open-multi-agent-kit 0.98.2 → 0.98.4

package/dist/core/compaction/compaction-headroom.js

@@ -0,0 +1,48 @@
+import { computeReservedTokenBudget } from "../context-budget-reserved-tokens.js";
+export const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;
+export const DEFAULT_COMPACTION_SETTINGS = {
+    enabled: true,
+    reserveTokens: 16384,
+    keepRecentTokens: 20000,
+    maxUsageRatio: DEFAULT_COMPACTION_MAX_USAGE_RATIO,
+};
+/** Return the earliest compaction threshold from ratio-based headroom and absolute reserve. */
+export function getCompactionHeadroomThreshold(contextWindow, settings) {
+    if (!settings.enabled || !Number.isFinite(contextWindow) || contextWindow <= 0) {
+        return undefined;
+    }
+    const windowTokens = Math.floor(contextWindow);
+    const reserveTokens = Number.isFinite(settings.reserveTokens) ? Math.max(0, Math.floor(settings.reserveTokens)) : 0;
+    const configuredMaxUsageRatio = settings.maxUsageRatio ?? DEFAULT_COMPACTION_MAX_USAGE_RATIO;
+    const maxUsageRatio = Number.isFinite(configuredMaxUsageRatio) && configuredMaxUsageRatio > 0 && configuredMaxUsageRatio < 1
+        ? configuredMaxUsageRatio
+        : DEFAULT_COMPACTION_MAX_USAGE_RATIO;
+    const maxUsageRatioTokens = Math.max(1, Math.floor(windowTokens * maxUsageRatio));
+    const reservedBudget = computeReservedTokenBudget({
+        modelContextWindow: windowTokens,
+        systemPromptTokens: 0,
+        reservedOutputTokens: settings.reservedOutputTokens ?? reserveTokens,
+        reservedToolResultTokens: settings.reservedToolResultTokens ?? 0,
+        safetyMarginTokens: settings.safetyMarginTokens ?? 0,
+        imageReserveTokens: settings.imageReserveTokens ?? 0,
+    });
+    const reserveBoundaryTokens = reservedBudget.overflow || reservedBudget.effectiveBudget === 0
+        ? maxUsageRatioTokens
+        : Math.max(1, reservedBudget.effectiveBudget);
+    const triggerTokens = Math.min(maxUsageRatioTokens, reserveBoundaryTokens);
+    return {
+        triggerTokens,
+        headroomTokens: windowTokens - triggerTokens,
+        maxUsageRatioTokens,
+        reserveBoundaryTokens,
+        limitedBy: reserveBoundaryTokens <= maxUsageRatioTokens ? "reserve_tokens" : "max_usage_ratio",
+    };
+}
+/**
+ * Check if compaction should trigger based on context usage.
+ */
+export function shouldCompact(contextTokens, contextWindow, settings) {
+    const threshold = getCompactionHeadroomThreshold(contextWindow, settings);
+    return threshold !== undefined && Number.isFinite(contextTokens) && contextTokens >= threshold.triggerTokens;
+}
+//# sourceMappingURL=compaction-headroom.js.map

package/dist/core/compaction/compaction-headroom.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"compaction-headroom.js","sourceRoot":"","sources":["../../../src/core/compaction/compaction-headroom.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,0BAA0B,EAAE,MAAM,sCAAsC,CAAC;AAgBlF,MAAM,CAAC,MAAM,kCAAkC,GAAG,GAAG,CAAC;AAEtD,MAAM,CAAC,MAAM,2BAA2B,GAAuB;IAC9D,OAAO,EAAE,IAAI;IACb,aAAa,EAAE,KAAK;IACpB,gBAAgB,EAAE,KAAK;IACvB,aAAa,EAAE,kCAAkC;CACjD,CAAC;AAYF,+FAA+F;AAC/F,MAAM,UAAU,8BAA8B,CAC7C,aAAqB,EACrB,QAA4B,EACc;IAC1C,IAAI,CAAC,QAAQ,CAAC,OAAO,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAC,IAAI,aAAa,IAAI,CAAC,EAAE,CAAC;QAChF,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,MAAM,YAAY,GAAG,IAAI,CAAC,KAAK,CAAC,aAAa,CAAC,CAAC;IAC/C,MAAM,aAAa,GAAG,MAAM,CAAC,QAAQ,CAAC,QAAQ,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACpH,MAAM,uBAAuB,GAAG,QAAQ,CAAC,aAAa,IAAI,kCAAkC,CAAC;IAC7F,MAAM,aAAa,GAClB,MAAM,CAAC,QAAQ,CAAC,uBAAuB,CAAC,IAAI,uBAAuB,GAAG,CAAC,IAAI,uBAAuB,GAAG,CAAC;QACrG,CAAC,CAAC,uBAAuB;QACzB,CAAC,CAAC,kCAAkC,CAAC;IACvC,MAAM,mBAAmB,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,YAAY,GAAG,aAAa,CAAC,CAAC,CAAC;IAClF,MAAM,cAAc,GAAG,0BAA0B,CAAC;QACjD,kBAAkB,EAAE,YAAY;QAChC,kBAAkB,EAAE,CAAC;QACrB,oBAAoB,EAAE,QAAQ,CAAC,oBAAoB,IAAI,aAAa;QACpE,wBAAwB,EAAE,QAAQ,CAAC,wBAAwB,IAAI,CAAC;QAChE,kBAAkB,EAAE,QAAQ,CAAC,kBAAkB,IAAI,CAAC;QACpD,kBAAkB,EAAE,QAAQ,CAAC,kBAAkB,IAAI,CAAC;KACpD,CAAC,CAAC;IACH,MAAM,qBAAqB,GAC1B,cAAc,CAAC,QAAQ,IAAI,cAAc,CAAC,eAAe,KAAK,CAAC;QAC9D,CAAC,CAAC,mBAAmB;QACrB,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,cAAc,CAAC,eAAe,CAAC,CAAC;IAChD,MAAM,aAAa,GAAG,IAAI,CAAC,GAAG,CAAC,mBAAmB,EAAE,qBAAqB,CAAC,CAAC;IAE3E,OAAO;QACN,aAAa;QACb,cAAc,EAAE,YAAY,GAAG,aAAa;QAC5C,mBAAmB;QACnB,qBAAqB;QACrB,SAAS,EAAE,qBAAqB,IAAI,mBAAmB,CAAC,CAAC,CAAC,gBAAgB,CAAC,CAAC,CAAC,iBAAiB;KAC9F,CAAC;AAAA,CACF;AAED;;GAEG;AACH,MAAM,UAAU,aAAa,CAAC,aAAqB,EAAE,aAAqB,EAAE,QAA4B,EAAW;IAClH,MAAM,SAAS,GAAG,8BAA8B,CAAC,aAAa,EAAE,QAAQ,CAAC,CAAC;IAC1E,OAAO,SAAS,KAAK,SAAS,IAAI,MAAM,CAAC,QAAQ,CAAC,aAAa,CAAC,IAAI,aAAa,IAAI,SAAS,CAAC,aAAa,CAAC;AAAA,CAC7G","sourcesContent":["import { computeReservedTokenBudget } from \"../context-budget-reserved-tokens.ts\";\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\treservedOutputTokens?: number;\n\treservedToolResultTokens?: number;\n\tsafetyMarginTokens?: number;\n\timageReserveTokens?: number;\n\tkeepRecentTokens: number;\n\t/** Maximum fraction of the context window to use before compaction. Default: 0.9. */\n\tmaxUsageRatio?: number;\n\trearmRatio?: number;\n\temergencyRatio?: number;\n}\n\nexport const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tkeepRecentTokens: 20000,\n\tmaxUsageRatio: DEFAULT_COMPACTION_MAX_USAGE_RATIO,\n};\n\nexport type CompactionHeadroomLimit = \"max_usage_ratio\" | \"reserve_tokens\";\n\nexport interface CompactionHeadroomThreshold {\n\ttriggerTokens: number;\n\theadroomTokens: number;\n\tmaxUsageRatioTokens: number;\n\treserveBoundaryTokens: number;\n\tlimitedBy: CompactionHeadroomLimit;\n}\n\n/** Return the earliest compaction threshold from ratio-based headroom and absolute reserve. */\nexport function getCompactionHeadroomThreshold(\n\tcontextWindow: number,\n\tsettings: CompactionSettings,\n): CompactionHeadroomThreshold | undefined {\n\tif (!settings.enabled || !Number.isFinite(contextWindow) || contextWindow <= 0) {\n\t\treturn undefined;\n\t}\n\n\tconst windowTokens = Math.floor(contextWindow);\n\tconst reserveTokens = Number.isFinite(settings.reserveTokens) ? Math.max(0, Math.floor(settings.reserveTokens)) : 0;\n\tconst configuredMaxUsageRatio = settings.maxUsageRatio ?? DEFAULT_COMPACTION_MAX_USAGE_RATIO;\n\tconst maxUsageRatio =\n\t\tNumber.isFinite(configuredMaxUsageRatio) && configuredMaxUsageRatio > 0 && configuredMaxUsageRatio < 1\n\t\t\t? configuredMaxUsageRatio\n\t\t\t: DEFAULT_COMPACTION_MAX_USAGE_RATIO;\n\tconst maxUsageRatioTokens = Math.max(1, Math.floor(windowTokens * maxUsageRatio));\n\tconst reservedBudget = computeReservedTokenBudget({\n\t\tmodelContextWindow: windowTokens,\n\t\tsystemPromptTokens: 0,\n\t\treservedOutputTokens: settings.reservedOutputTokens ?? reserveTokens,\n\t\treservedToolResultTokens: settings.reservedToolResultTokens ?? 0,\n\t\tsafetyMarginTokens: settings.safetyMarginTokens ?? 0,\n\t\timageReserveTokens: settings.imageReserveTokens ?? 0,\n\t});\n\tconst reserveBoundaryTokens =\n\t\treservedBudget.overflow || reservedBudget.effectiveBudget === 0\n\t\t\t? maxUsageRatioTokens\n\t\t\t: Math.max(1, reservedBudget.effectiveBudget);\n\tconst triggerTokens = Math.min(maxUsageRatioTokens, reserveBoundaryTokens);\n\n\treturn {\n\t\ttriggerTokens,\n\t\theadroomTokens: windowTokens - triggerTokens,\n\t\tmaxUsageRatioTokens,\n\t\treserveBoundaryTokens,\n\t\tlimitedBy: reserveBoundaryTokens <= maxUsageRatioTokens ? \"reserve_tokens\" : \"max_usage_ratio\",\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tconst threshold = getCompactionHeadroomThreshold(contextWindow, settings);\n\treturn threshold !== undefined && Number.isFinite(contextTokens) && contextTokens >= threshold.triggerTokens;\n}\n"]}

package/dist/core/compaction/compaction.d.ts

@@ -7,6 +7,8 @@
 import type { AgentMessage, StreamFn, ThinkingLevel } from "omk-agent-core";
 import type { AssistantMessage, Context, Model, SimpleStreamOptions, Usage } from "omk-ai";
 import { type RetryCallbacks, type RetryPolicy } from "omk-ai";
+import type { CompactionSettings } from "./compaction-headroom.ts";
+export { type CompactionHeadroomLimit, type CompactionHeadroomThreshold, type CompactionSettings, DEFAULT_COMPACTION_MAX_USAGE_RATIO, DEFAULT_COMPACTION_SETTINGS, getCompactionHeadroomThreshold, shouldCompact, } from "./compaction-headroom.ts";
 import { type SessionEntry } from "../session-manager.ts";
 import { type CompactionRuleHistory, type PreservedRuleRecord } from "./knowledge-triage.ts";
 import { type FileOperations } from "./utils.ts";
@@ -25,21 +27,6 @@ export interface CompactionResult<T = unknown> {
     /** Extension-specific data (e.g., ArtifactIndex, version markers for structured compaction) */
     details?: T;
 }
-export interface CompactionSettings {
-    enabled: boolean;
-    reserveTokens: number;
-    reservedOutputTokens?: number;
-    reservedToolResultTokens?: number;
-    safetyMarginTokens?: number;
-    imageReserveTokens?: number;
-    keepRecentTokens: number;
-    /** Maximum fraction of the context window to use before compaction. Default: 0.9. */
-    maxUsageRatio?: number;
-    rearmRatio?: number;
-    emergencyRatio?: number;
-}
-export declare const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;
-export declare const DEFAULT_COMPACTION_SETTINGS: CompactionSettings;
 /**
  * Calculate total context tokens from usage.
  * Uses the native totalTokens field when available, falls back to computing from components.
@@ -62,20 +49,6 @@ export interface ContextUsageEstimate {
 export declare function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate;
 /** Estimate context tokens after adding messages that have not been sent yet. */
 export declare function estimateProjectedContextTokens(messages: AgentMessage[], pendingMessages: AgentMessage[]): ContextUsageEstimate;
-export type CompactionHeadroomLimit = "max_usage_ratio" | "reserve_tokens";
-export interface CompactionHeadroomThreshold {
-    triggerTokens: number;
-    headroomTokens: number;
-    maxUsageRatioTokens: number;
-    reserveBoundaryTokens: number;
-    limitedBy: CompactionHeadroomLimit;
-}
-/** Return the earliest compaction threshold from ratio-based headroom and absolute reserve. */
-export declare function getCompactionHeadroomThreshold(contextWindow: number, settings: CompactionSettings): CompactionHeadroomThreshold | undefined;
-/**
- * Check if compaction should trigger based on context usage.
- */
-export declare function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean;
 /**
  * Estimate token count for a message using chars/4 heuristic.
  * This is conservative (overestimates tokens).

package/dist/core/compaction/compaction.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAE5E,OAAO,KAAK,EAAE,gBAAgB,EAAE,OAAO,EAAE,KAAK,EAAE,mBAAmB,EAAE,KAAK,EAAE,MAAM,QAAQ,CAAC;AAC3F,OAAO,EAAkB,KAAK,cAAc,EAAE,KAAK,WAAW,EAAsB,MAAM,QAAQ,CAAC;AAInG,OAAO,EAA6C,KAAK,YAAY,EAAE,MAAM,uBAAuB,CAAC;AACrG,OAAO,EAGN,KAAK,qBAAqB,EAC1B,KAAK,mBAAmB,EACxB,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAIN,KAAK,cAAc,EAKnB,MAAM,YAAY,CAAC;AAMpB,kEAAkE;AAClE,MAAM,WAAW,iBAAiB;IACjC,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,aAAa,EAAE,MAAM,EAAE,CAAC;IACxB,0FAA0F;IAC1F,cAAc,CAAC,EAAE,mBAAmB,EAAE,CAAC;CACvC;AAMD,8EAA8E;AAC9E,MAAM,WAAW,gBAAgB,CAAC,CAAC,GAAG,OAAO;IAC5C,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,EAAE,MAAM,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;IACrB,+FAA+F;IAC/F,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AAMD,MAAM,WAAW,kBAAkB;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,aAAa,EAAE,MAAM,CAAC;IACtB,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,wBAAwB,CAAC,EAAE,MAAM,CAAC;IAClC,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,gBAAgB,EAAE,MAAM,CAAC;IACzB,qFAAqF;IACrF,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,cAAc,CAAC,EAAE,MAAM,CAAC;CACxB;AAED,eAAO,MAAM,kCAAkC,MAAM,CAAC;AAEtD,eAAO,MAAM,2BAA2B,EAAE,kBAKzC,CAAC;AAMF;;;GAGG;AACH,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAgBD;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,KAAK,GAAG,SAAS,CAShF;AAED,MAAM,WAAW,oBAAoB;IACpC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,oBAAoB,CA4BpF;AAED,iFAAiF;AACjF,wBAAgB,8BAA8B,CAC7C,QAAQ,EAAE,YAAY,EAAE,EACxB,eAAe,EAAE,YAAY,EAAE,GAC7B,oBAAoB,CAEtB;AAED,MAAM,MAAM,uBAAuB,GAAG,iBAAiB,GAAG,gBAAgB,CAAC;AAE3E,MAAM,WAAW,2BAA2B;IAC3C,aAAa,EAAE,MAAM,CAAC;IACtB,cAAc,EAAE,MAAM,CAAC;IACvB,mBAAmB,EAAE,MAAM,CAAC;IAC5B,qBAAqB,EAAE,MAAM,CAAC;IAC9B,SAAS,EAAE,uBAAuB,CAAC;CACnC;AAED,+FAA+F;AAC/F,wBAAgB,8BAA8B,CAC7C,aAAa,EAAE,MAAM,EACrB,QAAQ,EAAE,kBAAkB,GAC1B,2BAA2B,GAAG,SAAS,CAkCzC;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAwBD;;;GAGG;AACH,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAwC5D;AAiDD;;;;GAIG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAe1G;AAED,MAAM,WAAW,cAAc;IAC9B,mCAAmC;IACnC,mBAAmB,EAAE,MAAM,CAAC;IAC5B,qFAAqF;IACrF,cAAc,EAAE,MAAM,CAAC;IACvB,uEAAuE;IACvE,WAAW,EAAE,OAAO,CAAC;CACrB;AAED;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,YAAY,CAC3B,OAAO,EAAE,YAAY,EAAE,EACvB,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CAyDhB;AA6FD;;;;;;GAMG;AACH,wBAAsB,qBAAqB,CAC1C,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,EAC5B,QAAQ,CAAC,EAAE,QAAQ,EACnB,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CAY3B;AAED;;;GAGG;AACH,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,MAAM,CAAC,CAqDjB;AAMD,MAAM,WAAW,qBAAqB;IACrC,kCAAkC;IAClC,gBAAgB,EAAE,MAAM,CAAC;IACzB,qDAAqD;IACrD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,iEAAiE;IACjE,WAAW,EAAE,OAAO,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,6DAA6D;IAC7D,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,sEAAsE;IACtE,mBAAmB,CAAC,EAAE,qBAAqB,CAAC;IAC5C,0EAA0E;IAC1E,kBAAkB,CAAC,EAAE,SAAS,YAAY,EAAE,CAAC;IAC7C,yDAAyD;IACzD,OAAO,EAAE,cAAc,CAAC;IACxB,8CAA8C;IAC9C,QAAQ,EAAE,kBAAkB,CAAC;CAC7B;AAED,wBAAgB,iBAAiB,CAChC,WAAW,EAAE,YAAY,EAAE,EAC3B,QAAQ,EAAE,kBAAkB,GAC1B,qBAAqB,GAAG,SAAS,CAgFnC;AAqBD;;;;;;;;;;;GAWG;AACH,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,EAC1B,cAAc,CAAC,EAAE,SAAS,KAAK,CAAC,GAAG,CAAC,EAAE,GACpC,OAAO,CAAC,gBAAgB,CAAC,CA0C3B","sourcesContent":["/**\n * Context compaction for long sessions.\n *\n * Pure functions for compaction logic. The session manager handles I/O,\n * and after compaction the session is reloaded.\n */\n\nimport type { AgentMessage, StreamFn, ThinkingLevel } from \"omk-agent-core\";\nimport { uuidv7 } from \"omk-agent-core\";\nimport type { AssistantMessage, Context, Model, SimpleStreamOptions, Usage } from \"omk-ai\";\nimport { completeSimple, type RetryCallbacks, type RetryPolicy, retryAssistantCall } from \"omk-ai\";\nimport { computeReservedTokenBudget } from \"../context-budget-reserved-tokens.ts\";\nimport { convertToLlm } from \"../messages.ts\";\nimport { isQuotaExhaustionMessage } from \"../provider-resilience.ts\";\nimport { buildSessionContext, type CompactionEntry, type SessionEntry } from \"../session-manager.ts\";\nimport {\n\tapplyCompactionKnowledgeTriage,\n\tbuildCompactionRuleHistory,\n\ttype CompactionRuleHistory,\n\ttype PreservedRuleRecord,\n} from \"./knowledge-triage.ts\";\nimport {\n\tcomputeFileLists,\n\textractCompactionFileOperations,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tgetMessageFromEntryForCompaction,\n\tSUMMARIZATION_SYSTEM_PROMPT,\n\tserializeConversation,\n} from \"./utils.ts\";\n\n// ============================================================================\n// File Operation Tracking\n// ============================================================================\n\n/** Details stored in CompactionEntry.details for file tracking */\nexport interface CompactionDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n\t/** Deterministically extracted, source-bound user rules carried outside LLM rewriting. */\n\tpreservedRules?: PreservedRuleRecord[];\n}\n\n// ============================================================================\n// Message Extraction\n// ============================================================================\n\n/** Result from compact() - SessionManager adds uuid/parentUuid when saving */\nexport interface CompactionResult<T = unknown> {\n\tsummary: string;\n\tfirstKeptEntryId: string;\n\ttokensBefore: number;\n\t/** Extension-specific data (e.g., ArtifactIndex, version markers for structured compaction) */\n\tdetails?: T;\n}\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\treservedOutputTokens?: number;\n\treservedToolResultTokens?: number;\n\tsafetyMarginTokens?: number;\n\timageReserveTokens?: number;\n\tkeepRecentTokens: number;\n\t/** Maximum fraction of the context window to use before compaction. Default: 0.9. */\n\tmaxUsageRatio?: number;\n\trearmRatio?: number;\n\temergencyRatio?: number;\n}\n\nexport const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tkeepRecentTokens: 20000,\n\tmaxUsageRatio: DEFAULT_COMPACTION_MAX_USAGE_RATIO,\n};\n\n// ============================================================================\n// Token calculation\n// ============================================================================\n\n/**\n * Calculate total context tokens from usage.\n * Uses the native totalTokens field when available, falls back to computing from components.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted and error messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (assistantMsg.stopReason !== \"aborted\" && assistantMsg.stopReason !== \"error\" && assistantMsg.usage) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/**\n * Find the last non-aborted assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/** Estimate context tokens after adding messages that have not been sent yet. */\nexport function estimateProjectedContextTokens(\n\tmessages: AgentMessage[],\n\tpendingMessages: AgentMessage[],\n): ContextUsageEstimate {\n\treturn estimateContextTokens([...messages, ...pendingMessages]);\n}\n\nexport type CompactionHeadroomLimit = \"max_usage_ratio\" | \"reserve_tokens\";\n\nexport interface CompactionHeadroomThreshold {\n\ttriggerTokens: number;\n\theadroomTokens: number;\n\tmaxUsageRatioTokens: number;\n\treserveBoundaryTokens: number;\n\tlimitedBy: CompactionHeadroomLimit;\n}\n\n/** Return the earliest compaction threshold from ratio-based headroom and absolute reserve. */\nexport function getCompactionHeadroomThreshold(\n\tcontextWindow: number,\n\tsettings: CompactionSettings,\n): CompactionHeadroomThreshold | undefined {\n\tif (!settings.enabled || !Number.isFinite(contextWindow) || contextWindow <= 0) {\n\t\treturn undefined;\n\t}\n\n\tconst windowTokens = Math.floor(contextWindow);\n\tconst reserveTokens = Number.isFinite(settings.reserveTokens) ? Math.max(0, Math.floor(settings.reserveTokens)) : 0;\n\tconst configuredMaxUsageRatio = settings.maxUsageRatio ?? DEFAULT_COMPACTION_MAX_USAGE_RATIO;\n\tconst maxUsageRatio =\n\t\tNumber.isFinite(configuredMaxUsageRatio) && configuredMaxUsageRatio > 0 && configuredMaxUsageRatio < 1\n\t\t\t? configuredMaxUsageRatio\n\t\t\t: DEFAULT_COMPACTION_MAX_USAGE_RATIO;\n\tconst maxUsageRatioTokens = Math.max(1, Math.floor(windowTokens * maxUsageRatio));\n\tconst reservedBudget = computeReservedTokenBudget({\n\t\tmodelContextWindow: windowTokens,\n\t\tsystemPromptTokens: 0,\n\t\treservedOutputTokens: settings.reservedOutputTokens ?? reserveTokens,\n\t\treservedToolResultTokens: settings.reservedToolResultTokens ?? 0,\n\t\tsafetyMarginTokens: settings.safetyMarginTokens ?? 0,\n\t\timageReserveTokens: settings.imageReserveTokens ?? 0,\n\t});\n\tconst reserveBoundaryTokens =\n\t\treservedBudget.overflow || reservedBudget.effectiveBudget === 0\n\t\t\t? maxUsageRatioTokens\n\t\t\t: Math.max(1, reservedBudget.effectiveBudget);\n\tconst triggerTokens = Math.min(maxUsageRatioTokens, reserveBoundaryTokens);\n\n\treturn {\n\t\ttriggerTokens,\n\t\theadroomTokens: windowTokens - triggerTokens,\n\t\tmaxUsageRatioTokens,\n\t\treserveBoundaryTokens,\n\t\tlimitedBy: reserveBoundaryTokens <= maxUsageRatioTokens ? \"reserve_tokens\" : \"max_usage_ratio\",\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tconst threshold = getCompactionHeadroomThreshold(contextWindow, settings);\n\treturn threshold !== undefined && Number.isFinite(contextTokens) && contextTokens >= threshold.triggerTokens;\n}\n\n// ============================================================================\n// Cut point detection\n// ============================================================================\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n\n/**\n * Find valid cut points: indices of user, assistant, custom, or bashExecution messages.\n * Never cut at tool results (they must follow their tool call).\n * When we cut at an assistant message with tool calls, its tool results follow it\n * and will be kept.\n * BashExecutionMessage is treated like a user message (user-initiated context).\n */\nfunction findValidCutPoints(entries: SessionEntry[], startIndex: number, endIndex: number): number[] {\n\tconst cutPoints: number[] = [];\n\tfor (let i = startIndex; i < endIndex; i++) {\n\t\tconst entry = entries[i];\n\t\tswitch (entry.type) {\n\t\t\tcase \"message\": {\n\t\t\t\tconst role = entry.message.role;\n\t\t\t\tswitch (role) {\n\t\t\t\t\tcase \"bashExecution\":\n\t\t\t\t\tcase \"custom\":\n\t\t\t\t\tcase \"branchSummary\":\n\t\t\t\t\tcase \"compactionSummary\":\n\t\t\t\t\tcase \"user\":\n\t\t\t\t\tcase \"assistant\":\n\t\t\t\t\t\tcutPoints.push(i);\n\t\t\t\t\t\tbreak;\n\t\t\t\t\tcase \"toolResult\":\n\t\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"thinking_level_change\":\n\t\t\tcase \"model_change\":\n\t\t\tcase \"compaction\":\n\t\t\tcase \"branch_summary\":\n\t\t\tcase \"custom\":\n\t\t\tcase \"custom_message\":\n\t\t\tcase \"label\":\n\t\t\tcase \"session_info\":\n\t\t\t\tbreak;\n\t\t}\n\n\t\t// branch_summary and custom_message are user-role messages, valid cut points\n\t\tif (entry.type === \"branch_summary\" || entry.type === \"custom_message\") {\n\t\t\tcutPoints.push(i);\n\t\t}\n\t}\n\treturn cutPoints;\n}\n\n/**\n * Find the user message (or bashExecution) that starts the turn containing the given entry index.\n * Returns -1 if no turn start found before the index.\n * BashExecutionMessage is treated like a user message for turn boundaries.\n */\nexport function findTurnStartIndex(entries: SessionEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let i = entryIndex; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\t// branch_summary and custom_message are user-role messages, can start a turn\n\t\tif (entry.type === \"branch_summary\" || entry.type === \"custom_message\") {\n\t\t\treturn i;\n\t\t}\n\t\tif (entry.type === \"message\") {\n\t\t\tconst role = entry.message.role;\n\t\t\tif (role === \"user\" || role === \"bashExecution\") {\n\t\t\t\treturn i;\n\t\t\t}\n\t\t}\n\t}\n\treturn -1;\n}\n\nexport interface CutPointResult {\n\t/** Index of first entry to keep */\n\tfirstKeptEntryIndex: number;\n\t/** Index of user message that starts the turn being split, or -1 if not splitting */\n\tturnStartIndex: number;\n\t/** Whether this cut splits a turn (cut point is not a user message) */\n\tisSplitTurn: boolean;\n}\n\n/**\n * Find the cut point in session entries that keeps approximately `keepRecentTokens`.\n *\n * Algorithm: Walk backwards from newest, accumulating estimated message sizes.\n * Stop when we've accumulated >= keepRecentTokens. Cut at that point.\n *\n * Can cut at user OR assistant messages (never tool results). When cutting at an\n * assistant message with tool calls, its tool results come after and will be kept.\n *\n * Returns CutPointResult with:\n * - firstKeptEntryIndex: the entry index to start keeping from\n * - turnStartIndex: if cutting mid-turn, the user message that started that turn\n * - isSplitTurn: whether we're cutting in the middle of a turn\n *\n * Only considers entries between `startIndex` and `endIndex` (exclusive).\n */\nexport function findCutPoint(\n\tentries: SessionEntry[],\n\tstartIndex: number,\n\tendIndex: number,\n\tkeepRecentTokens: number,\n): CutPointResult {\n\tconst cutPoints = findValidCutPoints(entries, startIndex, endIndex);\n\n\tif (cutPoints.length === 0) {\n\t\treturn { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };\n\t}\n\n\t// Walk backwards from newest, accumulating estimated message sizes\n\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0]; // Default: keep from first message (not header)\n\n\tfor (let i = endIndex - 1; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type !== \"message\") continue;\n\n\t\t// Estimate this message's size\n\t\tconst messageTokens = estimateTokens(entry.message);\n\t\taccumulatedTokens += messageTokens;\n\n\t\t// Check if we've exceeded the budget\n\t\tif (accumulatedTokens >= keepRecentTokens) {\n\t\t\t// Find the closest valid cut point at or after this entry\n\t\t\tfor (let c = 0; c < cutPoints.length; c++) {\n\t\t\t\tif (cutPoints[c] >= i) {\n\t\t\t\t\tcutIndex = cutPoints[c];\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t}\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Scan backwards from cutIndex to include any non-message entries (bash, settings, etc.)\n\twhile (cutIndex > startIndex) {\n\t\tconst prevEntry = entries[cutIndex - 1];\n\t\t// Stop at session header or compaction boundaries\n\t\tif (prevEntry.type === \"compaction\") {\n\t\t\tbreak;\n\t\t}\n\t\tif (prevEntry.type === \"message\") {\n\t\t\t// Stop if we hit any message\n\t\t\tbreak;\n\t\t}\n\t\t// Include this non-message entry (bash, settings change, etc.)\n\t\tcutIndex--;\n\t}\n\n\t// Determine if this is a split turn\n\tconst cutEntry = entries[cutIndex];\n\tconst isUserMessage = cutEntry.type === \"message\" && cutEntry.message.role === \"user\";\n\tconst turnStartIndex = isUserMessage ? -1 : findTurnStartIndex(entries, cutIndex, startIndex);\n\n\treturn {\n\t\tfirstKeptEntryIndex: cutIndex,\n\t\tturnStartIndex,\n\t\tisSplitTurn: !isUserMessage && turnStartIndex !== -1,\n\t};\n}\n\n// ============================================================================\n// Summarization\n// ============================================================================\n\nconst SUMMARIZATION_PROMPT = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.\n\nUse this EXACT format:\n\n## Goal\n[What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]\n\n## Constraints & Preferences\n- [Any constraints, preferences, or requirements mentioned by user]\n- [Or \"(none)\" if none were mentioned]\n\n## Progress\n### Done\n- [x] [Completed tasks/changes]\n\n### In Progress\n- [ ] [Current work]\n\n### Blocked\n- [Issues preventing progress, if any]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale]\n\n## Next Steps\n1. [Ordered list of what should happen next]\n\n## Critical Context\n- [Any data, examples, or references needed to continue]\n- [Or \"(none)\" if not applicable]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages. Do not create a Preserved Rules & Invariants section; OMK injects it deterministically after generation.`;\n\nconst UPDATE_SUMMARIZATION_PROMPT = `The messages above are NEW conversation messages to incorporate into the existing summary provided in <previous-summary> tags.\n\nUpdate the existing structured summary with new information. RULES:\n- PRESERVE all existing information from the previous summary\n- ADD new progress, decisions, and context from the new messages\n- UPDATE the Progress section: move items from \"In Progress\" to \"Done\" when completed\n- UPDATE \"Next Steps\" based on what was accomplished\n- PRESERVE exact file paths, function names, and error messages\n- If something is no longer relevant, you may remove it\n\nUse this EXACT format:\n\n## Goal\n[Preserve existing goals, add new ones if the task expanded]\n\n## Constraints & Preferences\n- [Preserve existing, add new ones discovered]\n\n## Progress\n### Done\n- [x] [Include previously done items AND newly completed items]\n\n### In Progress\n- [ ] [Current work - update based on progress]\n\n### Blocked\n- [Current blockers - remove if resolved]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale] (preserve all previous, add new)\n\n## Next Steps\n1. [Update based on current state]\n\n## Critical Context\n- [Preserve important context, add new if needed]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages. Do not create a Preserved Rules & Invariants section; OMK injects it deterministically after generation.`;\n\nfunction createSummarizationOptions(\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n): SimpleStreamOptions {\n\tconst options: SimpleStreamOptions = { maxTokens, signal, apiKey, headers };\n\tif (model.reasoning && thinkingLevel && thinkingLevel !== \"off\") {\n\t\toptions.reasoning = thinkingLevel;\n\t}\n\treturn options;\n}\n\n/**\n * Shared choke point for every compaction/branch-summary summarization call. Wraps the\n * single LLM call in {@link retryAssistantCall} so transient stream drops (e.g.\n * `terminated`, socket close) honor the configured retry policy instead of failing\n * the whole compaction on the first attempt. Deterministic errors and aborts return\n * immediately (see {@link retryAssistantCall}).\n */\nexport async function completeSummarization(\n\tmodel: Model<any>,\n\tcontext: Context,\n\toptions: SimpleStreamOptions,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\t// Summaries are standalone requests, so isolate routing and avoid cache writes that cannot be reused.\n\tconst requestOptions: SimpleStreamOptions = {\n\t\t...options,\n\t\tcacheRetention: \"none\",\n\t\tsessionId: uuidv7(),\n\t};\n\tconst produce = async (): Promise<AssistantMessage> =>\n\t\tstreamFn\n\t\t\t? (await streamFn(model, context, requestOptions)).result()\n\t\t\t: completeSimple(model, context, requestOptions);\n\treturn retryAssistantCall(produce, retry, requestOptions.signal, callbacks);\n}\n\n/**\n * Generate a summary of the conversation using the LLM.\n * If previousSummary is provided, uses the update prompt to merge.\n */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<string> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.8 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t);\n\n\t// Use update prompt if we have a previous summary, otherwise initial prompt\n\tlet basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tbasePrompt = `${basePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\n\t// Serialize conversation to text so model doesn't try to continue it\n\t// Convert to LLM messages first (handles custom types like bashExecution, custom, etc.)\n\tconst llmMessages = convertToLlm(currentMessages);\n\tconst conversationText = serializeConversation(llmMessages);\n\n\t// Build the prompt with conversation wrapped in tags\n\tlet promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n`;\n\tif (previousSummary) {\n\t\tpromptText += `<previous-summary>\\n${previousSummary}\\n</previous-summary>\\n\\n`;\n\t}\n\tpromptText += basePrompt;\n\n\tconst summarizationMessages = [\n\t\t{\n\t\t\trole: \"user\" as const,\n\t\t\tcontent: [{ type: \"text\" as const, text: promptText }],\n\t\t\ttimestamp: Date.now(),\n\t\t},\n\t];\n\n\tconst completionOptions = createSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel);\n\n\tconst response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcompletionOptions,\n\t\tstreamFn,\n\t\tretry,\n\t\tcallbacks,\n\t);\n\n\tif (response.stopReason === \"error\") {\n\t\tthrow new Error(`Summarization failed: ${response.errorMessage || \"Unknown error\"}`);\n\t}\n\n\tconst textContent = response.content\n\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n\n\treturn textContent;\n}\n\n// ============================================================================\n// Compaction Preparation (for extensions)\n// ============================================================================\n\nexport interface CompactionPreparation {\n\t/** UUID of first entry to keep */\n\tfirstKeptEntryId: string;\n\t/** Messages that will be summarized and discarded */\n\tmessagesToSummarize: AgentMessage[];\n\t/** Messages that will be turned into turn prefix summary (if splitting) */\n\tturnPrefixMessages: AgentMessage[];\n\t/** Whether this is a split turn (cut point in middle of turn) */\n\tisSplitTurn: boolean;\n\ttokensBefore: number;\n\t/** Summary from previous compaction, for iterative update */\n\tpreviousSummary?: string;\n\t/** Source-verified rule history from the prior default compaction. */\n\tpreviousRuleHistory?: CompactionRuleHistory;\n\t/** Exact session entries covered by this compaction's rule extraction. */\n\tcurrentRuleEntries?: readonly SessionEntry[];\n\t/** File operations extracted from messagesToSummarize */\n\tfileOps: FileOperations;\n\t/** Compaction settions from settings.jsonl\t*/\n\tsettings: CompactionSettings;\n}\n\nexport function prepareCompaction(\n\tpathEntries: SessionEntry[],\n\tsettings: CompactionSettings,\n): CompactionPreparation | undefined {\n\tif (pathEntries.length > 0 && pathEntries[pathEntries.length - 1].type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\n\tlet prevCompactionIndex = -1;\n\tfor (let i = pathEntries.length - 1; i >= 0; i--) {\n\t\tif (pathEntries[i].type === \"compaction\") {\n\t\t\tprevCompactionIndex = i;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\tlet previousSummary: string | undefined;\n\tlet previousRuleHistory: CompactionRuleHistory | undefined;\n\tlet boundaryStart = 0;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tpreviousRuleHistory = buildCompactionRuleHistory(prevCompaction, pathEntries);\n\t\tconst firstKeptEntryIndex = pathEntries.findIndex((entry) => entry.id === prevCompaction.firstKeptEntryId);\n\t\tboundaryStart = firstKeptEntryIndex >= 0 ? firstKeptEntryIndex : prevCompactionIndex + 1;\n\t}\n\tconst boundaryEnd = pathEntries.length;\n\n\tconst tokensBefore = estimateContextTokens(buildSessionContext(pathEntries).messages).tokens;\n\n\tconst cutPoint = findCutPoint(pathEntries, boundaryStart, boundaryEnd, settings.keepRecentTokens);\n\n\t// Get UUID of first kept entry\n\tconst firstKeptEntry = pathEntries[cutPoint.firstKeptEntryIndex];\n\tif (!firstKeptEntry?.id) {\n\t\treturn undefined; // Session needs migration\n\t}\n\tconst firstKeptEntryId = firstKeptEntry.id;\n\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\n\t// Messages to summarize (will be discarded after summary)\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let i = boundaryStart; i < historyEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\tif (msg) messagesToSummarize.push(msg);\n\t}\n\n\t// Messages for turn prefix summary (if splitting a turn)\n\tconst turnPrefixMessages: AgentMessage[] = [];\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (let i = cutPoint.turnStartIndex; i < cutPoint.firstKeptEntryIndex; i++) {\n\t\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\t\tif (msg) turnPrefixMessages.push(msg);\n\t\t}\n\t}\n\n\t// Extract file operations from messages and previous compaction\n\tconst fileOps = extractCompactionFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\n\tconst currentRuleEntries = [\n\t\t...pathEntries.slice(boundaryStart, historyEnd),\n\t\t...(cutPoint.isSplitTurn ? pathEntries.slice(cutPoint.turnStartIndex, cutPoint.firstKeptEntryIndex) : []),\n\t];\n\n\t// Also extract file ops from turn prefix if splitting\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (const msg of turnPrefixMessages) {\n\t\t\textractFileOpsFromMessage(msg, fileOps);\n\t\t}\n\t}\n\n\treturn {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tpreviousRuleHistory,\n\t\tcurrentRuleEntries,\n\t\tfileOps,\n\t\tsettings,\n\t};\n}\n\n// ============================================================================\n// Main compaction function\n// ============================================================================\n\nconst TURN_PREFIX_SUMMARIZATION_PROMPT = `This is the PREFIX of a turn that was too large to keep. The SUFFIX (recent work) is retained.\n\nSummarize the prefix to provide context for the retained suffix:\n\n## Original Request\n[What did the user ask for in this turn?]\n\n## Early Progress\n- [Key decisions and work done in the prefix]\n\n## Context for Suffix\n- [Information needed to understand the retained recent work]\n\nBe concise. Focus on what's needed to understand the kept suffix.`;\n\n/**\n * Generate summaries for compaction using prepared data.\n * Returns CompactionResult - SessionManager adds uuid/parentUuid when saving.\n *\n * When {@link failoverModels} is provided and the primary model fails with a\n * quota/billing exhaustion error (same-model retry is useless until reset),\n * the whole summarization is retried once with each candidate in order. Only\n * quota-class errors are absorbed; any other failure surfaces immediately.\n *\n * @param preparation - Pre-calculated preparation from prepareCompaction()\n * @param customInstructions - Optional custom focus for the summary\n */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tfailoverModels?: readonly Model<any>[],\n): Promise<CompactionResult> {\n\ttry {\n\t\treturn await compactWithModel(\n\t\t\tpreparation,\n\t\t\tmodel,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tcustomInstructions,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tif (!isQuotaExhaustionMessage(message) || !failoverModels || failoverModels.length === 0) throw error;\n\t\tfor (const candidate of failoverModels) {\n\t\t\tif (candidate.provider === model.provider && candidate.id === model.id) continue;\n\t\t\tif (signal?.aborted) throw error;\n\t\t\ttry {\n\t\t\t\treturn await compactWithModel(\n\t\t\t\t\tpreparation,\n\t\t\t\t\tcandidate,\n\t\t\t\t\tapiKey,\n\t\t\t\t\theaders,\n\t\t\t\t\tcustomInstructions,\n\t\t\t\t\tsignal,\n\t\t\t\t\tthinkingLevel,\n\t\t\t\t\tstreamFn,\n\t\t\t\t\tretry,\n\t\t\t\t\tcallbacks,\n\t\t\t\t);\n\t\t\t} catch (candidateError) {\n\t\t\t\tconst candidateMessage = candidateError instanceof Error ? candidateError.message : String(candidateError);\n\t\t\t\tif (!isQuotaExhaustionMessage(candidateMessage)) throw candidateError;\n\t\t\t\t// Candidate is also quota-blocked — try the next one.\n\t\t\t}\n\t\t}\n\t\t// Every candidate is quota-blocked too: surface the primary failure.\n\t\tthrow error;\n\t}\n}\n\n/** Single-model compaction attempt (no failover). Implementation of {@link compact}. */\nasync function compactWithModel(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<CompactionResult> {\n\tconst {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tpreviousRuleHistory,\n\t\tcurrentRuleEntries,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\t// Generate summaries (can be parallel if both needed) and merge into one\n\tlet summary: string;\n\n\tif (isSplitTurn && turnPrefixMessages.length > 0) {\n\t\t// Generate both summaries in parallel\n\t\tconst [historyResult, turnPrefixResult] = await Promise.all([\n\t\t\tmessagesToSummarize.length > 0\n\t\t\t\t? generateSummary(\n\t\t\t\t\t\tmessagesToSummarize,\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tsettings.reserveTokens,\n\t\t\t\t\t\tapiKey,\n\t\t\t\t\t\theaders,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\tcustomInstructions,\n\t\t\t\t\t\tpreviousSummary,\n\t\t\t\t\t\tthinkingLevel,\n\t\t\t\t\t\tstreamFn,\n\t\t\t\t\t\tretry,\n\t\t\t\t\t\tcallbacks,\n\t\t\t\t\t)\n\t\t\t\t: Promise.resolve(\"No prior history.\"),\n\t\t\tgenerateTurnPrefixSummary(\n\t\t\t\tturnPrefixMessages,\n\t\t\t\tmodel,\n\t\t\t\tsettings.reserveTokens,\n\t\t\t\tapiKey,\n\t\t\t\theaders,\n\t\t\t\tsignal,\n\t\t\t\tthinkingLevel,\n\t\t\t\tstreamFn,\n\t\t\t\tretry,\n\t\t\t\tcallbacks,\n\t\t\t),\n\t\t]);\n\t\t// Merge into single summary\n\t\tsummary = `${historyResult}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult}`;\n\t} else {\n\t\t// Just generate history summary\n\t\tsummary = await generateSummary(\n\t\t\tmessagesToSummarize,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t}\n\n\tconst triage = applyCompactionKnowledgeTriage({\n\t\tgeneratedSummary: summary,\n\t\tcurrentMessages: [...messagesToSummarize, ...turnPrefixMessages],\n\t\tcurrentEntries: currentRuleEntries,\n\t\tpreviousRules: previousRuleHistory?.rules,\n\t\tpreviousSummary,\n\t\tpreviousEntries: previousRuleHistory?.entries,\n\t});\n\tsummary = triage.summary;\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\tif (!firstKeptEntryId) {\n\t\tthrow new Error(\"First kept entry has no UUID - session may need migration\");\n\t}\n\n\treturn {\n\t\tsummary,\n\t\tfirstKeptEntryId,\n\t\ttokensBefore,\n\t\tdetails: { readFiles, modifiedFiles, preservedRules: [...triage.preservedRules] } as CompactionDetails,\n\t};\n}\n\n/**\n * Generate a summary for a turn prefix (when splitting a turn).\n */\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<string> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.5 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t); // Smaller budget for turn prefix\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversation(llmMessages);\n\tconst promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;\n\tconst summarizationMessages = [\n\t\t{\n\t\t\trole: \"user\" as const,\n\t\t\tcontent: [{ type: \"text\" as const, text: promptText }],\n\t\t\ttimestamp: Date.now(),\n\t\t},\n\t];\n\n\tconst response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcreateSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel),\n\t\tstreamFn,\n\t\tretry,\n\t\tcallbacks,\n\t);\n\n\tif (response.stopReason === \"error\") {\n\t\tthrow new Error(`Turn prefix summarization failed: ${response.errorMessage || \"Unknown error\"}`);\n\t}\n\n\treturn response.content\n\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n}\n"]}
+{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAE5E,OAAO,KAAK,EAAE,gBAAgB,EAAE,OAAO,EAAE,KAAK,EAAE,mBAAmB,EAAE,KAAK,EAAE,MAAM,QAAQ,CAAC;AAC3F,OAAO,EAAkB,KAAK,cAAc,EAAE,KAAK,WAAW,EAAsB,MAAM,QAAQ,CAAC;AACnG,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,0BAA0B,CAAC;AAEnE,OAAO,EACN,KAAK,uBAAuB,EAC5B,KAAK,2BAA2B,EAChC,KAAK,kBAAkB,EACvB,kCAAkC,EAClC,2BAA2B,EAC3B,8BAA8B,EAC9B,aAAa,GACb,MAAM,0BAA0B,CAAC;AAIlC,OAAO,EAA6C,KAAK,YAAY,EAAE,MAAM,uBAAuB,CAAC;AACrG,OAAO,EAGN,KAAK,qBAAqB,EAC1B,KAAK,mBAAmB,EACxB,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAIN,KAAK,cAAc,EAKnB,MAAM,YAAY,CAAC;AAMpB,kEAAkE;AAClE,MAAM,WAAW,iBAAiB;IACjC,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,aAAa,EAAE,MAAM,EAAE,CAAC;IACxB,0FAA0F;IAC1F,cAAc,CAAC,EAAE,mBAAmB,EAAE,CAAC;CACvC;AAMD,8EAA8E;AAC9E,MAAM,WAAW,gBAAgB,CAAC,CAAC,GAAG,OAAO;IAC5C,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,EAAE,MAAM,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;IACrB,+FAA+F;IAC/F,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AAUD;;;GAGG;AACH,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAgBD;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,KAAK,GAAG,SAAS,CAShF;AAED,MAAM,WAAW,oBAAoB;IACpC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,oBAAoB,CA4BpF;AAED,iFAAiF;AACjF,wBAAgB,8BAA8B,CAC7C,QAAQ,EAAE,YAAY,EAAE,EACxB,eAAe,EAAE,YAAY,EAAE,GAC7B,oBAAoB,CAEtB;AAwBD;;;GAGG;AACH,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAwC5D;AAiDD;;;;GAIG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAe1G;AAED,MAAM,WAAW,cAAc;IAC9B,mCAAmC;IACnC,mBAAmB,EAAE,MAAM,CAAC;IAC5B,qFAAqF;IACrF,cAAc,EAAE,MAAM,CAAC;IACvB,uEAAuE;IACvE,WAAW,EAAE,OAAO,CAAC;CACrB;AAED;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,YAAY,CAC3B,OAAO,EAAE,YAAY,EAAE,EACvB,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CAwEhB;AA6FD;;;;;;GAMG;AACH,wBAAsB,qBAAqB,CAC1C,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,EAC5B,QAAQ,CAAC,EAAE,QAAQ,EACnB,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CAY3B;AAED;;;GAGG;AACH,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,MAAM,CAAC,CAqDjB;AAMD,MAAM,WAAW,qBAAqB;IACrC,kCAAkC;IAClC,gBAAgB,EAAE,MAAM,CAAC;IACzB,qDAAqD;IACrD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,iEAAiE;IACjE,WAAW,EAAE,OAAO,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,6DAA6D;IAC7D,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,sEAAsE;IACtE,mBAAmB,CAAC,EAAE,qBAAqB,CAAC;IAC5C,0EAA0E;IAC1E,kBAAkB,CAAC,EAAE,SAAS,YAAY,EAAE,CAAC;IAC7C,yDAAyD;IACzD,OAAO,EAAE,cAAc,CAAC;IACxB,8CAA8C;IAC9C,QAAQ,EAAE,kBAAkB,CAAC;CAC7B;AAED,wBAAgB,iBAAiB,CAChC,WAAW,EAAE,YAAY,EAAE,EAC3B,QAAQ,EAAE,kBAAkB,GAC1B,qBAAqB,GAAG,SAAS,CAgFnC;AAqBD;;;;;;;;;;;GAWG;AACH,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,EAC1B,cAAc,CAAC,EAAE,SAAS,KAAK,CAAC,GAAG,CAAC,EAAE,GACpC,OAAO,CAAC,gBAAgB,CAAC,CA0C3B","sourcesContent":["/**\n * Context compaction for long sessions.\n *\n * Pure functions for compaction logic. The session manager handles I/O,\n * and after compaction the session is reloaded.\n */\n\nimport type { AgentMessage, StreamFn, ThinkingLevel } from \"omk-agent-core\";\nimport { uuidv7 } from \"omk-agent-core\";\nimport type { AssistantMessage, Context, Model, SimpleStreamOptions, Usage } from \"omk-ai\";\nimport { completeSimple, type RetryCallbacks, type RetryPolicy, retryAssistantCall } from \"omk-ai\";\nimport type { CompactionSettings } from \"./compaction-headroom.ts\";\n\nexport {\n\ttype CompactionHeadroomLimit,\n\ttype CompactionHeadroomThreshold,\n\ttype CompactionSettings,\n\tDEFAULT_COMPACTION_MAX_USAGE_RATIO,\n\tDEFAULT_COMPACTION_SETTINGS,\n\tgetCompactionHeadroomThreshold,\n\tshouldCompact,\n} from \"./compaction-headroom.ts\";\n\nimport { convertToLlm } from \"../messages.ts\";\nimport { isQuotaExhaustionMessage } from \"../provider-resilience.ts\";\nimport { buildSessionContext, type CompactionEntry, type SessionEntry } from \"../session-manager.ts\";\nimport {\n\tapplyCompactionKnowledgeTriage,\n\tbuildCompactionRuleHistory,\n\ttype CompactionRuleHistory,\n\ttype PreservedRuleRecord,\n} from \"./knowledge-triage.ts\";\nimport {\n\tcomputeFileLists,\n\textractCompactionFileOperations,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tgetMessageFromEntryForCompaction,\n\tSUMMARIZATION_SYSTEM_PROMPT,\n\tserializeConversation,\n} from \"./utils.ts\";\n\n// ============================================================================\n// File Operation Tracking\n// ============================================================================\n\n/** Details stored in CompactionEntry.details for file tracking */\nexport interface CompactionDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n\t/** Deterministically extracted, source-bound user rules carried outside LLM rewriting. */\n\tpreservedRules?: PreservedRuleRecord[];\n}\n\n// ============================================================================\n// Message Extraction\n// ============================================================================\n\n/** Result from compact() - SessionManager adds uuid/parentUuid when saving */\nexport interface CompactionResult<T = unknown> {\n\tsummary: string;\n\tfirstKeptEntryId: string;\n\ttokensBefore: number;\n\t/** Extension-specific data (e.g., ArtifactIndex, version markers for structured compaction) */\n\tdetails?: T;\n}\n\n// ============================================================================\n// Types\n// ============================================================================\n\n// ============================================================================\n// Token calculation\n// ============================================================================\n\n/**\n * Calculate total context tokens from usage.\n * Uses the native totalTokens field when available, falls back to computing from components.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted and error messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (assistantMsg.stopReason !== \"aborted\" && assistantMsg.stopReason !== \"error\" && assistantMsg.usage) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/**\n * Find the last non-aborted assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/** Estimate context tokens after adding messages that have not been sent yet. */\nexport function estimateProjectedContextTokens(\n\tmessages: AgentMessage[],\n\tpendingMessages: AgentMessage[],\n): ContextUsageEstimate {\n\treturn estimateContextTokens([...messages, ...pendingMessages]);\n}\n\n// ============================================================================\n// Cut point detection\n// ============================================================================\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n\n/**\n * Find valid cut points: indices of user, assistant, custom, or bashExecution messages.\n * Never cut at tool results (they must follow their tool call).\n * When we cut at an assistant message with tool calls, its tool results follow it\n * and will be kept.\n * BashExecutionMessage is treated like a user message (user-initiated context).\n */\nfunction findValidCutPoints(entries: SessionEntry[], startIndex: number, endIndex: number): number[] {\n\tconst cutPoints: number[] = [];\n\tfor (let i = startIndex; i < endIndex; i++) {\n\t\tconst entry = entries[i];\n\t\tswitch (entry.type) {\n\t\t\tcase \"message\": {\n\t\t\t\tconst role = entry.message.role;\n\t\t\t\tswitch (role) {\n\t\t\t\t\tcase \"bashExecution\":\n\t\t\t\t\tcase \"custom\":\n\t\t\t\t\tcase \"branchSummary\":\n\t\t\t\t\tcase \"compactionSummary\":\n\t\t\t\t\tcase \"user\":\n\t\t\t\t\tcase \"assistant\":\n\t\t\t\t\t\tcutPoints.push(i);\n\t\t\t\t\t\tbreak;\n\t\t\t\t\tcase \"toolResult\":\n\t\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"thinking_level_change\":\n\t\t\tcase \"model_change\":\n\t\t\tcase \"compaction\":\n\t\t\tcase \"branch_summary\":\n\t\t\tcase \"custom\":\n\t\t\tcase \"custom_message\":\n\t\t\tcase \"label\":\n\t\t\tcase \"session_info\":\n\t\t\t\tbreak;\n\t\t}\n\n\t\t// branch_summary and custom_message are user-role messages, valid cut points\n\t\tif (entry.type === \"branch_summary\" || entry.type === \"custom_message\") {\n\t\t\tcutPoints.push(i);\n\t\t}\n\t}\n\treturn cutPoints;\n}\n\n/**\n * Find the user message (or bashExecution) that starts the turn containing the given entry index.\n * Returns -1 if no turn start found before the index.\n * BashExecutionMessage is treated like a user message for turn boundaries.\n */\nexport function findTurnStartIndex(entries: SessionEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let i = entryIndex; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\t// branch_summary and custom_message are user-role messages, can start a turn\n\t\tif (entry.type === \"branch_summary\" || entry.type === \"custom_message\") {\n\t\t\treturn i;\n\t\t}\n\t\tif (entry.type === \"message\") {\n\t\t\tconst role = entry.message.role;\n\t\t\tif (role === \"user\" || role === \"bashExecution\") {\n\t\t\t\treturn i;\n\t\t\t}\n\t\t}\n\t}\n\treturn -1;\n}\n\nexport interface CutPointResult {\n\t/** Index of first entry to keep */\n\tfirstKeptEntryIndex: number;\n\t/** Index of user message that starts the turn being split, or -1 if not splitting */\n\tturnStartIndex: number;\n\t/** Whether this cut splits a turn (cut point is not a user message) */\n\tisSplitTurn: boolean;\n}\n\n/**\n * Find the cut point in session entries that keeps approximately `keepRecentTokens`.\n *\n * Algorithm: Walk backwards from newest, accumulating estimated message sizes.\n * Stop when we've accumulated >= keepRecentTokens. Cut at that point.\n *\n * Can cut at user OR assistant messages (never tool results). When cutting at an\n * assistant message with tool calls, its tool results come after and will be kept.\n *\n * Returns CutPointResult with:\n * - firstKeptEntryIndex: the entry index to start keeping from\n * - turnStartIndex: if cutting mid-turn, the user message that started that turn\n * - isSplitTurn: whether we're cutting in the middle of a turn\n *\n * Only considers entries between `startIndex` and `endIndex` (exclusive).\n */\nexport function findCutPoint(\n\tentries: SessionEntry[],\n\tstartIndex: number,\n\tendIndex: number,\n\tkeepRecentTokens: number,\n): CutPointResult {\n\tconst cutPoints = findValidCutPoints(entries, startIndex, endIndex);\n\n\tif (cutPoints.length === 0) {\n\t\treturn { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };\n\t}\n\n\t// Walk backwards from newest, accumulating estimated message sizes\n\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0]; // Default: keep from first message (not header)\n\n\tfor (let i = endIndex - 1; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type !== \"message\") continue;\n\n\t\t// Estimate this message's size\n\t\tconst messageTokens = estimateTokens(entry.message);\n\t\taccumulatedTokens += messageTokens;\n\n\t\t// Check if we've exceeded the budget\n\t\tif (accumulatedTokens >= keepRecentTokens) {\n\t\t\t// Find the closest valid cut point at or after this entry\n\t\t\tlet matched = false;\n\t\t\tfor (let c = 0; c < cutPoints.length; c++) {\n\t\t\t\tif (cutPoints[c] >= i) {\n\t\t\t\t\tcutIndex = cutPoints[c];\n\t\t\t\t\tmatched = true;\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (!matched) {\n\t\t\t\t// The budget-breaking entry is not itself cuttable (e.g. a huge\n\t\t\t\t// trailing tool result): fall back to its issuing assistant\n\t\t\t\t// message so compaction still makes progress and the call\n\t\t\t\t// stays with its result as one tail turn.\n\t\t\t\tfor (let j = i; j >= startIndex; j--) {\n\t\t\t\t\tconst candidate = entries[j];\n\t\t\t\t\tif (candidate.type === \"message\" && candidate.message.role === \"assistant\") {\n\t\t\t\t\t\tcutIndex = j;\n\t\t\t\t\t\tbreak;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Scan backwards from cutIndex to include any non-message entries (bash, settings, etc.)\n\twhile (cutIndex > startIndex) {\n\t\tconst prevEntry = entries[cutIndex - 1];\n\t\t// Stop at session header or compaction boundaries\n\t\tif (prevEntry.type === \"compaction\") {\n\t\t\tbreak;\n\t\t}\n\t\tif (prevEntry.type === \"message\") {\n\t\t\t// Stop if we hit any message\n\t\t\tbreak;\n\t\t}\n\t\t// Include this non-message entry (bash, settings change, etc.)\n\t\tcutIndex--;\n\t}\n\n\t// Determine if this is a split turn\n\tconst cutEntry = entries[cutIndex];\n\tconst isUserMessage = cutEntry.type === \"message\" && cutEntry.message.role === \"user\";\n\tconst turnStartIndex = isUserMessage ? -1 : findTurnStartIndex(entries, cutIndex, startIndex);\n\n\treturn {\n\t\tfirstKeptEntryIndex: cutIndex,\n\t\tturnStartIndex,\n\t\tisSplitTurn: !isUserMessage && turnStartIndex !== -1,\n\t};\n}\n\n// ============================================================================\n// Summarization\n// ============================================================================\n\nconst SUMMARIZATION_PROMPT = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.\n\nUse this EXACT format:\n\n## Goal\n[What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]\n\n## Constraints & Preferences\n- [Any constraints, preferences, or requirements mentioned by user]\n- [Or \"(none)\" if none were mentioned]\n\n## Progress\n### Done\n- [x] [Completed tasks/changes]\n\n### In Progress\n- [ ] [Current work]\n\n### Blocked\n- [Issues preventing progress, if any]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale]\n\n## Next Steps\n1. [Ordered list of what should happen next]\n\n## Critical Context\n- [Any data, examples, or references needed to continue]\n- [Or \"(none)\" if not applicable]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages. Do not create a Preserved Rules & Invariants section; OMK injects it deterministically after generation.`;\n\nconst UPDATE_SUMMARIZATION_PROMPT = `The messages above are NEW conversation messages to incorporate into the existing summary provided in <previous-summary> tags.\n\nUpdate the existing structured summary with new information. RULES:\n- PRESERVE all existing information from the previous summary\n- ADD new progress, decisions, and context from the new messages\n- UPDATE the Progress section: move items from \"In Progress\" to \"Done\" when completed\n- UPDATE \"Next Steps\" based on what was accomplished\n- PRESERVE exact file paths, function names, and error messages\n- If something is no longer relevant, you may remove it\n\nUse this EXACT format:\n\n## Goal\n[Preserve existing goals, add new ones if the task expanded]\n\n## Constraints & Preferences\n- [Preserve existing, add new ones discovered]\n\n## Progress\n### Done\n- [x] [Include previously done items AND newly completed items]\n\n### In Progress\n- [ ] [Current work - update based on progress]\n\n### Blocked\n- [Current blockers - remove if resolved]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale] (preserve all previous, add new)\n\n## Next Steps\n1. [Update based on current state]\n\n## Critical Context\n- [Preserve important context, add new if needed]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages. Do not create a Preserved Rules & Invariants section; OMK injects it deterministically after generation.`;\n\nfunction createSummarizationOptions(\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n): SimpleStreamOptions {\n\tconst options: SimpleStreamOptions = { maxTokens, signal, apiKey, headers };\n\tif (model.reasoning && thinkingLevel && thinkingLevel !== \"off\") {\n\t\toptions.reasoning = thinkingLevel;\n\t}\n\treturn options;\n}\n\n/**\n * Shared choke point for every compaction/branch-summary summarization call. Wraps the\n * single LLM call in {@link retryAssistantCall} so transient stream drops (e.g.\n * `terminated`, socket close) honor the configured retry policy instead of failing\n * the whole compaction on the first attempt. Deterministic errors and aborts return\n * immediately (see {@link retryAssistantCall}).\n */\nexport async function completeSummarization(\n\tmodel: Model<any>,\n\tcontext: Context,\n\toptions: SimpleStreamOptions,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\t// Summaries are standalone requests, so isolate routing and avoid cache writes that cannot be reused.\n\tconst requestOptions: SimpleStreamOptions = {\n\t\t...options,\n\t\tcacheRetention: \"none\",\n\t\tsessionId: uuidv7(),\n\t};\n\tconst produce = async (): Promise<AssistantMessage> =>\n\t\tstreamFn\n\t\t\t? (await streamFn(model, context, requestOptions)).result()\n\t\t\t: completeSimple(model, context, requestOptions);\n\treturn retryAssistantCall(produce, retry, requestOptions.signal, callbacks);\n}\n\n/**\n * Generate a summary of the conversation using the LLM.\n * If previousSummary is provided, uses the update prompt to merge.\n */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<string> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.8 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t);\n\n\t// Use update prompt if we have a previous summary, otherwise initial prompt\n\tlet basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tbasePrompt = `${basePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\n\t// Serialize conversation to text so model doesn't try to continue it\n\t// Convert to LLM messages first (handles custom types like bashExecution, custom, etc.)\n\tconst llmMessages = convertToLlm(currentMessages);\n\tconst conversationText = serializeConversation(llmMessages);\n\n\t// Build the prompt with conversation wrapped in tags\n\tlet promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n`;\n\tif (previousSummary) {\n\t\tpromptText += `<previous-summary>\\n${previousSummary}\\n</previous-summary>\\n\\n`;\n\t}\n\tpromptText += basePrompt;\n\n\tconst summarizationMessages = [\n\t\t{\n\t\t\trole: \"user\" as const,\n\t\t\tcontent: [{ type: \"text\" as const, text: promptText }],\n\t\t\ttimestamp: Date.now(),\n\t\t},\n\t];\n\n\tconst completionOptions = createSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel);\n\n\tconst response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcompletionOptions,\n\t\tstreamFn,\n\t\tretry,\n\t\tcallbacks,\n\t);\n\n\tif (response.stopReason === \"error\") {\n\t\tthrow new Error(`Summarization failed: ${response.errorMessage || \"Unknown error\"}`);\n\t}\n\n\tconst textContent = response.content\n\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n\n\treturn textContent;\n}\n\n// ============================================================================\n// Compaction Preparation (for extensions)\n// ============================================================================\n\nexport interface CompactionPreparation {\n\t/** UUID of first entry to keep */\n\tfirstKeptEntryId: string;\n\t/** Messages that will be summarized and discarded */\n\tmessagesToSummarize: AgentMessage[];\n\t/** Messages that will be turned into turn prefix summary (if splitting) */\n\tturnPrefixMessages: AgentMessage[];\n\t/** Whether this is a split turn (cut point in middle of turn) */\n\tisSplitTurn: boolean;\n\ttokensBefore: number;\n\t/** Summary from previous compaction, for iterative update */\n\tpreviousSummary?: string;\n\t/** Source-verified rule history from the prior default compaction. */\n\tpreviousRuleHistory?: CompactionRuleHistory;\n\t/** Exact session entries covered by this compaction's rule extraction. */\n\tcurrentRuleEntries?: readonly SessionEntry[];\n\t/** File operations extracted from messagesToSummarize */\n\tfileOps: FileOperations;\n\t/** Compaction settions from settings.jsonl\t*/\n\tsettings: CompactionSettings;\n}\n\nexport function prepareCompaction(\n\tpathEntries: SessionEntry[],\n\tsettings: CompactionSettings,\n): CompactionPreparation | undefined {\n\tif (pathEntries.length > 0 && pathEntries[pathEntries.length - 1].type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\n\tlet prevCompactionIndex = -1;\n\tfor (let i = pathEntries.length - 1; i >= 0; i--) {\n\t\tif (pathEntries[i].type === \"compaction\") {\n\t\t\tprevCompactionIndex = i;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\tlet previousSummary: string | undefined;\n\tlet previousRuleHistory: CompactionRuleHistory | undefined;\n\tlet boundaryStart = 0;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tpreviousRuleHistory = buildCompactionRuleHistory(prevCompaction, pathEntries);\n\t\tconst firstKeptEntryIndex = pathEntries.findIndex((entry) => entry.id === prevCompaction.firstKeptEntryId);\n\t\tboundaryStart = firstKeptEntryIndex >= 0 ? firstKeptEntryIndex : prevCompactionIndex + 1;\n\t}\n\tconst boundaryEnd = pathEntries.length;\n\n\tconst tokensBefore = estimateContextTokens(buildSessionContext(pathEntries).messages).tokens;\n\n\tconst cutPoint = findCutPoint(pathEntries, boundaryStart, boundaryEnd, settings.keepRecentTokens);\n\n\t// Get UUID of first kept entry\n\tconst firstKeptEntry = pathEntries[cutPoint.firstKeptEntryIndex];\n\tif (!firstKeptEntry?.id) {\n\t\treturn undefined; // Session needs migration\n\t}\n\tconst firstKeptEntryId = firstKeptEntry.id;\n\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\n\t// Messages to summarize (will be discarded after summary)\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let i = boundaryStart; i < historyEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\tif (msg) messagesToSummarize.push(msg);\n\t}\n\n\t// Messages for turn prefix summary (if splitting a turn)\n\tconst turnPrefixMessages: AgentMessage[] = [];\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (let i = cutPoint.turnStartIndex; i < cutPoint.firstKeptEntryIndex; i++) {\n\t\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\t\tif (msg) turnPrefixMessages.push(msg);\n\t\t}\n\t}\n\n\t// Extract file operations from messages and previous compaction\n\tconst fileOps = extractCompactionFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\n\tconst currentRuleEntries = [\n\t\t...pathEntries.slice(boundaryStart, historyEnd),\n\t\t...(cutPoint.isSplitTurn ? pathEntries.slice(cutPoint.turnStartIndex, cutPoint.firstKeptEntryIndex) : []),\n\t];\n\n\t// Also extract file ops from turn prefix if splitting\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (const msg of turnPrefixMessages) {\n\t\t\textractFileOpsFromMessage(msg, fileOps);\n\t\t}\n\t}\n\n\treturn {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tpreviousRuleHistory,\n\t\tcurrentRuleEntries,\n\t\tfileOps,\n\t\tsettings,\n\t};\n}\n\n// ============================================================================\n// Main compaction function\n// ============================================================================\n\nconst TURN_PREFIX_SUMMARIZATION_PROMPT = `This is the PREFIX of a turn that was too large to keep. The SUFFIX (recent work) is retained.\n\nSummarize the prefix to provide context for the retained suffix:\n\n## Original Request\n[What did the user ask for in this turn?]\n\n## Early Progress\n- [Key decisions and work done in the prefix]\n\n## Context for Suffix\n- [Information needed to understand the retained recent work]\n\nBe concise. Focus on what's needed to understand the kept suffix.`;\n\n/**\n * Generate summaries for compaction using prepared data.\n * Returns CompactionResult - SessionManager adds uuid/parentUuid when saving.\n *\n * When {@link failoverModels} is provided and the primary model fails with a\n * quota/billing exhaustion error (same-model retry is useless until reset),\n * the whole summarization is retried once with each candidate in order. Only\n * quota-class errors are absorbed; any other failure surfaces immediately.\n *\n * @param preparation - Pre-calculated preparation from prepareCompaction()\n * @param customInstructions - Optional custom focus for the summary\n */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tfailoverModels?: readonly Model<any>[],\n): Promise<CompactionResult> {\n\ttry {\n\t\treturn await compactWithModel(\n\t\t\tpreparation,\n\t\t\tmodel,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tcustomInstructions,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tif (!isQuotaExhaustionMessage(message) || !failoverModels || failoverModels.length === 0) throw error;\n\t\tfor (const candidate of failoverModels) {\n\t\t\tif (candidate.provider === model.provider && candidate.id === model.id) continue;\n\t\t\tif (signal?.aborted) throw error;\n\t\t\ttry {\n\t\t\t\treturn await compactWithModel(\n\t\t\t\t\tpreparation,\n\t\t\t\t\tcandidate,\n\t\t\t\t\tapiKey,\n\t\t\t\t\theaders,\n\t\t\t\t\tcustomInstructions,\n\t\t\t\t\tsignal,\n\t\t\t\t\tthinkingLevel,\n\t\t\t\t\tstreamFn,\n\t\t\t\t\tretry,\n\t\t\t\t\tcallbacks,\n\t\t\t\t);\n\t\t\t} catch (candidateError) {\n\t\t\t\tconst candidateMessage = candidateError instanceof Error ? candidateError.message : String(candidateError);\n\t\t\t\tif (!isQuotaExhaustionMessage(candidateMessage)) throw candidateError;\n\t\t\t\t// Candidate is also quota-blocked — try the next one.\n\t\t\t}\n\t\t}\n\t\t// Every candidate is quota-blocked too: surface the primary failure.\n\t\tthrow error;\n\t}\n}\n\n/** Single-model compaction attempt (no failover). Implementation of {@link compact}. */\nasync function compactWithModel(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<CompactionResult> {\n\tconst {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tpreviousRuleHistory,\n\t\tcurrentRuleEntries,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\t// Generate summaries (can be parallel if both needed) and merge into one\n\tlet summary: string;\n\n\tif (isSplitTurn && turnPrefixMessages.length > 0) {\n\t\t// Generate both summaries in parallel\n\t\tconst [historyResult, turnPrefixResult] = await Promise.all([\n\t\t\tmessagesToSummarize.length > 0\n\t\t\t\t? generateSummary(\n\t\t\t\t\t\tmessagesToSummarize,\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tsettings.reserveTokens,\n\t\t\t\t\t\tapiKey,\n\t\t\t\t\t\theaders,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\tcustomInstructions,\n\t\t\t\t\t\tpreviousSummary,\n\t\t\t\t\t\tthinkingLevel,\n\t\t\t\t\t\tstreamFn,\n\t\t\t\t\t\tretry,\n\t\t\t\t\t\tcallbacks,\n\t\t\t\t\t)\n\t\t\t\t: Promise.resolve(\"No prior history.\"),\n\t\t\tgenerateTurnPrefixSummary(\n\t\t\t\tturnPrefixMessages,\n\t\t\t\tmodel,\n\t\t\t\tsettings.reserveTokens,\n\t\t\t\tapiKey,\n\t\t\t\theaders,\n\t\t\t\tsignal,\n\t\t\t\tthinkingLevel,\n\t\t\t\tstreamFn,\n\t\t\t\tretry,\n\t\t\t\tcallbacks,\n\t\t\t),\n\t\t]);\n\t\t// Merge into single summary\n\t\tsummary = `${historyResult}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult}`;\n\t} else {\n\t\t// Just generate history summary\n\t\tsummary = await generateSummary(\n\t\t\tmessagesToSummarize,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t}\n\n\tconst triage = applyCompactionKnowledgeTriage({\n\t\tgeneratedSummary: summary,\n\t\tcurrentMessages: [...messagesToSummarize, ...turnPrefixMessages],\n\t\tcurrentEntries: currentRuleEntries,\n\t\tpreviousRules: previousRuleHistory?.rules,\n\t\tpreviousSummary,\n\t\tpreviousEntries: previousRuleHistory?.entries,\n\t});\n\tsummary = triage.summary;\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\tif (!firstKeptEntryId) {\n\t\tthrow new Error(\"First kept entry has no UUID - session may need migration\");\n\t}\n\n\treturn {\n\t\tsummary,\n\t\tfirstKeptEntryId,\n\t\ttokensBefore,\n\t\tdetails: { readFiles, modifiedFiles, preservedRules: [...triage.preservedRules] } as CompactionDetails,\n\t};\n}\n\n/**\n * Generate a summary for a turn prefix (when splitting a turn).\n */\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<string> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.5 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t); // Smaller budget for turn prefix\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversation(llmMessages);\n\tconst promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;\n\tconst summarizationMessages = [\n\t\t{\n\t\t\trole: \"user\" as const,\n\t\t\tcontent: [{ type: \"text\" as const, text: promptText }],\n\t\t\ttimestamp: Date.now(),\n\t\t},\n\t];\n\n\tconst response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcreateSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel),\n\t\tstreamFn,\n\t\tretry,\n\t\tcallbacks,\n\t);\n\n\tif (response.stopReason === \"error\") {\n\t\tthrow new Error(`Turn prefix summarization failed: ${response.errorMessage || \"Unknown error\"}`);\n\t}\n\n\treturn response.content\n\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n}\n"]}

package/dist/core/compaction/compaction.js

@@ -6,19 +6,15 @@
  */
 import { uuidv7 } from "omk-agent-core";
 import { completeSimple, retryAssistantCall } from "omk-ai";
-import { computeReservedTokenBudget } from "../context-budget-reserved-tokens.js";
+export { DEFAULT_COMPACTION_MAX_USAGE_RATIO, DEFAULT_COMPACTION_SETTINGS, getCompactionHeadroomThreshold, shouldCompact, } from "./compaction-headroom.js";
 import { convertToLlm } from "../messages.js";
 import { isQuotaExhaustionMessage } from "../provider-resilience.js";
 import { buildSessionContext } from "../session-manager.js";
 import { applyCompactionKnowledgeTriage, buildCompactionRuleHistory, } from "./knowledge-triage.js";
 import { computeFileLists, extractCompactionFileOperations, extractFileOpsFromMessage, formatFileOperations, getMessageFromEntryForCompaction, SUMMARIZATION_SYSTEM_PROMPT, serializeConversation, } from "./utils.js";
-export const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;
-export const DEFAULT_COMPACTION_SETTINGS = {
-    enabled: true,
-    reserveTokens: 16384,
-    keepRecentTokens: 20000,
-    maxUsageRatio: DEFAULT_COMPACTION_MAX_USAGE_RATIO,
-};
+// ============================================================================
+// Types
+// ============================================================================
 // ============================================================================
 // Token calculation
 // ============================================================================
@@ -98,45 +94,6 @@ export function estimateContextTokens(messages) {
 export function estimateProjectedContextTokens(messages, pendingMessages) {
     return estimateContextTokens([...messages, ...pendingMessages]);
 }
-/** Return the earliest compaction threshold from ratio-based headroom and absolute reserve. */
-export function getCompactionHeadroomThreshold(contextWindow, settings) {
-    if (!settings.enabled || !Number.isFinite(contextWindow) || contextWindow <= 0) {
-        return undefined;
-    }
-    const windowTokens = Math.floor(contextWindow);
-    const reserveTokens = Number.isFinite(settings.reserveTokens) ? Math.max(0, Math.floor(settings.reserveTokens)) : 0;
-    const configuredMaxUsageRatio = settings.maxUsageRatio ?? DEFAULT_COMPACTION_MAX_USAGE_RATIO;
-    const maxUsageRatio = Number.isFinite(configuredMaxUsageRatio) && configuredMaxUsageRatio > 0 && configuredMaxUsageRatio < 1
-        ? configuredMaxUsageRatio
-        : DEFAULT_COMPACTION_MAX_USAGE_RATIO;
-    const maxUsageRatioTokens = Math.max(1, Math.floor(windowTokens * maxUsageRatio));
-    const reservedBudget = computeReservedTokenBudget({
-        modelContextWindow: windowTokens,
-        systemPromptTokens: 0,
-        reservedOutputTokens: settings.reservedOutputTokens ?? reserveTokens,
-        reservedToolResultTokens: settings.reservedToolResultTokens ?? 0,
-        safetyMarginTokens: settings.safetyMarginTokens ?? 0,
-        imageReserveTokens: settings.imageReserveTokens ?? 0,
-    });
-    const reserveBoundaryTokens = reservedBudget.overflow || reservedBudget.effectiveBudget === 0
-        ? maxUsageRatioTokens
-        : Math.max(1, reservedBudget.effectiveBudget);
-    const triggerTokens = Math.min(maxUsageRatioTokens, reserveBoundaryTokens);
-    return {
-        triggerTokens,
-        headroomTokens: windowTokens - triggerTokens,
-        maxUsageRatioTokens,
-        reserveBoundaryTokens,
-        limitedBy: reserveBoundaryTokens <= maxUsageRatioTokens ? "reserve_tokens" : "max_usage_ratio",
-    };
-}
-/**
- * Check if compaction should trigger based on context usage.
- */
-export function shouldCompact(contextTokens, contextWindow, settings) {
-    const threshold = getCompactionHeadroomThreshold(contextWindow, settings);
-    return threshold !== undefined && Number.isFinite(contextTokens) && contextTokens >= threshold.triggerTokens;
-}
 // ============================================================================
 // Cut point detection
 // ============================================================================
@@ -299,12 +256,27 @@ export function findCutPoint(entries, startIndex, endIndex, keepRecentTokens) {
         // Check if we've exceeded the budget
         if (accumulatedTokens >= keepRecentTokens) {
             // Find the closest valid cut point at or after this entry
+            let matched = false;
             for (let c = 0; c < cutPoints.length; c++) {
                 if (cutPoints[c] >= i) {
                     cutIndex = cutPoints[c];
+                    matched = true;
                     break;
                 }
             }
+            if (!matched) {
+                // The budget-breaking entry is not itself cuttable (e.g. a huge
+                // trailing tool result): fall back to its issuing assistant
+                // message so compaction still makes progress and the call
+                // stays with its result as one tail turn.
+                for (let j = i; j >= startIndex; j--) {
+                    const candidate = entries[j];
+                    if (candidate.type === "message" && candidate.message.role === "assistant") {
+                        cutIndex = j;
+                        break;
+                    }
+                }
+            }
             break;
         }
     }