@averyyy/pi-agent-core 0.84.4-piclient.2 → 0.85.1-piclient.2

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,EACN,KAAK,GAAG,EACR,KAAK,gBAAgB,EACrB,KAAK,OAAO,EAEZ,KAAK,KAAK,EACV,KAAK,MAAM,EACX,KAAK,cAAc,EACnB,KAAK,WAAW,EAEhB,KAAK,mBAAmB,EACxB,KAAK,KAAK,EAEV,MAAM,uBAAuB,CAAC;AAC/B,OAAO,KAAK,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAClE,OAAO,EAAyD,KAAK,gBAAgB,EAAE,MAAM,sBAAsB,CAAC;AAQpH,OAAO,KAAK,EAAmB,KAAK,EAAE,MAAM,qBAAqB,CAAC;AAClE,OAAO,EAAE,eAAe,EAAW,KAAK,MAAM,EAAE,MAAM,aAAa,CAAC;AACpE,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;AA6FD,6EAA6E;AAC7E,MAAM,WAAW,aAAa,CAAC,CAAC,GAAG,OAAO;IACzC,sEAAsE;IACtE,OAAO,EAAE,MAAM,CAAC;IAChB,kDAAkD;IAClD,YAAY,EAAE,MAAM,CAAC;IACrB,4EAA4E;IAC5E,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,wEAAwE;IACxE,YAAY,EAAE,YAAY,EAAE,CAAC;IAC7B,iFAAiF;IACjF,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AAED,+EAA+E;AAC/E,MAAM,WAAW,mBAAmB,CAAC,CAAC,GAAG,OAAO,CAAE,SAAQ,aAAa,CAAC,CAAC,CAAC;IACzE,8CAA8C;IAC9C,gBAAgB,EAAE,MAAM,CAAC;CACzB;AAED,wBAAsB,yBAAyB,CAC9C,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,EAC5B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CAa3B;AAyBD,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;CACzB;AAED,uDAAuD;AACvD,eAAO,MAAM,2BAA2B,EAAE,kBAIzC,CAAC;AAEF,0DAA0D;AAC1D,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAgBD,6EAA6E;AAC7E,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,KAAK,EAAE,GAAG,KAAK,GAAG,SAAS,CASzE;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,gFAAgF;AAChF,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAoBD,qFAAqF;AACrF,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAwC5D;AAmCD,8EAA8E;AAC9E,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,KAAK,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAcnG;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,KAAK,EAAE,EAChB,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CA2ChB;AAgFD,eAAO,MAAM,2BAA2B,6TAEmF,CAAC;AA4S5H,gEAAgE;AAChE,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,MAAM,CAAC,MAAM,EAAE,eAAe,CAAC,CAAC,CAc1C;AAED,+EAA+E;AAC/E,wBAAsB,wBAAwB,CAC7C,eAAe,EAAE,YAAY,EAAE,EAC/B,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,MAAM,CAAC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,KAAK,CAAA;CAAE,EAAE,eAAe,CAAC,CAAC,CAqBlE;AAED,4CAA4C;AAC5C,MAAM,WAAW,qBAAqB;IACrC,oDAAoD;IACpD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,oFAAoF;IACpF,YAAY,EAAE,YAAY,EAAE,CAAC;IAC7B,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,KAAK,EAAE,EACpB,QAAQ,EAAE,kBAAkB,GAC1B,MAAM,CAAC,qBAAqB,GAAG,SAAS,EAAE,eAAe,CAAC,CAmE5D;AAED,mFAAmF;AACnF,MAAM,WAAW,4BAA4B;IAC5C,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,gGAAgG;AAChG,MAAM,WAAW,2BAA4B,SAAQ,qBAAqB;IACzE,gBAAgB,EAAE,MAAM,CAAC;CACzB;AAED,sGAAsG;AACtG,wBAAgB,uBAAuB,CACtC,WAAW,EAAE,gBAAgB,EAAE,EAC/B,QAAQ,EAAE,kBAAkB,EAC5B,OAAO,GAAE,4BAAiC,GACxC,MAAM,CAAC,2BAA2B,GAAG,SAAS,EAAE,eAAe,CAAC,CA8ElE;AA2BD,OAAO,EAAE,qBAAqB,EAAE,MAAM,YAAY,CAAC;AAEnD,sEAAsE;AACtE,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,EAC7B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,MAAM,CAAC,aAAa,EAAE,eAAe,CAAC,CAAC,CA8EjD;AAED,uFAAuF;AACvF,wBAAsB,aAAa,CAClC,WAAW,EAAE,2BAA2B,EACxC,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,EAC7B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,MAAM,CAAC,mBAAmB,EAAE,eAAe,CAAC,CAAC,CAYvD","sourcesContent":["import {\n\ttype Api,\n\ttype AssistantMessage,\n\ttype Context,\n\tisContextOverflow,\n\ttype Model,\n\ttype Models,\n\ttype RetryCallbacks,\n\ttype RetryPolicy,\n\tretryAssistantCall,\n\ttype SimpleStreamOptions,\n\ttype Usage,\n\tuuidv7,\n} from \"@earendil-works/pi-ai\";\nimport type { AgentMessage, ThinkingLevel } from \"../../types.ts\";\nimport { buildLegacySessionContext, type LegacyCompactionEntry, type SessionTreeEntry } from \"../legacy-session.ts\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport { buildSessionContext } from \"../session/context.ts\";\nimport type { CompactionEntry, Entry } from \"../session/types.ts\";\nimport { CompactionError, err, ok, type Result } 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: Entry[],\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.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: Entry): AgentMessage | undefined {\n\tif (entry.type === \"message\") {\n\t\treturn entry.message as AgentMessage;\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: Entry): AgentMessage | undefined {\n\tif (entry.type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\treturn getMessageFromEntry(entry);\n}\n\nfunction extractLegacyFileOperations(\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 LegacyCompactionEntry;\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 file of details.readFiles) fileOps.read.add(file);\n\t\t\t}\n\t\t\tif (Array.isArray(details.modifiedFiles)) {\n\t\t\t\tfor (const file of details.modifiedFiles) fileOps.edited.add(file);\n\t\t\t}\n\t\t}\n\t}\n\tfor (const message of messages) extractFileOpsFromMessage(message, fileOps);\n\treturn fileOps;\n}\n\nfunction getMessageFromLegacyEntry(entry: SessionTreeEntry): AgentMessage | undefined {\n\tif (entry.type === \"message\") return entry.message;\n\tif (entry.type === \"custom_message\") {\n\t\treturn createCustomMessage(entry.customType, entry.content, entry.display, entry.details, entry.timestamp);\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 getMessageFromLegacyEntryForCompaction(entry: SessionTreeEntry): AgentMessage | undefined {\n\treturn entry.type === \"compaction\" ? undefined : getMessageFromLegacyEntry(entry);\n}\n\n/** Generated compaction data ready to be persisted as a compaction entry. */\nexport interface CompactResult<T = unknown> {\n\t/** Summary text that replaces compacted history in future context. */\n\tsummary: string;\n\t/** Estimated context tokens before compaction. */\n\ttokensBefore: number;\n\t/** Usage from the LLM call(s) that generated this summary, if available. */\n\tusage?: Usage;\n\t/** Retained recent messages stored directly on the compaction entry. */\n\tretainedTail: AgentMessage[];\n\t/** Optional implementation-specific details stored with the compaction entry. */\n\tdetails?: T;\n}\n\n/** Compaction result used by the pi-client/pi-server session-tree protocol. */\nexport interface LegacyCompactResult<T = unknown> extends CompactResult<T> {\n\t/** Entry id where retained history starts. */\n\tfirstKeptEntryId: string;\n}\n\nexport async function completeSimpleWithRetries(\n\tmodels: Models,\n\tmodel: Model<Api>,\n\tcontext: Context,\n\toptions: SimpleStreamOptions,\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\treturn retryAssistantCall(\n\t\t() => models.completeSimple(model, context, requestOptions),\n\t\tretry,\n\t\trequestOptions.signal,\n\t\tcallbacks,\n\t);\n}\n\nfunction combineUsage(first: Usage, second: Usage): Usage {\n\treturn {\n\t\tinput: first.input + second.input,\n\t\toutput: first.output + second.output,\n\t\tcacheRead: first.cacheRead + second.cacheRead,\n\t\tcacheWrite: first.cacheWrite + second.cacheWrite,\n\t\t...(first.cacheWrite1h !== undefined || second.cacheWrite1h !== undefined\n\t\t\t? { cacheWrite1h: (first.cacheWrite1h ?? 0) + (second.cacheWrite1h ?? 0) }\n\t\t\t: {}),\n\t\t...(first.reasoning !== undefined || second.reasoning !== undefined\n\t\t\t? { reasoning: (first.reasoning ?? 0) + (second.reasoning ?? 0) }\n\t\t\t: {}),\n\t\ttotalTokens: first.totalTokens + second.totalTokens,\n\t\tcost: {\n\t\t\tinput: first.cost.input + second.cost.input,\n\t\t\toutput: first.cost.output + second.cost.output,\n\t\t\tcacheRead: first.cost.cacheRead + second.cost.cacheRead,\n\t\t\tcacheWrite: first.cost.cacheWrite + second.cost.cacheWrite,\n\t\t\ttotal: first.cost.total + second.cost.total,\n\t\t},\n\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}\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};\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 (\n\t\t\tassistantMsg.stopReason !== \"aborted\" &&\n\t\t\tassistantMsg.stopReason !== \"error\" &&\n\t\t\tassistantMsg.usage &&\n\t\t\tcalculateContextTokens(assistantMsg.usage) > 0\n\t\t) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/** Return usage from the last valid assistant message in session entries. */\nexport function getLastAssistantUsage(entries: Entry[]): 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/** Return whether context usage exceeds the configured compaction threshold. */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tif (!settings.enabled) return false;\n\treturn contextTokens > contextWindow - settings.reserveTokens;\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/** Estimate token count for one message using a conservative character heuristic. */\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 + safeJsonStringify(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}\nfunction findValidCutPoints(entries: Entry[], 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 \"active_tools_change\":\n\t\t\tcase \"compaction\":\n\t\t\tcase \"branch_summary\":\n\t\t\tcase \"custom\":\n\t\t\t\tbreak;\n\t\t}\n\t\tif (entry.type === \"branch_summary\") cutPoints.push(i);\n\t}\n\treturn cutPoints;\n}\n\n/** Find the user-visible message that starts the turn containing an entry. */\nexport function findTurnStartIndex(entries: Entry[], 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\") {\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: Entry[],\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\nfunction findLegacyValidCutPoints(entries: SessionTreeEntry[], startIndex: number, endIndex: number): number[] {\n\tconst cutPoints: number[] = [];\n\tfor (let index = startIndex; index < endIndex; index++) {\n\t\tconst entry = entries[index];\n\t\tif (entry.type === \"message\") {\n\t\t\tswitch (entry.message.role) {\n\t\t\t\tcase \"bashExecution\":\n\t\t\t\tcase \"custom\":\n\t\t\t\tcase \"branchSummary\":\n\t\t\t\tcase \"compactionSummary\":\n\t\t\t\tcase \"user\":\n\t\t\t\tcase \"assistant\":\n\t\t\t\t\tcutPoints.push(index);\n\t\t\t\t\tbreak;\n\t\t\t\tcase \"toolResult\":\n\t\t\t\t\tbreak;\n\t\t\t}\n\t\t} else if (entry.type === \"branch_summary\") {\n\t\t\tcutPoints.push(index);\n\t\t}\n\t}\n\treturn cutPoints;\n}\n\nfunction findLegacyTurnStartIndex(entries: SessionTreeEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let index = entryIndex; index >= startIndex; index--) {\n\t\tconst entry = entries[index];\n\t\tif (entry.type === \"branch_summary\") return index;\n\t\tif (entry.type === \"message\") {\n\t\t\tconst role = entry.message.role;\n\t\t\tif (role === \"user\" || role === \"bashExecution\") return index;\n\t\t}\n\t}\n\treturn -1;\n}\n\nfunction findLegacyCutPoint(\n\tentries: SessionTreeEntry[],\n\tstartIndex: number,\n\tendIndex: number,\n\tkeepRecentTokens: number,\n): CutPointResult {\n\tconst cutPoints = findLegacyValidCutPoints(entries, startIndex, endIndex);\n\tif (cutPoints.length === 0) {\n\t\treturn { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };\n\t}\n\n\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0];\n\tfor (let index = endIndex - 1; index >= startIndex; index--) {\n\t\tconst entry = entries[index];\n\t\tif (entry.type !== \"message\") continue;\n\t\taccumulatedTokens += estimateTokens(entry.message);\n\t\tif (accumulatedTokens >= keepRecentTokens) {\n\t\t\tfor (const cutPoint of cutPoints) {\n\t\t\t\tif (cutPoint >= index) {\n\t\t\t\t\tcutIndex = cutPoint;\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 previousEntry = entries[cutIndex - 1];\n\t\tif (previousEntry.type === \"compaction\" || previousEntry.type === \"message\") break;\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 : findLegacyTurnStartIndex(entries, cutIndex, startIndex);\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 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## Operational State\n- Modified files: [Exact paths, or \"(none)\"]\n- Read files: [Exact paths, or \"(none)\"]\n- Open failures: [Current unresolved failures/errors, or \"(none)\"]\n- Last command: [Most recent test/build/shell command and exit code, or \"(none)\"]\n- Last failing assertion/error: [Exact assertion, stack frame, or failing lines if available, or \"(none)\"]\n- Pending TODO: [Concrete remaining work, or \"(none)\"]\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## Operational State\n- Modified files: [Preserve/update exact paths, or \"(none)\"]\n- Read files: [Preserve/update exact paths, or \"(none)\"]\n- Open failures: [Current unresolved failures/errors, or \"(none)\"]\n- Last command: [Most recent test/build/shell command and exit code, or \"(none)\"]\n- Last failing assertion/error: [Exact assertion, stack frame, or failing lines if available, or \"(none)\"]\n- Pending TODO: [Concrete remaining work, or \"(none)\"]\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\nfunction getSummaryMaxTokens(model: Model<any>, reserveTokens: number): number {\n\treturn Math.min(Math.floor(0.8 * reserveTokens), model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY);\n}\n\nfunction estimateTextTokens(text: string): number {\n\treturn Math.ceil(text.length / 4);\n}\n\nfunction buildSummaryPrompt(conversationText: string, basePrompt: string, previousSummary: string | undefined): string {\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\treturn promptText + basePrompt;\n}\n\nfunction getSummaryInputBudget(\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tbasePrompt: string,\n\tpreviousSummary: string | undefined,\n): number {\n\tif (model.contextWindow <= 0) return Number.POSITIVE_INFINITY;\n\tconst overheadTokens = estimateTextTokens(buildSummaryPrompt(\"\", basePrompt, previousSummary));\n\treturn Math.max(1, Math.floor(model.contextWindow - maxTokens - overheadTokens));\n}\n\nfunction toSummarySegments(messages: AgentMessage[]): string[] {\n\treturn messages\n\t\t.map((message) => serializeConversation(convertToLlm([message])))\n\t\t.filter((segment) => segment.length > 0);\n}\n\nfunction takeSummaryChunk(segments: string[], tokenBudget: number): { chunk: string; remaining: string[] } {\n\tif (!Number.isFinite(tokenBudget)) return { chunk: segments.join(\"\\n\\n\"), remaining: [] };\n\n\tconst chunk: string[] = [];\n\tlet chunkTokens = 0;\n\n\tfor (let i = 0; i < segments.length; i++) {\n\t\tconst segment = segments[i];\n\t\tconst messageTokens = Math.max(1, estimateTextTokens(segment));\n\t\tif (chunk.length === 0 && messageTokens > tokenBudget) {\n\t\t\tconst chunkChars = Math.max(1, Math.floor(tokenBudget * 4));\n\t\t\tconst remainingSegment = segment.slice(chunkChars);\n\t\t\tconst remaining = remainingSegment ? [remainingSegment, ...segments.slice(i + 1)] : segments.slice(i + 1);\n\t\t\treturn { chunk: segment.slice(0, chunkChars), remaining };\n\t\t}\n\t\tif (chunk.length > 0 && chunkTokens + messageTokens > tokenBudget) {\n\t\t\treturn { chunk: chunk.join(\"\\n\\n\"), remaining: segments.slice(i) };\n\t\t}\n\t\tchunk.push(segment);\n\t\tchunkTokens += messageTokens;\n\t}\n\n\treturn { chunk: chunk.join(\"\\n\\n\"), remaining: [] };\n}\n\nfunction getResponseText(response: AssistantMessage): string {\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\nasync function summarizeChunk(\n\tconversationText: string,\n\tmodels: Models,\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tinitialPrompt: string,\n\tupdatePrompt: string,\n\tpreviousSummary: string | undefined,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<Result<{ text: string; usage: Usage }, CompactionError>> {\n\tconst basePrompt = previousSummary ? updatePrompt : initialPrompt;\n\tconst promptText = buildSummaryPrompt(conversationText, basePrompt, previousSummary);\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, reasoning: thinkingLevel }\n\t\t\t: { maxTokens, signal };\n\n\tconst response = await completeSimpleWithRetries(\n\t\tmodels,\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcompletionOptions,\n\t\tretry,\n\t\tcallbacks,\n\t);\n\n\tif (isContextOverflow(response, model.contextWindow) && conversationText.length > 1) {\n\t\tconst middle = Math.ceil(conversationText.length / 2);\n\t\tconst leftSummary = await summarizeChunk(\n\t\t\tconversationText.slice(0, middle),\n\t\t\tmodels,\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tpreviousSummary,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t\tif (!leftSummary.ok) return leftSummary;\n\t\tconst rightSummary = await summarizeChunk(\n\t\t\tconversationText.slice(middle),\n\t\t\tmodels,\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tleftSummary.value.text,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t\tif (!rightSummary.ok) return rightSummary;\n\t\treturn ok({\n\t\t\ttext: rightSummary.value.text,\n\t\t\tusage: combineUsage(leftSummary.value.usage, rightSummary.value.usage),\n\t\t});\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\treturn ok({ text: getResponseText(response), usage: response.usage });\n}\n\nasync function summarizeMessages(\n\tmessages: AgentMessage[],\n\tmodels: Models,\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tinitialPrompt: string,\n\tupdatePrompt: string,\n\tpreviousSummary: string | undefined,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<Result<{ text: string; usage: Usage }, CompactionError>> {\n\tlet remaining = toSummarySegments(messages);\n\tif (remaining.length === 0) {\n\t\treturn summarizeChunk(\n\t\t\t\"\",\n\t\t\tmodels,\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tpreviousSummary,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t}\n\tlet summary = previousSummary;\n\tlet summaryUsage: Usage | undefined;\n\twhile (remaining.length > 0) {\n\t\tconst basePrompt = summary ? updatePrompt : initialPrompt;\n\t\tconst next = takeSummaryChunk(remaining, getSummaryInputBudget(model, maxTokens, basePrompt, summary));\n\t\tconst chunkSummary = await summarizeChunk(\n\t\t\tnext.chunk,\n\t\t\tmodels,\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tsummary,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t\tif (!chunkSummary.ok) return chunkSummary;\n\t\tsummary = chunkSummary.value.text;\n\t\tsummaryUsage = summaryUsage ? combineUsage(summaryUsage, chunkSummary.value.usage) : chunkSummary.value.usage;\n\t\tremaining = next.remaining;\n\t}\n\tif (!summaryUsage) {\n\t\treturn err(new CompactionError(\"summarization_failed\", \"Summarization returned no usage data\"));\n\t}\n\treturn ok({ text: summary ?? \"\", usage: summaryUsage });\n}\n\n/** Generate or update a conversation summary for compaction. */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodels: Models,\n\tmodel: Model<Api>,\n\treserveTokens: number,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<Result<string, CompactionError>> {\n\tconst result = await generateSummaryWithUsage(\n\t\tcurrentMessages,\n\t\tmodels,\n\t\tmodel,\n\t\treserveTokens,\n\t\tsignal,\n\t\tcustomInstructions,\n\t\tpreviousSummary,\n\t\tthinkingLevel,\n\t\tretry,\n\t\tcallbacks,\n\t);\n\treturn result.ok ? ok(result.value.text) : err(result.error);\n}\n\n/** Generate or update a conversation summary and return its provider usage. */\nexport async function generateSummaryWithUsage(\n\tcurrentMessages: AgentMessage[],\n\tmodels: Models,\n\tmodel: Model<Api>,\n\treserveTokens: number,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<Result<{ text: string; usage: Usage }, CompactionError>> {\n\tconst maxTokens = getSummaryMaxTokens(model, reserveTokens);\n\tlet initialPrompt = SUMMARIZATION_PROMPT;\n\tlet updatePrompt = UPDATE_SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tinitialPrompt = `${initialPrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t\tupdatePrompt = `${updatePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\treturn summarizeMessages(\n\t\tcurrentMessages,\n\t\tmodels,\n\t\tmodel,\n\t\tmaxTokens,\n\t\tsignal,\n\t\tthinkingLevel,\n\t\tinitialPrompt,\n\t\tupdatePrompt,\n\t\tpreviousSummary,\n\t\tretry,\n\t\tcallbacks,\n\t);\n}\n\n/** Prepared inputs for a compaction run. */\nexport interface CompactionPreparation {\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/** Recent messages retained after compaction and stored on the compaction entry. */\n\tretainedTail: 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: Entry[],\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 compactableEntries = pathEntries;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tconst virtualRetainedEntries: Entry[] = prevCompaction.retainedTail.map((message, index) => ({\n\t\t\ttype: \"message\",\n\t\t\tid: `${prevCompaction.id}:retained:${index}`,\n\t\t\tparentId: index === 0 ? prevCompaction.id : `${prevCompaction.id}:retained:${index - 1}`,\n\t\t\tseq: prevCompaction.seq,\n\t\t\ttimestamp: message.timestamp,\n\t\t\tmessage,\n\t\t}));\n\t\tcompactableEntries = [...virtualRetainedEntries, ...pathEntries.slice(prevCompactionIndex + 1)];\n\t}\n\tconst boundaryEnd = compactableEntries.length;\n\n\tconst tokensBefore = estimateContextTokens(buildSessionContext(pathEntries).messages).tokens;\n\tconst cutPoint = findCutPoint(compactableEntries, 0, boundaryEnd, settings.keepRecentTokens);\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let i = 0; i < historyEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(compactableEntries[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(compactableEntries[i]);\n\t\t\tif (msg) turnPrefixMessages.push(msg);\n\t\t}\n\t}\n\tconst retainedTail: AgentMessage[] = [];\n\tfor (let i = cutPoint.firstKeptEntryIndex; i < boundaryEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(compactableEntries[i]);\n\t\tif (msg) retainedTail.push(msg);\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\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tretainedTail,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t});\n}\n\n/** Optional cut-point override for the pi-client/pi-server legacy session tree. */\nexport interface CompactionPreparationOptions {\n\tfirstKeptEntryId?: string;\n}\n\n/** Prepared compaction inputs with the retained entry id required by the pi-server protocol. */\nexport interface LegacyCompactionPreparation extends CompactionPreparation {\n\tfirstKeptEntryId: string;\n}\n\n/** Prepare a pi-client/pi-server session branch without changing its durable entry representation. */\nexport function prepareLegacyCompaction(\n\tpathEntries: SessionTreeEntry[],\n\tsettings: CompactionSettings,\n\toptions: CompactionPreparationOptions = {},\n): Result<LegacyCompactionPreparation | 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 index = pathEntries.length - 1; index >= 0; index--) {\n\t\tif (pathEntries[index].type === \"compaction\") {\n\t\t\tprevCompactionIndex = index;\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 LegacyCompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tconst firstKeptEntryIndex = prevCompaction.firstKeptEntryId\n\t\t\t? pathEntries.findIndex((entry) => entry.id === prevCompaction.firstKeptEntryId)\n\t\t\t: -1;\n\t\tboundaryStart = firstKeptEntryIndex >= 0 ? firstKeptEntryIndex : prevCompactionIndex + 1;\n\t}\n\tconst boundaryEnd = pathEntries.length;\n\tconst tokensBefore = estimateContextTokens(buildLegacySessionContext(pathEntries).messages).tokens;\n\n\tconst forcedFirstKeptEntryIndex = options.firstKeptEntryId\n\t\t? pathEntries.findIndex((entry) => entry.id === options.firstKeptEntryId)\n\t\t: -1;\n\tif (\n\t\toptions.firstKeptEntryId &&\n\t\t(forcedFirstKeptEntryIndex < boundaryStart || forcedFirstKeptEntryIndex >= boundaryEnd)\n\t) {\n\t\treturn err(new CompactionError(\"invalid_session\", \"Forced first kept entry was not found\"));\n\t}\n\tconst cutPoint =\n\t\tforcedFirstKeptEntryIndex >= boundaryStart && forcedFirstKeptEntryIndex < boundaryEnd\n\t\t\t? { firstKeptEntryIndex: forcedFirstKeptEntryIndex, turnStartIndex: -1, isSplitTurn: false }\n\t\t\t: findLegacyCutPoint(pathEntries, boundaryStart, boundaryEnd, settings.keepRecentTokens);\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\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let index = boundaryStart; index < historyEnd; index++) {\n\t\tconst message = getMessageFromLegacyEntryForCompaction(pathEntries[index]);\n\t\tif (message) messagesToSummarize.push(message);\n\t}\n\tconst turnPrefixMessages: AgentMessage[] = [];\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (let index = cutPoint.turnStartIndex; index < cutPoint.firstKeptEntryIndex; index++) {\n\t\t\tconst message = getMessageFromLegacyEntryForCompaction(pathEntries[index]);\n\t\t\tif (message) turnPrefixMessages.push(message);\n\t\t}\n\t}\n\tconst retainedTail: AgentMessage[] = [];\n\tfor (let index = cutPoint.firstKeptEntryIndex; index < boundaryEnd; index++) {\n\t\tconst message = getMessageFromLegacyEntryForCompaction(pathEntries[index]);\n\t\tif (message) retainedTail.push(message);\n\t}\n\tconst fileOps = extractLegacyFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (const message of turnPrefixMessages) extractFileOpsFromMessage(message, fileOps);\n\t}\n\n\treturn ok({\n\t\tfirstKeptEntryId: firstKeptEntry.id,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tretainedTail,\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\nconst UPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT = `The messages above are the next part of the same oversized turn prefix.\n\nUpdate the existing turn-prefix summary in <previous-summary> tags. Keep the same sections:\n\n## Original Request\n## Early Progress\n## Context for Suffix\n\nBe concise. Preserve exact file paths, function names, and error messages.`;\n\nexport { serializeConversation } from \"./utils.ts\";\n\n/** Generate compaction summary data from prepared session history. */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tmodels: Models,\n\tmodel: Model<Api>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<Result<CompactResult, CompactionError>> {\n\tconst {\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tretainedTail,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\tlet summary: string;\n\tlet summaryUsage: Usage;\n\n\tif (isSplitTurn && turnPrefixMessages.length > 0) {\n\t\tlet historyText = \"No prior history.\";\n\t\tlet historyUsage: Usage | undefined;\n\t\tif (messagesToSummarize.length > 0) {\n\t\t\tconst historyResult = await generateSummaryWithUsage(\n\t\t\t\tmessagesToSummarize,\n\t\t\t\tmodels,\n\t\t\t\tmodel,\n\t\t\t\tsettings.reserveTokens,\n\t\t\t\tsignal,\n\t\t\t\tcustomInstructions,\n\t\t\t\tpreviousSummary,\n\t\t\t\tthinkingLevel,\n\t\t\t\tretry,\n\t\t\t\tcallbacks,\n\t\t\t);\n\t\t\tif (!historyResult.ok) return err(historyResult.error);\n\t\t\thistoryText = historyResult.value.text;\n\t\t\thistoryUsage = historyResult.value.usage;\n\t\t}\n\t\tconst turnPrefixResult = await generateTurnPrefixSummary(\n\t\t\tturnPrefixMessages,\n\t\t\tmodels,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t\tif (!turnPrefixResult.ok) return err(turnPrefixResult.error);\n\t\tsummary = `${historyText}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult.value.text}`;\n\t\tsummaryUsage = historyUsage\n\t\t\t? combineUsage(historyUsage, turnPrefixResult.value.usage)\n\t\t\t: turnPrefixResult.value.usage;\n\t} else {\n\t\tconst summaryResult = await generateSummaryWithUsage(\n\t\t\tmessagesToSummarize,\n\t\t\tmodels,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t\tif (!summaryResult.ok) return err(summaryResult.error);\n\t\tsummary = summaryResult.value.text;\n\t\tsummaryUsage = summaryResult.value.usage;\n\t}\n\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\treturn ok({\n\t\tsummary,\n\t\ttokensBefore,\n\t\tusage: summaryUsage,\n\t\tretainedTail,\n\t\tdetails: { readFiles, modifiedFiles } as CompactionDetails,\n\t});\n}\n\n/** Compact a pi-client/pi-server session branch and preserve its retained entry id. */\nexport async function compactLegacy(\n\tpreparation: LegacyCompactionPreparation,\n\tmodels: Models,\n\tmodel: Model<Api>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<Result<LegacyCompactResult, CompactionError>> {\n\tconst result = await compact(\n\t\tpreparation,\n\t\tmodels,\n\t\tmodel,\n\t\tcustomInstructions,\n\t\tsignal,\n\t\tthinkingLevel,\n\t\tretry,\n\t\tcallbacks,\n\t);\n\treturn result.ok ? ok({ ...result.value, firstKeptEntryId: preparation.firstKeptEntryId }) : err(result.error);\n}\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodels: Models,\n\tmodel: Model<Api>,\n\treserveTokens: number,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<Result<{ text: string; usage: Usage }, 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\n\tconst result = await summarizeMessages(\n\t\tmessages,\n\t\tmodels,\n\t\tmodel,\n\t\tmaxTokens,\n\t\tsignal,\n\t\tthinkingLevel,\n\t\tTURN_PREFIX_SUMMARIZATION_PROMPT,\n\t\tUPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT,\n\t\tundefined,\n\t\tretry,\n\t\tcallbacks,\n\t);\n\tif (!result.ok && result.error.code === \"summarization_failed\") {\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: ${result.error.message.replace(/^Summarization failed: /, \"\")}`,\n\t\t\t),\n\t\t);\n\t}\n\treturn result;\n}\n"]}
+{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/harness/compaction/compaction.ts"],"names":[],"mappings":"AAAA,OAAO,EACN,KAAK,OAAO,IAAI,SAAS,EACzB,KAAK,GAAG,EACR,KAAK,gBAAgB,EAGrB,KAAK,KAAK,EACV,KAAK,MAAM,EACX,KAAK,cAAc,EACnB,KAAK,WAAW,EAEhB,KAAK,mBAAmB,EACxB,KAAK,KAAK,EAEV,MAAM,uBAAuB,CAAC;AAC/B,OAAO,KAAK,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAClE,OAAO,EAAsB,KAAK,OAAO,EAAwC,MAAM,eAAe,CAAC;AACvG,OAAO,EAAyD,KAAK,gBAAgB,EAAE,MAAM,sBAAsB,CAAC;AAQpH,OAAO,KAAK,EAAmB,KAAK,EAAE,SAAS,EAAE,MAAM,qBAAqB,CAAC;AAC7E,OAAO,EAAE,eAAe,EAAW,KAAK,MAAM,EAAE,MAAM,aAAa,CAAC;AAEpE,OAAO,EAIN,KAAK,cAAc,EAGnB,MAAM,YAAY,CAAC;AAEpB,qEAAqE;AACrE,MAAM,WAAW,iBAAkB,SAAQ,MAAM,CAAC,MAAM,EAAE,SAAS,CAAC;IACnE,2CAA2C;IAC3C,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,+CAA+C;IAC/C,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AAoGD,6EAA6E;AAC7E,MAAM,WAAW,aAAa,CAAC,CAAC,GAAG,SAAS;IAC3C,sEAAsE;IACtE,OAAO,EAAE,MAAM,CAAC;IAChB,kDAAkD;IAClD,YAAY,EAAE,MAAM,CAAC;IACrB,4EAA4E;IAC5E,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,wEAAwE;IACxE,YAAY,EAAE,YAAY,EAAE,CAAC;IAC7B,iFAAiF;IACjF,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AAED,+EAA+E;AAC/E,MAAM,WAAW,mBAAmB,CAAC,CAAC,GAAG,OAAO,CAAE,SAAQ,aAAa,CAAC,CAAC,CAAC;IACzE,8CAA8C;IAC9C,gBAAgB,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,MAAM,cAAc,GAAG,CAC5B,SAAS,EAAE,SAAS,EACpB,OAAO,EAAE,mBAAmB,EAC5B,OAAO,EAAE,OAAO,KACZ,OAAO,CAAC,gBAAgB,CAAC,CAAC;AAE/B,wBAAgB,2BAA2B,CAAC,OAAO,EAAE,mBAAmB,EAAE,OAAO,EAAE,OAAO,GAAG,mBAAmB,CAQ/G;AAED,wBAAsB,yBAAyB,CAC9C,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,SAAS,EAAE,SAAS,EACpB,OAAO,EAAE,mBAAmB,EAC5B,KAAK,EAAE,WAAW,GAAG,SAAS,EAC9B,SAAS,EAAE,cAAc,GAAG,SAAS,EACrC,OAAO,EAAE,OAAO,GACd,OAAO,CAAC,gBAAgB,CAAC,CAS3B;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;CACzB;AAED,uDAAuD;AACvD,eAAO,MAAM,2BAA2B,EAAE,kBAIzC,CAAC;AAEF,0DAA0D;AAC1D,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAgBD,6EAA6E;AAC7E,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,KAAK,EAAE,GAAG,KAAK,GAAG,SAAS,CASzE;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,gFAAgF;AAChF,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAoBD,qFAAqF;AACrF,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAwC5D;AAgCD,8EAA8E;AAC9E,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,KAAK,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAcnG;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,KAAK,EAAE,EAChB,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CA2ChB;AAgFD,eAAO,MAAM,2BAA2B,6TAEmF,CAAC;AAsR5H,gEAAgE;AAChE,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,EAC1B,OAAO,GAAE,OAA4B,GACnC,OAAO,CAAC,MAAM,CAAC,MAAM,EAAE,eAAe,CAAC,CAAC,CAc1C;AAED,+EAA+E;AAC/E,wBAAgB,wBAAwB,CACvC,eAAe,EAAE,YAAY,EAAE,EAC/B,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,EAC1B,OAAO,GAAE,OAA4B,GACnC,OAAO,CAAC,MAAM,CAAC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,KAAK,CAAA;CAAE,EAAE,eAAe,CAAC,CAAC,CAqBlE;AAED,MAAM,WAAW,wBAAwB;IACxC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IAClB,aAAa,EAAE,MAAM,CAAC;IACtB,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,aAAa,CAAC,EAAE,aAAa,CAAC;CAC9B;AAED,wEAAwE;AACxE,wBAAgB,0BAA0B,CACzC,eAAe,EAAE,YAAY,EAAE,EAC/B,OAAO,EAAE,wBAAwB,EACjC,OAAO,EAAE,cAAc,EACvB,OAAO,EAAE,OAAO,GACd,OAAO,CAAC,MAAM,CAAC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,KAAK,CAAA;CAAE,EAAE,eAAe,CAAC,CAAC,CAoBlE;AAED,4CAA4C;AAC5C,MAAM,WAAW,qBAAqB;IACrC,oDAAoD;IACpD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,oFAAoF;IACpF,YAAY,EAAE,YAAY,EAAE,CAAC;IAC7B,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,KAAK,EAAE,EACpB,QAAQ,EAAE,kBAAkB,GAC1B,MAAM,CAAC,qBAAqB,GAAG,SAAS,EAAE,eAAe,CAAC,CAqE5D;AAED,mFAAmF;AACnF,MAAM,WAAW,4BAA4B;IAC5C,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,gGAAgG;AAChG,MAAM,WAAW,2BAA4B,SAAQ,qBAAqB;IACzE,gBAAgB,EAAE,MAAM,CAAC;CACzB;AAED,sGAAsG;AACtG,wBAAgB,uBAAuB,CACtC,WAAW,EAAE,gBAAgB,EAAE,EAC/B,QAAQ,EAAE,kBAAkB,EAC5B,OAAO,GAAE,4BAAiC,GACxC,MAAM,CAAC,2BAA2B,GAAG,SAAS,EAAE,eAAe,CAAC,CA8ElE;AA2BD,OAAO,EAAE,qBAAqB,EAAE,MAAM,YAAY,CAAC;AAEnD,sEAAsE;AACtE,wBAAgB,OAAO,CACtB,WAAW,EAAE,qBAAqB,EAClC,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,kBAAkB,EAAE,MAAM,GAAG,SAAS,EACtC,aAAa,EAAE,aAAa,GAAG,SAAS,EACxC,KAAK,EAAE,WAAW,GAAG,SAAS,EAC9B,SAAS,EAAE,cAAc,GAAG,SAAS,EACrC,OAAO,EAAE,OAAO,GACd,OAAO,CAAC,MAAM,CAAC,aAAa,EAAE,eAAe,CAAC,CAAC,CAQjD;AAED,MAAM,WAAW,wBAAwB;IACxC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IAClB,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,aAAa,CAAC,EAAE,aAAa,CAAC;CAC9B;AAED,0FAA0F;AAC1F,wBAAsB,kBAAkB,CACvC,WAAW,EAAE,qBAAqB,EAClC,OAAO,EAAE,wBAAwB,EACjC,OAAO,EAAE,cAAc,EACvB,OAAO,EAAE,OAAO,GACd,OAAO,CAAC,MAAM,CAAC,aAAa,EAAE,eAAe,CAAC,CAAC,CA0DjD;AAED,uFAAuF;AACvF,wBAAsB,aAAa,CAClC,WAAW,EAAE,2BAA2B,EACxC,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,EAC7B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,MAAM,CAAC,mBAAmB,EAAE,eAAe,CAAC,CAAC,CAMvD","sourcesContent":["import {\n\ttype Context as AiContext,\n\ttype Api,\n\ttype AssistantMessage,\n\tcontentText,\n\tisContextOverflow,\n\ttype Model,\n\ttype Models,\n\ttype RetryCallbacks,\n\ttype RetryPolicy,\n\tretryAssistantCall,\n\ttype SimpleStreamOptions,\n\ttype Usage,\n\tuuidv7,\n} from \"@earendil-works/pi-ai\";\nimport type { AgentMessage, ThinkingLevel } from \"../../types.ts\";\nimport { BACKGROUND_CONTEXT, type Context, getTelemetryContext, withAbortSignal } from \"../context.ts\";\nimport { buildLegacySessionContext, type LegacyCompactionEntry, type SessionTreeEntry } from \"../legacy-session.ts\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport { buildContextEntries, sessionEntryToContextMessages } from \"../session/context.ts\";\nimport type { CompactionEntry, Entry, JsonValue } from \"../session/types.ts\";\nimport { CompactionError, err, ok, type Result } from \"../types.ts\";\nimport { addUsage } from \"../utils/usage.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 extends Record<string, JsonValue> {\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: Entry[],\n\tprevCompactionIndex: number,\n): FileOperations {\n\tconst fileOps = createFileOps();\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = entries[prevCompactionIndex] as CompactionEntry;\n\t\tif (\n\t\t\ttypeof prevCompaction.details === \"object\" &&\n\t\t\tprevCompaction.details !== null &&\n\t\t\t!Array.isArray(prevCompaction.details)\n\t\t) {\n\t\t\tif (Array.isArray(prevCompaction.details.readFiles)) {\n\t\t\t\tfor (const path of prevCompaction.details.readFiles) {\n\t\t\t\t\tif (typeof path === \"string\") fileOps.read.add(path);\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (Array.isArray(prevCompaction.details.modifiedFiles)) {\n\t\t\t\tfor (const path of prevCompaction.details.modifiedFiles) {\n\t\t\t\t\tif (typeof path === \"string\") fileOps.edited.add(path);\n\t\t\t\t}\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: Entry): AgentMessage | undefined {\n\tif (entry.type === \"message\") {\n\t\treturn entry.message as AgentMessage;\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: Entry): AgentMessage | undefined {\n\tif (entry.type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\treturn getMessageFromEntry(entry);\n}\n\nfunction extractLegacyFileOperations(\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 LegacyCompactionEntry;\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 file of details.readFiles) fileOps.read.add(file);\n\t\t\t}\n\t\t\tif (Array.isArray(details.modifiedFiles)) {\n\t\t\t\tfor (const file of details.modifiedFiles) fileOps.edited.add(file);\n\t\t\t}\n\t\t}\n\t}\n\tfor (const message of messages) extractFileOpsFromMessage(message, fileOps);\n\treturn fileOps;\n}\n\nfunction getMessageFromLegacyEntry(entry: SessionTreeEntry): AgentMessage | undefined {\n\tif (entry.type === \"message\") return entry.message;\n\tif (entry.type === \"custom_message\") {\n\t\treturn createCustomMessage(entry.customType, entry.content, entry.display, entry.details, entry.timestamp);\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 getMessageFromLegacyEntryForCompaction(entry: SessionTreeEntry): AgentMessage | undefined {\n\treturn entry.type === \"compaction\" ? undefined : getMessageFromLegacyEntry(entry);\n}\n\n/** Generated compaction data ready to be persisted as a compaction entry. */\nexport interface CompactResult<T = JsonValue> {\n\t/** Summary text that replaces compacted history in future context. */\n\tsummary: string;\n\t/** Estimated context tokens before compaction. */\n\ttokensBefore: number;\n\t/** Usage from the LLM call(s) that generated this summary, if available. */\n\tusage?: Usage;\n\t/** Retained recent messages stored directly on the compaction entry. */\n\tretainedTail: AgentMessage[];\n\t/** Optional implementation-specific details stored with the compaction entry. */\n\tdetails?: T;\n}\n\n/** Compaction result used by the pi-client/pi-server session-tree protocol. */\nexport interface LegacyCompactResult<T = unknown> extends CompactResult<T> {\n\t/** Entry id where retained history starts. */\n\tfirstKeptEntryId: string;\n}\n\nexport type SummaryRequest = (\n\taiContext: AiContext,\n\toptions: SimpleStreamOptions,\n\tcontext: Context,\n) => Promise<AssistantMessage>;\n\nexport function createSummaryRequestOptions(options: SimpleStreamOptions, context: Context): SimpleStreamOptions {\n\treturn {\n\t\t...options,\n\t\tsignal: context.abortSignal,\n\t\ttelemetryContext: getTelemetryContext(context),\n\t\tcacheRetention: \"none\",\n\t\tsessionId: options.sessionId ?? uuidv7(),\n\t};\n}\n\nexport async function completeSimpleWithRetries(\n\tmodels: Models,\n\tmodel: Model<Api>,\n\taiContext: AiContext,\n\toptions: SimpleStreamOptions,\n\tretry: RetryPolicy | undefined,\n\tcallbacks: RetryCallbacks | undefined,\n\tcontext: Context,\n): Promise<AssistantMessage> {\n\t// Summaries are standalone requests, so isolate routing and avoid cache writes that cannot be reused.\n\tconst requestOptions = createSummaryRequestOptions(options, context);\n\treturn retryAssistantCall(\n\t\t() => models.completeSimple(model, aiContext, requestOptions),\n\t\tretry,\n\t\trequestOptions.signal,\n\t\tcallbacks,\n\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}\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};\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 (\n\t\t\tassistantMsg.stopReason !== \"aborted\" &&\n\t\t\tassistantMsg.stopReason !== \"error\" &&\n\t\t\tassistantMsg.usage &&\n\t\t\tcalculateContextTokens(assistantMsg.usage) > 0\n\t\t) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/** Return usage from the last valid assistant message in session entries. */\nexport function getLastAssistantUsage(entries: Entry[]): 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/** Return whether context usage exceeds the configured compaction threshold. */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tif (!settings.enabled) return false;\n\treturn contextTokens > contextWindow - settings.reserveTokens;\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/** Estimate token count for one message using a conservative character heuristic. */\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 + safeJsonStringify(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}\nfunction findValidCutPoints(entries: Entry[], 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 \"compaction\":\n\t\t\tcase \"branch_summary\":\n\t\t\tcase \"custom\":\n\t\t\t\tbreak;\n\t\t}\n\t\tif (entry.type === \"branch_summary\") cutPoints.push(i);\n\t}\n\treturn cutPoints;\n}\n\n/** Find the user-visible message that starts the turn containing an entry. */\nexport function findTurnStartIndex(entries: Entry[], 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\") {\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: Entry[],\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\nfunction findLegacyValidCutPoints(entries: SessionTreeEntry[], startIndex: number, endIndex: number): number[] {\n\tconst cutPoints: number[] = [];\n\tfor (let index = startIndex; index < endIndex; index++) {\n\t\tconst entry = entries[index];\n\t\tif (entry.type === \"message\") {\n\t\t\tswitch (entry.message.role) {\n\t\t\t\tcase \"bashExecution\":\n\t\t\t\tcase \"custom\":\n\t\t\t\tcase \"branchSummary\":\n\t\t\t\tcase \"compactionSummary\":\n\t\t\t\tcase \"user\":\n\t\t\t\tcase \"assistant\":\n\t\t\t\t\tcutPoints.push(index);\n\t\t\t\t\tbreak;\n\t\t\t\tcase \"toolResult\":\n\t\t\t\t\tbreak;\n\t\t\t}\n\t\t} else if (entry.type === \"branch_summary\") {\n\t\t\tcutPoints.push(index);\n\t\t}\n\t}\n\treturn cutPoints;\n}\n\nfunction findLegacyTurnStartIndex(entries: SessionTreeEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let index = entryIndex; index >= startIndex; index--) {\n\t\tconst entry = entries[index];\n\t\tif (entry.type === \"branch_summary\") return index;\n\t\tif (entry.type === \"message\") {\n\t\t\tconst role = entry.message.role;\n\t\t\tif (role === \"user\" || role === \"bashExecution\") return index;\n\t\t}\n\t}\n\treturn -1;\n}\n\nfunction findLegacyCutPoint(\n\tentries: SessionTreeEntry[],\n\tstartIndex: number,\n\tendIndex: number,\n\tkeepRecentTokens: number,\n): CutPointResult {\n\tconst cutPoints = findLegacyValidCutPoints(entries, startIndex, endIndex);\n\tif (cutPoints.length === 0) {\n\t\treturn { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };\n\t}\n\n\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0];\n\tfor (let index = endIndex - 1; index >= startIndex; index--) {\n\t\tconst entry = entries[index];\n\t\tif (entry.type !== \"message\") continue;\n\t\taccumulatedTokens += estimateTokens(entry.message);\n\t\tif (accumulatedTokens >= keepRecentTokens) {\n\t\t\tfor (const cutPoint of cutPoints) {\n\t\t\t\tif (cutPoint >= index) {\n\t\t\t\t\tcutIndex = cutPoint;\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 previousEntry = entries[cutIndex - 1];\n\t\tif (previousEntry.type === \"compaction\" || previousEntry.type === \"message\") break;\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 : findLegacyTurnStartIndex(entries, cutIndex, startIndex);\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 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## Operational State\n- Modified files: [Exact paths, or \"(none)\"]\n- Read files: [Exact paths, or \"(none)\"]\n- Open failures: [Current unresolved failures/errors, or \"(none)\"]\n- Last command: [Most recent test/build/shell command and exit code, or \"(none)\"]\n- Last failing assertion/error: [Exact assertion, stack frame, or failing lines if available, or \"(none)\"]\n- Pending TODO: [Concrete remaining work, or \"(none)\"]\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## Operational State\n- Modified files: [Preserve/update exact paths, or \"(none)\"]\n- Read files: [Preserve/update exact paths, or \"(none)\"]\n- Open failures: [Current unresolved failures/errors, or \"(none)\"]\n- Last command: [Most recent test/build/shell command and exit code, or \"(none)\"]\n- Last failing assertion/error: [Exact assertion, stack frame, or failing lines if available, or \"(none)\"]\n- Pending TODO: [Concrete remaining work, or \"(none)\"]\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\nfunction getSummaryMaxTokens(model: Model<any>, reserveTokens: number): number {\n\treturn Math.min(Math.floor(0.8 * reserveTokens), model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY);\n}\n\nfunction estimateTextTokens(text: string): number {\n\treturn Math.ceil(text.length / 4);\n}\n\nfunction buildSummaryPrompt(conversationText: string, basePrompt: string, previousSummary: string | undefined): string {\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\treturn promptText + basePrompt;\n}\n\nfunction getSummaryInputBudget(\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tbasePrompt: string,\n\tpreviousSummary: string | undefined,\n): number {\n\tif (model.contextWindow <= 0) return Number.POSITIVE_INFINITY;\n\tconst overheadTokens = estimateTextTokens(buildSummaryPrompt(\"\", basePrompt, previousSummary));\n\treturn Math.max(1, Math.floor(model.contextWindow - maxTokens - overheadTokens));\n}\n\nfunction toSummarySegments(messages: AgentMessage[]): string[] {\n\treturn messages\n\t\t.map((message) => serializeConversation(convertToLlm([message])))\n\t\t.filter((segment) => segment.length > 0);\n}\n\nfunction takeSummaryChunk(segments: string[], tokenBudget: number): { chunk: string; remaining: string[] } {\n\tif (!Number.isFinite(tokenBudget)) return { chunk: segments.join(\"\\n\\n\"), remaining: [] };\n\n\tconst chunk: string[] = [];\n\tlet chunkTokens = 0;\n\n\tfor (let i = 0; i < segments.length; i++) {\n\t\tconst segment = segments[i];\n\t\tconst messageTokens = Math.max(1, estimateTextTokens(segment));\n\t\tif (chunk.length === 0 && messageTokens > tokenBudget) {\n\t\t\tconst chunkChars = Math.max(1, Math.floor(tokenBudget * 4));\n\t\t\tconst remainingSegment = segment.slice(chunkChars);\n\t\t\tconst remaining = remainingSegment ? [remainingSegment, ...segments.slice(i + 1)] : segments.slice(i + 1);\n\t\t\treturn { chunk: segment.slice(0, chunkChars), remaining };\n\t\t}\n\t\tif (chunk.length > 0 && chunkTokens + messageTokens > tokenBudget) {\n\t\t\treturn { chunk: chunk.join(\"\\n\\n\"), remaining: segments.slice(i) };\n\t\t}\n\t\tchunk.push(segment);\n\t\tchunkTokens += messageTokens;\n\t}\n\n\treturn { chunk: chunk.join(\"\\n\\n\"), remaining: [] };\n}\n\nasync function summarizeChunk(\n\tconversationText: string,\n\tmodel: Model<Api>,\n\tmaxTokens: number,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tinitialPrompt: string,\n\tupdatePrompt: string,\n\tpreviousSummary: string | undefined,\n\trequest: SummaryRequest,\n\tcontext: Context,\n): Promise<Result<{ text: string; usage: Usage }, CompactionError>> {\n\tconst basePrompt = previousSummary ? updatePrompt : initialPrompt;\n\tconst promptText = buildSummaryPrompt(conversationText, basePrompt, previousSummary);\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, reasoning: thinkingLevel }\n\t\t\t: { maxTokens };\n\n\tconst response = await request(\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcreateSummaryRequestOptions(completionOptions, context),\n\t\tcontext,\n\t);\n\n\tif (isContextOverflow(response, model.contextWindow) && conversationText.length > 1) {\n\t\tconst middle = Math.ceil(conversationText.length / 2);\n\t\tconst leftSummary = await summarizeChunk(\n\t\t\tconversationText.slice(0, middle),\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tthinkingLevel,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tpreviousSummary,\n\t\t\trequest,\n\t\t\tcontext,\n\t\t);\n\t\tif (!leftSummary.ok) return leftSummary;\n\t\tconst rightSummary = await summarizeChunk(\n\t\t\tconversationText.slice(middle),\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tthinkingLevel,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tleftSummary.value.text,\n\t\t\trequest,\n\t\t\tcontext,\n\t\t);\n\t\tif (!rightSummary.ok) return rightSummary;\n\t\treturn ok({\n\t\t\ttext: rightSummary.value.text,\n\t\t\tusage: addUsage(leftSummary.value.usage, rightSummary.value.usage),\n\t\t});\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\treturn ok({ text: contentText(response.content), usage: response.usage });\n}\n\nasync function summarizeMessages(\n\tmessages: AgentMessage[],\n\tmodel: Model<Api>,\n\tmaxTokens: number,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tinitialPrompt: string,\n\tupdatePrompt: string,\n\tpreviousSummary: string | undefined,\n\trequest: SummaryRequest,\n\tcontext: Context,\n): Promise<Result<{ text: string; usage: Usage }, CompactionError>> {\n\tlet remaining = toSummarySegments(messages);\n\tif (remaining.length === 0) {\n\t\treturn summarizeChunk(\n\t\t\t\"\",\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tthinkingLevel,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tpreviousSummary,\n\t\t\trequest,\n\t\t\tcontext,\n\t\t);\n\t}\n\tlet summary = previousSummary;\n\tlet summaryUsage: Usage | undefined;\n\twhile (remaining.length > 0) {\n\t\tconst basePrompt = summary ? updatePrompt : initialPrompt;\n\t\tconst next = takeSummaryChunk(remaining, getSummaryInputBudget(model, maxTokens, basePrompt, summary));\n\t\tconst chunkSummary = await summarizeChunk(\n\t\t\tnext.chunk,\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tthinkingLevel,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tsummary,\n\t\t\trequest,\n\t\t\tcontext,\n\t\t);\n\t\tif (!chunkSummary.ok) return chunkSummary;\n\t\tsummary = chunkSummary.value.text;\n\t\tsummaryUsage = summaryUsage ? addUsage(summaryUsage, chunkSummary.value.usage) : chunkSummary.value.usage;\n\t\tremaining = next.remaining;\n\t}\n\tif (!summaryUsage) {\n\t\treturn err(new CompactionError(\"summarization_failed\", \"Summarization returned no usage data\"));\n\t}\n\treturn ok({ text: summary ?? \"\", usage: summaryUsage });\n}\n\n/** Generate or update a conversation summary for compaction. */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodels: Models,\n\tmodel: Model<Api>,\n\treserveTokens: number,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tcontext: Context = BACKGROUND_CONTEXT,\n): Promise<Result<string, CompactionError>> {\n\tconst result = await generateSummaryWithUsage(\n\t\tcurrentMessages,\n\t\tmodels,\n\t\tmodel,\n\t\treserveTokens,\n\t\tcustomInstructions,\n\t\tpreviousSummary,\n\t\tthinkingLevel,\n\t\tretry,\n\t\tcallbacks,\n\t\tcontext,\n\t);\n\treturn result.ok ? ok(result.value.text) : err(result.error);\n}\n\n/** Generate or update a conversation summary and return its provider usage. */\nexport function generateSummaryWithUsage(\n\tcurrentMessages: AgentMessage[],\n\tmodels: Models,\n\tmodel: Model<Api>,\n\treserveTokens: number,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tcontext: Context = BACKGROUND_CONTEXT,\n): Promise<Result<{ text: string; usage: Usage }, CompactionError>> {\n\tconst maxTokens = getSummaryMaxTokens(model, reserveTokens);\n\tlet initialPrompt = SUMMARIZATION_PROMPT;\n\tlet updatePrompt = UPDATE_SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tinitialPrompt = `${initialPrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t\tupdatePrompt = `${updatePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\tconst request: SummaryRequest = (aiContext, options, requestContext) =>\n\t\tcompleteSimpleWithRetries(models, model, aiContext, options, retry, callbacks, requestContext);\n\treturn summarizeMessages(\n\t\tcurrentMessages,\n\t\tmodel,\n\t\tmaxTokens,\n\t\tthinkingLevel,\n\t\tinitialPrompt,\n\t\tupdatePrompt,\n\t\tpreviousSummary,\n\t\trequest,\n\t\tcontext,\n\t);\n}\n\nexport interface SummaryGenerationOptions {\n\tmodel: Model<Api>;\n\treserveTokens: number;\n\tcustomInstructions?: string;\n\tpreviousSummary?: string;\n\tthinkingLevel?: ThinkingLevel;\n}\n\n/** Generate one summary through a caller-owned one-request boundary. */\nexport function generateSummaryWithRequest(\n\tcurrentMessages: AgentMessage[],\n\toptions: SummaryGenerationOptions,\n\trequest: SummaryRequest,\n\tcontext: Context,\n): Promise<Result<{ text: string; usage: Usage }, CompactionError>> {\n\tconst { model, reserveTokens, customInstructions, previousSummary, thinkingLevel } = options;\n\tconst maxTokens = getSummaryMaxTokens(model, reserveTokens);\n\tlet initialPrompt = SUMMARIZATION_PROMPT;\n\tlet updatePrompt = UPDATE_SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tinitialPrompt = `${initialPrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t\tupdatePrompt = `${updatePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\treturn summarizeMessages(\n\t\tcurrentMessages,\n\t\tmodel,\n\t\tmaxTokens,\n\t\tthinkingLevel,\n\t\tinitialPrompt,\n\t\tupdatePrompt,\n\t\tpreviousSummary,\n\t\trequest,\n\t\tcontext,\n\t);\n}\n\n/** Prepared inputs for a compaction run. */\nexport interface CompactionPreparation {\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/** Recent messages retained after compaction and stored on the compaction entry. */\n\tretainedTail: 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: Entry[],\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 compactableEntries = pathEntries;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tconst virtualRetainedEntries: Entry[] = prevCompaction.retainedTail.map((message, index) => ({\n\t\t\ttype: \"message\",\n\t\t\tid: `${prevCompaction.id}:retained:${index}`,\n\t\t\tparentId: index === 0 ? prevCompaction.id : `${prevCompaction.id}:retained:${index - 1}`,\n\t\t\tseq: prevCompaction.seq,\n\t\t\ttimestamp: message.timestamp,\n\t\t\tmessage,\n\t\t}));\n\t\tcompactableEntries = [...virtualRetainedEntries, ...pathEntries.slice(prevCompactionIndex + 1)];\n\t}\n\tconst boundaryEnd = compactableEntries.length;\n\n\tconst tokensBefore = estimateContextTokens(\n\t\tbuildContextEntries(pathEntries).flatMap(sessionEntryToContextMessages),\n\t).tokens;\n\tconst cutPoint = findCutPoint(compactableEntries, 0, boundaryEnd, settings.keepRecentTokens);\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let i = 0; i < historyEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(compactableEntries[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(compactableEntries[i]);\n\t\t\tif (msg) turnPrefixMessages.push(msg);\n\t\t}\n\t}\n\tconst retainedTail: AgentMessage[] = [];\n\tfor (let i = cutPoint.firstKeptEntryIndex; i < boundaryEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(compactableEntries[i]);\n\t\tif (msg) retainedTail.push(msg);\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\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tretainedTail,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t});\n}\n\n/** Optional cut-point override for the pi-client/pi-server legacy session tree. */\nexport interface CompactionPreparationOptions {\n\tfirstKeptEntryId?: string;\n}\n\n/** Prepared compaction inputs with the retained entry id required by the pi-server protocol. */\nexport interface LegacyCompactionPreparation extends CompactionPreparation {\n\tfirstKeptEntryId: string;\n}\n\n/** Prepare a pi-client/pi-server session branch without changing its durable entry representation. */\nexport function prepareLegacyCompaction(\n\tpathEntries: SessionTreeEntry[],\n\tsettings: CompactionSettings,\n\toptions: CompactionPreparationOptions = {},\n): Result<LegacyCompactionPreparation | 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 index = pathEntries.length - 1; index >= 0; index--) {\n\t\tif (pathEntries[index].type === \"compaction\") {\n\t\t\tprevCompactionIndex = index;\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 LegacyCompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tconst firstKeptEntryIndex = prevCompaction.firstKeptEntryId\n\t\t\t? pathEntries.findIndex((entry) => entry.id === prevCompaction.firstKeptEntryId)\n\t\t\t: -1;\n\t\tboundaryStart = firstKeptEntryIndex >= 0 ? firstKeptEntryIndex : prevCompactionIndex + 1;\n\t}\n\tconst boundaryEnd = pathEntries.length;\n\tconst tokensBefore = estimateContextTokens(buildLegacySessionContext(pathEntries).messages).tokens;\n\n\tconst forcedFirstKeptEntryIndex = options.firstKeptEntryId\n\t\t? pathEntries.findIndex((entry) => entry.id === options.firstKeptEntryId)\n\t\t: -1;\n\tif (\n\t\toptions.firstKeptEntryId &&\n\t\t(forcedFirstKeptEntryIndex < boundaryStart || forcedFirstKeptEntryIndex >= boundaryEnd)\n\t) {\n\t\treturn err(new CompactionError(\"invalid_session\", \"Forced first kept entry was not found\"));\n\t}\n\tconst cutPoint =\n\t\tforcedFirstKeptEntryIndex >= boundaryStart && forcedFirstKeptEntryIndex < boundaryEnd\n\t\t\t? { firstKeptEntryIndex: forcedFirstKeptEntryIndex, turnStartIndex: -1, isSplitTurn: false }\n\t\t\t: findLegacyCutPoint(pathEntries, boundaryStart, boundaryEnd, settings.keepRecentTokens);\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\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let index = boundaryStart; index < historyEnd; index++) {\n\t\tconst message = getMessageFromLegacyEntryForCompaction(pathEntries[index]);\n\t\tif (message) messagesToSummarize.push(message);\n\t}\n\tconst turnPrefixMessages: AgentMessage[] = [];\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (let index = cutPoint.turnStartIndex; index < cutPoint.firstKeptEntryIndex; index++) {\n\t\t\tconst message = getMessageFromLegacyEntryForCompaction(pathEntries[index]);\n\t\t\tif (message) turnPrefixMessages.push(message);\n\t\t}\n\t}\n\tconst retainedTail: AgentMessage[] = [];\n\tfor (let index = cutPoint.firstKeptEntryIndex; index < boundaryEnd; index++) {\n\t\tconst message = getMessageFromLegacyEntryForCompaction(pathEntries[index]);\n\t\tif (message) retainedTail.push(message);\n\t}\n\tconst fileOps = extractLegacyFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (const message of turnPrefixMessages) extractFileOpsFromMessage(message, fileOps);\n\t}\n\n\treturn ok({\n\t\tfirstKeptEntryId: firstKeptEntry.id,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tretainedTail,\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\nconst UPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT = `The messages above are the next part of the same oversized turn prefix.\n\nUpdate the existing turn-prefix summary in <previous-summary> tags. Keep the same sections:\n\n## Original Request\n## Early Progress\n## Context for Suffix\n\nBe concise. Preserve exact file paths, function names, and error messages.`;\n\nexport { serializeConversation } from \"./utils.ts\";\n\n/** Generate compaction summary data from prepared session history. */\nexport function compact(\n\tpreparation: CompactionPreparation,\n\tmodels: Models,\n\tmodel: Model<Api>,\n\tcustomInstructions: string | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tretry: RetryPolicy | undefined,\n\tcallbacks: RetryCallbacks | undefined,\n\tcontext: Context,\n): Promise<Result<CompactResult, CompactionError>> {\n\treturn compactWithRequest(\n\t\tpreparation,\n\t\t{ model, customInstructions, thinkingLevel },\n\t\t(aiContext, options, requestContext) =>\n\t\t\tcompleteSimpleWithRetries(models, model, aiContext, options, retry, callbacks, requestContext),\n\t\tcontext,\n\t);\n}\n\nexport interface CompactGenerationOptions {\n\tmodel: Model<Api>;\n\tcustomInstructions?: string;\n\tthinkingLevel?: ThinkingLevel;\n}\n\n/** Generate compaction data through a caller-owned boundary for each provider request. */\nexport async function compactWithRequest(\n\tpreparation: CompactionPreparation,\n\toptions: CompactGenerationOptions,\n\trequest: SummaryRequest,\n\tcontext: Context,\n): Promise<Result<CompactResult, CompactionError>> {\n\tconst { model, customInstructions, thinkingLevel } = options;\n\tconst {\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tretainedTail,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\tlet summary: string;\n\tlet summaryUsage: Usage;\n\n\tif (isSplitTurn && turnPrefixMessages.length > 0) {\n\t\tlet historyText = \"No prior history.\";\n\t\tlet historyUsage: Usage | undefined;\n\t\tif (messagesToSummarize.length > 0) {\n\t\t\tconst historyResult = await generateSummaryWithRequest(\n\t\t\t\tmessagesToSummarize,\n\t\t\t\t{ model, reserveTokens: settings.reserveTokens, customInstructions, previousSummary, thinkingLevel },\n\t\t\t\trequest,\n\t\t\t\tcontext,\n\t\t\t);\n\t\t\tif (!historyResult.ok) return err(historyResult.error);\n\t\t\thistoryText = historyResult.value.text;\n\t\t\thistoryUsage = historyResult.value.usage;\n\t\t}\n\t\tconst turnPrefixResult = await generateTurnPrefixSummary(\n\t\t\tturnPrefixMessages,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tthinkingLevel,\n\t\t\trequest,\n\t\t\tcontext,\n\t\t);\n\t\tif (!turnPrefixResult.ok) return err(turnPrefixResult.error);\n\t\tsummary = `${historyText}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult.value.text}`;\n\t\tsummaryUsage = historyUsage ? addUsage(historyUsage, turnPrefixResult.value.usage) : turnPrefixResult.value.usage;\n\t} else {\n\t\tconst summaryResult = await generateSummaryWithRequest(\n\t\t\tmessagesToSummarize,\n\t\t\t{ model, reserveTokens: settings.reserveTokens, customInstructions, previousSummary, thinkingLevel },\n\t\t\trequest,\n\t\t\tcontext,\n\t\t);\n\t\tif (!summaryResult.ok) return err(summaryResult.error);\n\t\tsummary = summaryResult.value.text;\n\t\tsummaryUsage = summaryResult.value.usage;\n\t}\n\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\tconst details: CompactionDetails = { readFiles, modifiedFiles };\n\n\treturn ok({ summary, tokensBefore, usage: summaryUsage, retainedTail, details });\n}\n\n/** Compact a pi-client/pi-server session branch and preserve its retained entry id. */\nexport async function compactLegacy(\n\tpreparation: LegacyCompactionPreparation,\n\tmodels: Models,\n\tmodel: Model<Api>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<Result<LegacyCompactResult, CompactionError>> {\n\tconst context = signal ? withAbortSignal(signal, BACKGROUND_CONTEXT) : BACKGROUND_CONTEXT;\n\tconst request: SummaryRequest = (aiContext, options, requestContext) =>\n\t\tcompleteSimpleWithRetries(models, model, aiContext, options, retry, callbacks, requestContext);\n\tconst result = await compactWithRequest(preparation, { model, customInstructions, thinkingLevel }, request, context);\n\treturn result.ok ? ok({ ...result.value, firstKeptEntryId: preparation.firstKeptEntryId }) : err(result.error);\n}\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodel: Model<Api>,\n\treserveTokens: number,\n\tthinkingLevel: ThinkingLevel | undefined,\n\trequest: SummaryRequest,\n\tcontext: Context,\n): Promise<Result<{ text: string; usage: Usage }, 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 result = await summarizeMessages(\n\t\tmessages,\n\t\tmodel,\n\t\tmaxTokens,\n\t\tthinkingLevel,\n\t\tTURN_PREFIX_SUMMARIZATION_PROMPT,\n\t\tUPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT,\n\t\tundefined,\n\t\trequest,\n\t\tcontext,\n\t);\n\tif (!result.ok && result.error.code === \"summarization_failed\") {\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: ${result.error.message.replace(/^Summarization failed: /, \"\")}`,\n\t\t\t),\n\t\t);\n\t}\n\treturn result;\n}\n"]}

package/dist/harness/compaction/compaction.js

@@ -1,8 +1,10 @@
-import { isContextOverflow, retryAssistantCall, uuidv7, } from "@earendil-works/pi-ai";
+import { contentText, isContextOverflow, retryAssistantCall, uuidv7, } from "@earendil-works/pi-ai";
+import { BACKGROUND_CONTEXT, getTelemetryContext, withAbortSignal } from "../context.js";
 import { buildLegacySessionContext } from "../legacy-session.js";
 import { convertToLlm, createBranchSummaryMessage, createCompactionSummaryMessage, createCustomMessage, } from "../messages.js";
-import { buildSessionContext } from "../session/context.js";
+import { buildContextEntries, sessionEntryToContextMessages } from "../session/context.js";
 import { CompactionError, err, ok } from "../types.js";
+import { addUsage } from "../utils/usage.js";
 import { computeFileLists, createFileOps, extractFileOpsFromMessage, formatFileOperations, serializeConversation, } from "./utils.js";
 function safeJsonStringify(value) {
     try {
@@ -16,15 +18,20 @@ function extractFileOperations(messages, entries, prevCompactionIndex) {
     const fileOps = createFileOps();
     if (prevCompactionIndex >= 0) {
         const prevCompaction = entries[prevCompactionIndex];
-        if (prevCompaction.details) {
-            const details = prevCompaction.details;
-            if (Array.isArray(details.readFiles)) {
-                for (const f of details.readFiles)
-                    fileOps.read.add(f);
+        if (typeof prevCompaction.details === "object" &&
+            prevCompaction.details !== null &&
+            !Array.isArray(prevCompaction.details)) {
+            if (Array.isArray(prevCompaction.details.readFiles)) {
+                for (const path of prevCompaction.details.readFiles) {
+                    if (typeof path === "string")
+                        fileOps.read.add(path);
+                }
             }
-            if (Array.isArray(details.modifiedFiles)) {
-                for (const f of details.modifiedFiles)
-                    fileOps.edited.add(f);
+            if (Array.isArray(prevCompaction.details.modifiedFiles)) {
+                for (const path of prevCompaction.details.modifiedFiles) {
+                    if (typeof path === "string")
+                        fileOps.edited.add(path);
+                }
             }
         }
     }
@@ -88,36 +95,19 @@ function getMessageFromLegacyEntry(entry) {
 function getMessageFromLegacyEntryForCompaction(entry) {
     return entry.type === "compaction" ? undefined : getMessageFromLegacyEntry(entry);
 }
-export async function completeSimpleWithRetries(models, model, context, options, retry, callbacks) {
-    // Summaries are standalone requests, so isolate routing and avoid cache writes that cannot be reused.
-    const requestOptions = {
+export function createSummaryRequestOptions(options, context) {
+    return {
         ...options,
+        signal: context.abortSignal,
+        telemetryContext: getTelemetryContext(context),
         cacheRetention: "none",
-        sessionId: uuidv7(),
+        sessionId: options.sessionId ?? uuidv7(),
     };
-    return retryAssistantCall(() => models.completeSimple(model, context, requestOptions), retry, requestOptions.signal, callbacks);
 }
-function combineUsage(first, second) {
-    return {
-        input: first.input + second.input,
-        output: first.output + second.output,
-        cacheRead: first.cacheRead + second.cacheRead,
-        cacheWrite: first.cacheWrite + second.cacheWrite,
-        ...(first.cacheWrite1h !== undefined || second.cacheWrite1h !== undefined
-            ? { cacheWrite1h: (first.cacheWrite1h ?? 0) + (second.cacheWrite1h ?? 0) }
-            : {}),
-        ...(first.reasoning !== undefined || second.reasoning !== undefined
-            ? { reasoning: (first.reasoning ?? 0) + (second.reasoning ?? 0) }
-            : {}),
-        totalTokens: first.totalTokens + second.totalTokens,
-        cost: {
-            input: first.cost.input + second.cost.input,
-            output: first.cost.output + second.cost.output,
-            cacheRead: first.cost.cacheRead + second.cost.cacheRead,
-            cacheWrite: first.cost.cacheWrite + second.cost.cacheWrite,
-            total: first.cost.total + second.cost.total,
-        },
-    };
+export async function completeSimpleWithRetries(models, model, aiContext, options, retry, callbacks, context) {
+    // Summaries are standalone requests, so isolate routing and avoid cache writes that cannot be reused.
+    const requestOptions = createSummaryRequestOptions(options, context);
+    return retryAssistantCall(() => models.completeSimple(model, aiContext, requestOptions), retry, requestOptions.signal, callbacks);
 }
 /** Default compaction settings used by the harness. */
 export const DEFAULT_COMPACTION_SETTINGS = {
@@ -271,9 +261,6 @@ function findValidCutPoints(entries, startIndex, endIndex) {
                 }
                 break;
             }
-            case "thinking_level_change":
-            case "model_change":
-            case "active_tools_change":
             case "compaction":
             case "branch_summary":
             case "custom":
@@ -552,13 +539,7 @@ function takeSummaryChunk(segments, tokenBudget) {
     }
     return { chunk: chunk.join("\n\n"), remaining: [] };
 }
-function getResponseText(response) {
-    return response.content
-        .filter((c) => c.type === "text")
-        .map((c) => c.text)
-        .join("\n");
-}
-async function summarizeChunk(conversationText, models, model, maxTokens, signal, thinkingLevel, initialPrompt, updatePrompt, previousSummary, retry, callbacks) {
+async function summarizeChunk(conversationText, model, maxTokens, thinkingLevel, initialPrompt, updatePrompt, previousSummary, request, context) {
     const basePrompt = previousSummary ? updatePrompt : initialPrompt;
     const promptText = buildSummaryPrompt(conversationText, basePrompt, previousSummary);
     const summarizationMessages = [
@@ -569,20 +550,20 @@ async function summarizeChunk(conversationText, models, model, maxTokens, signal
         },
     ];
     const completionOptions = model.reasoning && thinkingLevel && thinkingLevel !== "off"
-        ? { maxTokens, signal, reasoning: thinkingLevel }
-        : { maxTokens, signal };
-    const response = await completeSimpleWithRetries(models, model, { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages }, completionOptions, retry, callbacks);
+        ? { maxTokens, reasoning: thinkingLevel }
+        : { maxTokens };
+    const response = await request({ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages }, createSummaryRequestOptions(completionOptions, context), context);
     if (isContextOverflow(response, model.contextWindow) && conversationText.length > 1) {
         const middle = Math.ceil(conversationText.length / 2);
-        const leftSummary = await summarizeChunk(conversationText.slice(0, middle), models, model, maxTokens, signal, thinkingLevel, initialPrompt, updatePrompt, previousSummary, retry, callbacks);
+        const leftSummary = await summarizeChunk(conversationText.slice(0, middle), model, maxTokens, thinkingLevel, initialPrompt, updatePrompt, previousSummary, request, context);
         if (!leftSummary.ok)
             return leftSummary;
-        const rightSummary = await summarizeChunk(conversationText.slice(middle), models, model, maxTokens, signal, thinkingLevel, initialPrompt, updatePrompt, leftSummary.value.text, retry, callbacks);
+        const rightSummary = await summarizeChunk(conversationText.slice(middle), model, maxTokens, thinkingLevel, initialPrompt, updatePrompt, leftSummary.value.text, request, context);
         if (!rightSummary.ok)
             return rightSummary;
         return ok({
             text: rightSummary.value.text,
-            usage: combineUsage(leftSummary.value.usage, rightSummary.value.usage),
+            usage: addUsage(leftSummary.value.usage, rightSummary.value.usage),
         });
     }
     if (response.stopReason === "aborted") {
@@ -591,23 +572,23 @@ async function summarizeChunk(conversationText, models, model, maxTokens, signal
     if (response.stopReason === "error") {
         return err(new CompactionError("summarization_failed", `Summarization failed: ${response.errorMessage || "Unknown error"}`));
     }
-    return ok({ text: getResponseText(response), usage: response.usage });
+    return ok({ text: contentText(response.content), usage: response.usage });
 }
-async function summarizeMessages(messages, models, model, maxTokens, signal, thinkingLevel, initialPrompt, updatePrompt, previousSummary, retry, callbacks) {
+async function summarizeMessages(messages, model, maxTokens, thinkingLevel, initialPrompt, updatePrompt, previousSummary, request, context) {
     let remaining = toSummarySegments(messages);
     if (remaining.length === 0) {
-        return summarizeChunk("", models, model, maxTokens, signal, thinkingLevel, initialPrompt, updatePrompt, previousSummary, retry, callbacks);
+        return summarizeChunk("", model, maxTokens, thinkingLevel, initialPrompt, updatePrompt, previousSummary, request, context);
     }
     let summary = previousSummary;
     let summaryUsage;
     while (remaining.length > 0) {
         const basePrompt = summary ? updatePrompt : initialPrompt;
         const next = takeSummaryChunk(remaining, getSummaryInputBudget(model, maxTokens, basePrompt, summary));
-        const chunkSummary = await summarizeChunk(next.chunk, models, model, maxTokens, signal, thinkingLevel, initialPrompt, updatePrompt, summary, retry, callbacks);
+        const chunkSummary = await summarizeChunk(next.chunk, model, maxTokens, thinkingLevel, initialPrompt, updatePrompt, summary, request, context);
         if (!chunkSummary.ok)
             return chunkSummary;
         summary = chunkSummary.value.text;
-        summaryUsage = summaryUsage ? combineUsage(summaryUsage, chunkSummary.value.usage) : chunkSummary.value.usage;
+        summaryUsage = summaryUsage ? addUsage(summaryUsage, chunkSummary.value.usage) : chunkSummary.value.usage;
         remaining = next.remaining;
     }
     if (!summaryUsage) {
@@ -616,12 +597,12 @@ async function summarizeMessages(messages, models, model, maxTokens, signal, thi
     return ok({ text: summary ?? "", usage: summaryUsage });
 }
 /** Generate or update a conversation summary for compaction. */
-export async function generateSummary(currentMessages, models, model, reserveTokens, signal, customInstructions, previousSummary, thinkingLevel, retry, callbacks) {
-    const result = await generateSummaryWithUsage(currentMessages, models, model, reserveTokens, signal, customInstructions, previousSummary, thinkingLevel, retry, callbacks);
+export async function generateSummary(currentMessages, models, model, reserveTokens, customInstructions, previousSummary, thinkingLevel, retry, callbacks, context = BACKGROUND_CONTEXT) {
+    const result = await generateSummaryWithUsage(currentMessages, models, model, reserveTokens, customInstructions, previousSummary, thinkingLevel, retry, callbacks, context);
     return result.ok ? ok(result.value.text) : err(result.error);
 }
 /** Generate or update a conversation summary and return its provider usage. */
-export async function generateSummaryWithUsage(currentMessages, models, model, reserveTokens, signal, customInstructions, previousSummary, thinkingLevel, retry, callbacks) {
+export function generateSummaryWithUsage(currentMessages, models, model, reserveTokens, customInstructions, previousSummary, thinkingLevel, retry, callbacks, context = BACKGROUND_CONTEXT) {
     const maxTokens = getSummaryMaxTokens(model, reserveTokens);
     let initialPrompt = SUMMARIZATION_PROMPT;
     let updatePrompt = UPDATE_SUMMARIZATION_PROMPT;
@@ -629,7 +610,20 @@ export async function generateSummaryWithUsage(currentMessages, models, model, r
         initialPrompt = `${initialPrompt}\n\nAdditional focus: ${customInstructions}`;
         updatePrompt = `${updatePrompt}\n\nAdditional focus: ${customInstructions}`;
     }
-    return summarizeMessages(currentMessages, models, model, maxTokens, signal, thinkingLevel, initialPrompt, updatePrompt, previousSummary, retry, callbacks);
+    const request = (aiContext, options, requestContext) => completeSimpleWithRetries(models, model, aiContext, options, retry, callbacks, requestContext);
+    return summarizeMessages(currentMessages, model, maxTokens, thinkingLevel, initialPrompt, updatePrompt, previousSummary, request, context);
+}
+/** Generate one summary through a caller-owned one-request boundary. */
+export function generateSummaryWithRequest(currentMessages, options, request, context) {
+    const { model, reserveTokens, customInstructions, previousSummary, thinkingLevel } = options;
+    const maxTokens = getSummaryMaxTokens(model, reserveTokens);
+    let initialPrompt = SUMMARIZATION_PROMPT;
+    let updatePrompt = UPDATE_SUMMARIZATION_PROMPT;
+    if (customInstructions) {
+        initialPrompt = `${initialPrompt}\n\nAdditional focus: ${customInstructions}`;
+        updatePrompt = `${updatePrompt}\n\nAdditional focus: ${customInstructions}`;
+    }
+    return summarizeMessages(currentMessages, model, maxTokens, thinkingLevel, initialPrompt, updatePrompt, previousSummary, request, context);
 }
 /** Prepare session entries for compaction, or return undefined when compaction is not applicable. */
 export function prepareCompaction(pathEntries, settings) {
@@ -659,7 +653,7 @@ export function prepareCompaction(pathEntries, settings) {
         compactableEntries = [...virtualRetainedEntries, ...pathEntries.slice(prevCompactionIndex + 1)];
     }
     const boundaryEnd = compactableEntries.length;
-    const tokensBefore = estimateContextTokens(buildSessionContext(pathEntries).messages).tokens;
+    const tokensBefore = estimateContextTokens(buildContextEntries(pathEntries).flatMap(sessionEntryToContextMessages)).tokens;
     const cutPoint = findCutPoint(compactableEntries, 0, boundaryEnd, settings.keepRecentTokens);
     const historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;
     const messagesToSummarize = [];
@@ -800,7 +794,12 @@ Update the existing turn-prefix summary in <previous-summary> tags. Keep the sam
 Be concise. Preserve exact file paths, function names, and error messages.`;
 export { serializeConversation } from "./utils.js";
 /** Generate compaction summary data from prepared session history. */
-export async function compact(preparation, models, model, customInstructions, signal, thinkingLevel, retry, callbacks) {
+export function compact(preparation, models, model, customInstructions, thinkingLevel, retry, callbacks, context) {
+    return compactWithRequest(preparation, { model, customInstructions, thinkingLevel }, (aiContext, options, requestContext) => completeSimpleWithRetries(models, model, aiContext, options, retry, callbacks, requestContext), context);
+}
+/** Generate compaction data through a caller-owned boundary for each provider request. */
+export async function compactWithRequest(preparation, options, request, context) {
+    const { model, customInstructions, thinkingLevel } = options;
     const { messagesToSummarize, turnPrefixMessages, retainedTail, isSplitTurn, tokensBefore, previousSummary, fileOps, settings, } = preparation;
     let summary;
     let summaryUsage;
@@ -808,22 +807,20 @@ export async function compact(preparation, models, model, customInstructions, si
         let historyText = "No prior history.";
         let historyUsage;
         if (messagesToSummarize.length > 0) {
-            const historyResult = await generateSummaryWithUsage(messagesToSummarize, models, model, settings.reserveTokens, signal, customInstructions, previousSummary, thinkingLevel, retry, callbacks);
+            const historyResult = await generateSummaryWithRequest(messagesToSummarize, { model, reserveTokens: settings.reserveTokens, customInstructions, previousSummary, thinkingLevel }, request, context);
             if (!historyResult.ok)
                 return err(historyResult.error);
             historyText = historyResult.value.text;
             historyUsage = historyResult.value.usage;
         }
-        const turnPrefixResult = await generateTurnPrefixSummary(turnPrefixMessages, models, model, settings.reserveTokens, signal, thinkingLevel, retry, callbacks);
+        const turnPrefixResult = await generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, thinkingLevel, request, context);
         if (!turnPrefixResult.ok)
             return err(turnPrefixResult.error);
         summary = `${historyText}\n\n---\n\n**Turn Context (split turn):**\n\n${turnPrefixResult.value.text}`;
-        summaryUsage = historyUsage
-            ? combineUsage(historyUsage, turnPrefixResult.value.usage)
-            : turnPrefixResult.value.usage;
+        summaryUsage = historyUsage ? addUsage(historyUsage, turnPrefixResult.value.usage) : turnPrefixResult.value.usage;
     }
     else {
-        const summaryResult = await generateSummaryWithUsage(messagesToSummarize, models, model, settings.reserveTokens, signal, customInstructions, previousSummary, thinkingLevel, retry, callbacks);
+        const summaryResult = await generateSummaryWithRequest(messagesToSummarize, { model, reserveTokens: settings.reserveTokens, customInstructions, previousSummary, thinkingLevel }, request, context);
         if (!summaryResult.ok)
             return err(summaryResult.error);
         summary = summaryResult.value.text;
@@ -831,22 +828,19 @@ export async function compact(preparation, models, model, customInstructions, si
     }
     const { readFiles, modifiedFiles } = computeFileLists(fileOps);
     summary += formatFileOperations(readFiles, modifiedFiles);
-    return ok({
-        summary,
-        tokensBefore,
-        usage: summaryUsage,
-        retainedTail,
-        details: { readFiles, modifiedFiles },
-    });
+    const details = { readFiles, modifiedFiles };
+    return ok({ summary, tokensBefore, usage: summaryUsage, retainedTail, details });
 }
 /** Compact a pi-client/pi-server session branch and preserve its retained entry id. */
 export async function compactLegacy(preparation, models, model, customInstructions, signal, thinkingLevel, retry, callbacks) {
-    const result = await compact(preparation, models, model, customInstructions, signal, thinkingLevel, retry, callbacks);
+    const context = signal ? withAbortSignal(signal, BACKGROUND_CONTEXT) : BACKGROUND_CONTEXT;
+    const request = (aiContext, options, requestContext) => completeSimpleWithRetries(models, model, aiContext, options, retry, callbacks, requestContext);
+    const result = await compactWithRequest(preparation, { model, customInstructions, thinkingLevel }, request, context);
     return result.ok ? ok({ ...result.value, firstKeptEntryId: preparation.firstKeptEntryId }) : err(result.error);
 }
-async function generateTurnPrefixSummary(messages, models, model, reserveTokens, signal, thinkingLevel, retry, callbacks) {
+async function generateTurnPrefixSummary(messages, model, reserveTokens, thinkingLevel, request, context) {
     const maxTokens = Math.min(Math.floor(0.5 * reserveTokens), model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY);
-    const result = await summarizeMessages(messages, models, model, maxTokens, signal, thinkingLevel, TURN_PREFIX_SUMMARIZATION_PROMPT, UPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT, undefined, retry, callbacks);
+    const result = await summarizeMessages(messages, model, maxTokens, thinkingLevel, TURN_PREFIX_SUMMARIZATION_PROMPT, UPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT, undefined, request, context);
     if (!result.ok && result.error.code === "summarization_failed") {
         return err(new CompactionError("summarization_failed", `Turn prefix summarization failed: ${result.error.message.replace(/^Summarization failed: /, "")}`));
     }