{ "schemaVersion": 2, "name": "deep-researcher", "description": "gpt-researcher-style deep research pipeline WITH live web grounding: sketch the landscape -> plan a sectioned outline -> review/approve the plan (headless auto gate; swap plan_reviewer for a human node for interactive approval) -> parallel per-section research via web_search (fan-out + collect) -> synthesize a cited draft -> fact-check draft claims against freshly retrieved web sources (catches fabricated DOIs/dates) -> publish a final report artifact. Grounding uses web_search's synthesized, cited answers (headless-reliable); control-flow handoffs (plan verdict, plan outline, fact-check verdict) are hardened with response.schema; thread nodes use maxAttempts:1 and rely on the review loops for revision.", "entry": "browser", "initialState": { "control": {}, "working": {}, "memory": {}, "result": {} }, "reducers": { "working.section_research": "collect" }, "nodes": { "browser": { "type": "agent", "description": "Initial broad scout: use web_search to sketch the real, current landscape with live sources.", "prompt": "You are the initial research Scout. Use the web_search tool (pass workflow:\"none\" and 2-3 varied queries) to ground your sketch in REAL, CURRENT sources. Rely on web_search's synthesized answers and the cited URLs it returns — do not call any other tool.\n\nQuery:\n{{input.task}}\n\nProduce a concise landscape sketch: what is known, the main sub-topics, the key contending positions or schools of thought, and the questions a reader would most want answered. Cite each major point with a real URL that web_search returned. Today's date is the current real date — use it.\n\nReturn 200-350 words of dense notes with inline source URLs. Do not write the final report here.", "readOnly": true, "tools": [ "read", "grep", "find", "ls", "web_search" ], "output": "working.sketch", "thinking": "off", "context": { "mode": "isolated" }, "retry": { "maxAttempts": 2, "backoffMs": 800 }, "idempotent": true, "limits": { "timeoutMs": 360000, "maxTurns": 10, "maxCostUsd": 0.6 }, "loadExtensions": true }, "planner": { "type": "agent", "description": "Editor: turn the sketch into a sectioned report outline. Re-runs on plan-review feedback (thread); relies on the plan_reviewer loop rather than node retry for revision.", "prompt": "You are the Editor/Planner. Using the landscape sketch and the query, produce a research-report outline: a concise title and EXACTLY 3 distinct section headers that a balanced report should cover. Do NOT include Introduction, Conclusion, or References as headers (those are written later). Focus only on substantive topical sections.\n\nQuery:\n{{input.task}}\n\nLandscape sketch:\n{{working.sketch}}\n\nPrior plan-review feedback (address it if present):\n{{control.plan_review}}\n\nReturn the outline through the node-output tool with the required shape: a title, exactly 3 section headers, and today's real date.", "readOnly": true, "tools": [], "output": "working.plan", "response": { "maxBytes": 16384, "schema": { "type": "object", "properties": { "title": { "type": "string", "maxLength": 200 }, "sections": { "type": "array", "minItems": 3, "maxItems": 3, "items": { "type": "string", "maxLength": 200 } }, "date": { "type": "string", "maxLength": 40 } }, "required": ["title", "sections", "date"], "additionalProperties": false } }, "thinking": "off", "context": { "mode": "thread", "threadKey": "planner" }, "idempotent": true, "limits": { "timeoutMs": 180000, "maxTurns": 6, "maxCostUsd": 0.3 } }, "plan_reviewer": { "type": "agent", "purpose": "reviewer", "description": "Independently judge whether the outline is balanced and sufficient. Routes back to planner when not.", "prompt": "You are an independent Plan Reviewer. Judge whether this report outline is balanced, covers the query, and is neither redundant nor missing an obvious angle.\n\nQuery:\n{{input.task}}\n\nOutline:\n{{working.plan}}\n\nReturn your verdict through the node-output tool with the required shape: approved (true if the outline is sound) and feedback. Set approved=false only with concrete, actionable, one-line feedback the planner can act on; do not nitpick cosmetics.", "readOnly": true, "tools": [], "output": "control.plan_review", "response": { "maxBytes": 16384, "schema": { "type": "object", "properties": { "approved": { "type": "boolean" }, "feedback": { "type": "string", "maxLength": 800 } }, "required": ["approved", "feedback"], "additionalProperties": false } }, "thinking": "off", "context": { "mode": "isolated" }, "retry": { "maxAttempts": 2, "backoffMs": 800 }, "idempotent": true, "limits": { "timeoutMs": 180000, "maxTurns": 6, "maxCostUsd": 0.3 } }, "section_a": { "type": "agent", "description": "Research the FIRST section of the outline in depth using live web sources.", "prompt": "You are an isolated section Researcher. Investigate ONLY the FIRST section header of the outline below.\n\nUse the web_search tool (pass workflow:\"none\" and 2-3 varied queries specific to this section) to gather CURRENT, REAL evidence. Rely on web_search's synthesized answers and the cited URLs it returns — do not call any other tool. Cite EVERY specific number, trial name, or claim with a real URL that web_search returned — never invent a DOI or URL, and never use placeholder text. If you cannot verify a fact, say so explicitly rather than fabricating a source.\n\nQuery:\n{{input.task}}\n\nOutline:\n{{working.plan}}\n\nLandscape sketch:\n{{working.sketch}}\n\nReturn 300-500 words for this one section only, with inline source URLs.", "readOnly": true, "tools": [ "read", "grep", "find", "ls", "web_search" ], "output": "working.section_research", "thinking": "off", "context": { "mode": "isolated" }, "retry": { "maxAttempts": 2, "backoffMs": 800 }, "idempotent": true, "limits": { "timeoutMs": 420000, "maxTurns": 10, "maxCostUsd": 0.7 }, "loadExtensions": true }, "section_b": { "type": "agent", "description": "Research the SECOND section of the outline in depth using live web sources.", "prompt": "You are an isolated section Researcher. Investigate ONLY the SECOND section header of the outline below.\n\nUse the web_search tool (pass workflow:\"none\" and 2-3 varied queries specific to this section) to gather CURRENT, REAL evidence. Rely on web_search's synthesized answers and the cited URLs it returns — do not call any other tool. Cite EVERY specific number, trial name, or claim with a real URL that web_search returned — never invent a DOI or URL, and never use placeholder text. If you cannot verify a fact, say so explicitly rather than fabricating a source.\n\nQuery:\n{{input.task}}\n\nOutline:\n{{working.plan}}\n\nLandscape sketch:\n{{working.sketch}}\n\nReturn 300-500 words for this one section only, with inline source URLs.", "readOnly": true, "tools": [ "read", "grep", "find", "ls", "web_search" ], "output": "working.section_research", "thinking": "off", "context": { "mode": "isolated" }, "retry": { "maxAttempts": 2, "backoffMs": 800 }, "idempotent": true, "limits": { "timeoutMs": 420000, "maxTurns": 10, "maxCostUsd": 0.7 }, "loadExtensions": true }, "section_c": { "type": "agent", "description": "Research the THIRD section of the outline in depth using live web sources.", "prompt": "You are an isolated section Researcher. Investigate ONLY the THIRD section header of the outline below.\n\nUse the web_search tool (pass workflow:\"none\" and 2-3 varied queries specific to this section) to gather CURRENT, REAL evidence. Rely on web_search's synthesized answers and the cited URLs it returns — do not call any other tool. Cite EVERY specific number, trial name, or claim with a real URL that web_search returned — never invent a DOI or URL, and never use placeholder text. If you cannot verify a fact, say so explicitly rather than fabricating a source.\n\nQuery:\n{{input.task}}\n\nOutline:\n{{working.plan}}\n\nLandscape sketch:\n{{working.sketch}}\n\nReturn 300-500 words for this one section only, with inline source URLs.", "readOnly": true, "tools": [ "read", "grep", "find", "ls", "web_search" ], "output": "working.section_research", "thinking": "off", "context": { "mode": "isolated" }, "retry": { "maxAttempts": 2, "backoffMs": 800 }, "idempotent": true, "limits": { "timeoutMs": 420000, "maxTurns": 10, "maxCostUsd": 0.7 }, "loadExtensions": true }, "writer": { "type": "agent", "description": "Synthesize a complete cited draft from the plan and per-section research. Re-runs on fact-check feedback (thread); relies on the fact_checker loop rather than node retry for revision.", "prompt": "You are the research Writer. Assemble a complete, well-written draft report from the outline and the per-section research. Structure: a title, a short Introduction, the 3 sections (each synthesized from its research, NOT copy-pasted, keeping inline source URLs), and a Conclusion. Synthesize across the three sections; do not just concatenate them.\n\nCRITICAL citation rule: only cite URLs that already appear in the per-section research below. NEVER invent a DOI, URL, or source. If the research lacks a source for a claim, state the claim without a citation or drop it. Using placeholder text like 'xxxxx' is forbidden. Use today's real date.\n\nQuery:\n{{input.task}}\n\nOutline:\n{{working.plan}}\n\nPer-section research:\n{{working.section_research}}\n\nPrior fact-check feedback (fix the flagged issues if present):\n{{working.fact_check}}\n\nReturn clean Markdown only.", "readOnly": true, "tools": [], "output": "working.draft", "response": { "format": "text", "maxBytes": 65536 }, "thinking": "off", "context": { "mode": "thread", "threadKey": "writer" }, "idempotent": true, "limits": { "timeoutMs": 300000, "maxTurns": 6, "maxCostUsd": 0.6 } }, "fact_checker": { "type": "agent", "purpose": "reviewer", "description": "Independently verify draft claims and citations against freshly retrieved web sources; route back to writer on failures.", "prompt": "You are a diligent Fact Checker. Your job is to CATCH fabrication, not rubber-stamp.\n\nMake AT MOST 3 web verifications total: target the 2-3 most doubtful claims or suspicious-looking citations (placeholder/odd DOIs, surprising numbers). Use the web_search tool (pass workflow:\"none\") to confirm or refute each target claim against a real retrieved source — do not call any other tool. Specifically check for and reject: (1) any DOI/URL containing placeholder text like 'xxxxx' or that does not resolve, (2) a wrong report date, (3) specific numbers or trial names that you cannot confirm against a real retrieved source, (4) URLs the writer invented that appear nowhere in the section research. Then stop and return your verdict — do not exhaustively re-check every line.\n\nDraft:\n{{working.draft}}\n\nPer-section research (the only legitimate citation pool):\n{{working.section_research}}\n\nReturn your verdict through the node-output tool with the required shape: approved (true ONLY when the key claims are confirmed against real retrieved sources AND no placeholder/fabricated citations remain) and issues (a list of concrete one-line fixes when approved is false). Do not loop forever on cosmetic style.", "readOnly": true, "tools": [ "read", "grep", "find", "ls", "web_search" ], "output": "working.fact_check", "response": { "maxBytes": 16384, "schema": { "type": "object", "properties": { "approved": { "type": "boolean" }, "issues": { "type": "array", "maxItems": 12, "items": { "type": "string", "maxLength": 400 } } }, "required": ["approved", "issues"], "additionalProperties": false } }, "thinking": "off", "context": { "mode": "isolated" }, "retry": { "maxAttempts": 2, "backoffMs": 800 }, "idempotent": true, "limits": { "timeoutMs": 420000, "maxTurns": 8, "maxCostUsd": 0.7 }, "loadExtensions": true }, "publisher": { "type": "agent", "description": "Produce the final polished report artifact: clean markdown, compiled sources list, and a TL;DR.", "prompt": "You are the Publisher. Produce the final, polished research report from the draft. Add: a one-line TL;DR at the top, clean and consistent Markdown formatting, and a consolidated Sources list at the end compiled from every inline citation in the draft (deduplicated). Only include URLs that actually appear in the draft. If any citation looks like a placeholder (e.g. 'xxxxx') or is obviously broken, omit it rather than ship a dead link.\n\nDraft:\n{{working.draft}}\n\nOutline:\n{{working.plan}}\n\nReturn the final report as Markdown.", "readOnly": true, "tools": [], "output": "result.report_artifact", "response": { "format": "text", "storage": "artifact", "maxBytes": 65536 }, "thinking": "off", "context": { "mode": "isolated" }, "retry": { "maxAttempts": 2, "backoffMs": 800 }, "idempotent": true, "limits": { "timeoutMs": 240000, "maxTurns": 5, "maxCostUsd": 0.5 } } }, "edges": [ { "from": "browser", "to": "planner" }, { "from": "planner", "to": "plan_reviewer" }, { "from": [ "section_a", "section_b", "section_c" ], "to": "writer" }, { "from": "writer", "to": "fact_checker" }, { "from": "publisher", "to": "__end__" }, { "from": "plan_reviewer", "cases": [ { "when": { "path": "control.plan_review.approved", "op": "eq", "value": true }, "to": [ "section_a", "section_b", "section_c" ] } ], "default": "planner" }, { "from": "fact_checker", "cases": [ { "when": { "path": "working.fact_check.approved", "op": "eq", "value": true }, "to": "publisher" } ], "default": "writer" } ], "limits": { "maxSteps": 18, "maxNodeRuns": 32, "maxConcurrency": 3, "maxCostUsd": 6, "timeoutMs": 2400000, "maxStateBytes": 131072, "maxPromptBytes": 32768 }, "result": { "paths": [ "result.report_artifact", "working.fact_check", "control.plan_review" ], "includeState": false, "maxBytes": 12288 }, "policy": { "allowNonInteractive": true, "allowNonInteractiveMutations": true, "confirmMutatingNodes": false } }