Agent halted: routing flag false.
", right() return instructions = "\n".join(payload["routing"]["agent_instructions"]) if payload.get("consensus") and payload["consensus"].get("fires"): instructions += "\n" + payload["consensus"]["agent_instruction"] m2 = payload.get("m2") or {} m5 = payload.get("m5") or {} deltas = m2.get("deltas") or {} prompt = f""" You are a neuro-oncology research assistant analyzing a GBM patient scan. PATIENT DATA: - Progression class (tentative): {m3.get("progression_class")} - M3 confidence: {m3.get("confidence")} (band: {m3.get("confidence_band")}) - Delta pattern flag: {m3.get("delta_pattern_flag")} - Biophysical deltas: delta_mu_d={deltas.get("delta_mu_d")}, delta_mu_r={deltas.get("delta_mu_r")}, delta_gamma={deltas.get("delta_gamma")}, over {deltas.get("delta_t_days")} days - cfDNA result: {m5.get("clinical_subtype")} (confidence {m5.get("detection_confidence")}) - MGMT status: {treatment.get("known_mgmt_status")} - IDH status: {treatment.get("known_idh_status")} - Regimen: {treatment.get("current_regimen")}, {treatment.get("days_since_rt_end")} days post-RT, {treatment.get("tmz_cycles_completed")} TMZ cycles completed AGENT INSTRUCTIONS FROM NEUROSIGHT: {instructions} TASK: Step 1 — Write your initial hypothesis based on the patient data above, before doing any research. Step 2 — Use the search tools to find evidence for or against it. You decide what to search and how many times. Step 3 — State your final hypothesis (revised if needed), confidence level, one alternative you considered and ruled out, and all sources. """ initial_state = { "messages": [HumanMessage(content=prompt)], "task_id": payload.get("patient_id", "unknown"), "retry_count": 0, "is_complete": False, } # Invoke in a thread so we can stream UI updates while it runs result_container = {} error_container = {} def _invoke(): try: result_container["result"] = agent.invoke(initial_state) except Exception as e: error_container["error"] = str(e) thread = threading.Thread(target=_invoke, daemon=True) thread.start() # Show animated states while backend runs tool_stages = [ ("PUBMED", "PubMed", "🔬", "pubmed"), ("INTERPRET", "bioRxiv", "📡", "biorxiv"), ("CLINICALTRIALS","ClinicalTrials", "🏥", "trials"), ("REVISION", "Internal RAG", "📚", "rag"), ] for node_key, tool_label, icon, icon_cls in tool_stages: if not thread.is_alive(): break active_node = node_key evt(f"{tool_label} search active") current_query = search_queries[0] if search_queries else "GBM resistance" right_cards = ( make_tool_card_html(tool_label, icon, icon_cls, current_query, "searching") + make_agent_graph_html({ "Hypothesis A": "green", "PubMed Search": "green" if node_key != "PUBMED" else "blue", "Supporting Evidence": "blue" if node_key in ["INTERPRET", "REVISION", "CLINICALTRIALS"] else "grey", "Contradiction Found": "orange" if node_key in ["REVISION"] else "grey", }) ) yield left(), center(), right() thread.join(timeout=6) # Wait for completion thread.join(timeout=120) # ── Error handling ── if "error" in error_container: err = error_container["error"] evt("⚠ Backend error encountered") warning = f"""{err[:200]}