTRACT_INPUTS_PROMPT = `\
-You are an input extraction engine. Given a user's natural-language request and a workflow's input schema, extract the structured inputs the workflow needs.
-
-Rules:
-- Extract ONLY values explicitly stated or clearly implied in the user's request.
-- Match each extracted value to the correct field in the input schema, paying attention to the field's **description** — not just its name.
-- Use the field descriptions to understand what each input represents and extract the semantically correct value from the request.
-- If a required field cannot be populated from the request, set "_extraction_failed" to true.
-- Do NOT invent, guess, or use default values for fields the user didn't mention.
-- Return ONLY a JSON object whose keys match the input schema's property names.
-- Include "_extraction_failed": true if any required field is missing, or "_extraction_failed": false if all required fields are satisfied.`;
+// ── Shared discovery prompts (re-exported for backward compatibility) ────────
+var prompts_1 = require("../shared/prompts");
+Object.defineProperty(exports, "WORKFLOW_MATCH_PROMPT", { enumerable: true, get: function () { return prompts_1.WORKFLOW_MATCH_PROMPT; } });
+Object.defineProperty(exports, "EXTRACT_INPUTS_PROMPT", { enumerable: true, get: function () { return prompts_1.EXTRACT_INPUTS_PROMPT; } });
 // ── Rounds-exhausted diagnostic prompt ──────────────────────────────────────
 exports.ROUNDS_EXHAUSTED_DIAGNOSTIC_PROMPT = `\
 You have run out of allowed tool rounds. Summarize:

package/build/system/workflows/mcp-query/strategy-advisors.d.ts

@@ -1,14 +1,3 @@
-/**
- * Strategy advisors — inspect loaded tool inventory and emit concrete,
- * tool-name-specific prompt sections so the LLM makes optimal tool choices.
- *
- * No tool names are hardcoded in the generic system prompt. Instead,
- * advisors run at tool-loading time and produce guidance referencing
- * the actual qualified names the LLM will see.
- */
-import type { ServerInfo } from './types';
-/**
- * Run all strategy advisors against the loaded server inventory
- * and return a combined strategy section for the LLM prompt.
- */
-export declare function generateStrategySection(servers: ServerInfo[]): string;
+/** Re-export from shared — this file exists for backward compatibility. */
+export { generateStrategySection } from '../shared/strategy-advisors';
+export type { ServerInfo } from '../shared/types';

package/build/system/workflows/mcp-query/strategy-advisors.js

@@ -1,109 +1,6 @@
 "use strict";
-/**
- * Strategy advisors — inspect loaded tool inventory and emit concrete,
- * tool-name-specific prompt sections so the LLM makes optimal tool choices.
- *
- * No tool names are hardcoded in the generic system prompt. Instead,
- * advisors run at tool-loading time and produce guidance referencing
- * the actual qualified names the LLM will see.
- */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.generateStrategySection = generateStrategySection;
-// ── Advisors ────────────────────────────────────────────────────────────────
-/**
- * Detects overlapping server categories (e.g., two browser-automation servers)
- * and emits concrete "USE THIS / AVOID THAT" guidance with actual tool names.
- */
-function overlappingCategoryAdvisor(servers) {
-    // Group by category
-    const byCategory = new Map();
-    for (const s of servers) {
-        const category = s.metadata?.category;
-        if (!category)
-            continue;
-        const group = byCategory.get(category) || [];
-        group.push(s);
-        byCategory.set(category, group);
-    }
-    const sections = [];
-    for (const [category, group] of byCategory) {
-        if (group.length < 2)
-            continue;
-        // Determine high vs low level: explicit metadata.level, then tool-count heuristic
-        let highLevel = null;
-        let lowLevel = null;
-        const withLevel = group.filter(s => s.metadata?.level);
-        if (withLevel.length >= 2) {
-            highLevel = withLevel.find(s => s.metadata?.level === 'high') || null;
-            lowLevel = withLevel.find(s => s.metadata?.level === 'low') || null;
-        }
-        if (!highLevel || !lowLevel) {
-            // Heuristic: fewer tools = higher-level (composite)
-            const sorted = [...group].sort((a, b) => a.toolCount - b.toolCount);
-            highLevel = highLevel || sorted[0];
-            lowLevel = lowLevel || sorted[sorted.length - 1];
-        }
-        if (highLevel === lowLevel)
-            continue;
-        const categoryLabel = category.replace(/-/g, ' ').replace(/\b\w/g, c => c.toUpperCase());
-        const hiSlug = highLevel.slug;
-        const loSlug = lowLevel.slug;
-        const hiTools = highLevel.toolNames.map(t => `${hiSlug}__${t}`);
-        const loTools = lowLevel.toolNames.map(t => `${loSlug}__${t}`);
-        sections.push(`## ${categoryLabel} — Tool Selection (CRITICAL)
-
-You have TWO ${categoryLabel.toLowerCase()} servers. One is high-level (composite), the other low-level (primitives).
-
-**USE: ${highLevel.name}** (${highLevel.toolCount} composite tools)
-Tools: ${hiTools.join(', ')}
-${highLevel.description ? `> ${highLevel.description}` : ''}
-Each tool handles a complete workflow in ONE call. Session management is automatic.
-
-**AVOID: ${lowLevel.name}** (${lowLevel.toolCount} primitive tools)
-Tools: ${loTools.join(', ')}
-These are individual primitives requiring manual session management. Using these when the composite tools can do the job causes 5-10x more calls and session failures.
-
-**RULE: ALWAYS use ${hiSlug}__* tools. NEVER use ${loSlug}__* tools unless no ${hiSlug}__* tool can accomplish the task.**`);
-    }
-    return sections.join('\n\n');
-}
-/**
- * Detects the human-queue server and emits guidance to use escalate_and_wait
- * instead of escalate_to_human + check_resolution polling.
- */
-function humanQueueAdvisor(servers) {
-    const hq = servers.find((s) => s.toolNames.includes('escalate_and_wait') && s.toolNames.includes('escalate_to_human'));
-    if (!hq)
-        return '';
-    const slug = hq.slug;
-    return (`## Human Queue — Tool Selection
-When you need input from a human before continuing (credentials, approval, data):
-USE: ${slug}__escalate_and_wait — pauses the workflow until the human responds. Specify a form_schema describing what fields you need (use format:"password" for sensitive values).
-AVOID: ${slug}__escalate_to_human + ${slug}__check_resolution polling loop — burns tool rounds and may exhaust your budget.`);
-}
-/**
- * Scans for batch-capable tools (accepting array inputs) and notes them
- * as preferred over repeated single-item calls.
- */
-function batchPreferenceAdvisor(servers) {
-    // This advisor is lightweight — it just identifies tools with array params
-    // from their names/descriptions. The overlapping category advisor handles
-    // the primary use case. This catches additional patterns.
-    return '';
-}
-// ── Public API ──────────────────────────────────────────────────────────────
-const advisors = [
-    overlappingCategoryAdvisor,
-    humanQueueAdvisor,
-    batchPreferenceAdvisor,
-];
-/**
- * Run all strategy advisors against the loaded server inventory
- * and return a combined strategy section for the LLM prompt.
- */
-function generateStrategySection(servers) {
-    const sections = advisors
-        .map(advisor => advisor(servers))
-        .filter(s => s.length > 0);
-    return sections.join('\n\n');
-}
+exports.generateStrategySection = void 0;
+/** Re-export from shared — this file exists for backward compatibility. */
+var strategy_advisors_1 = require("../shared/strategy-advisors");
+Object.defineProperty(exports, "generateStrategySection", { enumerable: true, get: function () { return strategy_advisors_1.generateStrategySection; } });

package/build/system/workflows/mcp-query/types.d.ts

@@ -1,10 +1,2 @@
-/** Type definitions for the mcp-query workflow. */
-export interface ServerInfo {
-    name: string;
-    description: string | null;
-    tags: string[];
-    metadata: Record<string, any> | null;
-    toolNames: string[];
-    toolCount: number;
-    slug: string;
-}
+/** Re-export shared types for backward compatibility. */
+export type { ServerInfo } from '../shared/types';

package/build/system/workflows/mcp-query/types.js

@@ -1,3 +1,2 @@
 "use strict";
-/** Type definitions for the mcp-query workflow. */
 Object.defineProperty(exports, "__esModule", { value: true });

package/build/system/workflows/mcp-query-router/index.js

@@ -39,7 +39,7 @@ const orchestrator_1 = require("../../../services/orchestrator");
 const activities = __importStar(require("./activities"));
 const { findCompiledWorkflows, evaluateWorkflowMatch, extractWorkflowInputs, } = hotmesh_1.Durable.workflow.proxyActivities({
     activities,
-    retryPolicy: {
+    retry: {
         maximumAttempts: 3,
         backoffCoefficient: 2,
         maximumInterval: '10 seconds',

package/build/system/workflows/mcp-triage/index.d.ts

@@ -7,8 +7,8 @@ import type { LTEnvelope, LTReturn } from '../../../types';
  * MCP tools scoped by tag affinity.
  *
  * Tool ecosystem grows over time:
- * - Built-in servers: document-vision, mcp-workflows, human-queue,
- *   workflow-compiler, db, telemetry
+ * - Built-in servers: translation, vision, mcp-workflows, human-queue,
+ *   workflow-compiler, db, http-fetch, file-storage, oauth, claude-code
  * - User-registered external MCP servers
  * - Compiled YAML workflows from past triage executions
  *

package/build/system/workflows/mcp-triage/index.js

@@ -40,9 +40,10 @@ const activities = __importStar(require("../../activities/triage"));
 const interceptorActivities = __importStar(require("../../../services/interceptor/activities"));
 const prompts_1 = require("./prompts");
 const response_1 = require("./response");
+const tool_result_guard_1 = require("../tool-result-guard");
 const { getUpstreamTasks, getEscalationHistory, getToolTags, loadTriageTools, callTriageTool, callTriageLLM, notifyEngineering, } = hotmesh_1.Durable.workflow.proxyActivities({
     activities,
-    retryPolicy: {
+    retry: {
         maximumAttempts: 3,
         backoffCoefficient: 2,
         maximumInterval: '10 seconds',
@@ -51,7 +52,7 @@ const { getUpstreamTasks, getEscalationHistory, getToolTags, loadTriageTools, ca
 const { ltCreateEscalation, ltCreateTask, ltGetTask, ltGetWorkflowConfig, ltStartWorkflow, ltEnrichEscalationRouting, } = hotmesh_1.Durable.workflow.proxyActivities({
     activities: interceptorActivities,
     taskQueue: 'lt-interceptor',
-    retryPolicy: { maximumAttempts: 3 },
+    retry: { maximumAttempts: 3 },
 });
 const MAX_TOOL_ROUNDS = defaults_1.TOOL_ROUNDS_TRIAGE;
 /** Proxied activity refs passed to response handlers */
@@ -72,8 +73,8 @@ const responseDeps = {
  * MCP tools scoped by tag affinity.
  *
  * Tool ecosystem grows over time:
- * - Built-in servers: document-vision, mcp-workflows, human-queue,
- *   workflow-compiler, db, telemetry
+ * - Built-in servers: translation, vision, mcp-workflows, human-queue,
+ *   workflow-compiler, db, http-fetch, file-storage, oauth, claude-code
  * - User-registered external MCP servers
  * - Compiled YAML workflows from past triage executions
  *
@@ -132,6 +133,10 @@ async function runTriageLLM(envelope, opts) {
         },
     ];
     let toolCallCount = 0;
+    // Repeated-error circuit breaker
+    const MAX_REPEATED_ERRORS = 2;
+    let lastErrorKey = '';
+    let repeatedErrorCount = 0;
     const BUDGET_WARNING_THRESHOLD = 3;
     for (let round = 0; round < MAX_TOOL_ROUNDS; round++) {
         const remaining = MAX_TOOL_ROUNDS - round;
@@ -155,6 +160,7 @@ async function runTriageLLM(envelope, opts) {
             content: response.content,
             tool_calls: fnCalls,
         });
+        let roundHadError = false;
         for (const toolCall of fnCalls) {
             toolCallCount++;
             let args = {};
@@ -168,7 +174,7 @@ async function runTriageLLM(envelope, opts) {
             const result = await callTriageTool(toolCall.function.name, args);
             // Durable waitFor: if the tool returns a signal, pause until human responds
             if (result?.type === 'waitFor' && result?.signalId) {
-                const ctx = hotmesh_1.Durable.workflow.getContext();
+                const ctx = hotmesh_1.Durable.workflow.workflowInfo();
                 const workflowType = ctx.workflowTopic.replace(`${ctx.taskQueue}-`, '');
                 await ltEnrichEscalationRouting({
                     escalationId: result.escalationId,
@@ -180,7 +186,7 @@ async function runTriageLLM(envelope, opts) {
                     },
                     claimForUserId: envelope.lt?.userId,
                 });
-                const signalData = await hotmesh_1.Durable.workflow.waitFor(result.signalId);
+                const signalData = await hotmesh_1.Durable.workflow.condition(result.signalId);
                 messages.push({
                     role: 'tool',
                     tool_call_id: toolCall.id,
@@ -188,11 +194,37 @@ async function runTriageLLM(envelope, opts) {
                 });
                 continue;
             }
+            const resultStr = (0, tool_result_guard_1.sanitizeToolResult)(result);
             messages.push({
                 role: 'tool',
                 tool_call_id: toolCall.id,
-                content: JSON.stringify(result),
+                content: resultStr,
             });
+            // Track repeated errors: same tool + same error = stuck
+            const isError = resultStr.includes('MCP error') || resultStr.includes('"error"');
+            if (isError) {
+                roundHadError = true;
+                const errorKey = `${toolCall.function.name}::${resultStr.slice(0, 200)}`;
+                if (errorKey === lastErrorKey) {
+                    repeatedErrorCount++;
+                }
+                else {
+                    lastErrorKey = errorKey;
+                    repeatedErrorCount = 1;
+                }
+                if (repeatedErrorCount >= MAX_REPEATED_ERRORS) {
+                    messages.push({
+                        role: 'user',
+                        content: `[CIRCUIT BREAKER] The tool "${toolCall.function.name}" has returned the same error ${repeatedErrorCount} times. Stop retrying this tool. Summarize your diagnosis and what failed.`,
+                    });
+                    const bailResponse = await callTriageLLM(messages, undefined);
+                    return (0, response_1.handleFinalResponse)(bailResponse.content || '', envelope, toolCallCount, responseDeps);
+                }
+            }
+        }
+        if (!roundHadError) {
+            lastErrorKey = '';
+            repeatedErrorCount = 0;
         }
     }
     // Exhausted rounds — ask for final synthesis without tools

package/build/system/workflows/mcp-triage/prompts.js

@@ -31,26 +31,19 @@ Tool names in function calls are prefixed with the server slug: \`server_slug__t
 
 ## Tool Categories
 
-**Compiled Workflows** (\`long_tail_mcp_workflows__*\`):
-Deterministic pipelines hardened from past triage successes. ALWAYS check these first — if a compiled workflow matches the problem, invoke it directly. This is the most efficient path.
+**Compiled workflows**: Deterministic pipelines hardened from past triage successes. ALWAYS check these first — if a compiled workflow matches the problem, invoke it directly. This is the most efficient path.
 
-**Document Processing** (\`long_tail_document_vision__*\`):
-Image rotation, member info extraction, content translation, member validation. Use for document-related issues.
+**Data & analytics**: Read-only queries against tasks, escalations, processes, workflow types, and system health. Use to investigate context.
 
-**Human Queue** (\`long_tail_human_queue__*\`):
-Create escalations within the triage vortex when you need human help (e.g., asking an engineer to install tools). Note: you can also escalate simply by returning \`correctedData: null\` — the orchestrator will create an engineer escalation automatically. Using human-queue tools gives you more control over routing but is usually unnecessary.
+**Text processing**: Content translation and language detection. Use for language-related issues.
 
-**Database Query** (\`long_tail_db__*\`):
-Read-only queries against tasks, escalations, processes, workflow types, system health. Use to investigate context.
+**Vision & image analysis**: Analyze images and extract structured data, text, or descriptions. Use for image-related issues.
 
-**Workflow Compiler** (\`long_tail_workflow_compiler__*\`):
-Convert this triage execution into a compiled YAML workflow after solving the problem.
+**Escalation management**: Create escalations within the triage vortex when you need human help. Note: you can also escalate simply by returning \`correctedData: null\` — the orchestrator will create an engineer escalation automatically.
 
-**HTTP Fetch** (\`long_tail_http_fetch__*\`):
-Fetch external data needed during investigation. Not for delivering results — your JSON response handles that.
+**Workflow compilation**: Convert this triage execution into a compiled YAML workflow after solving the problem.
 
-**Telemetry** (\`long_tail_telemetry__*\`):
-Honeycomb trace links for debugging.
+**Network & HTTP**: Fetch external data needed during investigation. Not for delivering results — your JSON response handles that.
 
 **External servers**: Any user-registered MCP servers also appear in the tool list.
 

package/build/system/workflows/mcp-triage-deterministic/index.js

@@ -38,7 +38,7 @@ const hotmesh_1 = require("@hotmeshio/hotmesh");
 const activities = __importStar(require("./activities"));
 const { invokeCompiledWorkflow } = hotmesh_1.Durable.workflow.proxyActivities({
     activities,
-    retryPolicy: {
+    retry: {
         maximumAttempts: 3,
         backoffCoefficient: 2,
         maximumInterval: '10 seconds',

package/build/system/workflows/mcp-triage-router/index.js

@@ -39,7 +39,7 @@ const orchestrator_1 = require("../../../services/orchestrator");
 const activities = __importStar(require("./activities"));
 const { findTriageWorkflows, evaluateTriageMatch, extractTriageInputs, } = hotmesh_1.Durable.workflow.proxyActivities({
     activities,
-    retryPolicy: {
+    retry: {
         maximumAttempts: 3,
         backoffCoefficient: 2,
         maximumInterval: '10 seconds',

package/build/system/workflows/mcp-workflow-builder/activities/caches.d.ts

@@ -0,0 +1,5 @@
+import type { ToolDefinition } from '../../../../services/llm';
+/** Maps qualified tool name -> MCP server name for routing calls */
+export declare const toolServerMap: Map<string, string>;
+/** Maps qualified tool name -> full tool definition */
+export declare const toolDefCache: Map<string, ToolDefinition>;

package/build/system/workflows/mcp-workflow-builder/activities/caches.js

@@ -0,0 +1,8 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.toolDefCache = exports.toolServerMap = void 0;
+// ── Tool caches (module-level, persist across proxy activity calls) ──
+/** Maps qualified tool name -> MCP server name for routing calls */
+exports.toolServerMap = new Map();
+/** Maps qualified tool name -> full tool definition */
+exports.toolDefCache = new Map();

package/build/system/workflows/mcp-workflow-builder/activities/index.d.ts

@@ -0,0 +1,2 @@
+export { loadBuilderTools } from './tool-loader';
+export { callBuilderLLM } from './llm';

package/build/system/workflows/mcp-workflow-builder/activities/index.js

@@ -0,0 +1,7 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.callBuilderLLM = exports.loadBuilderTools = void 0;
+var tool_loader_1 = require("./tool-loader");
+Object.defineProperty(exports, "loadBuilderTools", { enumerable: true, get: function () { return tool_loader_1.loadBuilderTools; } });
+var llm_1 = require("./llm");
+Object.defineProperty(exports, "callBuilderLLM", { enumerable: true, get: function () { return llm_1.callBuilderLLM; } });

package/build/system/workflows/mcp-workflow-builder/activities/llm.d.ts

@@ -0,0 +1,2 @@
+import { type LLMResponse } from '../../../../services/llm';
+export declare function callBuilderLLM(messages: any[], _toolIds?: string[]): Promise<LLMResponse>;

package/build/system/workflows/mcp-workflow-builder/activities/llm.js

@@ -0,0 +1,25 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.callBuilderLLM = callBuilderLLM;
+const llm_1 = require("../../../../services/llm");
+const defaults_1 = require("../../../../modules/defaults");
+const logger_1 = require("../../../../lib/logger");
+/**
+ * Call the LLM for workflow building.
+ *
+ * Uses a higher max_tokens than the shared caller because the builder
+ * returns complete YAML + manifest JSON (typically 2000-4000 tokens).
+ */
+const BUILDER_MAX_TOKENS = 4096;
+async function callBuilderLLM(messages, _toolIds) {
+    const t0 = Date.now();
+    const response = await (0, llm_1.callLLM)({
+        model: defaults_1.LLM_MODEL_PRIMARY,
+        messages,
+        temperature: 0,
+        max_tokens: BUILDER_MAX_TOKENS,
+    });
+    const usage = response.usage;
+    logger_1.loggerRegistry.info(`[workflowBuilder:callLLM] ${Date.now() - t0}ms | in=${usage?.prompt_tokens} out=${usage?.completion_tokens} total=${usage?.total_tokens} | tool_calls=${response.tool_calls?.length || 0}`);
+    return response;
+}

package/build/system/workflows/mcp-workflow-builder/activities/tool-loader.d.ts

@@ -0,0 +1,5 @@
+export declare function loadBuilderTools(tags?: string[]): Promise<{
+    toolIds: string[];
+    inventory: string;
+    strategy: string;
+}>;

package/build/system/workflows/mcp-workflow-builder/activities/tool-loader.js

@@ -0,0 +1,8 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.loadBuilderTools = loadBuilderTools;
+const tool_loader_1 = require("../../shared/tool-loader");
+const caches_1 = require("./caches");
+async function loadBuilderTools(tags) {
+    return (0, tool_loader_1.loadToolsFromServers)(tags, { toolServerMap: caches_1.toolServerMap, toolDefCache: caches_1.toolDefCache }, { logPrefix: 'workflowBuilder' });
+}

package/build/system/workflows/mcp-workflow-builder/index.d.ts

@@ -0,0 +1,16 @@
+import type { LTEnvelope, LTReturn } from '../../../types';
+/**
+ * MCP Workflow Builder — constructs HotMesh YAML DAGs directly.
+ *
+ * Unlike mcpQuery (which executes tools then compiles traces), this workflow
+ * reasons about tool schemas and builds the YAML declaratively. The LLM
+ * knows HotMesh's activity types, mapping syntax, and @pipe operators.
+ *
+ * Supports conversational clarification: when the prompt is ambiguous,
+ * the LLM returns questions instead of YAML. The user answers, and the
+ * exchange continues until the LLM has enough to build.
+ *
+ * Returns the generated YAML, input schema, and activity manifest —
+ * ready for deployment via the existing yaml-workflow infrastructure.
+ */
+export declare function mcpWorkflowBuilder(envelope: LTEnvelope): Promise<LTReturn>;