@deepagents/orchestrator 6.1.0 → 7.0.0

package/dist/lib/deepplan/replanner-agent.js

@@ -0,0 +1,418 @@
+// packages/orchestrator/src/lib/deepplan/replanner-agent.ts
+import { defaultSettingsMiddleware as defaultSettingsMiddleware2, wrapLanguageModel as wrapLanguageModel2 } from "ai";
+import z from "zod";
+import { lmstudio } from "@deepagents/agent";
+import {
+  ContextEngine,
+  InMemoryContextStore,
+  fragment,
+  persona,
+  structuredOutput,
+  user
+} from "@deepagents/context";
+
+// packages/orchestrator/src/lib/deepplan/executors/generic-executor.ts
+import { groq } from "@ai-sdk/groq";
+import { defaultSettingsMiddleware, wrapLanguageModel } from "ai";
+import { agent } from "@deepagents/agent";
+import { scratchpad_tool, search_content_tool } from "@deepagents/toolbox";
+var executorAgent = agent({
+  name: "executor_agent",
+  model: wrapLanguageModel({
+    model: groq("openai/gpt-oss-20b"),
+    middleware: defaultSettingsMiddleware({
+      settings: { temperature: 0 }
+    })
+  }),
+  prompt: `
+    <SystemContext>
+      You are a precise execution agent that follows plans and gathers information systematically.
+      You pay close attention to details and track what you find.
+    </SystemContext>
+
+    <Identity>
+      Your task is to execute a single step from a plan and report back with structured results.
+      You are part of an adaptive planning system where your findings will influence future steps.
+    </Identity>
+
+    <ExecutionPrinciples>
+      1. **Focus on the Objective**
+         - Understand exactly what this step is asking for
+         - Review the expected outcome
+         - Stay focused on gathering that specific information
+
+      2. **Extract Variables and Numerals**
+         - As you find information, extract key data points
+         - **CRITICAL: Extract numerical values separately from units**
+         - Store them as structured variables (e.g., {product_name: "iPhone 15", release_date: "Sep 2024", release_month: 9, release_year: 2024})
+         - Track counts, amounts, dates, and measurements explicitly
+         - These variables help track progress across steps
+
+      3. **Calculate Intermediate Results**
+         - Pay attention to correct numerical calculation and commonsense
+         - Perform any required calculations step-by-step
+         - Verify numerical accuracy (e.g., date ranges, counts, sums)
+         - Check results against common sense (e.g., dates in valid range, reasonable quantities)
+         - Show your calculation work when computing values
+
+      4. **Verify Results**
+         - Carefully review what you find for accuracy
+         - Ensure numerical values are consistent
+         - Check for logical coherence
+         - Note anything unexpected or inconsistent
+
+      5. **Make Observations**
+         - What patterns did you notice?
+         - What numerical trends are apparent?
+         - What might affect subsequent steps?
+
+      6. **Note Issues**
+         - Were there any calculation difficulties?
+         - Was the information hard to find?
+         - Were there any errors or ambiguities?
+    </ExecutionPrinciples>
+
+
+    <ContextAwareness>
+      You will be given:
+      - The current step to execute
+      - The original user request (for context)
+      - Variables gathered from previous steps
+      - Data collected so far
+
+      Use this context to:
+      - Avoid re-gathering information you already have
+      - Build on previous findings
+      - Connect new information to what's already known
+    </ContextAwareness>
+
+    <ReportingGuidelines>
+      Your report should include:
+
+      1. **Status**
+         - success: Step completed as expected
+         - partial: Got some but not all information
+         - failed: Could not complete the step
+
+      2. **Data Gathered**
+         - Summarize what you found
+         - Be specific and factual
+         - Include enough detail to be useful
+
+      3. **Extracted Variables and Numerals**
+         - Key data points you discovered
+         - Format as key-value pairs with numerical values separated
+         - Examples: {count: 5, average_price: 299, latest_date: "2024-01-15", date_month: 1, date_year: 2024}
+         - Include any intermediate calculations performed
+
+      4. **Observations**
+         - Notable patterns or insights
+         - Anything that might affect the plan
+         - Connections to the overall goal
+
+      5. **Issues**
+         - Any problems encountered
+         - Missing or unclear information
+         - Limitations of what you found
+    </ReportingGuidelines>
+
+    <Examples>
+      Step: "Find Apple's official announcements from the past 3 months"
+      Context: {company: "Apple", timeframe: "3 months"}
+
+      Good Execution:
+      - Use browser_search to find Apple newsroom/press releases
+      - Filter results to last 3 months
+      - List the announcements found
+      - Extract count, dates, types
+      - Note any gaps or issues
+
+      Report (with numerical extraction):
+      - Status: success
+      - Data: "Found 15 press releases from Apple newsroom (Oct-Dec 2024)"
+      - Variables: {
+          press_release_count: 15,
+          date_range: "Oct-Dec 2024",
+          start_month: 10,
+          end_month: 12,
+          year: 2024,
+          months_covered: 3
+        }
+      - Observations: ["Mix of product, service, and corporate announcements", "Exactly covers the requested 3-month timeframe"]
+      - Issues: []
+    </Examples>
+
+    <CriticalInstructions>
+      - Execute the step precisely as described
+      - **Extract numerical variables and their values separately**
+      - **Pay attention to correct numerical calculation and commonsense**
+      - Calculate and verify intermediate results step-by-step
+      - Report accurately with verified numbers - your results influence future steps
+      - Use the context from previous steps to work smarter
+      - If you can't find something, say so clearly
+      - Double-check all numerical values for accuracy
+    </CriticalInstructions>
+
+    <OutputFormat>
+      Provide a structured step result with:
+      - status (success/partial/failed)
+      - data_gathered (summary of findings)
+      - extracted_variables (key-value pairs)
+      - observations (notable insights)
+      - issues (any problems)
+    </OutputFormat>
+  `,
+  tools: {
+    scratchpad: scratchpad_tool,
+    search: search_content_tool
+    // browser_search: groq.tools.browserSearch({}),
+  }
+});
+
+// packages/orchestrator/src/lib/deepplan/replanner-agent.ts
+var PlanStepSchema = z.object({
+  description: z.string().describe("What information to find or action to take"),
+  expected_outcome: z.string().describe("What should be accomplished by this step")
+});
+var ReplanDecisionSchema = z.object({
+  reasoning: z.string().describe("Why this replanning decision was made"),
+  should_continue: z.boolean().describe("Should we continue executing?"),
+  remaining_steps: z.array(PlanStepSchema).max(4).describe("Updated list of remaining steps"),
+  new_insights: z.string().describe("New insights learned that affected the plan")
+});
+var replannerModel = wrapLanguageModel2({
+  // model: groq('moonshotai/kimi-k2-instruct-0905'),
+  model: lmstudio("google/gemma-3-12b"),
+  middleware: defaultSettingsMiddleware2({
+    settings: { temperature: 0.1 }
+  })
+});
+function createReplannerPrompt() {
+  return `
+    <SystemContext>
+      You are an adaptive replanning agent.
+      After each step execution, you review progress and adjust the plan as needed.
+    </SystemContext>
+
+    <Identity>
+      Your role is to ensure the execution stays on track and adapts to new information.
+      You analyze results, learn from findings, and modify the plan to achieve the goal efficiently.
+    </Identity>
+
+    <ReplanningPrinciples>
+      1. **Review Results**
+         - What was the step supposed to accomplish?
+         - What was actually accomplished?
+         - Was it successful, partial, or failed?
+
+      2. **Analyze Impact**
+         - How do these results affect the overall goal?
+         - What did we learn that we didn't know before?
+         - Does this change what we need to do next?
+
+      3. **Assess Remaining Plan**
+         - Are the remaining steps still appropriate?
+         - Do we need additional steps?
+         - Can we skip any steps?
+         - Should we change the order?
+
+      4. **Make Decisions**
+         - Should we continue with the plan?
+         - What adjustments are needed?
+         - How significant are the changes?
+
+      5. **Update Steps**
+         - Modify remaining steps based on learnings
+         - Add new steps if needed
+         - Remove unnecessary steps
+         - Reorder for efficiency
+    </ReplanningPrinciples>
+
+    <WhenToMakeChanges>
+      ## Add Steps When:
+      - Found more information than expected (need to filter/process)
+      - Discovered new relevant categories (need to explore)
+      - Hit limitations (need alternative approaches)
+      - Found partial data (need additional sources)
+
+      ## Remove Steps When:
+      - Already got the information another way
+      - The step is no longer relevant based on findings
+      - The goal can be achieved without it
+      - It's redundant with completed steps
+
+      ## Modify Steps When:
+      - Need to be more specific based on findings
+      - Scope needs adjustment (broader or narrower)
+      - Better approach became apparent
+      - Need to work around issues discovered
+
+      ## Keep Steps When:
+      - They're still relevant and needed
+      - Results so far support the original plan
+      - No better alternative is apparent
+    </WhenToMakeChanges>
+
+    <ChangeMagnitude>
+      Classify plan changes as:
+      - **none**: No changes needed, continue as planned
+      - **minor**: Small tweaks (wording, ordering, slight additions)
+      - **major**: Significant restructuring (many new steps, removed steps, different approach)
+    </ChangeMagnitude>
+
+    <DecisionMaking>
+      Decide if execution should continue:
+
+      \u2713 **Continue** when:
+      - More steps are needed to complete the goal
+      - Making progress toward the objective
+      - Can adapt the plan to handle issues
+
+      \u2717 **Stop** when:
+      - Goal has been fully achieved
+      - Discovered the answer isn't available
+      - Fundamental blocker that can't be worked around
+      - Success criteria have been met
+    </DecisionMaking>
+
+    <Examples>
+      Scenario 1: Found more data than expected
+
+      Step Result: Found 50 press releases instead of expected 10-15
+
+      Replanning Decision:
+      - Reasoning: Large volume needs filtering before analysis
+      - Changes: major
+      - Add step: "Filter press releases to product-related only"
+      - Modify next step: Focus on filtered subset
+      - Continue: true
+
+      ---
+
+      Scenario 2: Got exactly what was needed
+
+      Step Result: Successfully identified 3 products with all details
+
+      Replanning Decision:
+      - Reasoning: Already have complete information
+      - Changes: none or minor (maybe skip redundant verification step)
+      - Continue: Check if success criteria met, might stop
+
+      ---
+
+      Scenario 3: Information not available
+
+      Step Result: Official source doesn't have the data
+
+      Replanning Decision:
+      - Reasoning: Need alternative data source
+      - Changes: major
+      - Add step: "Search tech news sites for announcements"
+      - Continue: true (with new approach)
+    </Examples>
+
+    <ContextAwareness>
+      You have access to:
+      - Original user request
+      - Initial understanding and success criteria
+      - All previous step results
+      - Current variables and gathered data
+      - The remaining steps in the plan
+
+      Use this to make informed replanning decisions.
+    </ContextAwareness>
+
+    <CriticalInstructions>
+      - Be thoughtful but decisive
+      - Adapt the plan based on ACTUAL results, not assumptions
+      - Keep the end goal in mind (user's original request)
+      - Don't keep executing steps if the goal is already achieved
+      - Don't abandon the plan too easily - adapt when possible
+      - Make changes that improve efficiency and effectiveness
+      - Explain your reasoning clearly
+    </CriticalInstructions>
+
+    <OutputFormat>
+      Provide a replanning decision with:
+      - reasoning: Why you're making this decision
+      - should_continue: true/false
+      - remaining_steps: Updated list of steps to execute
+      - new_insights: What you learned that affected the plan
+    </OutputFormat>
+  `;
+}
+function formatReplannerPrompt(context, remainingSteps) {
+  const executionHistory = context.step_results.map((r, i) => `Step ${i}: ${r}`).join("\n");
+  return `
+<OriginalRequest>${context.original_request}</OriginalRequest>
+
+<Understanding>${context.understanding}</Understanding>
+
+<SuccessCriteria>
+Check if we have achieved the goal of the original request.
+</SuccessCriteria>
+
+<ExecutionHistory>
+${executionHistory}
+</ExecutionHistory>
+
+<MostRecentStep>
+${context.step_results.at(-1) ?? "No steps executed yet"}
+</MostRecentStep>
+
+<CurrentVariables>
+${JSON.stringify(context.variables, null, 2)}
+</CurrentVariables>
+
+<RemainingSteps>
+${remainingSteps.length > 0 ? remainingSteps.map((s, i) => `${i}. ${s.description} (Expected: ${s.expected_outcome})`).join("\n") : "No remaining steps"}
+</RemainingSteps>
+
+<YourTask>
+Review the most recent step result and determine:
+1. Should we continue execution?
+2. Does the plan need adjustment based on what we found?
+3. What should the remaining steps be?
+
+Consider:
+- Did we find what we were looking for?
+- Is the goal achievable with the current approach?
+- Do results suggest a better path forward?
+- Have we already met the success criteria?
+</YourTask>
+  `.trim();
+}
+async function replan(context) {
+  const remainingSteps = context.current_plan.slice(1);
+  const plannerContext = new ContextEngine({
+    store: new InMemoryContextStore(),
+    chatId: `replanner-${crypto.randomUUID()}`,
+    userId: "system"
+  });
+  plannerContext.set(
+    persona({
+      name: "replanner_agent",
+      role: "You are an adaptive replanning agent.",
+      objective: "Review execution progress and adjust remaining steps so the plan reaches the original goal efficiently."
+    }),
+    fragment("replanner_instructions", createReplannerPrompt()),
+    user(formatReplannerPrompt(context, remainingSteps))
+  );
+  const replannerOutput = structuredOutput({
+    model: replannerModel,
+    context: plannerContext,
+    schema: ReplanDecisionSchema
+  });
+  const output = await replannerOutput.generate();
+  const decision = output;
+  return {
+    ...context,
+    current_plan: decision.remaining_steps,
+    is_complete: !decision.should_continue
+  };
+}
+export {
+  ReplanDecisionSchema,
+  replan
+};
+//# sourceMappingURL=replanner-agent.js.map

package/dist/lib/deepplan/replanner-agent.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../../../src/lib/deepplan/replanner-agent.ts", "../../../src/lib/deepplan/executors/generic-executor.ts"],
+  "sourcesContent": ["import { defaultSettingsMiddleware, wrapLanguageModel } from 'ai';\nimport z from 'zod';\n\nimport { lmstudio } from '@deepagents/agent';\nimport {\n  ContextEngine,\n  InMemoryContextStore,\n  createBashTool,\n  fragment,\n  persona,\n  structuredOutput,\n  user,\n} from '@deepagents/context';\n\nimport { type ExecutionContext } from './executors/generic-executor.ts';\nimport type { PlanStep } from './planner-agent.ts';\n\nconst PlanStepSchema = z.object({\n  description: z\n    .string()\n    .describe('What information to find or action to take'),\n  expected_outcome: z\n    .string()\n    .describe('What should be accomplished by this step'),\n});\nexport const ReplanDecisionSchema = z.object({\n  reasoning: z.string().describe('Why this replanning decision was made'),\n  should_continue: z.boolean().describe('Should we continue executing?'),\n  remaining_steps: z\n    .array(PlanStepSchema)\n    .max(4)\n    .describe('Updated list of remaining steps'),\n  new_insights: z\n    .string()\n    .describe('New insights learned that affected the plan'),\n});\n\nexport type ReplanDecision = z.infer<typeof ReplanDecisionSchema>;\n\n/**\n * Replanner Agent\n *\n * After each step execution, the replanner:\n * 1. Reviews what was accomplished vs what was expected\n * 2. Analyzes how this affects the remaining plan\n * 3. Decides if the plan needs adjustment\n * 4. Updates remaining steps based on new information\n *\n * This creates an adaptive planning system that learns as it executes.\n */\nconst replannerModel = wrapLanguageModel({\n  // model: groq('moonshotai/kimi-k2-instruct-0905'),\n  model: lmstudio('google/gemma-3-12b'),\n  middleware: defaultSettingsMiddleware({\n    settings: { temperature: 0.1 },\n  }),\n});\n\nfunction createReplannerPrompt(): string {\n  return `\n    <SystemContext>\n      You are an adaptive replanning agent.\n      After each step execution, you review progress and adjust the plan as needed.\n    </SystemContext>\n\n    <Identity>\n      Your role is to ensure the execution stays on track and adapts to new information.\n      You analyze results, learn from findings, and modify the plan to achieve the goal efficiently.\n    </Identity>\n\n    <ReplanningPrinciples>\n      1. **Review Results**\n         - What was the step supposed to accomplish?\n         - What was actually accomplished?\n         - Was it successful, partial, or failed?\n\n      2. **Analyze Impact**\n         - How do these results affect the overall goal?\n         - What did we learn that we didn't know before?\n         - Does this change what we need to do next?\n\n      3. **Assess Remaining Plan**\n         - Are the remaining steps still appropriate?\n         - Do we need additional steps?\n         - Can we skip any steps?\n         - Should we change the order?\n\n      4. **Make Decisions**\n         - Should we continue with the plan?\n         - What adjustments are needed?\n         - How significant are the changes?\n\n      5. **Update Steps**\n         - Modify remaining steps based on learnings\n         - Add new steps if needed\n         - Remove unnecessary steps\n         - Reorder for efficiency\n    </ReplanningPrinciples>\n\n    <WhenToMakeChanges>\n      ## Add Steps When:\n      - Found more information than expected (need to filter/process)\n      - Discovered new relevant categories (need to explore)\n      - Hit limitations (need alternative approaches)\n      - Found partial data (need additional sources)\n\n      ## Remove Steps When:\n      - Already got the information another way\n      - The step is no longer relevant based on findings\n      - The goal can be achieved without it\n      - It's redundant with completed steps\n\n      ## Modify Steps When:\n      - Need to be more specific based on findings\n      - Scope needs adjustment (broader or narrower)\n      - Better approach became apparent\n      - Need to work around issues discovered\n\n      ## Keep Steps When:\n      - They're still relevant and needed\n      - Results so far support the original plan\n      - No better alternative is apparent\n    </WhenToMakeChanges>\n\n    <ChangeMagnitude>\n      Classify plan changes as:\n      - **none**: No changes needed, continue as planned\n      - **minor**: Small tweaks (wording, ordering, slight additions)\n      - **major**: Significant restructuring (many new steps, removed steps, different approach)\n    </ChangeMagnitude>\n\n    <DecisionMaking>\n      Decide if execution should continue:\n\n      \u2713 **Continue** when:\n      - More steps are needed to complete the goal\n      - Making progress toward the objective\n      - Can adapt the plan to handle issues\n\n      \u2717 **Stop** when:\n      - Goal has been fully achieved\n      - Discovered the answer isn't available\n      - Fundamental blocker that can't be worked around\n      - Success criteria have been met\n    </DecisionMaking>\n\n    <Examples>\n      Scenario 1: Found more data than expected\n\n      Step Result: Found 50 press releases instead of expected 10-15\n\n      Replanning Decision:\n      - Reasoning: Large volume needs filtering before analysis\n      - Changes: major\n      - Add step: \"Filter press releases to product-related only\"\n      - Modify next step: Focus on filtered subset\n      - Continue: true\n\n      ---\n\n      Scenario 2: Got exactly what was needed\n\n      Step Result: Successfully identified 3 products with all details\n\n      Replanning Decision:\n      - Reasoning: Already have complete information\n      - Changes: none or minor (maybe skip redundant verification step)\n      - Continue: Check if success criteria met, might stop\n\n      ---\n\n      Scenario 3: Information not available\n\n      Step Result: Official source doesn't have the data\n\n      Replanning Decision:\n      - Reasoning: Need alternative data source\n      - Changes: major\n      - Add step: \"Search tech news sites for announcements\"\n      - Continue: true (with new approach)\n    </Examples>\n\n    <ContextAwareness>\n      You have access to:\n      - Original user request\n      - Initial understanding and success criteria\n      - All previous step results\n      - Current variables and gathered data\n      - The remaining steps in the plan\n\n      Use this to make informed replanning decisions.\n    </ContextAwareness>\n\n    <CriticalInstructions>\n      - Be thoughtful but decisive\n      - Adapt the plan based on ACTUAL results, not assumptions\n      - Keep the end goal in mind (user's original request)\n      - Don't keep executing steps if the goal is already achieved\n      - Don't abandon the plan too easily - adapt when possible\n      - Make changes that improve efficiency and effectiveness\n      - Explain your reasoning clearly\n    </CriticalInstructions>\n\n    <OutputFormat>\n      Provide a replanning decision with:\n      - reasoning: Why you're making this decision\n      - should_continue: true/false\n      - remaining_steps: Updated list of steps to execute\n      - new_insights: What you learned that affected the plan\n    </OutputFormat>\n  `;\n}\n\nfunction formatReplannerPrompt(\n  context: ExecutionContext,\n  remainingSteps: PlanStep[],\n): string {\n  const executionHistory = context.step_results\n    .map((r, i) => `Step ${i}: ${r}`)\n    .join('\\n');\n\n  return `\n<OriginalRequest>${context.original_request}</OriginalRequest>\n\n<Understanding>${context.understanding}</Understanding>\n\n<SuccessCriteria>\nCheck if we have achieved the goal of the original request.\n</SuccessCriteria>\n\n<ExecutionHistory>\n${executionHistory}\n</ExecutionHistory>\n\n<MostRecentStep>\n${context.step_results.at(-1) ?? 'No steps executed yet'}\n</MostRecentStep>\n\n<CurrentVariables>\n${JSON.stringify(context.variables, null, 2)}\n</CurrentVariables>\n\n<RemainingSteps>\n${remainingSteps.length > 0 ? remainingSteps.map((s, i) => `${i}. ${s.description} (Expected: ${s.expected_outcome})`).join('\\n') : 'No remaining steps'}\n</RemainingSteps>\n\n<YourTask>\nReview the most recent step result and determine:\n1. Should we continue execution?\n2. Does the plan need adjustment based on what we found?\n3. What should the remaining steps be?\n\nConsider:\n- Did we find what we were looking for?\n- Is the goal achievable with the current approach?\n- Do results suggest a better path forward?\n- Have we already met the success criteria?\n</YourTask>\n  `.trim();\n}\n\nexport async function replan(context: ExecutionContext) {\n  const remainingSteps = context.current_plan.slice(1);\n  const plannerContext = new ContextEngine({\n    store: new InMemoryContextStore(),\n    chatId: `replanner-${crypto.randomUUID()}`,\n    userId: 'system',\n  });\n\n  plannerContext.set(\n    persona({\n      name: 'replanner_agent',\n      role: 'You are an adaptive replanning agent.',\n      objective:\n        'Review execution progress and adjust remaining steps so the plan reaches the original goal efficiently.',\n    }),\n    fragment('replanner_instructions', createReplannerPrompt()),\n    user(formatReplannerPrompt(context, remainingSteps)),\n  );\n\n  const replannerOutput = structuredOutput({\n    model: replannerModel,\n    context: plannerContext,\n    schema: ReplanDecisionSchema,\n  });\n\n  const output = await replannerOutput.generate();\n  const decision = output as {\n    remaining_steps: typeof context.current_plan;\n    should_continue: boolean;\n  };\n  return {\n    ...context,\n    current_plan: decision.remaining_steps,\n    is_complete: !decision.should_continue,\n  };\n}\n", "import { groq } from '@ai-sdk/groq';\nimport { defaultSettingsMiddleware, wrapLanguageModel } from 'ai';\n\nimport { agent } from '@deepagents/agent';\nimport { scratchpad_tool, search_content_tool } from '@deepagents/toolbox';\n\nimport type { PlanStep, PlannerOutput } from '../planner-agent.ts';\n\n/**\n * Executor Agent\n *\n * Executes individual plan steps while:\n * 1. Tracking what information is gathered\n * 2. Extracting numerical variables from results\n * 3. Calculating intermediate results with attention to accuracy\n * 4. Paying attention to correct numerical calculation and commonsense\n * 5. Making observations about findings\n * 6. Noting any issues or unexpected results\n */\nexport const executorAgent = agent({\n  name: 'executor_agent',\n  model: wrapLanguageModel({\n    model: groq('openai/gpt-oss-20b'),\n    middleware: defaultSettingsMiddleware({\n      settings: { temperature: 0 },\n    }),\n  }),\n  prompt: `\n    <SystemContext>\n      You are a precise execution agent that follows plans and gathers information systematically.\n      You pay close attention to details and track what you find.\n    </SystemContext>\n\n    <Identity>\n      Your task is to execute a single step from a plan and report back with structured results.\n      You are part of an adaptive planning system where your findings will influence future steps.\n    </Identity>\n\n    <ExecutionPrinciples>\n      1. **Focus on the Objective**\n         - Understand exactly what this step is asking for\n         - Review the expected outcome\n         - Stay focused on gathering that specific information\n\n      2. **Extract Variables and Numerals**\n         - As you find information, extract key data points\n         - **CRITICAL: Extract numerical values separately from units**\n         - Store them as structured variables (e.g., {product_name: \"iPhone 15\", release_date: \"Sep 2024\", release_month: 9, release_year: 2024})\n         - Track counts, amounts, dates, and measurements explicitly\n         - These variables help track progress across steps\n\n      3. **Calculate Intermediate Results**\n         - Pay attention to correct numerical calculation and commonsense\n         - Perform any required calculations step-by-step\n         - Verify numerical accuracy (e.g., date ranges, counts, sums)\n         - Check results against common sense (e.g., dates in valid range, reasonable quantities)\n         - Show your calculation work when computing values\n\n      4. **Verify Results**\n         - Carefully review what you find for accuracy\n         - Ensure numerical values are consistent\n         - Check for logical coherence\n         - Note anything unexpected or inconsistent\n\n      5. **Make Observations**\n         - What patterns did you notice?\n         - What numerical trends are apparent?\n         - What might affect subsequent steps?\n\n      6. **Note Issues**\n         - Were there any calculation difficulties?\n         - Was the information hard to find?\n         - Were there any errors or ambiguities?\n    </ExecutionPrinciples>\n\n\n    <ContextAwareness>\n      You will be given:\n      - The current step to execute\n      - The original user request (for context)\n      - Variables gathered from previous steps\n      - Data collected so far\n\n      Use this context to:\n      - Avoid re-gathering information you already have\n      - Build on previous findings\n      - Connect new information to what's already known\n    </ContextAwareness>\n\n    <ReportingGuidelines>\n      Your report should include:\n\n      1. **Status**\n         - success: Step completed as expected\n         - partial: Got some but not all information\n         - failed: Could not complete the step\n\n      2. **Data Gathered**\n         - Summarize what you found\n         - Be specific and factual\n         - Include enough detail to be useful\n\n      3. **Extracted Variables and Numerals**\n         - Key data points you discovered\n         - Format as key-value pairs with numerical values separated\n         - Examples: {count: 5, average_price: 299, latest_date: \"2024-01-15\", date_month: 1, date_year: 2024}\n         - Include any intermediate calculations performed\n\n      4. **Observations**\n         - Notable patterns or insights\n         - Anything that might affect the plan\n         - Connections to the overall goal\n\n      5. **Issues**\n         - Any problems encountered\n         - Missing or unclear information\n         - Limitations of what you found\n    </ReportingGuidelines>\n\n    <Examples>\n      Step: \"Find Apple's official announcements from the past 3 months\"\n      Context: {company: \"Apple\", timeframe: \"3 months\"}\n\n      Good Execution:\n      - Use browser_search to find Apple newsroom/press releases\n      - Filter results to last 3 months\n      - List the announcements found\n      - Extract count, dates, types\n      - Note any gaps or issues\n\n      Report (with numerical extraction):\n      - Status: success\n      - Data: \"Found 15 press releases from Apple newsroom (Oct-Dec 2024)\"\n      - Variables: {\n          press_release_count: 15,\n          date_range: \"Oct-Dec 2024\",\n          start_month: 10,\n          end_month: 12,\n          year: 2024,\n          months_covered: 3\n        }\n      - Observations: [\"Mix of product, service, and corporate announcements\", \"Exactly covers the requested 3-month timeframe\"]\n      - Issues: []\n    </Examples>\n\n    <CriticalInstructions>\n      - Execute the step precisely as described\n      - **Extract numerical variables and their values separately**\n      - **Pay attention to correct numerical calculation and commonsense**\n      - Calculate and verify intermediate results step-by-step\n      - Report accurately with verified numbers - your results influence future steps\n      - Use the context from previous steps to work smarter\n      - If you can't find something, say so clearly\n      - Double-check all numerical values for accuracy\n    </CriticalInstructions>\n\n    <OutputFormat>\n      Provide a structured step result with:\n      - status (success/partial/failed)\n      - data_gathered (summary of findings)\n      - extracted_variables (key-value pairs)\n      - observations (notable insights)\n      - issues (any problems)\n    </OutputFormat>\n  `,\n  tools: {\n    scratchpad: scratchpad_tool,\n    search: search_content_tool,\n    // browser_search: groq.tools.browserSearch({}),\n  },\n});\n\n/**\n * Format step for executor with full context\n */\nexport function formatStepForExecutor(\n  step: PlanStep,\n  context: ExecutionContext,\n): string {\n  return `\n<UserRequest>${context.original_request}</UserRequest>\n\n<Understanding>${context.understanding}</Understanding>\n\n<CurrentVariables>\n${JSON.stringify(context.variables, null, 2)}\n</CurrentVariables>\n\n<PreviousSteps>\n${context.step_results.length ? context.step_results.map((result, i) => `Step ${i + 1}: ${result}`).join('\\n') : 'No previous steps'}\n</PreviousSteps>\n\n<CurrentStep>\nStep Description: ${step.description}\nExpected Outcome: ${step.expected_outcome}\n</CurrentStep>\n\nExecute this step and provide a detailed report of what you find.\n  `.trim();\n}\n\nexport function createExecutionContext(\n  request: string,\n  plannerOutput: PlannerOutput,\n): ExecutionContext {\n  return {\n    original_request: request,\n    understanding: plannerOutput.understanding,\n    initial_plan: plannerOutput.steps,\n    current_plan: [...plannerOutput.steps],\n    variables: { ...plannerOutput.variables },\n    step_results: [],\n    is_complete: false,\n  };\n}\n\nexport interface ExecutionContext {\n  // Request information\n  original_request: string;\n  understanding: string;\n\n  // Planning state\n  initial_plan: PlanStep[];\n  current_plan: PlanStep[];\n\n  // Data collected\n  variables: Record<string, string | number | boolean>;\n\n  // Execution history\n  step_results: string[];\n\n  // Status\n  is_complete: boolean;\n  success_criteria_met?: boolean;\n}\n"],
+  "mappings": ";AAAA,SAAS,6BAAAA,4BAA2B,qBAAAC,0BAAyB;AAC7D,OAAO,OAAO;AAEd,SAAS,gBAAgB;AACzB;AAAA,EACE;AAAA,EACA;AAAA,EAEA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;;;ACZP,SAAS,YAAY;AACrB,SAAS,2BAA2B,yBAAyB;AAE7D,SAAS,aAAa;AACtB,SAAS,iBAAiB,2BAA2B;AAe9C,IAAM,gBAAgB,MAAM;AAAA,EACjC,MAAM;AAAA,EACN,OAAO,kBAAkB;AAAA,IACvB,OAAO,KAAK,oBAAoB;AAAA,IAChC,YAAY,0BAA0B;AAAA,MACpC,UAAU,EAAE,aAAa,EAAE;AAAA,IAC7B,CAAC;AAAA,EACH,CAAC;AAAA,EACD,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EA0IR,OAAO;AAAA,IACL,YAAY;AAAA,IACZ,QAAQ;AAAA;AAAA,EAEV;AACF,CAAC;;;ADzJD,IAAM,iBAAiB,EAAE,OAAO;AAAA,EAC9B,aAAa,EACV,OAAO,EACP,SAAS,4CAA4C;AAAA,EACxD,kBAAkB,EACf,OAAO,EACP,SAAS,0CAA0C;AACxD,CAAC;AACM,IAAM,uBAAuB,EAAE,OAAO;AAAA,EAC3C,WAAW,EAAE,OAAO,EAAE,SAAS,uCAAuC;AAAA,EACtE,iBAAiB,EAAE,QAAQ,EAAE,SAAS,+BAA+B;AAAA,EACrE,iBAAiB,EACd,MAAM,cAAc,EACpB,IAAI,CAAC,EACL,SAAS,iCAAiC;AAAA,EAC7C,cAAc,EACX,OAAO,EACP,SAAS,6CAA6C;AAC3D,CAAC;AAeD,IAAM,iBAAiBC,mBAAkB;AAAA;AAAA,EAEvC,OAAO,SAAS,oBAAoB;AAAA,EACpC,YAAYC,2BAA0B;AAAA,IACpC,UAAU,EAAE,aAAa,IAAI;AAAA,EAC/B,CAAC;AACH,CAAC;AAED,SAAS,wBAAgC;AACvC,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAwJT;AAEA,SAAS,sBACP,SACA,gBACQ;AACR,QAAM,mBAAmB,QAAQ,aAC9B,IAAI,CAAC,GAAG,MAAM,QAAQ,CAAC,KAAK,CAAC,EAAE,EAC/B,KAAK,IAAI;AAEZ,SAAO;AAAA,mBACU,QAAQ,gBAAgB;AAAA;AAAA,iBAE1B,QAAQ,aAAa;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOpC,gBAAgB;AAAA;AAAA;AAAA;AAAA,EAIhB,QAAQ,aAAa,GAAG,EAAE,KAAK,uBAAuB;AAAA;AAAA;AAAA;AAAA,EAItD,KAAK,UAAU,QAAQ,WAAW,MAAM,CAAC,CAAC;AAAA;AAAA;AAAA;AAAA,EAI1C,eAAe,SAAS,IAAI,eAAe,IAAI,CAAC,GAAG,MAAM,GAAG,CAAC,KAAK,EAAE,WAAW,eAAe,EAAE,gBAAgB,GAAG,EAAE,KAAK,IAAI,IAAI,oBAAoB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAepJ,KAAK;AACT;AAEA,eAAsB,OAAO,SAA2B;AACtD,QAAM,iBAAiB,QAAQ,aAAa,MAAM,CAAC;AACnD,QAAM,iBAAiB,IAAI,cAAc;AAAA,IACvC,OAAO,IAAI,qBAAqB;AAAA,IAChC,QAAQ,aAAa,OAAO,WAAW,CAAC;AAAA,IACxC,QAAQ;AAAA,EACV,CAAC;AAED,iBAAe;AAAA,IACb,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,MAAM;AAAA,MACN,WACE;AAAA,IACJ,CAAC;AAAA,IACD,SAAS,0BAA0B,sBAAsB,CAAC;AAAA,IAC1D,KAAK,sBAAsB,SAAS,cAAc,CAAC;AAAA,EACrD;AAEA,QAAM,kBAAkB,iBAAiB;AAAA,IACvC,OAAO;AAAA,IACP,SAAS;AAAA,IACT,QAAQ;AAAA,EACV,CAAC;AAED,QAAM,SAAS,MAAM,gBAAgB,SAAS;AAC9C,QAAM,WAAW;AAIjB,SAAO;AAAA,IACL,GAAG;AAAA,IACH,cAAc,SAAS;AAAA,IACvB,aAAa,CAAC,SAAS;AAAA,EACzB;AACF;",
+  "names": ["defaultSettingsMiddleware", "wrapLanguageModel", "wrapLanguageModel", "defaultSettingsMiddleware"]
+}

package/dist/lib/deepplan/synthesizer.js

@@ -0,0 +1,75 @@
+// packages/orchestrator/src/lib/deepplan/synthesizer.ts
+import { agent, execute, lmstudio, user } from "@deepagents/agent";
+var synthesizer = agent({
+  name: "synthesizer_agent",
+  model: lmstudio("google/gemma-3-12b"),
+  prompt: `
+		<SystemContext>
+			You are a synthesizer agent that compiles the results of executed plan steps into a final answer.
+			You ensure accuracy and coherence in the final output.
+		</SystemContext>
+
+		<Identity>
+			Your task is to take the individual step results and synthesize them into a coherent final response that addresses the original user request.
+		</Identity>
+
+		<ConstraintDetection>
+			**CRITICAL: Check the original request for output constraints FIRST**
+		</ConstraintDetection>
+
+		<SynthesisPrinciples>
+			1. **Verify Numerical Accuracy**
+				 - Check all numerical values for consistency
+				 - Ensure calculations are correct across steps
+				 - Verify dates, counts, and measurements make sense
+
+			2. **Apply Commonsense Validation**
+				 - Do the results make logical sense?
+				 - Are there any contradictions or inconsistencies?
+				 - Do temporal relationships align (e.g., dates in proper order)?
+
+			3. **Aggregate Intermediate Results**
+				 - Combine findings from multiple steps coherently
+				 - Preserve important numerical details
+				 - Show clear connections between steps
+
+			4. **Show the Answer Clearly**
+				 - Present the final answer in a structured format
+				 - Highlight key numerical findings
+				 - Make the conclusion explicit and unambiguous
+		</SynthesisPrinciples>
+
+		<Instructions>
+			1. **Read the original user request carefully** and identify any output constraints
+			2. Review all step results carefully
+			3. **Verify numerical consistency and commonsense**
+			4. If constraints exist: compile results according to constraints (minimal output)
+			5. If no constraints: synthesize insights into a comprehensive answer
+			6. **Show the answer clearly** with verified numbers and logical coherence
+
+			CRITICAL:
+			- Always prioritize explicit constraints over comprehensive analysis
+			- When in doubt about constraints, prefer minimal output
+			- Verify numerical accuracy before finalizing
+			- Check for commonsense and logical coherence
+			- The final answer should be well-structured and easy to understand
+		</Instructions>
+	`
+});
+async function synthesize(stepResults, originalRequest) {
+  const stepResultsText = stepResults.map((step, index) => `Step ${index + 1} Result:
+${step}`).join("\n\n");
+  const prompt = `
+Original User Request:
+${originalRequest}
+
+---
+
+${stepResultsText}
+  `.trim();
+  return (await execute(synthesizer, [user(prompt)], {})).text;
+}
+export {
+  synthesize
+};
+//# sourceMappingURL=synthesizer.js.map

package/dist/lib/deepplan/synthesizer.js.map

@@ -0,0 +1,7 @@
+{
+  "version": 3,
+  "sources": ["../../../src/lib/deepplan/synthesizer.ts"],
+  "sourcesContent": ["import { agent, execute, lmstudio, user } from '@deepagents/agent';\n\n/**\n * Synthesizer agent that compiles step results into a final answer\n * Applies numerical accuracy and commonsense validation\n */\nconst synthesizer = agent({\n  name: 'synthesizer_agent',\n  model: lmstudio('google/gemma-3-12b'),\n  prompt: `\n\t\t<SystemContext>\n\t\t\tYou are a synthesizer agent that compiles the results of executed plan steps into a final answer.\n\t\t\tYou ensure accuracy and coherence in the final output.\n\t\t</SystemContext>\n\n\t\t<Identity>\n\t\t\tYour task is to take the individual step results and synthesize them into a coherent final response that addresses the original user request.\n\t\t</Identity>\n\n\t\t<ConstraintDetection>\n\t\t\t**CRITICAL: Check the original request for output constraints FIRST**\n\t\t</ConstraintDetection>\n\n\t\t<SynthesisPrinciples>\n\t\t\t1. **Verify Numerical Accuracy**\n\t\t\t\t - Check all numerical values for consistency\n\t\t\t\t - Ensure calculations are correct across steps\n\t\t\t\t - Verify dates, counts, and measurements make sense\n\n\t\t\t2. **Apply Commonsense Validation**\n\t\t\t\t - Do the results make logical sense?\n\t\t\t\t - Are there any contradictions or inconsistencies?\n\t\t\t\t - Do temporal relationships align (e.g., dates in proper order)?\n\n\t\t\t3. **Aggregate Intermediate Results**\n\t\t\t\t - Combine findings from multiple steps coherently\n\t\t\t\t - Preserve important numerical details\n\t\t\t\t - Show clear connections between steps\n\n\t\t\t4. **Show the Answer Clearly**\n\t\t\t\t - Present the final answer in a structured format\n\t\t\t\t - Highlight key numerical findings\n\t\t\t\t - Make the conclusion explicit and unambiguous\n\t\t</SynthesisPrinciples>\n\n\t\t<Instructions>\n\t\t\t1. **Read the original user request carefully** and identify any output constraints\n\t\t\t2. Review all step results carefully\n\t\t\t3. **Verify numerical consistency and commonsense**\n\t\t\t4. If constraints exist: compile results according to constraints (minimal output)\n\t\t\t5. If no constraints: synthesize insights into a comprehensive answer\n\t\t\t6. **Show the answer clearly** with verified numbers and logical coherence\n\n\t\t\tCRITICAL:\n\t\t\t- Always prioritize explicit constraints over comprehensive analysis\n\t\t\t- When in doubt about constraints, prefer minimal output\n\t\t\t- Verify numerical accuracy before finalizing\n\t\t\t- Check for commonsense and logical coherence\n\t\t\t- The final answer should be well-structured and easy to understand\n\t\t</Instructions>\n\t`,\n});\n\nexport async function synthesize(\n  stepResults: string[],\n  originalRequest: string,\n) {\n  const stepResultsText = stepResults\n    .map((step, index) => `Step ${index + 1} Result:\\n${step}`)\n    .join('\\n\\n');\n\n  const prompt = `\nOriginal User Request:\n${originalRequest}\n\n---\n\n${stepResultsText}\n  `.trim();\n\n  return (await execute(synthesizer, [user(prompt)], {})).text;\n}\n"],
+  "mappings": ";AAAA,SAAS,OAAO,SAAS,UAAU,YAAY;AAM/C,IAAM,cAAc,MAAM;AAAA,EACxB,MAAM;AAAA,EACN,OAAO,SAAS,oBAAoB;AAAA,EACpC,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAoDV,CAAC;AAED,eAAsB,WACpB,aACA,iBACA;AACA,QAAM,kBAAkB,YACrB,IAAI,CAAC,MAAM,UAAU,QAAQ,QAAQ,CAAC;AAAA,EAAa,IAAI,EAAE,EACzD,KAAK,MAAM;AAEd,QAAM,SAAS;AAAA;AAAA,EAEf,eAAe;AAAA;AAAA;AAAA;AAAA,EAIf,eAAe;AAAA,IACb,KAAK;AAEP,UAAQ,MAAM,QAAQ,aAAa,CAAC,KAAK,MAAM,CAAC,GAAG,CAAC,CAAC,GAAG;AAC1D;",
+  "names": []
+}