@elizaos/agent 2.0.0-beta.1 → 2.0.0-beta.2

package/{dist/packages → packages}/agent/src/runtime/conversation-compactor-runtime.d.ts

@@ -26,7 +26,7 @@
  * here is best-effort and intentionally conservative.
  */
 import type { AgentRuntime } from "@elizaos/core";
-import { type CompactorModelCall, type CompactorTranscript } from "./conversation-compactor.types.ts";
+import { type CompactionStats, type CompactorMessage, type CompactorModelCall, type CompactorTranscript } from "./conversation-compactor.types.ts";
 export declare const STRATEGY_NAMES: readonly ["naive-summary", "structured-state", "hierarchical-summary", "hybrid-ledger"];
 export type StrategyName = (typeof STRATEGY_NAMES)[number];
 /**
@@ -89,6 +89,36 @@ export type ApplyConversationCompactionResult = {
     compactedTokens: number;
     latencyMs: number;
     strategy: StrategyName | null;
+    targetTokens: number;
+    replacementTargetTokens: number;
+    artifact?: {
+        replacementMessageCount: number;
+        stats: CompactionArtifactStats;
+    };
+};
+export type ApplyConversationMessageCompactionArgs = {
+    messages: CompactorMessage[];
+    strategy: StrategyName;
+    /** Current message token count (already estimated by caller when known). */
+    currentTokens: number;
+    /** Token budget the message list must fit into. */
+    targetTokens: number;
+    /** Wraps `runtime.useModel(TEXT_LARGE, ...)`; required for summarizers. */
+    callModel: CompactorModelCall;
+    /** Optional — message-count to preserve verbatim at the tail. */
+    preserveTailMessages?: number;
+};
+export type ApplyConversationMessageCompactionResult = Omit<ApplyConversationCompactionResult, "prompt"> & {
+    messages: CompactorMessage[];
+};
+type CompactionArtifactStats = {
+    originalMessageCount: CompactionStats["originalMessageCount"];
+    compactedMessageCount: CompactionStats["compactedMessageCount"];
+    originalTokens: CompactionStats["originalTokens"];
+    compactedTokens: CompactionStats["compactedTokens"];
+    summarizationModel?: CompactionStats["summarizationModel"];
+    latencyMs: CompactionStats["latencyMs"];
+    extra?: CompactionStats["extra"];
 };
 /**
  * Main entry point for runtime conversation-level compaction.
@@ -99,4 +129,13 @@ export type ApplyConversationCompactionResult = {
  * the original prompt with `didCompact: false`).
  */
 export declare function applyConversationCompaction(args: ApplyConversationCompactionArgs): Promise<ApplyConversationCompactionResult>;
+/**
+ * Message-level companion to `applyConversationCompaction`.
+ *
+ * v5 runtime model calls often pass OpenAI-style `messages` directly instead
+ * of a flattened prompt string. This path avoids lossy prompt parsing and
+ * lets the conversation compactor operate on the role-tagged transcript.
+ */
+export declare function applyConversationMessageCompaction(args: ApplyConversationMessageCompactionArgs): Promise<ApplyConversationMessageCompactionResult>;
+export {};
 //# sourceMappingURL=conversation-compactor-runtime.d.ts.map

package/packages/agent/src/runtime/conversation-compactor-runtime.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"conversation-compactor-runtime.d.ts","sourceRoot":"","sources":["../../../../../src/runtime/conversation-compactor-runtime.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,eAAe,CAAC;AAMlD,OAAO,EAEL,KAAK,eAAe,EACpB,KAAK,gBAAgB,EACrB,KAAK,kBAAkB,EACvB,KAAK,mBAAmB,EAEzB,MAAM,mCAAmC,CAAC;AAE3C,eAAO,MAAM,cAAc,yFAKjB,CAAC;AAEX,MAAM,MAAM,YAAY,GAAG,CAAC,OAAO,cAAc,CAAC,CAAC,MAAM,CAAC,CAAC;AAyB3D;;;;GAIG;AACH,wBAAgB,qBAAqB,IAAI,YAAY,GAAG,IAAI,CAY3D;AAuGD;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAgB,uBAAuB,CAAC,MAAM,EAAE,MAAM,GAAG,mBAAmB,CA8C3E;AAyBD;;;;;;;;;;;;GAYG;AACH,wBAAgB,2BAA2B,CACzC,QAAQ,EAAE,MAAM,EAChB,SAAS,EAAE,mBAAmB,GAC7B,MAAM,CAsCR;AAMD,MAAM,MAAM,+BAA+B,GAAG;IAC5C,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE,YAAY,CAAC;IACvB,oEAAoE;IACpE,aAAa,EAAE,MAAM,CAAC;IACtB,6CAA6C;IAC7C,YAAY,EAAE,MAAM,CAAC;IACrB,2EAA2E;IAC3E,SAAS,EAAE,kBAAkB,CAAC;IAC9B,0DAA0D;IAC1D,OAAO,CAAC,EAAE,YAAY,CAAC;IACvB,iEAAiE;IACjE,oBAAoB,CAAC,EAAE,MAAM,CAAC;CAC/B,CAAC;AAEF,MAAM,MAAM,iCAAiC,GAAG;IAC9C,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,OAAO,CAAC;IACpB,cAAc,EAAE,MAAM,CAAC;IACvB,eAAe,EAAE,MAAM,CAAC;IACxB,SAAS,EAAE,MAAM,CAAC;IAClB,QAAQ,EAAE,YAAY,GAAG,IAAI,CAAC;IAC9B,YAAY,EAAE,MAAM,CAAC;IACrB,uBAAuB,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,EAAE;QACT,uBAAuB,EAAE,MAAM,CAAC;QAChC,KAAK,EAAE,uBAAuB,CAAC;KAChC,CAAC;CACH,CAAC;AAEF,MAAM,MAAM,sCAAsC,GAAG;IACnD,QAAQ,EAAE,gBAAgB,EAAE,CAAC;IAC7B,QAAQ,EAAE,YAAY,CAAC;IACvB,4EAA4E;IAC5E,aAAa,EAAE,MAAM,CAAC;IACtB,mDAAmD;IACnD,YAAY,EAAE,MAAM,CAAC;IACrB,2EAA2E;IAC3E,SAAS,EAAE,kBAAkB,CAAC;IAC9B,iEAAiE;IACjE,oBAAoB,CAAC,EAAE,MAAM,CAAC;CAC/B,CAAC;AAEF,MAAM,MAAM,wCAAwC,GAAG,IAAI,CACzD,iCAAiC,EACjC,QAAQ,CACT,GAAG;IACF,QAAQ,EAAE,gBAAgB,EAAE,CAAC;CAC9B,CAAC;AAEF,KAAK,uBAAuB,GAAG;IAC7B,oBAAoB,EAAE,eAAe,CAAC,sBAAsB,CAAC,CAAC;IAC9D,qBAAqB,EAAE,eAAe,CAAC,uBAAuB,CAAC,CAAC;IAChE,cAAc,EAAE,eAAe,CAAC,gBAAgB,CAAC,CAAC;IAClD,eAAe,EAAE,eAAe,CAAC,iBAAiB,CAAC,CAAC;IACpD,kBAAkB,CAAC,EAAE,eAAe,CAAC,oBAAoB,CAAC,CAAC;IAC3D,SAAS,EAAE,eAAe,CAAC,WAAW,CAAC,CAAC;IACxC,KAAK,CAAC,EAAE,eAAe,CAAC,OAAO,CAAC,CAAC;CAClC,CAAC;AAEF;;;;;;;GAOG;AACH,wBAAsB,2BAA2B,CAC/C,IAAI,EAAE,+BAA+B,GACpC,OAAO,CAAC,iCAAiC,CAAC,CAmH5C;AAED;;;;;;GAMG;AACH,wBAAsB,kCAAkC,CACtD,IAAI,EAAE,sCAAsC,GAC3C,OAAO,CAAC,wCAAwC,CAAC,CAkGnD"}

package/{dist/packages → packages}/agent/src/runtime/conversation-compactor-runtime.js

@@ -25,8 +25,8 @@
  * before string assembly. Until that exists, the parse/serialize step
  * here is best-effort and intentionally conservative.
  */
-import { compactors, naiveSummaryCompactor, } from "./conversation-compactor.js";
-import { approxCountTokens, } from "./conversation-compactor.types.js";
+import { compactors, findSafeCompactionBoundary, naiveSummaryCompactor, } from "./conversation-compactor.js";
+import { approxCountTokens, countTranscriptTokens, } from "./conversation-compactor.types.js";
 export const STRATEGY_NAMES = [
     "naive-summary",
     "structured-state",
@@ -34,10 +34,19 @@ export const STRATEGY_NAMES = [
     "hybrid-ledger",
 ];
 const CONVERSATION_HEADER = "# Conversation Messages";
-const RECEIVED_HEADER_RE = /\n#{1,3}\s*Received Message\b/i;
-const CONVERSATION_LINE_RE = /^(\d{1,2}:\d{2})\s*\(([^)]*)\)\s*(?:\[([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})\])?\s*([^:]+):\s*(.*)$/i;
+const CONVERSATION_HEADER_RE = /^#{1,3}\s*Conversation Messages\b/gim;
+const RECEIVED_HEADER_RE = /\n#{1,3}\s*Received Message\b/gi;
+// Match any of:
+//   "12:53 (17 minutes ago) [uuid] Eliza: text"  ← canonical Eliza recorder
+//   "12:53 (17 minutes ago) Eliza: text"          ← without uuid
+//   "12:53 Eliza: text"                            ← minimal (dev/tests)
+// All four capture groups (time, relative, uuid, name, text) are returned, but
+// only `time` and the trailing `name: text` are required.
+const CONVERSATION_LINE_RE = /^(\d{1,2}:\d{2})(?:\s*\(([^)]*)\))?(?:\s*\[([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})\])?\s*([^:]+?):\s*(.*)$/i;
 const INTERNAL_THOUGHT_RE = /\([^)]*'s internal thought:[^)]*\)/i;
 const ACTIONS_LINE_RE = /\([^)]*'s actions:[^)]*\)/i;
+const USER_SPEAKER_RE = /^(?:user|operator|human|client|customer)$/i;
+const ASSISTANT_SPEAKER_RE = /^(?:eliza|milady|agent|assistant|system|bot|ai)(?:\s*\([^)]*\))?$/i;
 // ---------------------------------------------------------------------------
 // Env config
 // ---------------------------------------------------------------------------
@@ -60,14 +69,19 @@ export function selectStrategyFromEnv() {
         `Expected one of: ${STRATEGY_NAMES.join(", ")}`);
 }
 function locateConversationRegion(prompt) {
-    const start = prompt.indexOf(CONVERSATION_HEADER);
-    if (start === -1)
-        return null;
-    const tailSearch = prompt.slice(start);
-    const tailMatch = tailSearch.match(RECEIVED_HEADER_RE);
-    const endOffset = tailMatch && typeof tailMatch.index === "number"
-        ? start + tailMatch.index
+    const receivedMatches = [...prompt.matchAll(RECEIVED_HEADER_RE)];
+    const lastReceived = receivedMatches.at(-1);
+    const receivedStart = lastReceived && typeof lastReceived.index === "number"
+        ? lastReceived.index
         : prompt.length;
+    const conversationMatches = [...prompt.matchAll(CONVERSATION_HEADER_RE)];
+    const startMatch = conversationMatches
+        .filter((match) => typeof match.index === "number" && match.index < receivedStart)
+        .at(-1);
+    if (!startMatch || typeof startMatch.index !== "number")
+        return null;
+    const start = startMatch.index;
+    const endOffset = receivedStart > start ? receivedStart : prompt.length;
     return {
         prefix: prompt.slice(0, start),
         region: prompt.slice(start, endOffset),
@@ -79,7 +93,11 @@ function parseConversationBody(body) {
     const blocks = [];
     let current = null;
     for (const line of lines) {
-        if (CONVERSATION_LINE_RE.test(line.trim())) {
+        const match = line.trim().match(CONVERSATION_LINE_RE);
+        const speaker = match?.[4]?.trim() ?? "";
+        const canonicalTurn = Boolean(match?.[2] || match?.[3]);
+        const knownSpeaker = USER_SPEAKER_RE.test(speaker) || ASSISTANT_SPEAKER_RE.test(speaker);
+        if (match && (current === null || canonicalTurn || knownSpeaker)) {
             if (current)
                 blocks.push(current);
             current = [line];
@@ -100,13 +118,19 @@ function parseConversationBody(body) {
             continue;
         const [, time, , , name, text] = match;
         const blockText = block.join("\n");
-        const role = INTERNAL_THOUGHT_RE.test(blockText) || ACTIONS_LINE_RE.test(blockText)
+        const normalizedName = name.trim();
+        const isUserSpeaker = USER_SPEAKER_RE.test(normalizedName);
+        const isAssistantSpeaker = ASSISTANT_SPEAKER_RE.test(normalizedName);
+        const role = !isUserSpeaker &&
+            (isAssistantSpeaker ||
+                INTERNAL_THOUGHT_RE.test(blockText) ||
+                ACTIONS_LINE_RE.test(blockText))
             ? "assistant"
             : "user";
         messages.push({
             role,
             raw: blockText,
-            name: name.trim(),
+            name: normalizedName,
             text: text.trim(),
             time,
         });
@@ -139,9 +163,9 @@ export function parsePromptToTranscript(prompt) {
             metadata: { parseFallback: true },
         };
     }
-    // The conversation region always starts with the literal header. Strip
-    // the header line so the body parser doesn't have to special-case it.
-    const headerStripped = region.region.slice(CONVERSATION_HEADER.length);
+    // Strip the section header line so the body parser doesn't have to
+    // special-case it.
+    const headerStripped = region.region.replace(/^#{1,3}\s*Conversation Messages\b[^\n]*(?:\n)?/i, "");
     const parsed = parseConversationBody(headerStripped);
     const messages = [];
     if (region.prefix.trim().length > 0) {
@@ -263,6 +287,8 @@ export async function applyConversationCompaction(args) {
             compactedTokens: originalTokens,
             latencyMs: 0,
             strategy: args.strategy,
+            targetTokens: args.targetTokens,
+            replacementTargetTokens: args.targetTokens,
         };
     }
     const strategyImpl = compactors[args.strategy] ?? naiveSummaryCompactor;
@@ -277,28 +303,29 @@ export async function applyConversationCompaction(args) {
             compactedTokens: originalTokens,
             latencyMs: Date.now() - startedAt,
             strategy: args.strategy,
+            targetTokens: args.targetTokens,
+            replacementTargetTokens: args.targetTokens,
         };
     }
+    const systemOffset = transcript.messages[0]?.role === "system" ? 1 : 0;
+    const hasActiveSuffix = Boolean(locateConversationRegion(args.prompt)?.suffix.trim().length);
+    const preserveTail = Math.max(args.preserveTailMessages ?? 6, hasActiveSuffix ? 1 : 0);
+    const boundary = findSafeCompactionBoundary(transcript.messages, preserveTail);
+    const systemPrefix = systemOffset === 1 ? [transcript.messages[0]] : [];
+    const preservedTail = transcript.messages.slice(boundary);
+    const nonCompactableTokens = approxCountTokens([...systemPrefix, ...preservedTail].map((m) => m.content).join("\n"));
+    const minimumReplacementBudget = Math.min(64, args.targetTokens);
+    const replacementTargetTokens = Math.max(minimumReplacementBudget, Math.min(args.targetTokens, args.targetTokens - nonCompactableTokens - 32));
     const artifact = await strategyImpl.compact(transcript, {
-        targetTokens: args.targetTokens,
+        targetTokens: replacementTargetTokens,
         callModel: args.callModel,
         countTokens: approxCountTokens,
-        ...(args.preserveTailMessages !== undefined
-            ? { preserveTailMessages: args.preserveTailMessages }
-            : {}),
+        preserveTailMessages: preserveTail,
     });
     // Reconstruct a transcript = systemPrefix + replacement + preservedTail.
     // The compactor returned only the replacement; we need to combine with
     // the boundary it computed. Easiest: re-split the original transcript
     // and rebuild here.
-    const systemOffset = transcript.messages[0]?.role === "system" ? 1 : 0;
-    // The compactor uses findSafeCompactionBoundary internally; mirror its
-    // default tail size unless the caller overrode it.
-    const preserveTail = args.preserveTailMessages ?? 6;
-    const total = transcript.messages.length;
-    const naiveBoundary = Math.max(systemOffset, total - preserveTail);
-    const systemPrefix = systemOffset === 1 ? [transcript.messages[0]] : [];
-    const preservedTail = transcript.messages.slice(naiveBoundary);
     const compactedTranscript = {
         messages: [
             ...systemPrefix,
@@ -309,6 +336,22 @@ export async function applyConversationCompaction(args) {
     };
     const compactedPrompt = serializeTranscriptToPrompt(args.prompt, compactedTranscript);
     const compactedTokens = Math.ceil(compactedPrompt.length / 4);
+    if (compactedTokens >= originalTokens) {
+        return {
+            prompt: args.prompt,
+            didCompact: false,
+            originalTokens,
+            compactedTokens: originalTokens,
+            latencyMs: Date.now() - startedAt,
+            strategy: args.strategy,
+            targetTokens: args.targetTokens,
+            replacementTargetTokens,
+            artifact: {
+                replacementMessageCount: artifact.replacementMessages.length,
+                stats: artifact.stats,
+            },
+        };
+    }
     return {
         prompt: compactedPrompt,
         didCompact: compactedPrompt !== args.prompt,
@@ -316,5 +359,102 @@ export async function applyConversationCompaction(args) {
         compactedTokens,
         latencyMs: Date.now() - startedAt,
         strategy: args.strategy,
+        targetTokens: args.targetTokens,
+        replacementTargetTokens,
+        artifact: {
+            replacementMessageCount: artifact.replacementMessages.length,
+            stats: artifact.stats,
+        },
+    };
+}
+/**
+ * Message-level companion to `applyConversationCompaction`.
+ *
+ * v5 runtime model calls often pass OpenAI-style `messages` directly instead
+ * of a flattened prompt string. This path avoids lossy prompt parsing and
+ * lets the conversation compactor operate on the role-tagged transcript.
+ */
+export async function applyConversationMessageCompaction(args) {
+    const startedAt = Date.now();
+    const originalTokens = args.currentTokens;
+    if (args.currentTokens <= args.targetTokens) {
+        return {
+            messages: args.messages,
+            didCompact: false,
+            originalTokens,
+            compactedTokens: originalTokens,
+            latencyMs: 0,
+            strategy: args.strategy,
+            targetTokens: args.targetTokens,
+            replacementTargetTokens: args.targetTokens,
+        };
+    }
+    const strategyImpl = compactors[args.strategy] ?? naiveSummaryCompactor;
+    const transcript = { messages: args.messages };
+    const systemOffset = args.messages[0]?.role === "system" ? 1 : 0;
+    const preserveTail = args.preserveTailMessages ?? 6;
+    const boundary = findSafeCompactionBoundary(args.messages, preserveTail);
+    const systemPrefix = systemOffset === 1 ? [args.messages[0]] : [];
+    const preservedTail = args.messages.slice(boundary);
+    const nonCompactableTokens = approxCountTokens([...systemPrefix, ...preservedTail].map((m) => m.content).join("\n"));
+    const minimumReplacementBudget = Math.min(64, args.targetTokens);
+    const replacementTargetTokens = Math.max(minimumReplacementBudget, Math.min(args.targetTokens, args.targetTokens - nonCompactableTokens - 32));
+    const artifact = await strategyImpl.compact(transcript, {
+        targetTokens: replacementTargetTokens,
+        callModel: args.callModel,
+        countTokens: approxCountTokens,
+        preserveTailMessages: preserveTail,
+    });
+    if (artifact.replacementMessages.length === 0) {
+        return {
+            messages: args.messages,
+            didCompact: false,
+            originalTokens,
+            compactedTokens: originalTokens,
+            latencyMs: Date.now() - startedAt,
+            strategy: args.strategy,
+            targetTokens: args.targetTokens,
+            replacementTargetTokens,
+            artifact: {
+                replacementMessageCount: 0,
+                stats: artifact.stats,
+            },
+        };
+    }
+    const compactedMessages = [
+        ...systemPrefix,
+        ...artifact.replacementMessages,
+        ...preservedTail,
+    ];
+    const compactedTokens = countTranscriptTokens({ messages: compactedMessages }, approxCountTokens);
+    if (compactedTokens >= originalTokens) {
+        return {
+            messages: args.messages,
+            didCompact: false,
+            originalTokens,
+            compactedTokens: originalTokens,
+            latencyMs: Date.now() - startedAt,
+            strategy: args.strategy,
+            targetTokens: args.targetTokens,
+            replacementTargetTokens,
+            artifact: {
+                replacementMessageCount: artifact.replacementMessages.length,
+                stats: artifact.stats,
+            },
+        };
+    }
+    return {
+        messages: compactedMessages,
+        didCompact: true,
+        originalTokens,
+        compactedTokens,
+        latencyMs: Date.now() - startedAt,
+        strategy: args.strategy,
+        targetTokens: args.targetTokens,
+        replacementTargetTokens,
+        artifact: {
+            replacementMessageCount: artifact.replacementMessages.length,
+            stats: artifact.stats,
+        },
     };
 }

package/{dist/packages → packages}/agent/src/runtime/conversation-compactor.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"conversation-compactor.d.ts","sourceRoot":"","sources":["../../../../../src/runtime/conversation-compactor.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,EAGL,KAAK,SAAS,EACd,KAAK,gBAAgB,EAMtB,MAAM,mCAAmC,CAAC;AAQ3C;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,0BAA0B,CACxC,QAAQ,EAAE,gBAAgB,EAAE,EAC5B,oBAAoB,GAAE,MAA8B,GACnD,MAAM,CAkFR;AAiMD,eAAO,MAAM,qBAAqB,EAAE,SAInC,CAAC;AAoPF,eAAO,MAAM,wBAAwB,EAAE,SAItC,CAAC;AAmJF,eAAO,MAAM,4BAA4B,EAAE,SAI1C,CAAC;AA4MF,eAAO,MAAM,qBAAqB,EAAE,SAInC,CAAC;AAMF,eAAO,MAAM,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,SAAS,CAKhD,CAAC"}
+{"version":3,"file":"conversation-compactor.d.ts","sourceRoot":"","sources":["../../../../../src/runtime/conversation-compactor.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,EAGL,KAAK,SAAS,EACd,KAAK,gBAAgB,EAMtB,MAAM,mCAAmC,CAAC;AAQ3C;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,0BAA0B,CACxC,QAAQ,EAAE,gBAAgB,EAAE,EAC5B,oBAAoB,GAAE,MAA8B,GACnD,MAAM,CAqFR;AAgkBD,eAAO,MAAM,qBAAqB,EAAE,SAInC,CAAC;AAiQF,eAAO,MAAM,wBAAwB,EAAE,SAItC,CAAC;AAsJF,eAAO,MAAM,4BAA4B,EAAE,SAI1C,CAAC;AAmOF,eAAO,MAAM,qBAAqB,EAAE,SAInC,CAAC;AAMF,eAAO,MAAM,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,SAAS,CAKhD,CAAC"}