easyoref 1.13.1 → 1.14.0

package/dist/agent/graph.js

@@ -1,47 +1,32 @@
 /**
- * LangGraph.js enrichment pipeline — tiered validation + tool calling.
+ * LangGraph.js enrichment pipeline — phase-aware, time-validated.
  *
- * Design: minimize tokens, maximize confidence.
- * When confidence is low, offer tools — agent decides if they help.
+ * KEY DESIGN PRINCIPLES:
+ *   1. TIME IS KING — every post is validated against the alert time window.
+ *      LLM receives alert time + post time and scores time_relevance.
+ *      Posts about previous/different attacks are rejected.
+ *   2. PHASE-AWARE — each phase extracts only what's relevant:
+ *      - early_warning: origin, ETA, rocket count, cassette
+ *      - siren: carries early data + adds interception, impacts
+ *      - resolved: carries all + adds casualties, injuries, final stats
+ *   3. CARRY-FORWARD — results persist in Redis (EnrichmentData).
+ *      Each phase inherits previous phase's findings.
+ *   4. INLINE CITATIONS — no superscripts, no footer sources.
+ *      Format: [[1]](url) right after each data point.
+ *   5. DEDUP EDITS — hash-based check prevents "message not modified" spam.
  *
- * ┌──────────────────────────────────────────────────────────────┐
- * │ Tier 0: preFilter       (deterministic, 0 tokens)           │
- * │   → keyword + region check on raw post text                 │
- * │                                                              │
- * │ Tier 1: extractAndValidate  (1 LLM call per post)           │
- * │   → combined extraction + 3 validators in single JSON       │
- * │                                                              │
- * │ Tier 2: postFilter      (deterministic, 0 tokens)           │
- * │   → reject low relevance / trust / alarmist / empty         │
- * │                                                              │
- * │ Tier 3: vote            (deterministic, 0 tokens)           │
- * │   → majority consensus across validated sources             │
- * │                                                              │
- * │ Tier 3.5: shouldClarify (conditional edge)                  │
- * │   → if confidence < threshold AND tools enabled:            │
- * │     → clarify: LLM sees voted result + 4 tools              │
- * │       • read_telegram_sources (1-4 channel posts)           │
- * │       • alert_history (Oref history verification)           │
- * │       • resolve_area (defense-zone proximity check)         │
- * │       • betterstack_log (query recent pipeline logs)        │
- * │       LLM decides: call 0, 1, 2, or 3+ tools.              │
- * │     → revote with merged extractions                        │
- * │   → else: proceed to editMessage                            │
- * │                                                              │
- * │ Tier 4: editMessage     (deterministic, 0 tokens)           │
- * │   → inline update of existing key:value pairs               │
- * └──────────────────────────────────────────────────────────────┘
- *
- * Checkpointer: MemorySaver — session-level state persistence.
- * Total LLM cost: 1 call × N posts + (optional) 1 clarify call + 0-N tools.
+ * Pipeline:
+ *   preFilter → extractAndValidate → postFilter → vote → [clarify] → editMessage
  */
 import { Annotation, MemorySaver, StateGraph } from "@langchain/langgraph";
 import { ChatOpenAI } from "@langchain/openai";
 import { Bot } from "grammy";
+import { createHash } from "node:crypto";
 import { config } from "../config.js";
 import * as logger from "../logger.js";
 import { runClarify } from "./clarify.js";
-import { getChannelPosts } from "./store.js";
+import { getActiveSession, getChannelPosts, getEnrichmentData, saveEnrichmentData, } from "./store.js";
+import { emptyEnrichmentData } from "./types.js";
 // ── State ──────────────────────────────────────────────
 const AgentState = Annotation.Root({
     alertId: Annotation({ reducer: (_, b) => b }),
@@ -58,6 +43,12 @@ const AgentState = Annotation.Root({
     votedResult: Annotation({ reducer: (_, b) => b }),
     /** Tracks whether clarify has already run (prevents infinite loop) */
     clarifyAttempted: Annotation({ reducer: (_, b) => b }),
+    /** Cross-phase enrichment data loaded at start */
+    previousEnrichment: Annotation({ reducer: (_, b) => b }),
+    /** Session start timestamp for time window calculations */
+    sessionStartTs: Annotation({ reducer: (_, b) => b }),
+    /** Phase start timestamp */
+    phaseStartTs: Annotation({ reducer: (_, b) => b }),
 });
 // ── LLM ───────────────────────────────────────────────
 function getLLM() {
@@ -72,85 +63,220 @@ function getLLM() {
         },
         apiKey: config.agent.apiKey,
         temperature: 0,
-        maxTokens: 400,
+        maxTokens: 500,
     });
 }
 // ── Region keywords (Hebrew + transliterations) ────────
-/**
- * Build keyword list from config areas + area_labels.
- * Returns lowercased keywords for matching.
- */
 function buildRegionKeywords() {
     const keywords = [];
     for (const area of config.areas) {
         keywords.push(area.toLowerCase());
-        // First word often enough (e.g. "תל אביב" → "תל")
         const first = area.split(" ")[0];
         if (first && first.length >= 2)
             keywords.push(first.toLowerCase());
     }
     for (const [he, label] of Object.entries(config.agent.areaLabels)) {
         keywords.push(he.toLowerCase());
-        // Add transliterated label words (e.g. "Дан центр" → "дан", "центр")
         for (const word of label.split(/\s+/)) {
             if (word.length >= 3)
                 keywords.push(word.toLowerCase());
         }
     }
     // Common attack-related keywords (always relevant)
-    keywords.push("ישראל", "israel", "израиль", "ракет", "rocket", "missile", "iron dome", "כיפת ברזל", "жд", "перехват", "intercept", "siren", "азака", "צבע אדום", "red alert");
+    keywords.push("ישראל", "israel", "израиль", "ракет", "rocket", "missile", "iron dome", "כיפת ברזל", "перехват", "intercept", "צבע אדום", "red alert");
     return [...new Set(keywords)];
 }
+// ── Launch detection keywords (strict — early_warning only) ──
+const LAUNCH_KEYWORDS = [
+    "שיגור",
+    "שיגורים",
+    "שוגרו",
+    "נורו",
+    "зафиксированы запуски",
+    "обнаружены запуски",
+    "запуски ракет",
+    "запуск ракет",
+    "пуски ракет",
+    "ракетный обстрел",
+    "ракетная атака",
+    "missile launch",
+    "rocket launch",
+    "barrage",
+    "fired towards",
+    "launches detected",
+    "missiles fired",
+    "שיגורים לישראל",
+    "ירי טילים",
+    "ירי רקטות",
+    "إطلاق صواريخ",
+].map((kw) => kw.toLowerCase());
+// ── Time window per phase (ms before alertTs to accept posts) ──
+const TIME_WINDOW_MS = {
+    early_warning: 5 * 60 * 1000, // 5 min before alert
+    siren: 10 * 60 * 1000, // 10 min (includes early_warning period)
+    resolved: 30 * 60 * 1000, // 30 min (full session window)
+};
+// ── Helpers ────────────────────────────────────────────
+/** Format timestamp as HH:MM Israel time */
+function toIsraelTime(ts) {
+    return new Date(ts).toLocaleTimeString("he-IL", {
+        hour: "2-digit",
+        minute: "2-digit",
+        timeZone: "Asia/Jerusalem",
+    });
+}
+/** MD5 hash for edit dedup */
+function textHash(text) {
+    return createHash("md5").update(text).digest("hex");
+}
 // ─────────────────────────────────────────────────────────
-// Tier 0: Pre-filter (deterministic, 0 tokens)
+// Tier 0: Pre-filter (phase-aware, time-bounded, 0 tokens)
 // ─────────────────────────────────────────────────────────
 async function collectAndPreFilter(state) {
-    // Session-scoped: all posts belong to the current session already
     const posts = await getChannelPosts(state.alertId);
+    const prevEnrichment = await getEnrichmentData();
+    // Load session for time boundaries
+    const session = await getActiveSession();
+    const sessionStartTs = session?.sessionStartTs ?? state.alertTs;
+    const phaseStartTs = session?.phaseStartTs ?? state.alertTs;
     if (posts.length === 0) {
         logger.info("Agent: no posts in session", { alertId: state.alertId });
-        return { channelPosts: posts, filteredPosts: [] };
+        return {
+            channelPosts: posts,
+            filteredPosts: [],
+            previousEnrichment: prevEnrichment,
+            sessionStartTs,
+            phaseStartTs,
+        };
     }
     const keywords = buildRegionKeywords();
-    const filtered = posts.filter((post) => {
-        const text = post.text.toLowerCase();
-        // Must contain at least 1 region/attack keyword
-        return keywords.some((kw) => text.includes(kw));
-    });
-    logger.info("Agent: pre-filter", {
-        alertId: state.alertId,
-        total: posts.length,
-        after_keyword_filter: filtered.length,
-    });
-    return { channelPosts: posts, filteredPosts: filtered };
+    const alertType = state.alertType;
+    const alertTs = state.alertTs;
+    // Time window: reject posts older than window before alertTs
+    const windowMs = TIME_WINDOW_MS[alertType];
+    const cutoffTs = alertTs - windowMs;
+    let filtered;
+    if (alertType === "early_warning") {
+        // ── STRICT launch-only filter for early warning ──
+        // Step 1: Find posts with launch keywords, within time window
+        const launchPosts = posts.filter((post) => {
+            if (post.ts < cutoffTs)
+                return false;
+            const text = post.text.toLowerCase();
+            return LAUNCH_KEYWORDS.some((kw) => text.includes(kw));
+        });
+        // Step 2: Get channels that posted about launches
+        const channelFirstLaunchTs = new Map();
+        for (const post of launchPosts) {
+            const current = channelFirstLaunchTs.get(post.channel);
+            if (current === undefined || post.ts < current) {
+                channelFirstLaunchTs.set(post.channel, post.ts);
+            }
+        }
+        // Step 3: Accept follow-up posts from launch channels only (within window)
+        filtered = posts.filter((post) => {
+            if (post.ts < cutoffTs)
+                return false;
+            const text = post.text.toLowerCase();
+            if (!keywords.some((kw) => text.includes(kw)))
+                return false;
+            const firstLaunch = channelFirstLaunchTs.get(post.channel);
+            if (firstLaunch === undefined)
+                return false;
+            return post.ts >= firstLaunch;
+        });
+        logger.info("Agent: pre-filter (early_warning)", {
+            alertId: state.alertId,
+            total: posts.length,
+            launch_posts: launchPosts.length,
+            launch_channels: channelFirstLaunchTs.size,
+            after_filter: filtered.length,
+            cutoff: toIsraelTime(cutoffTs),
+        });
+    }
+    else {
+        // ── Siren & Resolved: broader filter, time-bounded ──
+        filtered = posts.filter((post) => {
+            if (post.ts < cutoffTs)
+                return false;
+            const text = post.text.toLowerCase();
+            return keywords.some((kw) => text.includes(kw));
+        });
+        logger.info("Agent: pre-filter", {
+            alertId: state.alertId,
+            alertType,
+            total: posts.length,
+            after_filter: filtered.length,
+            cutoff: toIsraelTime(cutoffTs),
+        });
+    }
+    return {
+        channelPosts: posts,
+        filteredPosts: filtered,
+        previousEnrichment: prevEnrichment,
+        sessionStartTs,
+        phaseStartTs,
+    };
 }
 // ─────────────────────────────────────────────────────────
 // Tier 1: Extract + validate (1 LLM call per post)
+// Phase-aware prompts — agent knows what to look for.
+// TIME CONTEXT — agent sees alert time + post time.
 // ─────────────────────────────────────────────────────────
 const QUAL_VALUES = '"all"|"most"|"many"|"few"|"exists"|"none"|"more_than"|"less_than"';
-const SYSTEM_PROMPT = `You analyze Telegram channel messages about a missile/rocket attack on Israel.
-Your job: extract factual data AND assess message quality. Be concise.
+/** Phase-specific extraction instructions */
+function getPhaseInstructions(alertType) {
+    switch (alertType) {
+        case "early_warning":
+            return `PHASE: EARLY WARNING (radar detected launches, sirens not yet).
+Focus on: country_origin (WHERE were rockets launched from?), eta_refined_minutes, rocket_count, is_cassette.
+Do NOT extract: intercepted, sea_impact, open_area_impact, hits_confirmed, casualties, injuries — these are IMPOSSIBLE at this stage.
+If a message discusses interception results, it is about a PREVIOUS attack — set time_relevance=0.`;
+        case "siren":
+            return `PHASE: SIREN (rockets incoming, impact imminent).
+Focus on: country_origin (if not known yet), rocket_count, intercepted, sea_impact, open_area_impact, is_cassette.
+Do NOT extract: hits_confirmed, casualties, injuries — too early for confirmed damage reports.
+If a message discusses casualties or confirmed hits, verify the timing carefully - it may be about a previous attack.`;
+        case "resolved":
+            return `PHASE: RESOLVED (incident over, assessing damage).
+Focus on: intercepted (final count), hits_confirmed, casualties, injuries, open_area_impact.
+All fields are valid at this stage. Prioritize confirmed official reports.`;
+    }
+}
+const SYSTEM_PROMPT_BASE = `You analyze Telegram channel messages about a missile/rocket attack on Israel.
+Your job: extract factual data, assess quality, AND validate temporal relevance.
+
+CRITICAL — TIME VALIDATION:
+You will receive the alert time and the post time. You MUST determine if this post
+is about the CURRENT attack or about a previous/different event.
+- If post discusses events clearly BEFORE the alert time → time_relevance=0
+- If post is generic military news not specific to this attack → time_relevance=0.2
+- If post discusses the current attack → time_relevance=1.0
+- If uncertain → time_relevance=0.5 (the system will use alert_history to verify)
 
 Return ONLY valid JSON (no markdown, no explanation):
 {
   "region_relevance": float,       // 0–1: does this message discuss the specified alert region?
   "source_trust": float,           // 0–1: factual reporting (1.0) vs unverified rumors/panic (0.0)
-  "tone": "calm"|"neutral"|"alarmist",  // message tone — reject alarmist content
+  "tone": "calm"|"neutral"|"alarmist",
+  "time_relevance": float,         // 0–1: is this post about the CURRENT attack? (see rules above)
   "country_origin": string|null,   // "Iran","Yemen","Lebanon","Gaza","Iraq","Syria" or null
-  "rocket_count": int|null,        // total rockets/missiles launched if mentioned
-  "is_cassette": bool|null,        // cluster/cassette munitions confirmed?
-  "intercepted": int|null,         // exact number intercepted by Iron Dome/air defense
-  "intercepted_qual": ${QUAL_VALUES}|null, // qualitative if no exact number; null if exact number given
-  "intercepted_qual_num": int|null, // reference number for more_than/less_than (e.g. 5 if "more than 5")
-  "sea_impact": int|null,          // exact number fell in sea/unpopulated area
+  "rocket_count": int|null,
+  "is_cassette": bool|null,
+  "intercepted": int|null,
+  "intercepted_qual": ${QUAL_VALUES}|null,
+  "intercepted_qual_num": int|null,
+  "sea_impact": int|null,
   "sea_impact_qual": ${QUAL_VALUES}|null,
   "sea_impact_qual_num": int|null,
-  "open_area_impact": int|null,    // exact number hit open/populated ground
+  "open_area_impact": int|null,
   "open_area_impact_qual": ${QUAL_VALUES}|null,
   "open_area_impact_qual_num": int|null,
-  "hits_confirmed": int|null,      // confirmed hits on structures/buildings
-  "eta_refined_minutes": int|null, // refined time-to-impact if mentioned
-  "confidence": float              // 0–1: overall confidence in this extraction
+  "hits_confirmed": int|null,
+  "casualties": int|null,
+  "injuries": int|null,
+  "eta_refined_minutes": int|null,
+  "confidence": float
 }
 
 Rules:
@@ -158,12 +284,12 @@ Rules:
 - If message is speculative/unconfirmed rumor, set source_trust < 0.4.
 - If message uses excessive caps, exclamation marks, panic language → tone="alarmist".
 - Only extract concrete numbers explicitly stated in the text. Never guess.
-- intercpted + sea_impact + open_area_impact should sum to rocket_count when all are known.
-- If partial breakdown known, set unknown sub-fields to null (not 0).
-- *_qual fields: use ONLY when the message explicitly states a qualitative descriptor WITHOUT an exact count.
-  If an exact number is given, set *_qual to null. Do NOT infer from absence.
-- NEVER extract qualitative descriptors for casualties or injuries — hits_confirmed handles structural hits only.
-- "none" qual is only valid if explicitly stated in the message (e.g., "все перехвачены", "не упало в море").`;
+- *_qual fields: use ONLY when NO exact count is given. If exact number present, set *_qual=null.
+- "none" qual is only valid if explicitly stated (e.g., "все перехвачены", "не упало в море").
+- For IDF (@idf_telegram) posts about ongoing operations (not this specific attack) → time_relevance=0.
+- LANGUAGE NEUTRALITY: Posts may be in Hebrew, Russian, Arabic, or English. The language of the post
+  MUST NOT affect source_trust or confidence. Russian-language Israeli channels are equally reliable
+  and often break news faster than Hebrew ones. Judge ONLY by factual content and tone.`;
 async function extractAndValidate(state) {
     if (state.filteredPosts.length === 0) {
         logger.info("Agent: no filtered posts to extract", {
@@ -176,31 +302,26 @@ async function extractAndValidate(state) {
     const regionHint = state.alertAreas.length > 0
         ? state.alertAreas.join(", ")
         : Object.keys(config.agent.areaLabels).join(", ") || "Israel";
-    // Format alert time in Israel timezone
-    const alertTimeIL = new Date(state.alertTs).toLocaleTimeString("he-IL", {
-        hour: "2-digit",
-        minute: "2-digit",
-        timeZone: "Asia/Jerusalem",
-    });
-    const nowIL = new Date().toLocaleTimeString("he-IL", {
-        hour: "2-digit",
-        minute: "2-digit",
-        timeZone: "Asia/Jerusalem",
-    });
-    const alertTypeLabel = state.alertType === "early_warning"
-        ? "early warning (radar detection)"
-        : state.alertType === "siren"
-            ? "siren (impact imminent)"
-            : state.alertType;
-    const contextHeader = `Alert type: ${alertTypeLabel}\n` +
-        `Alert time: ${alertTimeIL} (Israel)\n` +
-        `Current time: ${nowIL} (Israel)\n` +
-        `Alert region: ${regionHint}\n` +
-        `UI language: ${config.language}\n`;
+    const alertTimeIL = toIsraelTime(state.alertTs);
+    const nowIL = toIsraelTime(Date.now());
+    const phaseInstructions = getPhaseInstructions(state.alertType);
+    const systemPrompt = SYSTEM_PROMPT_BASE + "\n\n" + phaseInstructions;
     const results = await Promise.all(posts.map(async (post) => {
+        const postTimeIL = toIsraelTime(post.ts);
+        const postAgeMin = Math.round((state.alertTs - post.ts) / 60_000);
+        const postAgeSuffix = postAgeMin > 0
+            ? `(${postAgeMin} min BEFORE alert)`
+            : postAgeMin < 0
+                ? `(${Math.abs(postAgeMin)} min AFTER alert)`
+                : "(same time as alert)";
+        const contextHeader = `Alert time: ${alertTimeIL} (Israel)\n` +
+            `Post time:  ${postTimeIL} (Israel) ${postAgeSuffix}\n` +
+            `Current time: ${nowIL} (Israel)\n` +
+            `Alert region: ${regionHint}\n` +
+            `UI language: ${config.language}\n`;
         try {
             const response = await llm.invoke([
-                { role: "system", content: SYSTEM_PROMPT },
+                { role: "system", content: systemPrompt },
                 {
                     role: "user",
                     content: `${contextHeader}Channel: ${post.channel}\n\nMessage:\n${post.text.slice(0, 800)}`,
@@ -209,7 +330,6 @@ async function extractAndValidate(state) {
             const raw = typeof response.content === "string"
                 ? response.content
                 : JSON.stringify(response.content);
-            // Strip markdown code fences (```json ... ```) that some models wrap around JSON
             const text = raw
                 .replace(/^```(?:json)?\s*\n?/i, "")
                 .replace(/\n?```\s*$/i, "");
@@ -218,6 +338,7 @@ async function extractAndValidate(state) {
                 ...parsed,
                 channel: post.channel,
                 messageUrl: post.messageUrl,
+                time_relevance: parsed.time_relevance ?? 0.5,
                 valid: true,
             };
         }
@@ -231,6 +352,7 @@ async function extractAndValidate(state) {
                 region_relevance: 0,
                 source_trust: 0,
                 tone: "neutral",
+                time_relevance: 0,
                 country_origin: null,
                 rocket_count: null,
                 is_cassette: null,
@@ -244,6 +366,8 @@ async function extractAndValidate(state) {
                 open_area_impact_qual: null,
                 open_area_impact_qual_num: null,
                 hits_confirmed: null,
+                casualties: null,
+                injuries: null,
                 eta_refined_minutes: null,
                 confidence: 0,
                 valid: false,
@@ -254,14 +378,23 @@ async function extractAndValidate(state) {
     logger.info("Agent: extracted", {
         alertId: state.alertId,
         count: results.length,
+        timeRelevance: results.map((r) => ({
+            ch: r.channel,
+            tr: r.time_relevance,
+        })),
     });
     return { extractions: results };
 }
 // ─────────────────────────────────────────────────────────
 // Tier 2: Post-filter (deterministic, 0 tokens)
+// Now includes TIME RELEVANCE check.
 // ─────────────────────────────────────────────────────────
 function postFilter(state) {
     const validated = state.extractions.map((ext) => {
+        // V0: TIME RELEVANCE — the most important check
+        if (ext.time_relevance < 0.5) {
+            return { ...ext, valid: false, reject_reason: "stale_post" };
+        }
         // V1: region relevance
         if (ext.region_relevance < 0.5) {
             return { ...ext, valid: false, reject_reason: "region_irrelevant" };
@@ -270,7 +403,7 @@ function postFilter(state) {
         if (ext.source_trust < 0.4) {
             return { ...ext, valid: false, reject_reason: "untrusted_source" };
         }
-        // V3: tone — reject alarmist (бот для успокоения, не для паники)
+        // V3: tone — reject alarmist
         if (ext.tone === "alarmist") {
             return { ...ext, valid: false, reject_reason: "alarmist_tone" };
         }
@@ -278,7 +411,11 @@ function postFilter(state) {
         const hasData = ext.country_origin !== null ||
             ext.rocket_count !== null ||
             ext.is_cassette !== null ||
+            ext.intercepted !== null ||
+            ext.intercepted_qual !== null ||
             ext.hits_confirmed !== null ||
+            ext.casualties !== null ||
+            ext.injuries !== null ||
             ext.eta_refined_minutes !== null;
         if (!hasData) {
             return { ...ext, valid: false, reject_reason: "no_data" };
@@ -307,20 +444,19 @@ function vote(state) {
     if (valid.length === 0) {
         return { votedResult: null };
     }
-    // Assign 1-based citation indices to valid extractions
+    // Assign 1-based citation indices
     const indexed = valid.map((e, i) => ({ ...e, idx: i + 1 }));
-    // All valid sources become cited sources
     const citedSources = indexed.map((e) => ({
         index: e.idx,
         channel: e.channel,
         messageUrl: e.messageUrl ?? null,
     }));
-    // ETA: highest confidence source that has eta
+    // ETA: highest confidence source
     const withEta = indexed
         .filter((e) => e.eta_refined_minutes !== null)
         .sort((a, b) => b.confidence - a.confidence);
     const bestEta = withEta[0] ?? null;
-    // Country: group unique values, each with their source indices
+    // Country: group unique values
     const countryMap = new Map();
     for (const e of indexed) {
         if (e.country_origin) {
@@ -335,35 +471,19 @@ function vote(state) {
             citations,
         }))
         : null;
-    // Rocket count: range across sources (min … max)
+    // Rocket count: range
     const rocketSrcs = indexed.filter((e) => e.rocket_count !== null);
     const rocketVals = rocketSrcs.map((e) => e.rocket_count);
     const rocket_count_min = rocketVals.length > 0 ? Math.min(...rocketVals) : null;
     const rocket_count_max = rocketVals.length > 0 ? Math.max(...rocketVals) : null;
     const rocket_citations = rocketSrcs.map((e) => e.idx);
-    // Helper: avg weighted confidence for a set of sources
+    // Helper: avg weighted confidence
     function fieldConf(srcs) {
         if (srcs.length === 0)
             return 0;
         return (srcs.reduce((s, e) => s + e.source_trust * e.confidence, 0) / srcs.length);
     }
-    // Cassette: majority
-    const cassSrcs = indexed.filter((e) => e.is_cassette !== null);
-    const cassVals = cassSrcs.map((e) => e.is_cassette);
-    const is_cassette = cassVals.length > 0
-        ? cassVals.filter(Boolean).length > cassVals.length / 2
-        : null;
-    const is_cassette_confidence = fieldConf(cassSrcs);
-    // Hits: median
-    const hitsSrcs = indexed.filter((e) => e.hits_confirmed !== null && e.hits_confirmed > 0);
-    const hitsVals = indexed
-        .filter((e) => e.hits_confirmed !== null)
-        .map((e) => e.hits_confirmed)
-        .sort((a, b) => a - b);
-    const hits_confirmed = hitsVals.length > 0 ? hitsVals[Math.floor(hitsVals.length / 2)] : null;
-    const hits_citations = hitsSrcs.map((e) => e.idx);
-    const hits_confidence = fieldConf(hitsSrcs);
-    // Helper: mode (most frequent non-null value) for QualCount aggregation
+    // Helper: mode for QualCount
     function modeQual(srcs, key) {
         const vals = srcs
             .map((e) => e[key])
@@ -382,7 +502,23 @@ function vote(state) {
             .sort((a, b) => a - b);
         return vals.length > 0 ? vals[Math.floor(vals.length / 2)] : null;
     }
-    // Intercepted: median across sources that reported exact number; mode for qual
+    // Cassette: majority
+    const cassSrcs = indexed.filter((e) => e.is_cassette !== null);
+    const cassVals = cassSrcs.map((e) => e.is_cassette);
+    const is_cassette = cassVals.length > 0
+        ? cassVals.filter(Boolean).length > cassVals.length / 2
+        : null;
+    const is_cassette_confidence = fieldConf(cassSrcs);
+    // Hits: median
+    const hitsSrcs = indexed.filter((e) => e.hits_confirmed !== null && e.hits_confirmed > 0);
+    const hitsVals = indexed
+        .filter((e) => e.hits_confirmed !== null)
+        .map((e) => e.hits_confirmed)
+        .sort((a, b) => a - b);
+    const hits_confirmed = hitsVals.length > 0 ? hitsVals[Math.floor(hitsVals.length / 2)] : null;
+    const hits_citations = hitsSrcs.map((e) => e.idx);
+    const hits_confidence = fieldConf(hitsSrcs);
+    // Intercepted: median / qual
     const interceptedSrcs = indexed.filter((e) => e.intercepted !== null);
     const interceptedQualSrcs = indexed.filter((e) => e.intercepted_qual !== null);
     const interceptedVals = interceptedSrcs
@@ -424,6 +560,26 @@ function vote(state) {
         ? medianQualNum(openQualSrcs, "open_area_impact_qual_num")
         : null;
     const open_area_confidence = fieldConf(openSrcs.length > 0 ? openSrcs : openQualSrcs);
+    // Casualties
+    const casualtySrcs = indexed.filter((e) => e.casualties !== null && e.casualties > 0);
+    const casualtyVals = casualtySrcs
+        .map((e) => e.casualties)
+        .sort((a, b) => a - b);
+    const casualties = casualtyVals.length > 0
+        ? casualtyVals[Math.floor(casualtyVals.length / 2)]
+        : null;
+    const casualties_citations = casualtySrcs.map((e) => e.idx);
+    const casualties_confidence = fieldConf(casualtySrcs);
+    // Injuries
+    const injurySrcs = indexed.filter((e) => e.injuries !== null && e.injuries > 0);
+    const injuryVals = injurySrcs
+        .map((e) => e.injuries)
+        .sort((a, b) => a - b);
+    const injuries = injuryVals.length > 0
+        ? injuryVals[Math.floor(injuryVals.length / 2)]
+        : null;
+    const injuries_citations = injurySrcs.map((e) => e.idx);
+    const injuries_confidence = fieldConf(injurySrcs);
     // Rocket confidence
     const rocket_confidence = fieldConf(rocketSrcs);
     // Overall weighted confidence
@@ -454,6 +610,12 @@ function vote(state) {
         hits_confirmed,
         hits_citations,
         hits_confidence,
+        casualties,
+        casualties_citations,
+        casualties_confidence,
+        injuries,
+        injuries_citations,
+        injuries_confidence,
         confidence: Math.round(weightedConf * 100) / 100,
         sources_count: indexed.length,
         citedSources,
@@ -462,7 +624,7 @@ function vote(state) {
     return { votedResult: voted };
 }
 // ─────────────────────────────────────────────────────────
-// Tier 4: Edit message — inline update (0 tokens)
+// Tier 4: Edit message — inline citations, carry-forward
 // ─────────────────────────────────────────────────────────
 /** EN country name → Russian */
 const COUNTRY_RU = {
@@ -474,174 +636,305 @@ const COUNTRY_RU = {
     Syria: "Сирия",
     Hezbollah: "Хезболла",
 };
-/** Convert index to Unicode superscript string: 1 → ¹, 13 → ¹³ */
-const SUPERSCRIPTS = ["⁰", "¹", "²", "³", "⁴", "⁵", "⁶", "⁷", "⁸", "⁹"];
-function sup(indices) {
-    return indices
-        .map((n) => String(n)
-        .split("")
-        .map((d) => SUPERSCRIPTS[Number(d)])
-        .join(""))
-        .join("");
+/** Format inline citations: [[1]](url), [[2]](url) */
+function inlineCites(indices, citedSources) {
+    const parts = [];
+    for (const idx of indices) {
+        const src = citedSources.find((s) => s.index === idx);
+        if (src?.messageUrl) {
+            parts.push(`<a href="${src.messageUrl}">[${idx}]</a>`);
+        }
+    }
+    return parts.length > 0 ? " " + parts.join(", ") : "";
+}
+/** Get InlineCite[] from citation indices */
+function extractCites(indices, citedSources) {
+    const cites = [];
+    for (const idx of indices) {
+        const src = citedSources.find((s) => s.index === idx);
+        if (src?.messageUrl) {
+            cites.push({ url: src.messageUrl, channel: src.channel });
+        }
+    }
+    return cites;
+}
+/** Format inline citations from InlineCite[] (for carry-forward data) */
+function inlineCitesFromData(cites) {
+    if (cites.length === 0)
+        return "";
+    return (" " + cites.map((c, i) => `<a href="${c.url}">[${i + 1}]</a>`).join(", "));
+}
+// Confidence thresholds
+const SKIP = 0.6;
+const UNCERTAIN = 0.75;
+const CERTAIN = 0.95;
+function qualDisplay(qual, qualNum, conf) {
+    if (qual === null)
+        return null;
+    if (qual === "none")
+        return conf >= CERTAIN ? "нет" : null;
+    const map = {
+        all: "все",
+        most: "большинство",
+        many: "много",
+        few: "несколько",
+        exists: "есть",
+        none: "нет",
+        more_than: qualNum != null ? `>${qualNum}` : ">1",
+        less_than: qualNum != null ? `<${qualNum}` : "<нескольких",
+    };
+    return map[qual];
+}
+function breakdownItem(label, num, qual, qualNum, conf) {
+    if (conf < SKIP)
+        return null;
+    const u = conf < UNCERTAIN ? " (?)" : "";
+    if (num !== null)
+        return `${label} — ${num}${u}`;
+    const qs = qualDisplay(qual, qualNum, conf);
+    if (qs === null)
+        return null;
+    return `${label} — ${qs}${u}`;
 }
 /**
- * Merge enrichment data INTO the existing key:value message.
- * Format:
- *   Подлётное время: ~00:21¹          ← ETA as absolute clock time
- *
- *   Откуда: Иран¹³ + Ливан²           ← blank line before intel block
- *   Ракет: ~5-7
- *   Попадания (Дан центр): 2¹
- *   Время оповещения: 03:47
- *   —
- *   Источники: [1](url) [2](url) [3](url)
+ * Build enrichment data from current vote + previous enrichment (carry-forward).
+ * Returns updated EnrichmentData for Redis persistence.
  */
-function buildEnrichedMessage(currentText, alertType, alertTs, r) {
-    let text = currentText;
-    // Refine ETA in-place (early/siren only)
+function buildEnrichmentFromVote(r, prev, alertType, alertTs) {
+    const data = { ...prev };
+    // Origin — update if voted has it
+    if (r.country_origins && r.country_origins.length > 0) {
+        data.origin = r.country_origins
+            .map((c) => COUNTRY_RU[c.name] ?? c.name)
+            .join(" + ");
+        data.originCites = r.country_origins.flatMap((c) => extractCites(c.citations, r.citedSources));
+    }
+    // ETA — only for early_warning/siren
     if (r.eta_refined_minutes !== null &&
-        r.eta_citations.length > 0 &&
         (alertType === "early_warning" || alertType === "siren")) {
-        text = refineEtaInPlace(text, r.eta_refined_minutes, alertTs, r.eta_citations);
-    }
-    // Insert "Откуда" before time line (with leading blank line for visual separation)
-    if (r.country_origins && r.country_origins.length > 0) {
-        const parts = r.country_origins.map((c) => {
-            const ru = COUNTRY_RU[c.name] ?? c.name;
-            return `${ru}${sup(c.citations)}`;
+        const absTime = new Date(alertTs + r.eta_refined_minutes * 60_000).toLocaleTimeString("he-IL", {
+            hour: "2-digit",
+            minute: "2-digit",
+            timeZone: "Asia/Jerusalem",
         });
-        text = insertBeforeTimeLine(text, `\n<b>Откуда:</b> ${parts.join(" + ")}`);
-    }
-    // Confidence thresholds for uncertainty markers
-    const SKIP = 0.6; // below this → skip field entirely
-    const UNCERTAIN = 0.75; // below this (but ≥ SKIP) → add (?)
-    const CERTAIN = 0.95; // "none" qual requires this level
-    // Convert QualCount to Russian display string.
-    // Returns null if the qual should be suppressed (e.g. "none" below CERTAIN).
-    function qualDisplay(qual, qualNum, conf) {
-        if (qual === null)
-            return null;
-        if (qual === "none")
-            return conf >= CERTAIN ? "нет" : null;
-        const map = {
-            all: "все",
-            most: "большинство",
-            many: "много",
-            few: "несколько",
-            exists: "есть",
-            none: "нет",
-            more_than: qualNum != null ? `>​${qualNum}` : ">​1",
-            less_than: qualNum != null ? `<​${qualNum}` : "<​нескольких",
-        };
-        return map[qual];
+        data.etaAbsolute = `~${absTime}`;
+        data.etaCites = extractCites(r.eta_citations, r.citedSources);
     }
-    // Format one breakdown item: prefer exact number, fall back to qual.
-    // Returns null if nothing to show (below threshold or not reported).
-    function breakdownItem(label, num, qual, qualNum, conf) {
-        if (conf < SKIP)
-            return null;
-        const u = conf < UNCERTAIN ? " (?)" : "";
-        if (num !== null)
-            return `${label} — ${num}${u}`;
-        const qs = qualDisplay(qual, qualNum, conf);
-        if (qs === null)
-            return null;
-        return `${label} — ${qs}${u}`;
-    }
-    // Rocket count with breakdown and uncertainty markers
-    if (r.rocket_count_min !== null &&
-        r.rocket_count_max !== null &&
-        r.rocket_confidence >= SKIP) {
-        const rocketUncertain = r.rocket_confidence < UNCERTAIN ? " (?)" : "";
-        const countStr = r.rocket_count_min === r.rocket_count_max
-            ? `${r.rocket_count_min}`
-            : `~${r.rocket_count_min}–${r.rocket_count_max}`;
-        const bParts = [];
-        const bi = breakdownItem("перехвачено", r.intercepted, r.intercepted_qual, r.intercepted_qual_num, r.intercepted_confidence);
-        if (bi)
-            bParts.push(bi);
-        const bs = breakdownItem("упали в море", r.sea_impact, r.sea_impact_qual, r.sea_impact_qual_num, r.sea_confidence);
-        if (bs)
-            bParts.push(bs);
-        const bo = breakdownItem("открытая местность", r.open_area_impact, r.open_area_impact_qual, r.open_area_impact_qual_num, r.open_area_confidence);
-        if (bo)
-            bParts.push(bo);
-        const breakdown = bParts.length > 0 ? `, из них: ${bParts.join(", ")}` : "";
-        const cassetteU = r.is_cassette_confidence < UNCERTAIN ? " (?)" : "";
-        const cassette = r.is_cassette && r.is_cassette_confidence >= SKIP
-            ? `, есть кассетные${cassetteU}`
-            : "";
-        text = insertBeforeTimeLine(text, `<b>Ракет:</b> ${countStr}${rocketUncertain}${breakdown}${cassette}`);
-    }
-    // Hits: есть прямое попадание/-ия в <area>: N — only if confidence ≥ SKIP
+    // Rocket count
+    if (r.rocket_count_min !== null && r.rocket_count_max !== null) {
+        const u = r.rocket_confidence < UNCERTAIN ? " (?)" : "";
+        data.rocketCount =
+            r.rocket_count_min === r.rocket_count_max
+                ? `${r.rocket_count_min}${u}`
+                : `~${r.rocket_count_min}–${r.rocket_count_max}${u}`;
+        data.rocketCites = extractCites(r.rocket_citations, r.citedSources);
+    }
+    // Cassette
+    if (r.is_cassette !== null && r.is_cassette_confidence >= SKIP) {
+        data.isCassette = r.is_cassette;
+    }
+    // Intercepted
+    if (r.intercepted !== null && r.intercepted_confidence >= SKIP) {
+        const u = r.intercepted_confidence < UNCERTAIN ? " (?)" : "";
+        data.intercepted = `${r.intercepted}${u}`;
+        data.interceptedCites = extractCites(r.citedSources
+            .filter((s) => {
+            const ext = r.citedSources.find((cs) => cs.index === s.index);
+            return ext !== undefined;
+        })
+            .map((s) => s.index), r.citedSources);
+    }
+    else if (r.intercepted_qual !== null && r.intercepted_confidence >= SKIP) {
+        const qs = qualDisplay(r.intercepted_qual, r.intercepted_qual_num, r.intercepted_confidence);
+        if (qs)
+            data.intercepted = qs;
+    }
+    // Hits
     if (r.hits_confirmed !== null &&
         r.hits_confirmed > 0 &&
         r.hits_confidence >= SKIP) {
+        const u = r.hits_confidence < UNCERTAIN ? " (?)" : "";
+        data.hitsConfirmed = `${r.hits_confirmed}${u}`;
+        data.hitsCites = extractCites(r.hits_citations, r.citedSources);
+    }
+    // Casualties
+    if (r.casualties !== null &&
+        r.casualties > 0 &&
+        r.casualties_confidence >= SKIP) {
+        const u = r.casualties_confidence < UNCERTAIN ? " (?)" : "";
+        data.casualties = `${r.casualties}${u}`;
+        data.casualtiesCites = extractCites(r.casualties_citations, r.citedSources);
+    }
+    // Injuries
+    if (r.injuries !== null && r.injuries > 0 && r.injuries_confidence >= SKIP) {
+        const u = r.injuries_confidence < UNCERTAIN ? " (?)" : "";
+        data.injuries = `${r.injuries}${u}`;
+        data.injuriesCites = extractCites(r.injuries_citations, r.citedSources);
+    }
+    // Early warning time — record when first early_warning was received
+    if (alertType === "early_warning" && !data.earlyWarningTime) {
+        data.earlyWarningTime = toIsraelTime(alertTs);
+    }
+    return data;
+}
+/**
+ * Build the enriched message text from current message + enrichment data.
+ * Uses inline [[1]](url) citations. No superscripts. No footer sources.
+ */
+function buildEnrichedMessage(currentText, alertType, alertTs, enrichment) {
+    let text = currentText;
+    // ── Refine ETA in-place ──
+    if (enrichment.etaAbsolute &&
+        (alertType === "early_warning" || alertType === "siren")) {
+        const etaCiteStr = inlineCitesFromData(enrichment.etaCites);
+        const refined = `${enrichment.etaAbsolute}${etaCiteStr}`;
+        const etaPatterns = [
+            /~\d+[–-]\d+\s*мин/, // ~5–12 мин
+            /~\d+[–-]\d+\s*min/, // ~5–12 min
+            /~\d+[–-]\d+\s*דקות/, // ~5–12 דקות
+            /~\d+[–-]\d+\s*دقيقة/, // ~5–12 دقيقة
+            /1\.5\s*мин/, // 1.5 мин (siren)
+            /1\.5\s*min/,
+            /1\.5\s*דקות/,
+            /1\.5\s*دقيقة/,
+        ];
+        for (const pattern of etaPatterns) {
+            if (pattern.test(text)) {
+                text = text.replace(pattern, refined);
+                break;
+            }
+        }
+    }
+    // ── Siren: show "Раннее предупреждение: было в HH:MM" ──
+    if (alertType === "siren" && enrichment.earlyWarningTime) {
+        text = insertBeforeTimeLine(text, `<b>Раннее предупреждение:</b> было в ${enrichment.earlyWarningTime}`);
+    }
+    // ── Origin ──
+    if (enrichment.origin) {
+        const citeStr = inlineCitesFromData(enrichment.originCites);
+        text = insertBeforeTimeLine(text, `\n<b>Откуда:</b> ${enrichment.origin}${citeStr}`);
+    }
+    // ── Rocket count + breakdown ──
+    if (enrichment.rocketCount) {
+        const citeStr = inlineCitesFromData(enrichment.rocketCites);
+        const cassette = enrichment.isCassette ? ", есть кассетные" : "";
+        let breakdown = "";
+        const bParts = [];
+        if (enrichment.intercepted) {
+            bParts.push(`перехвачено — ${enrichment.intercepted}`);
+        }
+        if (enrichment.seaImpact) {
+            bParts.push(`упали в море — ${enrichment.seaImpact}`);
+        }
+        if (enrichment.openAreaImpact) {
+            bParts.push(`открытая местность — ${enrichment.openAreaImpact}`);
+        }
+        if (bParts.length > 0)
+            breakdown = `, из них: ${bParts.join(", ")}`;
+        text = insertBeforeTimeLine(text, `<b>Ракет:</b> ${enrichment.rocketCount}${breakdown}${cassette}${citeStr}`);
+    }
+    else if (enrichment.intercepted && alertType !== "early_warning") {
+        // No rocket count but have interception data
+        const citeStr = inlineCitesFromData(enrichment.interceptedCites);
+        text = insertBeforeTimeLine(text, `<b>Перехвачено:</b> ${enrichment.intercepted}${citeStr}`);
+    }
+    // ── Hits ──
+    if (enrichment.hitsConfirmed && alertType !== "early_warning") {
         const areaLabel = Object.values(config.agent.areaLabels)[0] ?? "район";
-        const hitWord = r.hits_confirmed === 1 ? "попадание" : "попадания";
-        const hitsCite = r.hits_citations.length > 0 ? sup(r.hits_citations) : "";
-        const hitsU = r.hits_confidence < UNCERTAIN ? " (?)" : "";
-        text = insertBeforeTimeLine(text, `есть прямое ${hitWord} в ${areaLabel}: ${r.hits_confirmed}${hitsCite}${hitsU}`);
-    }
-    // Sources footer: [1](url)   [2](url)   ...
-    const sourcesWithUrl = r.citedSources.filter((s) => s.messageUrl);
-    if (sourcesWithUrl.length > 0) {
-        const links = sourcesWithUrl
-            .map((s) => `<a href="${s.messageUrl}">[${s.index}]</a>`)
-            .join("  ");
-        text += `\n—\n<i>Источники: ${links}</i>`;
+        const citeStr = inlineCitesFromData(enrichment.hitsCites);
+        text = insertBeforeTimeLine(text, `<b>Попадания (${areaLabel}):</b> ${enrichment.hitsConfirmed}${citeStr}`);
+    }
+    // ── Casualties / Injuries (resolved only) ──
+    if (enrichment.casualties && alertType === "resolved") {
+        const citeStr = inlineCitesFromData(enrichment.casualtiesCites);
+        text = insertBeforeTimeLine(text, `<b>Погибшие:</b> ${enrichment.casualties}${citeStr}`);
+    }
+    if (enrichment.injuries && alertType === "resolved") {
+        const citeStr = inlineCitesFromData(enrichment.injuriesCites);
+        text = insertBeforeTimeLine(text, `<b>Пострадавшие:</b> ${enrichment.injuries}${citeStr}`);
     }
     return text;
 }
 /**
  * Insert a line before the time line (last "Время" / "Time" / "שעת" line).
- * This keeps new data visually grouped with existing fields.
  */
 function insertBeforeTimeLine(text, line) {
-    // Match "Время оповещения" / "Alert time" / "שעת ההתרעה" / "وقت الإنذار"
     const timePattern = /(<b>(?:Время оповещения|Alert time|שעת ההתרעה|وقت الإنذار):<\/b>)/;
     const match = text.match(timePattern);
     if (match?.index !== undefined) {
         return text.slice(0, match.index) + line + "\n" + text.slice(match.index);
     }
-    // Fallback: append before last line
     const lines = text.split("\n");
     lines.splice(Math.max(lines.length - 1, 0), 0, line);
     return lines.join("\n");
 }
-/**
- * Replace the default ETA range with absolute impact time + superscript citation.
- * "~5–12 мин" → "~00:21¹"
- */
-function refineEtaInPlace(text, minutes, alertTs, citations) {
-    // Compute absolute impact time in Israel timezone
-    const absTime = new Date(alertTs + minutes * 60_000).toLocaleTimeString("he-IL", { hour: "2-digit", minute: "2-digit", timeZone: "Asia/Jerusalem" });
-    const refined = `~${absTime}${sup(citations)}`;
-    const etaPatterns = [
-        /~\d+[–-]\d+\s*мин/, // ~5–12 мин
-        /~\d+[–-]\d+\s*min/, // ~5–12 min
-        /~\d+[–-]\d+\s*דקות/, // ~5–12 דקות
-        /~\d+[–-]\d+\s*دقائق/, // ~5–12 دقائق
-        /1\.5\s*мин/, // 1.5 мин (siren)
-        /1\.5\s*min/, // 1.5 min
-        /1\.5\s*דקות/, // 1.5 דקות
-        /1\.5\s*دقائق/, // 1.5 دقائق
-    ];
-    for (const pattern of etaPatterns) {
-        if (pattern.test(text)) {
-            return text.replace(pattern, refined);
-        }
-    }
-    return text;
-}
 async function editMessage(state) {
     const { votedResult } = state;
     if (!config.botToken)
         return {};
     const tgBot = new Bot(config.botToken);
-    // No valid sources found — silently skip (don't touch the message)
+    // No valid sources — carry forward previous data only
+    const prevEnrichment = state.previousEnrichment ?? emptyEnrichmentData();
     if (!votedResult) {
-        logger.info("Agent: no voted result — skipping edit", {
+        // No new data from channels — still try to build message from carry-forward
+        if (prevEnrichment.origin || prevEnrichment.intercepted) {
+            // Have carry-forward data, build message
+            const newText = buildEnrichedMessage(state.currentText, state.alertType, state.alertTs, prevEnrichment);
+            const hash = textHash(newText);
+            if (hash === prevEnrichment.lastEditHash) {
+                logger.info("Agent: no change in message (dedup) — skipping edit", {
+                    alertId: state.alertId,
+                });
+                return {};
+            }
+            try {
+                if (state.isCaption) {
+                    await tgBot.api.editMessageCaption(state.chatId, state.messageId, {
+                        caption: newText,
+                        parse_mode: "HTML",
+                    });
+                }
+                else {
+                    await tgBot.api.editMessageText(state.chatId, state.messageId, newText, { parse_mode: "HTML" });
+                }
+                prevEnrichment.lastEditHash = hash;
+                await saveEnrichmentData(prevEnrichment);
+                logger.info("Agent: message enriched (carry-forward only)", {
+                    alertId: state.alertId,
+                    messageId: state.messageId,
+                });
+            }
+            catch (err) {
+                const errStr = String(err);
+                if (errStr.includes("message is not modified")) {
+                    prevEnrichment.lastEditHash = hash;
+                    await saveEnrichmentData(prevEnrichment);
+                    logger.info("Agent: message already up-to-date (dedup)", {
+                        alertId: state.alertId,
+                    });
+                }
+                else {
+                    logger.error("Agent: failed to edit message", {
+                        alertId: state.alertId,
+                        error: errStr,
+                    });
+                }
+            }
+        }
+        else {
+            logger.info("Agent: no voted result — skipping edit", {
+                alertId: state.alertId,
+            });
+        }
+        return {};
+    }
+    // Build enrichment data: merge vote + previous
+    const enrichment = buildEnrichmentFromVote(votedResult, prevEnrichment, state.alertType, state.alertTs);
+    const newText = buildEnrichedMessage(state.currentText, state.alertType, state.alertTs, enrichment);
+    // Dedup: skip if text hasn't changed
+    const hash = textHash(newText);
+    if (hash === enrichment.lastEditHash) {
+        logger.info("Agent: no change in message (dedup) — skipping edit", {
             alertId: state.alertId,
         });
         return {};
@@ -654,7 +947,6 @@ async function editMessage(state) {
             threshold: config.agent.confidenceThreshold,
         });
     }
-    const newText = buildEnrichedMessage(state.currentText, state.alertType, state.alertTs, votedResult);
     try {
         if (state.isCaption) {
             await tgBot.api.editMessageCaption(state.chatId, state.messageId, {
@@ -667,18 +959,31 @@ async function editMessage(state) {
                 parse_mode: "HTML",
             });
         }
+        enrichment.lastEditHash = hash;
+        await saveEnrichmentData(enrichment);
         logger.info("Agent: message enriched", {
             alertId: state.alertId,
             messageId: state.messageId,
             confidence: votedResult.confidence,
             sources: votedResult.sources_count,
+            phase: state.alertType,
         });
     }
     catch (err) {
-        logger.error("Agent: failed to edit message", {
-            alertId: state.alertId,
-            error: String(err),
-        });
+        const errStr = String(err);
+        if (errStr.includes("message is not modified")) {
+            enrichment.lastEditHash = hash;
+            await saveEnrichmentData(enrichment);
+            logger.info("Agent: message already up-to-date (dedup)", {
+                alertId: state.alertId,
+            });
+        }
+        else {
+            logger.error("Agent: failed to edit message", {
+                alertId: state.alertId,
+                error: errStr,
+            });
+        }
     }
     return {};
 }
@@ -686,7 +991,7 @@ async function editMessage(state) {
 // Clarify Node — MCP tool calling via ReAct (conditional)
 // ─────────────────────────────────────────────────────────
 async function clarifyNode(state) {
-    const { votedResult, extractions, alertId, alertAreas, alertType, messageId, currentText, } = state;
+    const { votedResult, extractions, alertId, alertAreas, alertType, alertTs, messageId, currentText, } = state;
     if (!votedResult) {
         logger.info("Agent: clarify skipped — no voted result", { alertId });
         return { clarifyAttempted: true };
@@ -695,18 +1000,19 @@ async function clarifyNode(state) {
         alertId,
         confidence: votedResult.confidence,
         threshold: config.agent.confidenceThreshold,
+        phase: alertType,
     });
     try {
         const result = await runClarify({
             alertId,
             alertAreas,
             alertType,
+            alertTs,
             messageId,
             currentText,
             extractions,
             votedResult,
         });
-        // Merge new extractions with existing valid ones
         const mergedExtractions = [...extractions, ...result.newExtractions];
         logger.info("Agent: clarify completed", {
             alertId,
@@ -717,7 +1023,6 @@ async function clarifyNode(state) {
         });
         return {
             extractions: mergedExtractions,
-            // Reset votedResult so vote() re-runs with merged data
             votedResult: null,
             clarifyAttempted: true,
         };
@@ -732,19 +1037,13 @@ async function clarifyNode(state) {
 }
 // ── Conditional routing after vote ─────────────────────
 function shouldClarify(state) {
-    // Only clarify once per pipeline run (prevents infinite loop)
-    if (state.clarifyAttempted) {
+    if (state.clarifyAttempted)
         return "editMessage";
-    }
-    // MCP tools must be enabled
-    if (!config.agent.mcpTools) {
+    if (!config.agent.mcpTools)
         return "editMessage";
-    }
-    // No voted result → nothing to clarify
-    if (!state.votedResult) {
+    if (!state.votedResult)
         return "editMessage";
-    }
-    // Confidence below threshold → clarify
+    // Low confidence → clarify (may use Oref tool for time validation)
     if (state.votedResult.confidence < config.agent.confidenceThreshold) {
         logger.info("Agent: routing to clarify (low confidence)", {
             confidence: state.votedResult.confidence,
@@ -752,10 +1051,27 @@ function shouldClarify(state) {
         });
         return "clarify";
     }
+    // Suspicious time: if the only country is unexpected for the region, verify
+    // This catches cases like "Lebanon" appearing on a Tel Aviv alert
+    // when the real attack is from Iran/Yemen
+    const origins = state.votedResult.country_origins;
+    if (origins &&
+        origins.length === 1 &&
+        state.votedResult.sources_count === 1) {
+        const singleOrigin = origins[0].name;
+        // Lebanon attacks typically don't reach central Israel
+        if (singleOrigin === "Lebanon" &&
+            state.alertAreas.some((a) => a.includes("תל אביב") ||
+                a.includes("גוש דן") ||
+                a.includes("שרון") ||
+                a.includes("מרכז"))) {
+            logger.info("Agent: routing to clarify (suspicious single source: Lebanon for central Israel)", { origin: singleOrigin });
+            return "clarify";
+        }
+    }
     return "editMessage";
 }
 // ── Build graph ────────────────────────────────────────
-/** MemorySaver checkpointer — session-level state persistence */
 const checkpointer = new MemorySaver();
 function buildGraph() {
     const graph = new StateGraph(AgentState)
@@ -764,13 +1080,12 @@ function buildGraph() {
         .addNode("postFilter", postFilter)
         .addNode("vote", vote)
         .addNode("clarify", clarifyNode)
-        .addNode("revote", vote) // Re-run vote after clarify with merged data
+        .addNode("revote", vote)
         .addNode("editMessage", editMessage)
         .addEdge("__start__", "collectAndPreFilter")
         .addEdge("collectAndPreFilter", "extractAndValidate")
         .addEdge("extractAndValidate", "postFilter")
         .addEdge("postFilter", "vote")
-        // Conditional edge: vote → clarify (low conf) or editMessage (high conf)
         .addConditionalEdges("vote", shouldClarify, {
         clarify: "clarify",
         editMessage: "editMessage",
@@ -796,8 +1111,32 @@ export async function runEnrichment(input) {
         extractions: [],
         votedResult: null,
         clarifyAttempted: false,
-    }, 
-    // Thread ID for MemorySaver — enables session-level state persistence
-    { configurable: { thread_id: input.alertId } });
+        previousEnrichment: emptyEnrichmentData(),
+        sessionStartTs: 0,
+        phaseStartTs: 0,
+    }, { configurable: { thread_id: input.alertId } });
 }
+// ── Exported for testing ───────────────────────────────
+export const _test = {
+    getLLM,
+    buildRegionKeywords,
+    LAUNCH_KEYWORDS,
+    TIME_WINDOW_MS,
+    toIsraelTime,
+    textHash,
+    postFilter,
+    vote,
+    buildEnrichmentFromVote,
+    buildEnrichedMessage,
+    insertBeforeTimeLine,
+    inlineCites,
+    inlineCitesFromData,
+    extractCites,
+    COUNTRY_RU,
+    SYSTEM_PROMPT_BASE,
+    getPhaseInstructions,
+    SKIP,
+    UNCERTAIN,
+    CERTAIN,
+};
 //# sourceMappingURL=graph.js.map