omk-agent-core 0.98.0 → 0.98.1

package/dist/harness/compaction/compaction.d.ts

@@ -1,7 +1,8 @@
-import type { Message, Model, Usage } from "omk-ai";
+import type { Message, Model, RetryCallbacks, RetryPolicy, Usage } from "omk-ai";
 import type { AgentMessage, ThinkingLevel } from "../../types.ts";
 import { CompactionError, type Result, type SessionTreeEntry } from "../types.ts";
 import { type FileOperations } from "./utils.ts";
+export { SUMMARIZATION_SYSTEM_PROMPT } from "./summarization-prompts.ts";
 /** File-operation details stored on generated compaction entries. */
 export interface CompactionDetails {
     /** Files read in the compacted history. */
@@ -20,6 +21,10 @@ export interface CompactionResult<T = unknown> {
     /** Optional implementation-specific details stored with the compaction entry. */
     details?: T;
 }
+export interface SummarizationRetryOptions {
+    readonly retry?: RetryPolicy;
+    readonly callbacks?: RetryCallbacks;
+}
 /** Compaction thresholds and retention settings. */
 export interface CompactionSettings {
     /** Enable automatic compaction decisions. */
@@ -87,9 +92,8 @@ export interface CutPointResult {
 }
 /** Find the compaction cut point that keeps approximately the requested recent-token budget. */
 export declare function findCutPoint(entries: SessionTreeEntry[], startIndex: number, endIndex: number, keepRecentTokens: number): CutPointResult;
-export declare const SUMMARIZATION_SYSTEM_PROMPT = "You are a context summarization assistant. Your task is to read a conversation between a user and an AI coding assistant, then produce a structured summary following the exact format specified.\n\nDo NOT continue the conversation. Do NOT respond to any questions in the conversation. ONLY output the structured summary.";
 /** Generate or update a conversation summary for compaction. */
-export declare function generateSummary(currentMessages: AgentMessage[], model: Model<any>, reserveTokens: number, apiKey: string, headers?: Record<string, string>, signal?: AbortSignal, customInstructions?: string, previousSummary?: string, thinkingLevel?: ThinkingLevel): Promise<Result<string, CompactionError>>;
+export declare function generateSummary(currentMessages: AgentMessage[], model: Model<any>, reserveTokens: number, apiKey: string, headers?: Record<string, string>, signal?: AbortSignal, customInstructions?: string, previousSummary?: string, thinkingLevel?: ThinkingLevel, retryOptions?: SummarizationRetryOptions): Promise<Result<string, CompactionError>>;
 /** Prepared inputs for a compaction run. */
 export interface CompactionPreparation {
     /** Entry id where retained history starts. */
@@ -126,5 +130,5 @@ export { serializeConversation } from "./utils.ts";
  */
 export declare function serializeConversationBounded(messages: Message[], model: Model<any>, maxOutputTokens: number): string;
 /** Generate compaction summary data from prepared session history. */
-export declare function compact(preparation: CompactionPreparation, model: Model<any>, apiKey: string, headers?: Record<string, string>, customInstructions?: string, signal?: AbortSignal, thinkingLevel?: ThinkingLevel): Promise<Result<CompactionResult, CompactionError>>;
+export declare function compact(preparation: CompactionPreparation, model: Model<any>, apiKey: string, headers?: Record<string, string>, customInstructions?: string, signal?: AbortSignal, thinkingLevel?: ThinkingLevel, retryOptions?: SummarizationRetryOptions): Promise<Result<CompactionResult, CompactionError>>;
 //# sourceMappingURL=compaction.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/harness/compaction/compaction.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAkC,OAAO,EAAE,KAAK,EAAe,KAAK,EAAE,MAAM,QAAQ,CAAC;AAEjG,OAAO,KAAK,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAQlE,OAAO,EAAwB,eAAe,EAAW,KAAK,MAAM,EAAE,KAAK,gBAAgB,EAAE,MAAM,aAAa,CAAC;AACjH,OAAO,EAIN,KAAK,cAAc,EAGnB,MAAM,YAAY,CAAC;AAEpB,qEAAqE;AACrE,MAAM,WAAW,iBAAiB;IACjC,2CAA2C;IAC3C,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,+CAA+C;IAC/C,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AA8DD,6EAA6E;AAC7E,MAAM,WAAW,gBAAgB,CAAC,CAAC,GAAG,OAAO;IAC5C,sEAAsE;IACtE,OAAO,EAAE,MAAM,CAAC;IAChB,8CAA8C;IAC9C,gBAAgB,EAAE,MAAM,CAAC;IACzB,kDAAkD;IAClD,YAAY,EAAE,MAAM,CAAC;IACrB,iFAAiF;IACjF,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AAED,oDAAoD;AACpD,MAAM,WAAW,kBAAkB;IAClC,6CAA6C;IAC7C,OAAO,EAAE,OAAO,CAAC;IACjB,qDAAqD;IACrD,aAAa,EAAE,MAAM,CAAC;IACtB,kEAAkE;IAClE,gBAAgB,EAAE,MAAM,CAAC;IACzB,qFAAqF;IACrF,aAAa,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,oEAAoE;AACpE,eAAO,MAAM,kCAAkC,MAAM,CAAC;AAEtD,uDAAuD;AACvD,eAAO,MAAM,2BAA2B,EAAE,kBAKzC,CAAC;AAEF,0DAA0D;AAC1D,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAWD,kFAAkF;AAClF,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,gBAAgB,EAAE,GAAG,KAAK,GAAG,SAAS,CASpF;AAED,wDAAwD;AACxD,MAAM,WAAW,oBAAoB;IACpC,sCAAsC;IACtC,MAAM,EAAE,MAAM,CAAC;IACf,gEAAgE;IAChE,WAAW,EAAE,MAAM,CAAC;IACpB,oEAAoE;IACpE,cAAc,EAAE,MAAM,CAAC;IACvB,0EAA0E;IAC1E,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD,gFAAgF;AAChF,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,8DAA8D;AAC9D,MAAM,WAAW,2BAA2B;IAC3C,oDAAoD;IACpD,aAAa,EAAE,MAAM,CAAC;IACtB,6CAA6C;IAC7C,cAAc,EAAE,MAAM,CAAC;IACvB,4DAA4D;IAC5D,mBAAmB,EAAE,MAAM,CAAC;IAC5B,gEAAgE;IAChE,qBAAqB,EAAE,MAAM,CAAC;IAC9B,wCAAwC;IACxC,SAAS,EAAE,uBAAuB,CAAC;CACnC;AAED,+FAA+F;AAC/F,wBAAgB,8BAA8B,CAC7C,aAAa,EAAE,MAAM,EACrB,QAAQ,EAAE,kBAAkB,GAC1B,2BAA2B,GAAG,SAAS,CAwBzC;AAED,gFAAgF;AAChF,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAsBD,qFAAqF;AACrF,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAsC5D;AA2CD,8EAA8E;AAC9E,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,gBAAgB,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAc9G;AAED,yCAAyC;AACzC,MAAM,WAAW,cAAc;IAC9B,0DAA0D;IAC1D,mBAAmB,EAAE,MAAM,CAAC;IAC5B,8EAA8E;IAC9E,cAAc,EAAE,MAAM,CAAC;IACvB,iEAAiE;IACjE,WAAW,EAAE,OAAO,CAAC;CACrB;AAED,gGAAgG;AAChG,wBAAgB,YAAY,CAC3B,OAAO,EAAE,gBAAgB,EAAE,EAC3B,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CA2ChB;AAED,eAAO,MAAM,2BAA2B,oUAEmF,CAAC;AA0E5H,gEAAgE;AAChE,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,EACd,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,GAC3B,OAAO,CAAC,MAAM,CAAC,MAAM,EAAE,eAAe,CAAC,CAAC,CAqD1C;AAED,4CAA4C;AAC5C,MAAM,WAAW,qBAAqB;IACrC,8CAA8C;IAC9C,gBAAgB,EAAE,MAAM,CAAC;IACzB,oDAAoD;IACpD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,wCAAwC;IACxC,WAAW,EAAE,OAAO,CAAC;IACrB,kDAAkD;IAClD,YAAY,EAAE,MAAM,CAAC;IACrB,8DAA8D;IAC9D,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,yDAAyD;IACzD,OAAO,EAAE,cAAc,CAAC;IACxB,2CAA2C;IAC3C,QAAQ,EAAE,kBAAkB,CAAC;CAC7B;AAED,qGAAqG;AACrG,wBAAgB,iBAAiB,CAChC,WAAW,EAAE,gBAAgB,EAAE,EAC/B,QAAQ,EAAE,kBAAkB,GAC1B,MAAM,CAAC,qBAAqB,GAAG,SAAS,EAAE,eAAe,CAAC,CAsE5D;AAiBD,OAAO,EAAE,qBAAqB,EAAE,MAAM,YAAY,CAAC;AAsBnD;;;;;;;;;;;GAWG;AACH,wBAAgB,4BAA4B,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EAAE,eAAe,EAAE,MAAM,GAAG,MAAM,CAwBpH;AAED,sEAAsE;AACtE,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,MAAM,EAAE,MAAM,EACd,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,GAC3B,OAAO,CAAC,MAAM,CAAC,gBAAgB,EAAE,eAAe,CAAC,CAAC,CAuEpD","sourcesContent":["import type { AssistantMessage, ImageContent, Message, Model, TextContent, Usage } from \"omk-ai\";\nimport { completeSimple } from \"omk-ai\";\nimport type { AgentMessage, ThinkingLevel } from \"../../types.ts\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport { buildSessionContext } from \"../session/session.ts\";\nimport { type CompactionEntry, CompactionError, err, ok, type Result, type SessionTreeEntry } from \"../types.ts\";\nimport {\n\tcomputeFileLists,\n\tcreateFileOps,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tserializeConversation,\n} from \"./utils.ts\";\n\n/** File-operation details stored on generated compaction entries. */\nexport interface CompactionDetails {\n\t/** Files read in the compacted history. */\n\treadFiles: string[];\n\t/** Files modified in the compacted history. */\n\tmodifiedFiles: string[];\n}\nfunction safeJsonStringify(value: unknown): string {\n\ttry {\n\t\treturn JSON.stringify(value) ?? \"undefined\";\n\t} catch {\n\t\treturn \"[unserializable]\";\n\t}\n}\n\nfunction extractFileOperations(\n\tmessages: AgentMessage[],\n\tentries: SessionTreeEntry[],\n\tprevCompactionIndex: number,\n): FileOperations {\n\tconst fileOps = createFileOps();\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = entries[prevCompactionIndex] as CompactionEntry;\n\t\tif (!prevCompaction.fromHook && prevCompaction.details) {\n\t\t\tconst details = prevCompaction.details as CompactionDetails;\n\t\t\tif (Array.isArray(details.readFiles)) {\n\t\t\t\tfor (const f of details.readFiles) fileOps.read.add(f);\n\t\t\t}\n\t\t\tif (Array.isArray(details.modifiedFiles)) {\n\t\t\t\tfor (const f of details.modifiedFiles) fileOps.edited.add(f);\n\t\t\t}\n\t\t}\n\t}\n\tfor (const msg of messages) {\n\t\textractFileOpsFromMessage(msg, fileOps);\n\t}\n\n\treturn fileOps;\n}\nfunction getMessageFromEntry(entry: SessionTreeEntry): AgentMessage | undefined {\n\tif (entry.type === \"message\") {\n\t\treturn entry.message as AgentMessage;\n\t}\n\tif (entry.type === \"custom_message\") {\n\t\treturn createCustomMessage(\n\t\t\tentry.customType,\n\t\t\tentry.content as string | (TextContent | ImageContent)[],\n\t\t\tentry.display,\n\t\t\tentry.details,\n\t\t\tentry.timestamp,\n\t\t);\n\t}\n\tif (entry.type === \"branch_summary\") {\n\t\treturn createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);\n\t}\n\tif (entry.type === \"compaction\") {\n\t\treturn createCompactionSummaryMessage(entry.summary, entry.tokensBefore, entry.timestamp);\n\t}\n\treturn undefined;\n}\n\nfunction getMessageFromEntryForCompaction(entry: SessionTreeEntry): AgentMessage | undefined {\n\tif (entry.type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\treturn getMessageFromEntry(entry);\n}\n\n/** Generated compaction data ready to be persisted as a compaction entry. */\nexport interface CompactionResult<T = unknown> {\n\t/** Summary text that replaces compacted history in future context. */\n\tsummary: string;\n\t/** Entry id where retained history starts. */\n\tfirstKeptEntryId: string;\n\t/** Estimated context tokens before compaction. */\n\ttokensBefore: number;\n\t/** Optional implementation-specific details stored with the compaction entry. */\n\tdetails?: T;\n}\n\n/** Compaction thresholds and retention settings. */\nexport interface CompactionSettings {\n\t/** Enable automatic compaction decisions. */\n\tenabled: boolean;\n\t/** Tokens reserved for summary prompt and output. */\n\treserveTokens: number;\n\t/** Approximate recent-context tokens to keep after compaction. */\n\tkeepRecentTokens: number;\n\t/** Maximum fraction of the context window to use before compaction. Default: 0.9. */\n\tmaxUsageRatio?: number;\n}\n\n/** Default context-window ratio used before compaction triggers. */\nexport const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;\n\n/** Default compaction settings used by the harness. */\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/** Calculate total context tokens from provider usage. */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\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/** Return usage from the last successful assistant message in session entries. */\nexport function getLastAssistantUsage(entries: SessionTreeEntry[]): 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 as AgentMessage);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/** Estimated context-token usage for a message list. */\nexport interface ContextUsageEstimate {\n\t/** Estimated total context tokens. */\n\ttokens: number;\n\t/** Tokens reported by the most recent assistant usage block. */\n\tusageTokens: number;\n\t/** Estimated tokens after the most recent assistant usage block. */\n\ttrailingTokens: number;\n\t/** Index of the message that provided usage, or null when none exists. */\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/** Estimate context tokens for messages using provider usage when available. */\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\n/** Threshold calculation details for automatic compaction. */\nexport interface CompactionHeadroomThreshold {\n\t/** Context-token count that triggers compaction. */\n\ttriggerTokens: number;\n\t/** Tokens kept free at the trigger point. */\n\theadroomTokens: number;\n\t/** Boundary derived from the configured max usage ratio. */\n\tmaxUsageRatioTokens: number;\n\t/** Boundary derived from the absolute reserve token setting. */\n\treserveBoundaryTokens: number;\n\t/** Which boundary triggered earlier. */\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 reserveBoundaryTokens =\n\t\treserveTokens >= windowTokens ? maxUsageRatioTokens : Math.max(1, windowTokens - reserveTokens);\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/** Return whether context usage exceeds the configured compaction threshold. */\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\nconst ESTIMATED_IMAGE_CHARS = 4800;\nconst ESTIMATED_MESSAGE_OVERHEAD_TOKENS = 4;\nconst ESTIMATED_UNKNOWN_MESSAGE_TOKENS = 4;\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/** Estimate token count for one message using a conservative character heuristic. */\nexport function estimateTokens(message: AgentMessage): number {\n\tswitch (message.role) {\n\t\tcase \"user\":\n\t\t\treturn (\n\t\t\t\tESTIMATED_MESSAGE_OVERHEAD_TOKENS +\n\t\t\t\tMath.ceil(\n\t\t\t\t\testimateTextAndImageContentChars(\n\t\t\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t\t\t) / 4,\n\t\t\t\t)\n\t\t\t);\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tlet chars = 0;\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 + safeJsonStringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn ESTIMATED_MESSAGE_OVERHEAD_TOKENS + Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\":\n\t\t\treturn ESTIMATED_MESSAGE_OVERHEAD_TOKENS + Math.ceil(estimateTextAndImageContentChars(message.content) / 4);\n\t\tcase \"bashExecution\":\n\t\t\treturn ESTIMATED_MESSAGE_OVERHEAD_TOKENS + Math.ceil((message.command.length + message.output.length) / 4);\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\":\n\t\t\treturn ESTIMATED_MESSAGE_OVERHEAD_TOKENS + Math.ceil(message.summary.length / 4);\n\t\tdefault:\n\t\t\t// Keep unknown runtime extensions conservative: compaction may trigger\n\t\t\t// earlier, never after the previous estimate would have triggered it.\n\t\t\treturn ESTIMATED_UNKNOWN_MESSAGE_TOKENS;\n\t}\n}\nfunction findValidCutPoints(entries: SessionTreeEntry[], 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\tdefault:\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 \"active_tools_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\tcase \"leaf\":\n\t\t\tdefault:\n\t\t\t\tbreak;\n\t\t}\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/** Find the user-visible message that starts the turn containing an entry. */\nexport function findTurnStartIndex(entries: SessionTreeEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let i = entryIndex; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\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\n/** Cut point selected for compaction. */\nexport interface CutPointResult {\n\t/** Index of the first entry retained after compaction. */\n\tfirstKeptEntryIndex: number;\n\t/** Index of the turn-start entry when the cut splits a turn, otherwise -1. */\n\tturnStartIndex: number;\n\t/** Whether the selected cut point splits an in-progress turn. */\n\tisSplitTurn: boolean;\n}\n\n/** Find the compaction cut point that keeps approximately the requested recent-token budget. */\nexport function findCutPoint(\n\tentries: SessionTreeEntry[],\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\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0];\n\n\tfor (let i = endIndex - 1; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type !== \"message\") continue;\n\t\tconst messageTokens = estimateTokens(entry.message as AgentMessage);\n\t\taccumulatedTokens += messageTokens;\n\t\tif (accumulatedTokens >= keepRecentTokens) {\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\twhile (cutIndex > startIndex) {\n\t\tconst prevEntry = entries[cutIndex - 1];\n\t\tif (prevEntry.type === \"compaction\") {\n\t\t\tbreak;\n\t\t}\n\t\tif (prevEntry.type === \"message\") {\n\t\t\tbreak;\n\t\t}\n\t\tcutIndex--;\n\t}\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\nexport const SUMMARIZATION_SYSTEM_PROMPT = `You are a context summarization assistant. Your task is to read a conversation between a user and an AI coding assistant, then produce a structured summary following the exact format specified.\n\nDo NOT continue the conversation. Do NOT respond to any questions in the conversation. ONLY output the structured summary.`;\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.`;\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.`;\n\n/** Generate or update a conversation summary for compaction. */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n): Promise<Result<string, CompactionError>> {\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\tlet basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tbasePrompt = `${basePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\tconst llmMessages = convertToLlm(currentMessages);\n\tconst conversationText = serializeConversationBounded(llmMessages, model, maxTokens);\n\tlet promptText = `<conversation>\\n${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 =\n\t\tmodel.reasoning && thinkingLevel && thinkingLevel !== \"off\"\n\t\t\t? { maxTokens, signal, apiKey, headers, reasoning: thinkingLevel }\n\t\t\t: { maxTokens, signal, apiKey, headers };\n\n\tconst response = await completeSimple(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcompletionOptions,\n\t);\n\tif (response.stopReason === \"aborted\") {\n\t\treturn err(new CompactionError(\"aborted\", response.errorMessage || \"Summarization aborted\"));\n\t}\n\tif (response.stopReason === \"error\") {\n\t\treturn err(\n\t\t\tnew CompactionError(\n\t\t\t\t\"summarization_failed\",\n\t\t\t\t`Summarization failed: ${response.errorMessage || \"Unknown error\"}`,\n\t\t\t),\n\t\t);\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 ok(textContent);\n}\n\n/** Prepared inputs for a compaction run. */\nexport interface CompactionPreparation {\n\t/** Entry id where retained history starts. */\n\tfirstKeptEntryId: string;\n\t/** Messages summarized into the history summary. */\n\tmessagesToSummarize: AgentMessage[];\n\t/** Prefix messages summarized separately when compaction splits a turn. */\n\tturnPrefixMessages: AgentMessage[];\n\t/** Whether compaction splits a turn. */\n\tisSplitTurn: boolean;\n\t/** Estimated context tokens before compaction. */\n\ttokensBefore: number;\n\t/** Previous compaction summary used for iterative updates. */\n\tpreviousSummary?: string;\n\t/** File operations extracted from summarized history. */\n\tfileOps: FileOperations;\n\t/** Settings used to prepare compaction. */\n\tsettings: CompactionSettings;\n}\n\n/** Prepare session entries for compaction, or return undefined when compaction is not applicable. */\nexport function prepareCompaction(\n\tpathEntries: SessionTreeEntry[],\n\tsettings: CompactionSettings,\n): Result<CompactionPreparation | undefined, CompactionError> {\n\tif (pathEntries.length === 0 || pathEntries[pathEntries.length - 1].type === \"compaction\") {\n\t\treturn ok(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 boundaryStart = 0;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\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\tconst dropsHistoryEntries = cutPoint.firstKeptEntryIndex > boundaryStart;\n\tconst summarizesSplitTurnPrefix = cutPoint.isSplitTurn && cutPoint.firstKeptEntryIndex > cutPoint.turnStartIndex;\n\t// Invariant: normal compactions that drop at least one entry, or summarize a\n\t// split-turn prefix, are unchanged. Only true no-ops are skipped.\n\tif (!dropsHistoryEntries && !summarizesSplitTurnPrefix) {\n\t\treturn ok(undefined);\n\t}\n\n\tconst firstKeptEntry = pathEntries[cutPoint.firstKeptEntryIndex];\n\tif (!firstKeptEntry?.id) {\n\t\treturn err(new CompactionError(\"invalid_session\", \"First kept entry has no UUID - session may need migration\"));\n\t}\n\tconst firstKeptEntryId = firstKeptEntry.id;\n\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\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\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\tconst fileOps = extractFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\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 ok({\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t});\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\nexport { serializeConversation } from \"./utils.ts\";\n\n/**\n * Rough chars-per-token estimate for bounding summarization input.\n * Conservative on purpose: serialized code/JSON runs below 4 chars/token.\n */\nconst SUMMARY_CHARS_PER_TOKEN = 3.5;\n/**\n * Headroom subtracted on top of the output reserve: the summarization system\n * prompt, the <conversation>/prompt wrapper, and tokenizer drift between the\n * chars-per-token estimate and the real tokenizer.\n */\nconst SUMMARY_INPUT_SAFETY_TOKENS = 2048;\n\nfunction summaryElisionMessage(omittedCount: number): Message {\n\treturn {\n\t\trole: \"user\",\n\t\tcontent: `[... ${omittedCount} earlier message(s) omitted to fit the summarization context window ...]`,\n\t\ttimestamp: Date.now(),\n\t} as Message;\n}\n\n/**\n * Serialize a conversation for summarization, bounded to the model's context\n * window. When the full serialization would exceed the window, middle\n * messages are dropped (keeping the goal-bearing head and the recent-work\n * tail) with an explicit elision marker; oversized single messages fall back\n * to hard text truncation keeping both ends.\n *\n * Without this bound, a session that already overflows the model cannot be\n * compacted: the summarization request itself exceeds the window, the\n * provider rejects it (e.g. Codex context_length_exceeded), and overflow\n * recovery is stuck permanently.\n */\nexport function serializeConversationBounded(messages: Message[], model: Model<any>, maxOutputTokens: number): string {\n\tconst serialized = serializeConversation(messages);\n\tconst contextWindow = model.contextWindow ?? 0;\n\tif (contextWindow <= 0) return serialized;\n\tconst budgetTokens = contextWindow - maxOutputTokens - SUMMARY_INPUT_SAFETY_TOKENS;\n\tconst budgetChars = Math.floor(Math.max(0, budgetTokens) * SUMMARY_CHARS_PER_TOKEN);\n\tif (serialized.length <= budgetChars) return serialized;\n\t// Window too small to bound meaningfully; leave it to the provider error path.\n\tif (budgetChars < 4096) return serialized;\n\n\t// Halve the kept head/tail spans until the elided conversation fits.\n\tfor (let keep = Math.floor(messages.length / 2); keep >= 1; keep = Math.floor(keep / 2)) {\n\t\tconst head = messages.slice(0, keep);\n\t\tconst tail = messages.slice(messages.length - keep);\n\t\tconst omitted = messages.length - head.length - tail.length;\n\t\tif (omitted <= 0) continue;\n\t\tconst bounded = serializeConversation([...head, summaryElisionMessage(omitted), ...tail]);\n\t\tif (bounded.length <= budgetChars) return bounded;\n\t}\n\n\t// Single oversized message(s): truncate the text itself, keeping both ends.\n\tconst headChars = Math.floor(budgetChars / 2);\n\tconst tailChars = Math.max(0, budgetChars - headChars - 128);\n\treturn `${serialized.slice(0, headChars)}\\n[... middle omitted to fit the summarization context window ...]\\n${serialized.slice(serialized.length - tailChars)}`;\n}\n\n/** Generate compaction summary data from prepared session history. */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n): Promise<Result<CompactionResult, CompactionError>> {\n\tconst {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\tif (!firstKeptEntryId) {\n\t\treturn err(new CompactionError(\"invalid_session\", \"First kept entry has no UUID - session may need migration\"));\n\t}\n\n\tlet summary: string;\n\n\tif (isSplitTurn && turnPrefixMessages.length > 0) {\n\t\tconst [historyResult, turnPrefixResult] = await Promise.all([\n\t\t\tmessagesToSummarize.length > 0\n\t\t\t\t? generateSummary(\n\t\t\t\t\t\tmessagesToSummarize,\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tsettings.reserveTokens,\n\t\t\t\t\t\tapiKey,\n\t\t\t\t\t\theaders,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\tcustomInstructions,\n\t\t\t\t\t\tpreviousSummary,\n\t\t\t\t\t\tthinkingLevel,\n\t\t\t\t\t)\n\t\t\t\t: Promise.resolve(ok<string, CompactionError>(\"No prior history.\")),\n\t\t\tgenerateTurnPrefixSummary(\n\t\t\t\tturnPrefixMessages,\n\t\t\t\tmodel,\n\t\t\t\tsettings.reserveTokens,\n\t\t\t\tapiKey,\n\t\t\t\theaders,\n\t\t\t\tsignal,\n\t\t\t\tthinkingLevel,\n\t\t\t),\n\t\t]);\n\t\tif (!historyResult.ok) return err(historyResult.error);\n\t\tif (!turnPrefixResult.ok) return err(turnPrefixResult.error);\n\t\tsummary = `${historyResult.value}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult.value}`;\n\t} else {\n\t\tconst summaryResult = await generateSummary(\n\t\t\tmessagesToSummarize,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t);\n\t\tif (!summaryResult.ok) return err(summaryResult.error);\n\t\tsummary = summaryResult.value;\n\t}\n\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\treturn ok({\n\t\tsummary,\n\t\tfirstKeptEntryId,\n\t\ttokensBefore,\n\t\tdetails: { readFiles, modifiedFiles } as CompactionDetails,\n\t});\n}\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n): Promise<Result<string, CompactionError>> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.5 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t);\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversationBounded(llmMessages, model, maxTokens);\n\tconst promptText = `<conversation>\\n${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 completeSimple(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tmodel.reasoning && thinkingLevel && thinkingLevel !== \"off\"\n\t\t\t? { maxTokens, signal, apiKey, headers, reasoning: thinkingLevel }\n\t\t\t: { maxTokens, signal, apiKey, headers },\n\t);\n\tif (response.stopReason === \"aborted\") {\n\t\treturn err(new CompactionError(\"aborted\", response.errorMessage || \"Turn prefix summarization aborted\"));\n\t}\n\tif (response.stopReason === \"error\") {\n\t\treturn err(\n\t\t\tnew CompactionError(\n\t\t\t\t\"summarization_failed\",\n\t\t\t\t`Turn prefix summarization failed: ${response.errorMessage || \"Unknown error\"}`,\n\t\t\t),\n\t\t);\n\t}\n\n\treturn ok(\n\t\tresponse.content\n\t\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t\t.map((c) => c.text)\n\t\t\t.join(\"\\n\"),\n\t);\n}\n"]}
+{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/harness/compaction/compaction.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAGX,OAAO,EACP,KAAK,EACL,cAAc,EACd,WAAW,EAEX,KAAK,EACL,MAAM,QAAQ,CAAC;AAEhB,OAAO,KAAK,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAQlE,OAAO,EAAwB,eAAe,EAAW,KAAK,MAAM,EAAE,KAAK,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAOjH,OAAO,EAIN,KAAK,cAAc,EAGnB,MAAM,YAAY,CAAC;AAEpB,OAAO,EAAE,2BAA2B,EAAE,MAAM,4BAA4B,CAAC;AAEzE,qEAAqE;AACrE,MAAM,WAAW,iBAAiB;IACjC,2CAA2C;IAC3C,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,+CAA+C;IAC/C,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AA8DD,6EAA6E;AAC7E,MAAM,WAAW,gBAAgB,CAAC,CAAC,GAAG,OAAO;IAC5C,sEAAsE;IACtE,OAAO,EAAE,MAAM,CAAC;IAChB,8CAA8C;IAC9C,gBAAgB,EAAE,MAAM,CAAC;IACzB,kDAAkD;IAClD,YAAY,EAAE,MAAM,CAAC;IACrB,iFAAiF;IACjF,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AAED,MAAM,WAAW,yBAAyB;IACzC,QAAQ,CAAC,KAAK,CAAC,EAAE,WAAW,CAAC;IAC7B,QAAQ,CAAC,SAAS,CAAC,EAAE,cAAc,CAAC;CACpC;AAED,oDAAoD;AACpD,MAAM,WAAW,kBAAkB;IAClC,6CAA6C;IAC7C,OAAO,EAAE,OAAO,CAAC;IACjB,qDAAqD;IACrD,aAAa,EAAE,MAAM,CAAC;IACtB,kEAAkE;IAClE,gBAAgB,EAAE,MAAM,CAAC;IACzB,qFAAqF;IACrF,aAAa,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,oEAAoE;AACpE,eAAO,MAAM,kCAAkC,MAAM,CAAC;AAEtD,uDAAuD;AACvD,eAAO,MAAM,2BAA2B,EAAE,kBAKzC,CAAC;AAEF,0DAA0D;AAC1D,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAWD,kFAAkF;AAClF,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,gBAAgB,EAAE,GAAG,KAAK,GAAG,SAAS,CASpF;AAED,wDAAwD;AACxD,MAAM,WAAW,oBAAoB;IACpC,sCAAsC;IACtC,MAAM,EAAE,MAAM,CAAC;IACf,gEAAgE;IAChE,WAAW,EAAE,MAAM,CAAC;IACpB,oEAAoE;IACpE,cAAc,EAAE,MAAM,CAAC;IACvB,0EAA0E;IAC1E,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD,gFAAgF;AAChF,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,8DAA8D;AAC9D,MAAM,WAAW,2BAA2B;IAC3C,oDAAoD;IACpD,aAAa,EAAE,MAAM,CAAC;IACtB,6CAA6C;IAC7C,cAAc,EAAE,MAAM,CAAC;IACvB,4DAA4D;IAC5D,mBAAmB,EAAE,MAAM,CAAC;IAC5B,gEAAgE;IAChE,qBAAqB,EAAE,MAAM,CAAC;IAC9B,wCAAwC;IACxC,SAAS,EAAE,uBAAuB,CAAC;CACnC;AAED,+FAA+F;AAC/F,wBAAgB,8BAA8B,CAC7C,aAAa,EAAE,MAAM,EACrB,QAAQ,EAAE,kBAAkB,GAC1B,2BAA2B,GAAG,SAAS,CAwBzC;AAED,gFAAgF;AAChF,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAsBD,qFAAqF;AACrF,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAsC5D;AA2CD,8EAA8E;AAC9E,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,gBAAgB,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAc9G;AAED,yCAAyC;AACzC,MAAM,WAAW,cAAc;IAC9B,0DAA0D;IAC1D,mBAAmB,EAAE,MAAM,CAAC;IAC5B,8EAA8E;IAC9E,cAAc,EAAE,MAAM,CAAC;IACvB,iEAAiE;IACjE,WAAW,EAAE,OAAO,CAAC;CACrB;AAED,gGAAgG;AAChG,wBAAgB,YAAY,CAC3B,OAAO,EAAE,gBAAgB,EAAE,EAC3B,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CA2ChB;AAED,gEAAgE;AAChE,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,EACd,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,YAAY,CAAC,EAAE,yBAAyB,GACtC,OAAO,CAAC,MAAM,CAAC,MAAM,EAAE,eAAe,CAAC,CAAC,CA2D1C;AAED,4CAA4C;AAC5C,MAAM,WAAW,qBAAqB;IACrC,8CAA8C;IAC9C,gBAAgB,EAAE,MAAM,CAAC;IACzB,oDAAoD;IACpD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,wCAAwC;IACxC,WAAW,EAAE,OAAO,CAAC;IACrB,kDAAkD;IAClD,YAAY,EAAE,MAAM,CAAC;IACrB,8DAA8D;IAC9D,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,yDAAyD;IACzD,OAAO,EAAE,cAAc,CAAC;IACxB,2CAA2C;IAC3C,QAAQ,EAAE,kBAAkB,CAAC;CAC7B;AAED,qGAAqG;AACrG,wBAAgB,iBAAiB,CAChC,WAAW,EAAE,gBAAgB,EAAE,EAC/B,QAAQ,EAAE,kBAAkB,GAC1B,MAAM,CAAC,qBAAqB,GAAG,SAAS,EAAE,eAAe,CAAC,CAsE5D;AAED,OAAO,EAAE,qBAAqB,EAAE,MAAM,YAAY,CAAC;AAsBnD;;;;;;;;;;;GAWG;AACH,wBAAgB,4BAA4B,CAAC,QAAQ,EAAE,OAAO,EAAE,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EAAE,eAAe,EAAE,MAAM,GAAG,MAAM,CAwBpH;AAED,sEAAsE;AACtE,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,MAAM,EAAE,MAAM,EACd,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,YAAY,CAAC,EAAE,yBAAyB,GACtC,OAAO,CAAC,MAAM,CAAC,gBAAgB,EAAE,eAAe,CAAC,CAAC,CA0EpD","sourcesContent":["import type {\n\tAssistantMessage,\n\tImageContent,\n\tMessage,\n\tModel,\n\tRetryCallbacks,\n\tRetryPolicy,\n\tTextContent,\n\tUsage,\n} from \"omk-ai\";\nimport { completeSimple, retryAssistantCall } from \"omk-ai\";\nimport type { AgentMessage, ThinkingLevel } from \"../../types.ts\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport { buildSessionContext } from \"../session/session.ts\";\nimport { type CompactionEntry, CompactionError, err, ok, type Result, type SessionTreeEntry } from \"../types.ts\";\nimport {\n\tSUMMARIZATION_PROMPT,\n\tSUMMARIZATION_SYSTEM_PROMPT,\n\tTURN_PREFIX_SUMMARIZATION_PROMPT,\n\tUPDATE_SUMMARIZATION_PROMPT,\n} from \"./summarization-prompts.ts\";\nimport {\n\tcomputeFileLists,\n\tcreateFileOps,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tserializeConversation,\n} from \"./utils.ts\";\n\nexport { SUMMARIZATION_SYSTEM_PROMPT } from \"./summarization-prompts.ts\";\n\n/** File-operation details stored on generated compaction entries. */\nexport interface CompactionDetails {\n\t/** Files read in the compacted history. */\n\treadFiles: string[];\n\t/** Files modified in the compacted history. */\n\tmodifiedFiles: string[];\n}\nfunction safeJsonStringify(value: unknown): string {\n\ttry {\n\t\treturn JSON.stringify(value) ?? \"undefined\";\n\t} catch {\n\t\treturn \"[unserializable]\";\n\t}\n}\n\nfunction extractFileOperations(\n\tmessages: AgentMessage[],\n\tentries: SessionTreeEntry[],\n\tprevCompactionIndex: number,\n): FileOperations {\n\tconst fileOps = createFileOps();\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = entries[prevCompactionIndex] as CompactionEntry;\n\t\tif (!prevCompaction.fromHook && prevCompaction.details) {\n\t\t\tconst details = prevCompaction.details as CompactionDetails;\n\t\t\tif (Array.isArray(details.readFiles)) {\n\t\t\t\tfor (const f of details.readFiles) fileOps.read.add(f);\n\t\t\t}\n\t\t\tif (Array.isArray(details.modifiedFiles)) {\n\t\t\t\tfor (const f of details.modifiedFiles) fileOps.edited.add(f);\n\t\t\t}\n\t\t}\n\t}\n\tfor (const msg of messages) {\n\t\textractFileOpsFromMessage(msg, fileOps);\n\t}\n\n\treturn fileOps;\n}\nfunction getMessageFromEntry(entry: SessionTreeEntry): AgentMessage | undefined {\n\tif (entry.type === \"message\") {\n\t\treturn entry.message as AgentMessage;\n\t}\n\tif (entry.type === \"custom_message\") {\n\t\treturn createCustomMessage(\n\t\t\tentry.customType,\n\t\t\tentry.content as string | (TextContent | ImageContent)[],\n\t\t\tentry.display,\n\t\t\tentry.details,\n\t\t\tentry.timestamp,\n\t\t);\n\t}\n\tif (entry.type === \"branch_summary\") {\n\t\treturn createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);\n\t}\n\tif (entry.type === \"compaction\") {\n\t\treturn createCompactionSummaryMessage(entry.summary, entry.tokensBefore, entry.timestamp);\n\t}\n\treturn undefined;\n}\n\nfunction getMessageFromEntryForCompaction(entry: SessionTreeEntry): AgentMessage | undefined {\n\tif (entry.type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\treturn getMessageFromEntry(entry);\n}\n\n/** Generated compaction data ready to be persisted as a compaction entry. */\nexport interface CompactionResult<T = unknown> {\n\t/** Summary text that replaces compacted history in future context. */\n\tsummary: string;\n\t/** Entry id where retained history starts. */\n\tfirstKeptEntryId: string;\n\t/** Estimated context tokens before compaction. */\n\ttokensBefore: number;\n\t/** Optional implementation-specific details stored with the compaction entry. */\n\tdetails?: T;\n}\n\nexport interface SummarizationRetryOptions {\n\treadonly retry?: RetryPolicy;\n\treadonly callbacks?: RetryCallbacks;\n}\n\n/** Compaction thresholds and retention settings. */\nexport interface CompactionSettings {\n\t/** Enable automatic compaction decisions. */\n\tenabled: boolean;\n\t/** Tokens reserved for summary prompt and output. */\n\treserveTokens: number;\n\t/** Approximate recent-context tokens to keep after compaction. */\n\tkeepRecentTokens: number;\n\t/** Maximum fraction of the context window to use before compaction. Default: 0.9. */\n\tmaxUsageRatio?: number;\n}\n\n/** Default context-window ratio used before compaction triggers. */\nexport const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;\n\n/** Default compaction settings used by the harness. */\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/** Calculate total context tokens from provider usage. */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\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/** Return usage from the last successful assistant message in session entries. */\nexport function getLastAssistantUsage(entries: SessionTreeEntry[]): 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 as AgentMessage);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/** Estimated context-token usage for a message list. */\nexport interface ContextUsageEstimate {\n\t/** Estimated total context tokens. */\n\ttokens: number;\n\t/** Tokens reported by the most recent assistant usage block. */\n\tusageTokens: number;\n\t/** Estimated tokens after the most recent assistant usage block. */\n\ttrailingTokens: number;\n\t/** Index of the message that provided usage, or null when none exists. */\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/** Estimate context tokens for messages using provider usage when available. */\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\n/** Threshold calculation details for automatic compaction. */\nexport interface CompactionHeadroomThreshold {\n\t/** Context-token count that triggers compaction. */\n\ttriggerTokens: number;\n\t/** Tokens kept free at the trigger point. */\n\theadroomTokens: number;\n\t/** Boundary derived from the configured max usage ratio. */\n\tmaxUsageRatioTokens: number;\n\t/** Boundary derived from the absolute reserve token setting. */\n\treserveBoundaryTokens: number;\n\t/** Which boundary triggered earlier. */\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 reserveBoundaryTokens =\n\t\treserveTokens >= windowTokens ? maxUsageRatioTokens : Math.max(1, windowTokens - reserveTokens);\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/** Return whether context usage exceeds the configured compaction threshold. */\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\nconst ESTIMATED_IMAGE_CHARS = 4800;\nconst ESTIMATED_MESSAGE_OVERHEAD_TOKENS = 4;\nconst ESTIMATED_UNKNOWN_MESSAGE_TOKENS = 4;\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/** Estimate token count for one message using a conservative character heuristic. */\nexport function estimateTokens(message: AgentMessage): number {\n\tswitch (message.role) {\n\t\tcase \"user\":\n\t\t\treturn (\n\t\t\t\tESTIMATED_MESSAGE_OVERHEAD_TOKENS +\n\t\t\t\tMath.ceil(\n\t\t\t\t\testimateTextAndImageContentChars(\n\t\t\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t\t\t) / 4,\n\t\t\t\t)\n\t\t\t);\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tlet chars = 0;\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 + safeJsonStringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn ESTIMATED_MESSAGE_OVERHEAD_TOKENS + Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\":\n\t\t\treturn ESTIMATED_MESSAGE_OVERHEAD_TOKENS + Math.ceil(estimateTextAndImageContentChars(message.content) / 4);\n\t\tcase \"bashExecution\":\n\t\t\treturn ESTIMATED_MESSAGE_OVERHEAD_TOKENS + Math.ceil((message.command.length + message.output.length) / 4);\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\":\n\t\t\treturn ESTIMATED_MESSAGE_OVERHEAD_TOKENS + Math.ceil(message.summary.length / 4);\n\t\tdefault:\n\t\t\t// Keep unknown runtime extensions conservative: compaction may trigger\n\t\t\t// earlier, never after the previous estimate would have triggered it.\n\t\t\treturn ESTIMATED_UNKNOWN_MESSAGE_TOKENS;\n\t}\n}\nfunction findValidCutPoints(entries: SessionTreeEntry[], 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\tdefault:\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 \"active_tools_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\tcase \"leaf\":\n\t\t\tdefault:\n\t\t\t\tbreak;\n\t\t}\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/** Find the user-visible message that starts the turn containing an entry. */\nexport function findTurnStartIndex(entries: SessionTreeEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let i = entryIndex; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\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\n/** Cut point selected for compaction. */\nexport interface CutPointResult {\n\t/** Index of the first entry retained after compaction. */\n\tfirstKeptEntryIndex: number;\n\t/** Index of the turn-start entry when the cut splits a turn, otherwise -1. */\n\tturnStartIndex: number;\n\t/** Whether the selected cut point splits an in-progress turn. */\n\tisSplitTurn: boolean;\n}\n\n/** Find the compaction cut point that keeps approximately the requested recent-token budget. */\nexport function findCutPoint(\n\tentries: SessionTreeEntry[],\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\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0];\n\n\tfor (let i = endIndex - 1; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type !== \"message\") continue;\n\t\tconst messageTokens = estimateTokens(entry.message as AgentMessage);\n\t\taccumulatedTokens += messageTokens;\n\t\tif (accumulatedTokens >= keepRecentTokens) {\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\twhile (cutIndex > startIndex) {\n\t\tconst prevEntry = entries[cutIndex - 1];\n\t\tif (prevEntry.type === \"compaction\") {\n\t\t\tbreak;\n\t\t}\n\t\tif (prevEntry.type === \"message\") {\n\t\t\tbreak;\n\t\t}\n\t\tcutIndex--;\n\t}\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/** Generate or update a conversation summary for compaction. */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tretryOptions?: SummarizationRetryOptions,\n): Promise<Result<string, CompactionError>> {\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\tlet basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tbasePrompt = `${basePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\tconst llmMessages = convertToLlm(currentMessages);\n\tconst conversationText = serializeConversationBounded(llmMessages, model, maxTokens);\n\tlet promptText = `<conversation>\\n${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 =\n\t\tmodel.reasoning && thinkingLevel && thinkingLevel !== \"off\"\n\t\t\t? { maxTokens, signal, apiKey, headers, reasoning: thinkingLevel }\n\t\t\t: { maxTokens, signal, apiKey, headers };\n\n\tconst response = await retryAssistantCall(\n\t\t() =>\n\t\t\tcompleteSimple(\n\t\t\t\tmodel,\n\t\t\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\t\t\tcompletionOptions,\n\t\t\t),\n\t\tretryOptions?.retry,\n\t\tsignal,\n\t\tretryOptions?.callbacks,\n\t);\n\tif (response.stopReason === \"aborted\") {\n\t\treturn err(new CompactionError(\"aborted\", response.errorMessage || \"Summarization aborted\"));\n\t}\n\tif (response.stopReason === \"error\") {\n\t\treturn err(\n\t\t\tnew CompactionError(\n\t\t\t\t\"summarization_failed\",\n\t\t\t\t`Summarization failed: ${response.errorMessage || \"Unknown error\"}`,\n\t\t\t),\n\t\t);\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 ok(textContent);\n}\n\n/** Prepared inputs for a compaction run. */\nexport interface CompactionPreparation {\n\t/** Entry id where retained history starts. */\n\tfirstKeptEntryId: string;\n\t/** Messages summarized into the history summary. */\n\tmessagesToSummarize: AgentMessage[];\n\t/** Prefix messages summarized separately when compaction splits a turn. */\n\tturnPrefixMessages: AgentMessage[];\n\t/** Whether compaction splits a turn. */\n\tisSplitTurn: boolean;\n\t/** Estimated context tokens before compaction. */\n\ttokensBefore: number;\n\t/** Previous compaction summary used for iterative updates. */\n\tpreviousSummary?: string;\n\t/** File operations extracted from summarized history. */\n\tfileOps: FileOperations;\n\t/** Settings used to prepare compaction. */\n\tsettings: CompactionSettings;\n}\n\n/** Prepare session entries for compaction, or return undefined when compaction is not applicable. */\nexport function prepareCompaction(\n\tpathEntries: SessionTreeEntry[],\n\tsettings: CompactionSettings,\n): Result<CompactionPreparation | undefined, CompactionError> {\n\tif (pathEntries.length === 0 || pathEntries[pathEntries.length - 1].type === \"compaction\") {\n\t\treturn ok(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 boundaryStart = 0;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\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\tconst dropsHistoryEntries = cutPoint.firstKeptEntryIndex > boundaryStart;\n\tconst summarizesSplitTurnPrefix = cutPoint.isSplitTurn && cutPoint.firstKeptEntryIndex > cutPoint.turnStartIndex;\n\t// Invariant: normal compactions that drop at least one entry, or summarize a\n\t// split-turn prefix, are unchanged. Only true no-ops are skipped.\n\tif (!dropsHistoryEntries && !summarizesSplitTurnPrefix) {\n\t\treturn ok(undefined);\n\t}\n\n\tconst firstKeptEntry = pathEntries[cutPoint.firstKeptEntryIndex];\n\tif (!firstKeptEntry?.id) {\n\t\treturn err(new CompactionError(\"invalid_session\", \"First kept entry has no UUID - session may need migration\"));\n\t}\n\tconst firstKeptEntryId = firstKeptEntry.id;\n\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\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\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\tconst fileOps = extractFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\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 ok({\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t});\n}\n\nexport { serializeConversation } from \"./utils.ts\";\n\n/**\n * Rough chars-per-token estimate for bounding summarization input.\n * Conservative on purpose: serialized code/JSON runs below 4 chars/token.\n */\nconst SUMMARY_CHARS_PER_TOKEN = 3.5;\n/**\n * Headroom subtracted on top of the output reserve: the summarization system\n * prompt, the <conversation>/prompt wrapper, and tokenizer drift between the\n * chars-per-token estimate and the real tokenizer.\n */\nconst SUMMARY_INPUT_SAFETY_TOKENS = 2048;\n\nfunction summaryElisionMessage(omittedCount: number): Message {\n\treturn {\n\t\trole: \"user\",\n\t\tcontent: `[... ${omittedCount} earlier message(s) omitted to fit the summarization context window ...]`,\n\t\ttimestamp: Date.now(),\n\t} as Message;\n}\n\n/**\n * Serialize a conversation for summarization, bounded to the model's context\n * window. When the full serialization would exceed the window, middle\n * messages are dropped (keeping the goal-bearing head and the recent-work\n * tail) with an explicit elision marker; oversized single messages fall back\n * to hard text truncation keeping both ends.\n *\n * Without this bound, a session that already overflows the model cannot be\n * compacted: the summarization request itself exceeds the window, the\n * provider rejects it (e.g. Codex context_length_exceeded), and overflow\n * recovery is stuck permanently.\n */\nexport function serializeConversationBounded(messages: Message[], model: Model<any>, maxOutputTokens: number): string {\n\tconst serialized = serializeConversation(messages);\n\tconst contextWindow = model.contextWindow ?? 0;\n\tif (contextWindow <= 0) return serialized;\n\tconst budgetTokens = contextWindow - maxOutputTokens - SUMMARY_INPUT_SAFETY_TOKENS;\n\tconst budgetChars = Math.floor(Math.max(0, budgetTokens) * SUMMARY_CHARS_PER_TOKEN);\n\tif (serialized.length <= budgetChars) return serialized;\n\t// Window too small to bound meaningfully; leave it to the provider error path.\n\tif (budgetChars < 4096) return serialized;\n\n\t// Halve the kept head/tail spans until the elided conversation fits.\n\tfor (let keep = Math.floor(messages.length / 2); keep >= 1; keep = Math.floor(keep / 2)) {\n\t\tconst head = messages.slice(0, keep);\n\t\tconst tail = messages.slice(messages.length - keep);\n\t\tconst omitted = messages.length - head.length - tail.length;\n\t\tif (omitted <= 0) continue;\n\t\tconst bounded = serializeConversation([...head, summaryElisionMessage(omitted), ...tail]);\n\t\tif (bounded.length <= budgetChars) return bounded;\n\t}\n\n\t// Single oversized message(s): truncate the text itself, keeping both ends.\n\tconst headChars = Math.floor(budgetChars / 2);\n\tconst tailChars = Math.max(0, budgetChars - headChars - 128);\n\treturn `${serialized.slice(0, headChars)}\\n[... middle omitted to fit the summarization context window ...]\\n${serialized.slice(serialized.length - tailChars)}`;\n}\n\n/** Generate compaction summary data from prepared session history. */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tretryOptions?: SummarizationRetryOptions,\n): Promise<Result<CompactionResult, CompactionError>> {\n\tconst {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\tif (!firstKeptEntryId) {\n\t\treturn err(new CompactionError(\"invalid_session\", \"First kept entry has no UUID - session may need migration\"));\n\t}\n\n\tlet summary: string;\n\n\tif (isSplitTurn && turnPrefixMessages.length > 0) {\n\t\tconst [historyResult, turnPrefixResult] = await Promise.all([\n\t\t\tmessagesToSummarize.length > 0\n\t\t\t\t? generateSummary(\n\t\t\t\t\t\tmessagesToSummarize,\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tsettings.reserveTokens,\n\t\t\t\t\t\tapiKey,\n\t\t\t\t\t\theaders,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\tcustomInstructions,\n\t\t\t\t\t\tpreviousSummary,\n\t\t\t\t\t\tthinkingLevel,\n\t\t\t\t\t\tretryOptions,\n\t\t\t\t\t)\n\t\t\t\t: Promise.resolve(ok<string, CompactionError>(\"No prior history.\")),\n\t\t\tgenerateTurnPrefixSummary(\n\t\t\t\tturnPrefixMessages,\n\t\t\t\tmodel,\n\t\t\t\tsettings.reserveTokens,\n\t\t\t\tapiKey,\n\t\t\t\theaders,\n\t\t\t\tsignal,\n\t\t\t\tthinkingLevel,\n\t\t\t\tretryOptions,\n\t\t\t),\n\t\t]);\n\t\tif (!historyResult.ok) return err(historyResult.error);\n\t\tif (!turnPrefixResult.ok) return err(turnPrefixResult.error);\n\t\tsummary = `${historyResult.value}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult.value}`;\n\t} else {\n\t\tconst summaryResult = await generateSummary(\n\t\t\tmessagesToSummarize,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tretryOptions,\n\t\t);\n\t\tif (!summaryResult.ok) return err(summaryResult.error);\n\t\tsummary = summaryResult.value;\n\t}\n\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\treturn ok({\n\t\tsummary,\n\t\tfirstKeptEntryId,\n\t\ttokensBefore,\n\t\tdetails: { readFiles, modifiedFiles } as CompactionDetails,\n\t});\n}\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tretryOptions?: SummarizationRetryOptions,\n): Promise<Result<string, CompactionError>> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.5 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t);\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversationBounded(llmMessages, model, maxTokens);\n\tconst promptText = `<conversation>\\n${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 retryAssistantCall(\n\t\t() =>\n\t\t\tcompleteSimple(\n\t\t\t\tmodel,\n\t\t\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\t\t\tmodel.reasoning && thinkingLevel && thinkingLevel !== \"off\"\n\t\t\t\t\t? { maxTokens, signal, apiKey, headers, reasoning: thinkingLevel }\n\t\t\t\t\t: { maxTokens, signal, apiKey, headers },\n\t\t\t),\n\t\tretryOptions?.retry,\n\t\tsignal,\n\t\tretryOptions?.callbacks,\n\t);\n\tif (response.stopReason === \"aborted\") {\n\t\treturn err(new CompactionError(\"aborted\", response.errorMessage || \"Turn prefix summarization aborted\"));\n\t}\n\tif (response.stopReason === \"error\") {\n\t\treturn err(\n\t\t\tnew CompactionError(\n\t\t\t\t\"summarization_failed\",\n\t\t\t\t`Turn prefix summarization failed: ${response.errorMessage || \"Unknown error\"}`,\n\t\t\t),\n\t\t);\n\t}\n\n\treturn ok(\n\t\tresponse.content\n\t\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t\t.map((c) => c.text)\n\t\t\t.join(\"\\n\"),\n\t);\n}\n"]}

package/dist/harness/compaction/compaction.js

@@ -1,8 +1,10 @@
-import { completeSimple } from "omk-ai";
+import { completeSimple, retryAssistantCall } from "omk-ai";
 import { convertToLlm, createBranchSummaryMessage, createCompactionSummaryMessage, createCustomMessage, } from "../messages.js";
 import { buildSessionContext } from "../session/session.js";
 import { CompactionError, err, ok } from "../types.js";
+import { SUMMARIZATION_PROMPT, SUMMARIZATION_SYSTEM_PROMPT, TURN_PREFIX_SUMMARIZATION_PROMPT, UPDATE_SUMMARIZATION_PROMPT, } from "./summarization-prompts.js";
 import { computeFileLists, createFileOps, extractFileOpsFromMessage, formatFileOperations, serializeConversation, } from "./utils.js";
+export { SUMMARIZATION_SYSTEM_PROMPT } from "./summarization-prompts.js";
 function safeJsonStringify(value) {
     try {
         return JSON.stringify(value) ?? "undefined";
@@ -307,81 +309,8 @@ export function findCutPoint(entries, startIndex, endIndex, keepRecentTokens) {
         isSplitTurn: !isUserMessage && turnStartIndex !== -1,
     };
 }
-export const SUMMARIZATION_SYSTEM_PROMPT = `You are a context summarization assistant. Your task is to read a conversation between a user and an AI coding assistant, then produce a structured summary following the exact format specified.
-
-Do NOT continue the conversation. Do NOT respond to any questions in the conversation. ONLY output the structured summary.`;
-const 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.
-
-Use this EXACT format:
-
-## Goal
-[What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]
-
-## Constraints & Preferences
-- [Any constraints, preferences, or requirements mentioned by user]
-- [Or "(none)" if none were mentioned]
-
-## Progress
-### Done
-- [x] [Completed tasks/changes]
-
-### In Progress
-- [ ] [Current work]
-
-### Blocked
-- [Issues preventing progress, if any]
-
-## Key Decisions
-- **[Decision]**: [Brief rationale]
-
-## Next Steps
-1. [Ordered list of what should happen next]
-
-## Critical Context
-- [Any data, examples, or references needed to continue]
-- [Or "(none)" if not applicable]
-
-Keep each section concise. Preserve exact file paths, function names, and error messages.`;
-const UPDATE_SUMMARIZATION_PROMPT = `The messages above are NEW conversation messages to incorporate into the existing summary provided in <previous-summary> tags.
-
-Update the existing structured summary with new information. RULES:
-- PRESERVE all existing information from the previous summary
-- ADD new progress, decisions, and context from the new messages
-- UPDATE the Progress section: move items from "In Progress" to "Done" when completed
-- UPDATE "Next Steps" based on what was accomplished
-- PRESERVE exact file paths, function names, and error messages
-- If something is no longer relevant, you may remove it
-
-Use this EXACT format:
-
-## Goal
-[Preserve existing goals, add new ones if the task expanded]
-
-## Constraints & Preferences
-- [Preserve existing, add new ones discovered]
-
-## Progress
-### Done
-- [x] [Include previously done items AND newly completed items]
-
-### In Progress
-- [ ] [Current work - update based on progress]
-
-### Blocked
-- [Current blockers - remove if resolved]
-
-## Key Decisions
-- **[Decision]**: [Brief rationale] (preserve all previous, add new)
-
-## Next Steps
-1. [Update based on current state]
-
-## Critical Context
-- [Preserve important context, add new if needed]
-
-Keep each section concise. Preserve exact file paths, function names, and error messages.`;
 /** Generate or update a conversation summary for compaction. */
-export async function generateSummary(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel) {
+export async function generateSummary(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, retryOptions) {
     const maxTokens = Math.min(Math.floor(0.8 * reserveTokens), model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY);
     let basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;
     if (customInstructions) {
@@ -404,7 +333,7 @@ export async function generateSummary(currentMessages, model, reserveTokens, api
     const completionOptions = model.reasoning && thinkingLevel && thinkingLevel !== "off"
         ? { maxTokens, signal, apiKey, headers, reasoning: thinkingLevel }
         : { maxTokens, signal, apiKey, headers };
-    const response = await completeSimple(model, { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages }, completionOptions);
+    const response = await retryAssistantCall(() => completeSimple(model, { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages }, completionOptions), retryOptions?.retry, signal, retryOptions?.callbacks);
     if (response.stopReason === "aborted") {
         return err(new CompactionError("aborted", response.errorMessage || "Summarization aborted"));
     }
@@ -484,20 +413,6 @@ export function prepareCompaction(pathEntries, settings) {
         settings,
     });
 }
-const TURN_PREFIX_SUMMARIZATION_PROMPT = `This is the PREFIX of a turn that was too large to keep. The SUFFIX (recent work) is retained.
-
-Summarize the prefix to provide context for the retained suffix:
-
-## Original Request
-[What did the user ask for in this turn?]
-
-## Early Progress
-- [Key decisions and work done in the prefix]
-
-## Context for Suffix
-- [Information needed to understand the retained recent work]
-
-Be concise. Focus on what's needed to understand the kept suffix.`;
 export { serializeConversation } from "./utils.js";
 /**
  * Rough chars-per-token estimate for bounding summarization input.
@@ -558,7 +473,7 @@ export function serializeConversationBounded(messages, model, maxOutputTokens) {
     return `${serialized.slice(0, headChars)}\n[... middle omitted to fit the summarization context window ...]\n${serialized.slice(serialized.length - tailChars)}`;
 }
 /** Generate compaction summary data from prepared session history. */
-export async function compact(preparation, model, apiKey, headers, customInstructions, signal, thinkingLevel) {
+export async function compact(preparation, model, apiKey, headers, customInstructions, signal, thinkingLevel, retryOptions) {
     const { firstKeptEntryId, messagesToSummarize, turnPrefixMessages, isSplitTurn, tokensBefore, previousSummary, fileOps, settings, } = preparation;
     if (!firstKeptEntryId) {
         return err(new CompactionError("invalid_session", "First kept entry has no UUID - session may need migration"));
@@ -567,9 +482,9 @@ export async function compact(preparation, model, apiKey, headers, customInstruc
     if (isSplitTurn && turnPrefixMessages.length > 0) {
         const [historyResult, turnPrefixResult] = await Promise.all([
             messagesToSummarize.length > 0
-                ? generateSummary(messagesToSummarize, model, settings.reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel)
+                ? generateSummary(messagesToSummarize, model, settings.reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, retryOptions)
                 : Promise.resolve(ok("No prior history.")),
-            generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, apiKey, headers, signal, thinkingLevel),
+            generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, apiKey, headers, signal, thinkingLevel, retryOptions),
         ]);
         if (!historyResult.ok)
             return err(historyResult.error);
@@ -578,7 +493,7 @@ export async function compact(preparation, model, apiKey, headers, customInstruc
         summary = `${historyResult.value}\n\n---\n\n**Turn Context (split turn):**\n\n${turnPrefixResult.value}`;
     }
     else {
-        const summaryResult = await generateSummary(messagesToSummarize, model, settings.reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel);
+        const summaryResult = await generateSummary(messagesToSummarize, model, settings.reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, retryOptions);
         if (!summaryResult.ok)
             return err(summaryResult.error);
         summary = summaryResult.value;
@@ -592,7 +507,7 @@ export async function compact(preparation, model, apiKey, headers, customInstruc
         details: { readFiles, modifiedFiles },
     });
 }
-async function generateTurnPrefixSummary(messages, model, reserveTokens, apiKey, headers, signal, thinkingLevel) {
+async function generateTurnPrefixSummary(messages, model, reserveTokens, apiKey, headers, signal, thinkingLevel, retryOptions) {
     const maxTokens = Math.min(Math.floor(0.5 * reserveTokens), model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY);
     const llmMessages = convertToLlm(messages);
     const conversationText = serializeConversationBounded(llmMessages, model, maxTokens);
@@ -604,9 +519,9 @@ async function generateTurnPrefixSummary(messages, model, reserveTokens, apiKey,
             timestamp: Date.now(),
         },
     ];
-    const response = await completeSimple(model, { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages }, model.reasoning && thinkingLevel && thinkingLevel !== "off"
+    const response = await retryAssistantCall(() => completeSimple(model, { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages }, model.reasoning && thinkingLevel && thinkingLevel !== "off"
         ? { maxTokens, signal, apiKey, headers, reasoning: thinkingLevel }
-        : { maxTokens, signal, apiKey, headers });
+        : { maxTokens, signal, apiKey, headers }), retryOptions?.retry, signal, retryOptions?.callbacks);
     if (response.stopReason === "aborted") {
         return err(new CompactionError("aborted", response.errorMessage || "Turn prefix summarization aborted"));
     }