@averyyy/pi-coding-agent 0.84.2-piclient.3 → 0.84.4-piclient.1

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

@@ -1 +1 @@
-{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,+BAA+B,CAAC;AAC3F,OAAO,EAAe,KAAK,cAAc,EAAE,KAAK,WAAW,EAA8B,MAAM,uBAAuB,CAAC;AACvH,OAAO,KAAK,EAAE,gBAAgB,EAAE,OAAO,EAAE,KAAK,EAAE,mBAAmB,EAAE,KAAK,EAAE,MAAM,8BAA8B,CAAC;AAGjH,OAAO,EAGN,KAAK,YAAY,EAEjB,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAIN,KAAK,cAAc,EAInB,MAAM,YAAY,CAAC;AAMpB,kEAAkE;AAClE,MAAM,WAAW,iBAAiB;IACjC,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AAkDD,8EAA8E;AAC9E,MAAM,WAAW,gBAAgB,CAAC,CAAC,GAAG,OAAO;IAC5C,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,EAAE,MAAM,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;IACrB,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,2EAA2E;IAC3E,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,+FAA+F;IAC/F,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AA6BD,MAAM,WAAW,kBAAkB;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,aAAa,EAAE,MAAM,CAAC;IACtB,gBAAgB,EAAE,MAAM,CAAC;CACzB;AAED,eAAO,MAAM,2BAA2B,EAAE,kBAIzC,CAAC;AAMF;;;GAGG;AACH,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAqBD;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,KAAK,GAAG,SAAS,CAShF;AAED,MAAM,WAAW,oBAAoB;IACpC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,oBAAoB,CA4BpF;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAwBD;;;GAGG;AACH,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAwC5D;AA2DD;;;GAGG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAO1G;AAED,MAAM,WAAW,cAAc;IAC9B,mCAAmC;IACnC,mBAAmB,EAAE,MAAM,CAAC;IAC5B,qFAAqF;IACrF,cAAc,EAAE,MAAM,CAAC;IACvB,uEAAuE;IACvE,WAAW,EAAE,OAAO,CAAC;CACrB;AAED;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,YAAY,CAC3B,OAAO,EAAE,YAAY,EAAE,EACvB,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CAqDhB;AAwKD;;;;;;GAMG;AACH,wBAAsB,qBAAqB,CAC1C,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,EAC5B,QAAQ,CAAC,EAAE,QAAQ,EACnB,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CAY3B;AAoFD,wBAAsB,iBAAiB,CACtC,QAAQ,EAAE,YAAY,EAAE,EACxB,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,SAAS,EAAE,MAAM,EACjB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,SAAS,EAC3C,GAAG,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,SAAS,EACvC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,aAAa,EAAE,aAAa,GAAG,SAAS,EACxC,QAAQ,EAAE,QAAQ,GAAG,SAAS,EAC9B,aAAa,EAAE,MAAM,EACrB,YAAY,EAAE,MAAM,EACpB,eAAe,EAAE,MAAM,GAAG,SAAS,EACnC,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,KAAK,CAAA;CAAE,CAAC,CA+CzC;AAED;;;GAGG;AACH,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAC5B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,MAAM,CAAC,CAkBjB;AAED,+EAA+E;AAC/E,wBAAsB,wBAAwB,CAC7C,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAC5B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,KAAK,CAAA;CAAE,CAAC,CA0BzC;AAMD,MAAM,WAAW,qBAAqB;IACrC,kCAAkC;IAClC,gBAAgB,EAAE,MAAM,CAAC;IACzB,qDAAqD;IACrD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,iEAAiE;IACjE,WAAW,EAAE,OAAO,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,6DAA6D;IAC7D,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,yDAAyD;IACzD,OAAO,EAAE,cAAc,CAAC;IACxB,8CAA8C;IAC9C,QAAQ,EAAE,kBAAkB,CAAC;CAC7B;AAED,MAAM,WAAW,4BAA4B;IAC5C,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,wBAAgB,iBAAiB,CAChC,WAAW,EAAE,YAAY,EAAE,EAC3B,QAAQ,EAAE,kBAAkB,EAC5B,OAAO,GAAE,4BAAiC,GACxC,qBAAqB,GAAG,SAAS,CAwFnC;AA+BD;;;;;;GAMG;AACH,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAC5B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CA0F3B","sourcesContent":["/**\n * Context compaction for long sessions.\n *\n * Pure functions for compaction logic. The session manager handles I/O,\n * and after compaction the session is reloaded.\n */\n\nimport type { AgentMessage, StreamFn, ThinkingLevel } from \"@earendil-works/pi-agent-core\";\nimport { contentText, type RetryCallbacks, type RetryPolicy, retryAssistantCall, uuidv7 } from \"@earendil-works/pi-ai\";\nimport type { AssistantMessage, Context, Model, SimpleStreamOptions, Usage } from \"@earendil-works/pi-ai/compat\";\nimport { completeSimple, isContextOverflow } from \"@earendil-works/pi-ai/compat\";\nimport { convertToLlm } from \"../messages.ts\";\nimport {\n\tbuildSessionContext,\n\ttype CompactionEntry,\n\ttype SessionEntry,\n\tsessionEntryToContextMessages,\n} from \"../session-manager.ts\";\nimport {\n\tcomputeFileLists,\n\tcreateFileOps,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tSUMMARIZATION_SYSTEM_PROMPT,\n\tserializeConversation,\n} from \"./utils.ts\";\n\n// ============================================================================\n// File Operation Tracking\n// ============================================================================\n\n/** Details stored in CompactionEntry.details for file tracking */\nexport interface CompactionDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n}\n\n/**\n * Extract file operations from messages and previous compaction entries.\n */\nfunction extractFileOperations(\n\tmessages: AgentMessage[],\n\tentries: SessionEntry[],\n\tprevCompactionIndex: number,\n): FileOperations {\n\tconst fileOps = createFileOps();\n\n\t// Collect from previous compaction's details (if pi-generated)\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = entries[prevCompactionIndex] as CompactionEntry;\n\t\tif (!prevCompaction.fromHook && prevCompaction.details) {\n\t\t\t// fromHook field kept for session file compatibility\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\n\t// Extract from tool calls in messages\n\tfor (const msg of messages) {\n\t\textractFileOpsFromMessage(msg, fileOps);\n\t}\n\n\treturn fileOps;\n}\n\n// ============================================================================\n// Message Extraction\n// ============================================================================\n\n/**\n * Extract AgentMessage from an entry if it produces one.\n * Returns undefined for entries that don't contribute to LLM context.\n */\nfunction getMessageFromEntryForCompaction(entry: SessionEntry): AgentMessage | undefined {\n\tif (entry.type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\treturn sessionEntryToContextMessages(entry)[0];\n}\n\n/** Result from compact() - SessionManager adds uuid/parentUuid when saving */\nexport interface CompactionResult<T = unknown> {\n\tsummary: string;\n\tfirstKeptEntryId: string;\n\ttokensBefore: number;\n\testimatedTokensAfter?: number;\n\t/** Usage from the LLM call(s) that generated this summary, if available */\n\tusage?: Usage;\n\t/** Extension-specific data (e.g., ArtifactIndex, version markers for structured compaction) */\n\tdetails?: 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// ============================================================================\n// Types\n// ============================================================================\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\tkeepRecentTokens: number;\n}\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tkeepRecentTokens: 20000,\n};\n\n// ============================================================================\n// Token calculation\n// ============================================================================\n\n/**\n * Calculate total context tokens from usage.\n * Uses the native totalTokens field when available, falls back to computing from components.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted, error, and all-zero usage messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (\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/**\n * Find the last valid assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tif (!settings.enabled) return false;\n\treturn contextTokens > contextWindow - settings.reserveTokens;\n}\n\n// ============================================================================\n// Cut point detection\n// ============================================================================\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n\nfunction isCutPointMessage(message: AgentMessage): boolean {\n\tswitch (message.role) {\n\t\tcase \"user\":\n\t\tcase \"assistant\":\n\t\tcase \"bashExecution\":\n\t\tcase \"custom\":\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\":\n\t\t\treturn true;\n\t\tcase \"toolResult\":\n\t\t\treturn false;\n\t}\n\treturn false;\n}\n\nfunction isTurnStartMessage(message: AgentMessage): boolean {\n\tswitch (message.role) {\n\t\tcase \"user\":\n\t\tcase \"bashExecution\":\n\t\tcase \"custom\":\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\":\n\t\t\treturn true;\n\t\tcase \"assistant\":\n\t\tcase \"toolResult\":\n\t\t\treturn false;\n\t}\n\treturn false;\n}\n\nfunction isTurnStartEntry(entry: SessionEntry): boolean {\n\tif (entry.type === \"compaction\") {\n\t\treturn false;\n\t}\n\treturn sessionEntryToContextMessages(entry).some(isTurnStartMessage);\n}\n\n/**\n * Find valid cut points: indices of context-visible user-like or assistant messages.\n * Never cut at tool results (they must follow their tool call).\n * When we cut at an assistant message with tool calls, its tool results follow it\n * and will be kept.\n */\nfunction findValidCutPoints(entries: SessionEntry[], startIndex: number, endIndex: number): number[] {\n\tconst cutPoints: number[] = [];\n\tfor (let i = startIndex; i < endIndex; i++) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"compaction\") {\n\t\t\tcontinue;\n\t\t}\n\t\tif (sessionEntryToContextMessages(entry).some(isCutPointMessage)) {\n\t\t\tcutPoints.push(i);\n\t\t}\n\t}\n\treturn cutPoints;\n}\n\n/**\n * Find the context-visible user-role message that starts the turn containing the given entry index.\n * Returns -1 if no turn start found before the index.\n */\nexport function findTurnStartIndex(entries: SessionEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let i = entryIndex; i >= startIndex; i--) {\n\t\tif (isTurnStartEntry(entries[i])) {\n\t\t\treturn i;\n\t\t}\n\t}\n\treturn -1;\n}\n\nexport interface CutPointResult {\n\t/** Index of first entry to keep */\n\tfirstKeptEntryIndex: number;\n\t/** Index of user message that starts the turn being split, or -1 if not splitting */\n\tturnStartIndex: number;\n\t/** Whether this cut splits a turn (cut point is not a user message) */\n\tisSplitTurn: boolean;\n}\n\n/**\n * Find the cut point in session entries that keeps approximately `keepRecentTokens`.\n *\n * Algorithm: Walk backwards from newest, accumulating estimated message sizes.\n * Stop when we've accumulated >= keepRecentTokens. Cut at that point.\n *\n * Can cut at user OR assistant messages (never tool results). When cutting at an\n * assistant message with tool calls, its tool results come after and will be kept.\n *\n * Returns CutPointResult with:\n * - firstKeptEntryIndex: the entry index to start keeping from\n * - turnStartIndex: if cutting mid-turn, the user message that started that turn\n * - isSplitTurn: whether we're cutting in the middle of a turn\n *\n * Only considers entries between `startIndex` and `endIndex` (exclusive).\n */\nexport function findCutPoint(\n\tentries: SessionEntry[],\n\tstartIndex: number,\n\tendIndex: number,\n\tkeepRecentTokens: number,\n): CutPointResult {\n\tconst cutPoints = findValidCutPoints(entries, startIndex, endIndex);\n\n\tif (cutPoints.length === 0) {\n\t\treturn { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };\n\t}\n\n\t// Walk backwards from newest, accumulating estimated message sizes\n\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0]; // Default: keep from first message (not header)\n\n\tfor (let i = endIndex - 1; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\tconst messageTokens = sessionEntryToContextMessages(entry).reduce(\n\t\t\t(sum, message) => sum + estimateTokens(message),\n\t\t\t0,\n\t\t);\n\t\tif (messageTokens === 0) continue;\n\t\taccumulatedTokens += messageTokens;\n\n\t\t// Check if we've exceeded the budget\n\t\tif (accumulatedTokens >= keepRecentTokens) {\n\t\t\t// Find the closest valid cut point at or after this entry\n\t\t\tfor (let c = 0; c < cutPoints.length; c++) {\n\t\t\t\tif (cutPoints[c] >= i) {\n\t\t\t\t\tcutIndex = cutPoints[c];\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t}\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Scan backwards from cutIndex to include adjacent metadata entries that do not affect context.\n\twhile (cutIndex > startIndex) {\n\t\tconst prevEntry = entries[cutIndex - 1];\n\t\t// Stop at compaction boundaries or context-visible entries.\n\t\tif (prevEntry.type === \"compaction\" || sessionEntryToContextMessages(prevEntry).length > 0) {\n\t\t\tbreak;\n\t\t}\n\t\tcutIndex--;\n\t}\n\n\t// Determine if this is a split turn\n\tconst cutEntry = entries[cutIndex];\n\tconst startsTurn = isTurnStartEntry(cutEntry);\n\tconst turnStartIndex = startsTurn ? -1 : findTurnStartIndex(entries, cutIndex, startIndex);\n\n\treturn {\n\t\tfirstKeptEntryIndex: cutIndex,\n\t\tturnStartIndex,\n\t\tisSplitTurn: !startsTurn && turnStartIndex !== -1,\n\t};\n}\n\n// ============================================================================\n// Summarization\n// ============================================================================\n\nconst SUMMARIZATION_PROMPT = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.\n\nUse this EXACT format:\n\n## Goal\n[What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]\n\n## Constraints & Preferences\n- [Any constraints, preferences, or requirements mentioned by user]\n- [Or \"(none)\" if none were mentioned]\n\n## Progress\n### Done\n- [x] [Completed tasks/changes]\n\n### In Progress\n- [ ] [Current work]\n\n### Blocked\n- [Issues preventing progress, if any]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale]\n\n## 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 createSummarizationOptions(\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tenv: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n): SimpleStreamOptions {\n\tconst options: SimpleStreamOptions = { maxTokens, signal, apiKey, headers, env };\n\tif (model.reasoning && thinkingLevel && thinkingLevel !== \"off\") {\n\t\toptions.reasoning = thinkingLevel;\n\t}\n\treturn options;\n}\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\n/**\n * Shared choke point for every compaction/branch-summary summarization call. Wraps the\n * single LLM call in {@link retryAssistantCall} so transient stream drops (e.g.\n * `terminated`, socket close) honor the configured retry policy instead of failing\n * the whole compaction on the first attempt. Deterministic errors and aborts return\n * immediately (see {@link retryAssistantCall}).\n */\nexport async function completeSummarization(\n\tmodel: Model<any>,\n\tcontext: Context,\n\toptions: SimpleStreamOptions,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\t// Summaries are standalone requests, so isolate routing and avoid cache writes that cannot be reused.\n\tconst requestOptions: SimpleStreamOptions = {\n\t\t...options,\n\t\tcacheRetention: \"none\",\n\t\tsessionId: uuidv7(),\n\t};\n\tconst produce = async (): Promise<AssistantMessage> =>\n\t\tstreamFn\n\t\t\t? (await streamFn(model, context, requestOptions)).result()\n\t\t\t: completeSimple(model, context, requestOptions);\n\treturn retryAssistantCall(produce, retry, requestOptions.signal, callbacks);\n}\n\nasync function summarizeChunk(\n\tconversationText: string,\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tenv: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tstreamFn: StreamFn | undefined,\n\tinitialPrompt: string,\n\tupdatePrompt: string,\n\tpreviousSummary: string | undefined,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<{ text: string; usage: Usage }> {\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 response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcreateSummarizationOptions(model, maxTokens, apiKey, headers, env, signal, thinkingLevel),\n\t\tstreamFn,\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\tmodel,\n\t\t\tmaxTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\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\tconst rightSummary = await summarizeChunk(\n\t\t\tconversationText.slice(middle),\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tleftSummary.text,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t\treturn {\n\t\t\ttext: rightSummary.text,\n\t\t\tusage: combineUsage(leftSummary.usage, rightSummary.usage),\n\t\t};\n\t}\n\n\tif (response.stopReason === \"error\") {\n\t\tthrow new Error(`Summarization failed: ${response.errorMessage || \"Unknown error\"}`);\n\t}\n\n\treturn { text: contentText(response.content), usage: response.usage };\n}\n\nexport async function summarizeMessages(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tenv: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tstreamFn: StreamFn | undefined,\n\tinitialPrompt: string,\n\tupdatePrompt: string,\n\tpreviousSummary: string | undefined,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<{ text: string; usage: Usage }> {\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\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\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\tmodel,\n\t\t\tmaxTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\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\tsummary = chunkSummary.text;\n\t\tsummaryUsage = summaryUsage ? combineUsage(summaryUsage, chunkSummary.usage) : chunkSummary.usage;\n\t\tremaining = next.remaining;\n\t}\n\tif (!summaryUsage) throw new Error(\"Summarization returned no usage data\");\n\treturn { text: summary ?? \"\", usage: summaryUsage };\n}\n\n/**\n * Generate a summary of the conversation using the LLM.\n * If previousSummary is provided, uses the update prompt to merge.\n */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tenv?: Record<string, string>,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<string> {\n\treturn (\n\t\tawait generateSummaryWithUsage(\n\t\t\tcurrentMessages,\n\t\t\tmodel,\n\t\t\treserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tenv,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t)\n\t).text;\n}\n\n/** Generate or update a conversation summary and return its provider usage. */\nexport async function generateSummaryWithUsage(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tenv?: Record<string, string>,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<{ text: string; usage: Usage }> {\n\tconst maxTokens = getSummaryMaxTokens(model, reserveTokens);\n\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\n\treturn summarizeMessages(\n\t\tcurrentMessages,\n\t\tmodel,\n\t\tmaxTokens,\n\t\tapiKey,\n\t\theaders,\n\t\tenv,\n\t\tsignal,\n\t\tthinkingLevel,\n\t\tstreamFn,\n\t\tinitialPrompt,\n\t\tupdatePrompt,\n\t\tpreviousSummary,\n\t\tretry,\n\t\tcallbacks,\n\t);\n}\n\n// ============================================================================\n// Compaction Preparation (for extensions)\n// ============================================================================\n\nexport interface CompactionPreparation {\n\t/** UUID of first entry to keep */\n\tfirstKeptEntryId: string;\n\t/** Messages that will be summarized and discarded */\n\tmessagesToSummarize: AgentMessage[];\n\t/** Messages that will be turned into turn prefix summary (if splitting) */\n\tturnPrefixMessages: AgentMessage[];\n\t/** Whether this is a split turn (cut point in middle of turn) */\n\tisSplitTurn: boolean;\n\ttokensBefore: number;\n\t/** Summary from previous compaction, for iterative update */\n\tpreviousSummary?: string;\n\t/** File operations extracted from messagesToSummarize */\n\tfileOps: FileOperations;\n\t/** Compaction settions from settings.jsonl\t*/\n\tsettings: CompactionSettings;\n}\n\nexport interface CompactionPreparationOptions {\n\tfirstKeptEntryId?: string;\n}\n\nexport function prepareCompaction(\n\tpathEntries: SessionEntry[],\n\tsettings: CompactionSettings,\n\toptions: CompactionPreparationOptions = {},\n): CompactionPreparation | undefined {\n\tif (pathEntries.length > 0 && pathEntries[pathEntries.length - 1].type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\n\tlet prevCompactionIndex = -1;\n\tfor (let i = pathEntries.length - 1; i >= 0; i--) {\n\t\tif (pathEntries[i].type === \"compaction\") {\n\t\t\tprevCompactionIndex = i;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\tlet previousSummary: string | undefined;\n\tlet boundaryStart = 0;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tconst firstKeptEntryIndex = pathEntries.findIndex((entry) => entry.id === prevCompaction.firstKeptEntryId);\n\t\tboundaryStart = firstKeptEntryIndex >= 0 ? firstKeptEntryIndex : prevCompactionIndex + 1;\n\t}\n\tconst boundaryEnd = pathEntries.length;\n\n\tconst tokensBefore = estimateContextTokens(buildSessionContext(pathEntries).messages).tokens;\n\n\tconst 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 undefined;\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: findCutPoint(pathEntries, boundaryStart, boundaryEnd, settings.keepRecentTokens);\n\n\t// Get UUID of first kept entry\n\tconst firstKeptEntry = pathEntries[cutPoint.firstKeptEntryIndex];\n\tif (!firstKeptEntry?.id) {\n\t\treturn undefined; // Session needs migration\n\t}\n\tconst firstKeptEntryId = firstKeptEntry.id;\n\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\n\t// Messages to summarize (will be discarded after summary)\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let i = boundaryStart; i < historyEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\tif (msg) messagesToSummarize.push(msg);\n\t}\n\n\t// Messages for turn prefix summary (if splitting a turn)\n\tconst turnPrefixMessages: AgentMessage[] = [];\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (let i = cutPoint.turnStartIndex; i < cutPoint.firstKeptEntryIndex; i++) {\n\t\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\t\tif (msg) turnPrefixMessages.push(msg);\n\t\t}\n\t}\n\n\tif (messagesToSummarize.length === 0 && turnPrefixMessages.length === 0) {\n\t\treturn undefined;\n\t}\n\n\t// Extract file operations from messages and previous compaction\n\tconst fileOps = extractFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\n\n\t// Also extract file ops from turn prefix if splitting\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (const msg of turnPrefixMessages) {\n\t\t\textractFileOpsFromMessage(msg, fileOps);\n\t\t}\n\t}\n\n\treturn {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t};\n}\n\n// ============================================================================\n// Main compaction function\n// ============================================================================\n\nconst TURN_PREFIX_SUMMARIZATION_PROMPT = `This is the PREFIX of a turn that was too large to keep. The SUFFIX (recent work) is retained.\n\nSummarize the prefix to provide context for the retained suffix:\n\n## Original Request\n[What did the user ask for in this turn?]\n\n## Early Progress\n- [Key decisions and work done in the prefix]\n\n## Context for Suffix\n- [Information needed to understand the retained recent work]\n\nBe concise. Focus on what's needed to understand the kept suffix.`;\n\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\n/**\n * Generate summaries for compaction using prepared data.\n * Returns CompactionResult - SessionManager adds uuid/parentUuid when saving.\n *\n * @param preparation - Pre-calculated preparation from prepareCompaction()\n * @param customInstructions - Optional custom focus for the summary\n */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tenv?: Record<string, string>,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<CompactionResult> {\n\tconst {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\t// Generate summaries and merge into one\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\tmodel,\n\t\t\t\tsettings.reserveTokens,\n\t\t\t\tapiKey,\n\t\t\t\theaders,\n\t\t\t\tsignal,\n\t\t\t\tcustomInstructions,\n\t\t\t\tpreviousSummary,\n\t\t\t\tthinkingLevel,\n\t\t\t\tstreamFn,\n\t\t\t\tenv,\n\t\t\t\tretry,\n\t\t\t\tcallbacks,\n\t\t\t);\n\t\t\thistoryText = historyResult.text;\n\t\t\thistoryUsage = historyResult.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\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t\t// Merge into single summary\n\t\tsummary = `${historyText}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult.text}`;\n\t\tsummaryUsage = historyUsage ? combineUsage(historyUsage, turnPrefixResult.usage) : turnPrefixResult.usage;\n\t} else {\n\t\t// Just generate history summary\n\t\tconst result = await generateSummaryWithUsage(\n\t\t\tmessagesToSummarize,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tenv,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t\tsummary = result.text;\n\t\tsummaryUsage = result.usage;\n\t}\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\tif (!firstKeptEntryId) {\n\t\tthrow new Error(\"First kept entry has no UUID - session may need migration\");\n\t}\n\n\treturn {\n\t\tsummary,\n\t\tfirstKeptEntryId,\n\t\ttokensBefore,\n\t\tusage: summaryUsage,\n\t\tdetails: { readFiles, modifiedFiles } as CompactionDetails,\n\t};\n}\n\n/**\n * Generate a summary for a turn prefix (when splitting a turn).\n */\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tenv?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<{ text: string; usage: Usage }> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.5 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t); // Smaller budget for turn prefix\n\n\ttry {\n\t\treturn await summarizeMessages(\n\t\t\tmessages,\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tTURN_PREFIX_SUMMARIZATION_PROMPT,\n\t\t\tUPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT,\n\t\t\tundefined,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t);\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tthrow new Error(`Turn prefix summarization failed: ${message.replace(/^Summarization failed: /, \"\")}`);\n\t}\n}\n"]}
+{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,+BAA+B,CAAC;AAC3F,OAAO,EAAe,KAAK,cAAc,EAAE,KAAK,WAAW,EAA8B,MAAM,uBAAuB,CAAC;AACvH,OAAO,KAAK,EAAE,gBAAgB,EAAE,OAAO,EAAE,KAAK,EAAE,mBAAmB,EAAE,KAAK,EAAE,MAAM,8BAA8B,CAAC;AAGjH,OAAO,EAGN,KAAK,YAAY,EAEjB,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAIN,KAAK,cAAc,EAInB,MAAM,YAAY,CAAC;AAMpB,kEAAkE;AAClE,MAAM,WAAW,iBAAiB;IACjC,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AAkDD,8EAA8E;AAC9E,MAAM,WAAW,gBAAgB,CAAC,CAAC,GAAG,OAAO;IAC5C,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,EAAE,MAAM,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;IACrB,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,2EAA2E;IAC3E,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,+FAA+F;IAC/F,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AA6BD,MAAM,WAAW,kBAAkB;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,aAAa,EAAE,MAAM,CAAC;IACtB,gBAAgB,EAAE,MAAM,CAAC;CACzB;AAED,eAAO,MAAM,2BAA2B,EAAE,kBAIzC,CAAC;AAMF;;;GAGG;AACH,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAqBD;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,KAAK,GAAG,SAAS,CAShF;AAED,MAAM,WAAW,oBAAoB;IACpC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,oBAAoB,CA4BpF;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAwBD;;;GAGG;AACH,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAwC5D;AA2DD;;;GAGG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAO1G;AAED,MAAM,WAAW,cAAc;IAC9B,mCAAmC;IACnC,mBAAmB,EAAE,MAAM,CAAC;IAC5B,qFAAqF;IACrF,cAAc,EAAE,MAAM,CAAC;IACvB,uEAAuE;IACvE,WAAW,EAAE,OAAO,CAAC;CACrB;AAED;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,YAAY,CAC3B,OAAO,EAAE,YAAY,EAAE,EACvB,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CAqDhB;AA8FD;;;GAGG;AACH,wBAAgB,uBAAuB,CAAC,QAAQ,EAAE,gBAAgB,EAAE,KAAK,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAQrG;AA6ED;;;;;;GAMG;AACH,wBAAsB,qBAAqB,CAC1C,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,OAAO,EAAE,OAAO,EAChB,OAAO,EAAE,mBAAmB,EAC5B,QAAQ,CAAC,EAAE,QAAQ,EACnB,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,GACxB,OAAO,CAAC,gBAAgB,CAAC,CAY3B;AA2FD,wBAAsB,iBAAiB,CACtC,QAAQ,EAAE,YAAY,EAAE,EACxB,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,SAAS,EAAE,MAAM,EACjB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,SAAS,EAC3C,GAAG,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,SAAS,EACvC,MAAM,EAAE,WAAW,GAAG,SAAS,EAC/B,aAAa,EAAE,aAAa,GAAG,SAAS,EACxC,QAAQ,EAAE,QAAQ,GAAG,SAAS,EAC9B,aAAa,EAAE,MAAM,EACrB,YAAY,EAAE,MAAM,EACpB,eAAe,EAAE,MAAM,GAAG,SAAS,EACnC,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,EAC1B,SAAS,CAAC,EAAE,MAAM,GAChB,OAAO,CAAC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,KAAK,CAAA;CAAE,CAAC,CAiDzC;AAED;;;GAGG;AACH,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAC5B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,EAC1B,SAAS,CAAC,EAAE,MAAM,GAChB,OAAO,CAAC,MAAM,CAAC,CAmBjB;AAED,+EAA+E;AAC/E,wBAAsB,wBAAwB,CAC7C,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAC5B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,EAC1B,SAAS,CAAC,EAAE,MAAM,GAChB,OAAO,CAAC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,KAAK,CAAA;CAAE,CAAC,CA2BzC;AAMD,MAAM,WAAW,qBAAqB;IACrC,kCAAkC;IAClC,gBAAgB,EAAE,MAAM,CAAC;IACzB,qDAAqD;IACrD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,iEAAiE;IACjE,WAAW,EAAE,OAAO,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,6DAA6D;IAC7D,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,yDAAyD;IACzD,OAAO,EAAE,cAAc,CAAC;IACxB,8CAA8C;IAC9C,QAAQ,EAAE,kBAAkB,CAAC;CAC7B;AAED,MAAM,WAAW,4BAA4B;IAC5C,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,wBAAgB,iBAAiB,CAChC,WAAW,EAAE,YAAY,EAAE,EAC3B,QAAQ,EAAE,kBAAkB,EAC5B,OAAO,GAAE,4BAAiC,GACxC,qBAAqB,GAAG,SAAS,CAwFnC;AA+BD;;;;;;GAMG;AACH,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,EACnB,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAC5B,KAAK,CAAC,EAAE,WAAW,EACnB,SAAS,CAAC,EAAE,cAAc,EAC1B,SAAS,CAAC,EAAE,MAAM,GAChB,OAAO,CAAC,gBAAgB,CAAC,CA6F3B","sourcesContent":["/**\n * Context compaction for long sessions.\n *\n * Pure functions for compaction logic. The session manager handles I/O,\n * and after compaction the session is reloaded.\n */\n\nimport type { AgentMessage, StreamFn, ThinkingLevel } from \"@earendil-works/pi-agent-core\";\nimport { contentText, type RetryCallbacks, type RetryPolicy, retryAssistantCall, uuidv7 } from \"@earendil-works/pi-ai\";\nimport type { AssistantMessage, Context, Model, SimpleStreamOptions, Usage } from \"@earendil-works/pi-ai/compat\";\nimport { completeSimple, isContextOverflow } from \"@earendil-works/pi-ai/compat\";\nimport { convertToLlm } from \"../messages.ts\";\nimport {\n\tbuildSessionContext,\n\ttype CompactionEntry,\n\ttype SessionEntry,\n\tsessionEntryToContextMessages,\n} from \"../session-manager.ts\";\nimport {\n\tcomputeFileLists,\n\tcreateFileOps,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tSUMMARIZATION_SYSTEM_PROMPT,\n\tserializeConversation,\n} from \"./utils.ts\";\n\n// ============================================================================\n// File Operation Tracking\n// ============================================================================\n\n/** Details stored in CompactionEntry.details for file tracking */\nexport interface CompactionDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n}\n\n/**\n * Extract file operations from messages and previous compaction entries.\n */\nfunction extractFileOperations(\n\tmessages: AgentMessage[],\n\tentries: SessionEntry[],\n\tprevCompactionIndex: number,\n): FileOperations {\n\tconst fileOps = createFileOps();\n\n\t// Collect from previous compaction's details (if pi-generated)\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = entries[prevCompactionIndex] as CompactionEntry;\n\t\tif (!prevCompaction.fromHook && prevCompaction.details) {\n\t\t\t// fromHook field kept for session file compatibility\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\n\t// Extract from tool calls in messages\n\tfor (const msg of messages) {\n\t\textractFileOpsFromMessage(msg, fileOps);\n\t}\n\n\treturn fileOps;\n}\n\n// ============================================================================\n// Message Extraction\n// ============================================================================\n\n/**\n * Extract AgentMessage from an entry if it produces one.\n * Returns undefined for entries that don't contribute to LLM context.\n */\nfunction getMessageFromEntryForCompaction(entry: SessionEntry): AgentMessage | undefined {\n\tif (entry.type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\treturn sessionEntryToContextMessages(entry)[0];\n}\n\n/** Result from compact() - SessionManager adds uuid/parentUuid when saving */\nexport interface CompactionResult<T = unknown> {\n\tsummary: string;\n\tfirstKeptEntryId: string;\n\ttokensBefore: number;\n\testimatedTokensAfter?: number;\n\t/** Usage from the LLM call(s) that generated this summary, if available */\n\tusage?: Usage;\n\t/** Extension-specific data (e.g., ArtifactIndex, version markers for structured compaction) */\n\tdetails?: 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// ============================================================================\n// Types\n// ============================================================================\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\tkeepRecentTokens: number;\n}\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tkeepRecentTokens: 20000,\n};\n\n// ============================================================================\n// Token calculation\n// ============================================================================\n\n/**\n * Calculate total context tokens from usage.\n * Uses the native totalTokens field when available, falls back to computing from components.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted, error, and all-zero usage messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (\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/**\n * Find the last valid assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tif (!settings.enabled) return false;\n\treturn contextTokens > contextWindow - settings.reserveTokens;\n}\n\n// ============================================================================\n// Cut point detection\n// ============================================================================\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n\nfunction isCutPointMessage(message: AgentMessage): boolean {\n\tswitch (message.role) {\n\t\tcase \"user\":\n\t\tcase \"assistant\":\n\t\tcase \"bashExecution\":\n\t\tcase \"custom\":\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\":\n\t\t\treturn true;\n\t\tcase \"toolResult\":\n\t\t\treturn false;\n\t}\n\treturn false;\n}\n\nfunction isTurnStartMessage(message: AgentMessage): boolean {\n\tswitch (message.role) {\n\t\tcase \"user\":\n\t\tcase \"bashExecution\":\n\t\tcase \"custom\":\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\":\n\t\t\treturn true;\n\t\tcase \"assistant\":\n\t\tcase \"toolResult\":\n\t\t\treturn false;\n\t}\n\treturn false;\n}\n\nfunction isTurnStartEntry(entry: SessionEntry): boolean {\n\tif (entry.type === \"compaction\") {\n\t\treturn false;\n\t}\n\treturn sessionEntryToContextMessages(entry).some(isTurnStartMessage);\n}\n\n/**\n * Find valid cut points: indices of context-visible user-like or assistant messages.\n * Never cut at tool results (they must follow their tool call).\n * When we cut at an assistant message with tool calls, its tool results follow it\n * and will be kept.\n */\nfunction findValidCutPoints(entries: SessionEntry[], startIndex: number, endIndex: number): number[] {\n\tconst cutPoints: number[] = [];\n\tfor (let i = startIndex; i < endIndex; i++) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"compaction\") {\n\t\t\tcontinue;\n\t\t}\n\t\tif (sessionEntryToContextMessages(entry).some(isCutPointMessage)) {\n\t\t\tcutPoints.push(i);\n\t\t}\n\t}\n\treturn cutPoints;\n}\n\n/**\n * Find the context-visible user-role message that starts the turn containing the given entry index.\n * Returns -1 if no turn start found before the index.\n */\nexport function findTurnStartIndex(entries: SessionEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let i = entryIndex; i >= startIndex; i--) {\n\t\tif (isTurnStartEntry(entries[i])) {\n\t\t\treturn i;\n\t\t}\n\t}\n\treturn -1;\n}\n\nexport interface CutPointResult {\n\t/** Index of first entry to keep */\n\tfirstKeptEntryIndex: number;\n\t/** Index of user message that starts the turn being split, or -1 if not splitting */\n\tturnStartIndex: number;\n\t/** Whether this cut splits a turn (cut point is not a user message) */\n\tisSplitTurn: boolean;\n}\n\n/**\n * Find the cut point in session entries that keeps approximately `keepRecentTokens`.\n *\n * Algorithm: Walk backwards from newest, accumulating estimated message sizes.\n * Stop when we've accumulated >= keepRecentTokens. Cut at that point.\n *\n * Can cut at user OR assistant messages (never tool results). When cutting at an\n * assistant message with tool calls, its tool results come after and will be kept.\n *\n * Returns CutPointResult with:\n * - firstKeptEntryIndex: the entry index to start keeping from\n * - turnStartIndex: if cutting mid-turn, the user message that started that turn\n * - isSplitTurn: whether we're cutting in the middle of a turn\n *\n * Only considers entries between `startIndex` and `endIndex` (exclusive).\n */\nexport function findCutPoint(\n\tentries: SessionEntry[],\n\tstartIndex: number,\n\tendIndex: number,\n\tkeepRecentTokens: number,\n): CutPointResult {\n\tconst cutPoints = findValidCutPoints(entries, startIndex, endIndex);\n\n\tif (cutPoints.length === 0) {\n\t\treturn { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };\n\t}\n\n\t// Walk backwards from newest, accumulating estimated message sizes\n\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0]; // Default: keep from first message (not header)\n\n\tfor (let i = endIndex - 1; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\tconst messageTokens = sessionEntryToContextMessages(entry).reduce(\n\t\t\t(sum, message) => sum + estimateTokens(message),\n\t\t\t0,\n\t\t);\n\t\tif (messageTokens === 0) continue;\n\t\taccumulatedTokens += messageTokens;\n\n\t\t// Check if we've exceeded the budget\n\t\tif (accumulatedTokens >= keepRecentTokens) {\n\t\t\t// Find the closest valid cut point at or after this entry\n\t\t\tfor (let c = 0; c < cutPoints.length; c++) {\n\t\t\t\tif (cutPoints[c] >= i) {\n\t\t\t\t\tcutIndex = cutPoints[c];\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t}\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Scan backwards from cutIndex to include adjacent metadata entries that do not affect context.\n\twhile (cutIndex > startIndex) {\n\t\tconst prevEntry = entries[cutIndex - 1];\n\t\t// Stop at compaction boundaries or context-visible entries.\n\t\tif (prevEntry.type === \"compaction\" || sessionEntryToContextMessages(prevEntry).length > 0) {\n\t\t\tbreak;\n\t\t}\n\t\tcutIndex--;\n\t}\n\n\t// Determine if this is a split turn\n\tconst cutEntry = entries[cutIndex];\n\tconst startsTurn = isTurnStartEntry(cutEntry);\n\tconst turnStartIndex = startsTurn ? -1 : findTurnStartIndex(entries, cutIndex, startIndex);\n\n\treturn {\n\t\tfirstKeptEntryIndex: cutIndex,\n\t\tturnStartIndex,\n\t\tisSplitTurn: !startsTurn && turnStartIndex !== -1,\n\t};\n}\n\n// ============================================================================\n// Summarization\n// ============================================================================\n\nconst SUMMARIZATION_PROMPT = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.\n\nUse this EXACT format:\n\n## Goal\n[What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]\n\n## Constraints & Preferences\n- [Any constraints, preferences, or requirements mentioned by user]\n- [Or \"(none)\" if none were mentioned]\n\n## Progress\n### Done\n- [x] [Completed tasks/changes]\n\n### In Progress\n- [ ] [Current work]\n\n### Blocked\n- [Issues preventing progress, if any]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale]\n\n## 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\n/**\n * Returns an error message when a summarization response cannot safely be persisted.\n * A length stop contains partial text and must not become a session checkpoint.\n */\nexport function getSummarizationFailure(response: AssistantMessage, label: string): string | undefined {\n\tif (response.stopReason === \"error\") {\n\t\treturn `${label} failed: ${response.errorMessage || \"Unknown error\"}`;\n\t}\n\tif (response.stopReason === \"length\") {\n\t\treturn `${label} failed: generation hit the token cap and the summary is incomplete`;\n\t}\n\treturn undefined;\n}\n\nfunction createSummarizationOptions(\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tenv: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tsessionId: string | undefined,\n): SimpleStreamOptions {\n\tconst options: SimpleStreamOptions = { maxTokens, signal, apiKey, headers, env, sessionId };\n\tif (model.reasoning && thinkingLevel && thinkingLevel !== \"off\") {\n\t\toptions.reasoning = thinkingLevel;\n\t}\n\treturn options;\n}\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\n/**\n * Shared choke point for every compaction/branch-summary summarization call. Wraps the\n * single LLM call in {@link retryAssistantCall} so transient stream drops (e.g.\n * `terminated`, socket close) honor the configured retry policy instead of failing\n * the whole compaction on the first attempt. Deterministic errors and aborts return\n * immediately (see {@link retryAssistantCall}).\n */\nexport async function completeSummarization(\n\tmodel: Model<any>,\n\tcontext: Context,\n\toptions: SimpleStreamOptions,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n): Promise<AssistantMessage> {\n\t// Summaries are standalone requests, so isolate routing and avoid cache writes that cannot be reused.\n\tconst requestOptions: SimpleStreamOptions = {\n\t\t...options,\n\t\tcacheRetention: \"none\",\n\t\tsessionId: options.sessionId ?? uuidv7(),\n\t};\n\tconst produce = async (): Promise<AssistantMessage> =>\n\t\tstreamFn\n\t\t\t? (await streamFn(model, context, requestOptions)).result()\n\t\t\t: completeSimple(model, context, requestOptions);\n\treturn retryAssistantCall(produce, retry, requestOptions.signal, callbacks);\n}\n\nasync function summarizeChunk(\n\tconversationText: string,\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tenv: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tstreamFn: StreamFn | undefined,\n\tinitialPrompt: string,\n\tupdatePrompt: string,\n\tpreviousSummary: string | undefined,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tsessionId?: string,\n): Promise<{ text: string; usage: Usage }> {\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 response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcreateSummarizationOptions(model, maxTokens, apiKey, headers, env, signal, thinkingLevel, sessionId),\n\t\tstreamFn,\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\tmodel,\n\t\t\tmaxTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tpreviousSummary,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t\tsessionId,\n\t\t);\n\t\tconst rightSummary = await summarizeChunk(\n\t\t\tconversationText.slice(middle),\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tleftSummary.text,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t\tsessionId,\n\t\t);\n\t\treturn {\n\t\t\ttext: rightSummary.text,\n\t\t\tusage: combineUsage(leftSummary.usage, rightSummary.usage),\n\t\t};\n\t}\n\n\tconst failure = getSummarizationFailure(response, \"Summarization\");\n\tif (failure) {\n\t\tthrow new Error(failure);\n\t}\n\tif (response.content.some((block) => block.type === \"toolCall\")) {\n\t\tthrow new Error(\"Summarization attempted to call a tool\");\n\t}\n\n\treturn { text: contentText(response.content), usage: response.usage };\n}\n\nexport async function summarizeMessages(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tenv: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n\tstreamFn: StreamFn | undefined,\n\tinitialPrompt: string,\n\tupdatePrompt: string,\n\tpreviousSummary: string | undefined,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tsessionId?: string,\n): Promise<{ text: string; usage: Usage }> {\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\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tpreviousSummary,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t\tsessionId,\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\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tinitialPrompt,\n\t\t\tupdatePrompt,\n\t\t\tsummary,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t\tsessionId,\n\t\t);\n\t\tsummary = chunkSummary.text;\n\t\tsummaryUsage = summaryUsage ? combineUsage(summaryUsage, chunkSummary.usage) : chunkSummary.usage;\n\t\tremaining = next.remaining;\n\t}\n\tif (!summaryUsage) throw new Error(\"Summarization returned no usage data\");\n\treturn { text: summary ?? \"\", usage: summaryUsage };\n}\n\n/**\n * Generate a summary of the conversation using the LLM.\n * If previousSummary is provided, uses the update prompt to merge.\n */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tenv?: Record<string, string>,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tsessionId?: string,\n): Promise<string> {\n\treturn (\n\t\tawait generateSummaryWithUsage(\n\t\t\tcurrentMessages,\n\t\t\tmodel,\n\t\t\treserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tenv,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t\tsessionId,\n\t\t)\n\t).text;\n}\n\n/** Generate or update a conversation summary and return its provider usage. */\nexport async function generateSummaryWithUsage(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tenv?: Record<string, string>,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tsessionId?: string,\n): Promise<{ text: string; usage: Usage }> {\n\tconst maxTokens = getSummaryMaxTokens(model, reserveTokens);\n\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\n\treturn summarizeMessages(\n\t\tcurrentMessages,\n\t\tmodel,\n\t\tmaxTokens,\n\t\tapiKey,\n\t\theaders,\n\t\tenv,\n\t\tsignal,\n\t\tthinkingLevel,\n\t\tstreamFn,\n\t\tinitialPrompt,\n\t\tupdatePrompt,\n\t\tpreviousSummary,\n\t\tretry,\n\t\tcallbacks,\n\t\tsessionId,\n\t);\n}\n\n// ============================================================================\n// Compaction Preparation (for extensions)\n// ============================================================================\n\nexport interface CompactionPreparation {\n\t/** UUID of first entry to keep */\n\tfirstKeptEntryId: string;\n\t/** Messages that will be summarized and discarded */\n\tmessagesToSummarize: AgentMessage[];\n\t/** Messages that will be turned into turn prefix summary (if splitting) */\n\tturnPrefixMessages: AgentMessage[];\n\t/** Whether this is a split turn (cut point in middle of turn) */\n\tisSplitTurn: boolean;\n\ttokensBefore: number;\n\t/** Summary from previous compaction, for iterative update */\n\tpreviousSummary?: string;\n\t/** File operations extracted from messagesToSummarize */\n\tfileOps: FileOperations;\n\t/** Compaction settions from settings.jsonl\t*/\n\tsettings: CompactionSettings;\n}\n\nexport interface CompactionPreparationOptions {\n\tfirstKeptEntryId?: string;\n}\n\nexport function prepareCompaction(\n\tpathEntries: SessionEntry[],\n\tsettings: CompactionSettings,\n\toptions: CompactionPreparationOptions = {},\n): CompactionPreparation | undefined {\n\tif (pathEntries.length > 0 && pathEntries[pathEntries.length - 1].type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\n\tlet prevCompactionIndex = -1;\n\tfor (let i = pathEntries.length - 1; i >= 0; i--) {\n\t\tif (pathEntries[i].type === \"compaction\") {\n\t\t\tprevCompactionIndex = i;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\tlet previousSummary: string | undefined;\n\tlet boundaryStart = 0;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tconst firstKeptEntryIndex = pathEntries.findIndex((entry) => entry.id === prevCompaction.firstKeptEntryId);\n\t\tboundaryStart = firstKeptEntryIndex >= 0 ? firstKeptEntryIndex : prevCompactionIndex + 1;\n\t}\n\tconst boundaryEnd = pathEntries.length;\n\n\tconst tokensBefore = estimateContextTokens(buildSessionContext(pathEntries).messages).tokens;\n\n\tconst 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 undefined;\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: findCutPoint(pathEntries, boundaryStart, boundaryEnd, settings.keepRecentTokens);\n\n\t// Get UUID of first kept entry\n\tconst firstKeptEntry = pathEntries[cutPoint.firstKeptEntryIndex];\n\tif (!firstKeptEntry?.id) {\n\t\treturn undefined; // Session needs migration\n\t}\n\tconst firstKeptEntryId = firstKeptEntry.id;\n\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\n\t// Messages to summarize (will be discarded after summary)\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let i = boundaryStart; i < historyEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\tif (msg) messagesToSummarize.push(msg);\n\t}\n\n\t// Messages for turn prefix summary (if splitting a turn)\n\tconst turnPrefixMessages: AgentMessage[] = [];\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (let i = cutPoint.turnStartIndex; i < cutPoint.firstKeptEntryIndex; i++) {\n\t\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\t\tif (msg) turnPrefixMessages.push(msg);\n\t\t}\n\t}\n\n\tif (messagesToSummarize.length === 0 && turnPrefixMessages.length === 0) {\n\t\treturn undefined;\n\t}\n\n\t// Extract file operations from messages and previous compaction\n\tconst fileOps = extractFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\n\n\t// Also extract file ops from turn prefix if splitting\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (const msg of turnPrefixMessages) {\n\t\t\textractFileOpsFromMessage(msg, fileOps);\n\t\t}\n\t}\n\n\treturn {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t};\n}\n\n// ============================================================================\n// Main compaction function\n// ============================================================================\n\nconst TURN_PREFIX_SUMMARIZATION_PROMPT = `This is the PREFIX of a turn that was too large to keep. The SUFFIX (recent work) is retained.\n\nSummarize the prefix to provide context for the retained suffix:\n\n## Original Request\n[What did the user ask for in this turn?]\n\n## Early Progress\n- [Key decisions and work done in the prefix]\n\n## Context for Suffix\n- [Information needed to understand the retained recent work]\n\nBe concise. Focus on what's needed to understand the kept suffix.`;\n\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\n/**\n * Generate summaries for compaction using prepared data.\n * Returns CompactionResult - SessionManager adds uuid/parentUuid when saving.\n *\n * @param preparation - Pre-calculated preparation from prepareCompaction()\n * @param customInstructions - Optional custom focus for the summary\n */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tenv?: Record<string, string>,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tsessionId?: string,\n): Promise<CompactionResult> {\n\tconst {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\t// Generate summaries and merge into one\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\tmodel,\n\t\t\t\tsettings.reserveTokens,\n\t\t\t\tapiKey,\n\t\t\t\theaders,\n\t\t\t\tsignal,\n\t\t\t\tcustomInstructions,\n\t\t\t\tpreviousSummary,\n\t\t\t\tthinkingLevel,\n\t\t\t\tstreamFn,\n\t\t\t\tenv,\n\t\t\t\tretry,\n\t\t\t\tcallbacks,\n\t\t\t\tsessionId,\n\t\t\t);\n\t\t\thistoryText = historyResult.text;\n\t\t\thistoryUsage = historyResult.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\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t\tsessionId,\n\t\t);\n\t\t// Merge into single summary\n\t\tsummary = `${historyText}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult.text}`;\n\t\tsummaryUsage = historyUsage ? combineUsage(historyUsage, turnPrefixResult.usage) : turnPrefixResult.usage;\n\t} else {\n\t\t// Just generate history summary\n\t\tconst result = await generateSummaryWithUsage(\n\t\t\tmessagesToSummarize,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tenv,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t\tsessionId,\n\t\t);\n\t\tsummary = result.text;\n\t\tsummaryUsage = result.usage;\n\t}\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\tif (!firstKeptEntryId) {\n\t\tthrow new Error(\"First kept entry has no UUID - session may need migration\");\n\t}\n\n\treturn {\n\t\tsummary,\n\t\tfirstKeptEntryId,\n\t\ttokensBefore,\n\t\tusage: summaryUsage,\n\t\tdetails: { readFiles, modifiedFiles } as CompactionDetails,\n\t};\n}\n\n/**\n * Generate a summary for a turn prefix (when splitting a turn).\n */\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tenv?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n\tretry?: RetryPolicy,\n\tcallbacks?: RetryCallbacks,\n\tsessionId?: string,\n): Promise<{ text: string; usage: Usage }> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.5 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t); // Smaller budget for turn prefix\n\n\ttry {\n\t\treturn await summarizeMessages(\n\t\t\tmessages,\n\t\t\tmodel,\n\t\t\tmaxTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tenv,\n\t\t\tsignal,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t\tTURN_PREFIX_SUMMARIZATION_PROMPT,\n\t\t\tUPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT,\n\t\t\tundefined,\n\t\t\tretry,\n\t\t\tcallbacks,\n\t\t\tsessionId,\n\t\t);\n\t} catch (error) {\n\t\tconst message = error instanceof Error ? error.message : String(error);\n\t\tif (message === \"Summarization attempted to call a tool\") {\n\t\t\tthrow new Error(\"Turn prefix summarization attempted to call a tool\");\n\t\t}\n\t\tthrow new Error(`Turn prefix summarization failed: ${message.replace(/^Summarization failed: /, \"\")}`);\n\t}\n}\n"]}

package/dist/core/compaction/compaction.js

@@ -439,8 +439,21 @@ Use this EXACT format:
 - [Preserve important context, add new if needed]
 
 Keep each section concise. Preserve exact file paths, function names, and error messages.`;
-function createSummarizationOptions(model, maxTokens, apiKey, headers, env, signal, thinkingLevel) {
-    const options = { maxTokens, signal, apiKey, headers, env };
+/**
+ * Returns an error message when a summarization response cannot safely be persisted.
+ * A length stop contains partial text and must not become a session checkpoint.
+ */
+export function getSummarizationFailure(response, label) {
+    if (response.stopReason === "error") {
+        return `${label} failed: ${response.errorMessage || "Unknown error"}`;
+    }
+    if (response.stopReason === "length") {
+        return `${label} failed: generation hit the token cap and the summary is incomplete`;
+    }
+    return undefined;
+}
+function createSummarizationOptions(model, maxTokens, apiKey, headers, env, signal, thinkingLevel, sessionId) {
+    const options = { maxTokens, signal, apiKey, headers, env, sessionId };
     if (model.reasoning && thinkingLevel && thinkingLevel !== "off") {
         options.reasoning = thinkingLevel;
     }
@@ -504,14 +517,14 @@ export async function completeSummarization(model, context, options, streamFn, r
     const requestOptions = {
         ...options,
         cacheRetention: "none",
-        sessionId: uuidv7(),
+        sessionId: options.sessionId ?? uuidv7(),
     };
     const produce = async () => streamFn
         ? (await streamFn(model, context, requestOptions)).result()
         : completeSimple(model, context, requestOptions);
     return retryAssistantCall(produce, retry, requestOptions.signal, callbacks);
 }
-async function summarizeChunk(conversationText, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks) {
+async function summarizeChunk(conversationText, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks, sessionId) {
     const basePrompt = previousSummary ? updatePrompt : initialPrompt;
     const promptText = buildSummaryPrompt(conversationText, basePrompt, previousSummary);
     const summarizationMessages = [
@@ -521,32 +534,36 @@ async function summarizeChunk(conversationText, model, maxTokens, apiKey, header
             timestamp: Date.now(),
         },
     ];
-    const response = await completeSummarization(model, { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages }, createSummarizationOptions(model, maxTokens, apiKey, headers, env, signal, thinkingLevel), streamFn, retry, callbacks);
+    const response = await completeSummarization(model, { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages }, createSummarizationOptions(model, maxTokens, apiKey, headers, env, signal, thinkingLevel, sessionId), streamFn, retry, callbacks);
     if (isContextOverflow(response, model.contextWindow) && conversationText.length > 1) {
         const middle = Math.ceil(conversationText.length / 2);
-        const leftSummary = await summarizeChunk(conversationText.slice(0, middle), model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks);
-        const rightSummary = await summarizeChunk(conversationText.slice(middle), model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, leftSummary.text, retry, callbacks);
+        const leftSummary = await summarizeChunk(conversationText.slice(0, middle), model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks, sessionId);
+        const rightSummary = await summarizeChunk(conversationText.slice(middle), model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, leftSummary.text, retry, callbacks, sessionId);
         return {
             text: rightSummary.text,
             usage: combineUsage(leftSummary.usage, rightSummary.usage),
         };
     }
-    if (response.stopReason === "error") {
-        throw new Error(`Summarization failed: ${response.errorMessage || "Unknown error"}`);
+    const failure = getSummarizationFailure(response, "Summarization");
+    if (failure) {
+        throw new Error(failure);
+    }
+    if (response.content.some((block) => block.type === "toolCall")) {
+        throw new Error("Summarization attempted to call a tool");
     }
     return { text: contentText(response.content), usage: response.usage };
 }
-export async function summarizeMessages(messages, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks) {
+export async function summarizeMessages(messages, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks, sessionId) {
     let remaining = toSummarySegments(messages);
     if (remaining.length === 0) {
-        return summarizeChunk("", model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks);
+        return summarizeChunk("", model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks, sessionId);
     }
     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, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, summary, retry, callbacks);
+        const chunkSummary = await summarizeChunk(next.chunk, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, summary, retry, callbacks, sessionId);
         summary = chunkSummary.text;
         summaryUsage = summaryUsage ? combineUsage(summaryUsage, chunkSummary.usage) : chunkSummary.usage;
         remaining = next.remaining;
@@ -559,11 +576,11 @@ export async function summarizeMessages(messages, model, maxTokens, apiKey, head
  * Generate a summary of the conversation using the LLM.
  * If previousSummary is provided, uses the update prompt to merge.
  */
-export async function generateSummary(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks) {
-    return (await generateSummaryWithUsage(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks)).text;
+export async function generateSummary(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks, sessionId) {
+    return (await generateSummaryWithUsage(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks, sessionId)).text;
 }
 /** Generate or update a conversation summary and return its provider usage. */
-export async function generateSummaryWithUsage(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks) {
+export async function generateSummaryWithUsage(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks, sessionId) {
     const maxTokens = getSummaryMaxTokens(model, reserveTokens);
     let initialPrompt = SUMMARIZATION_PROMPT;
     let updatePrompt = UPDATE_SUMMARIZATION_PROMPT;
@@ -571,7 +588,7 @@ export async function generateSummaryWithUsage(currentMessages, model, reserveTo
         initialPrompt = `${initialPrompt}\n\nAdditional focus: ${customInstructions}`;
         updatePrompt = `${updatePrompt}\n\nAdditional focus: ${customInstructions}`;
     }
-    return summarizeMessages(currentMessages, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks);
+    return summarizeMessages(currentMessages, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, initialPrompt, updatePrompt, previousSummary, retry, callbacks, sessionId);
 }
 export function prepareCompaction(pathEntries, settings, options = {}) {
     if (pathEntries.length > 0 && pathEntries[pathEntries.length - 1].type === "compaction") {
@@ -682,7 +699,7 @@ Be concise. Preserve exact file paths, function names, and error messages.`;
  * @param preparation - Pre-calculated preparation from prepareCompaction()
  * @param customInstructions - Optional custom focus for the summary
  */
-export async function compact(preparation, model, apiKey, headers, customInstructions, signal, thinkingLevel, streamFn, env, retry, callbacks) {
+export async function compact(preparation, model, apiKey, headers, customInstructions, signal, thinkingLevel, streamFn, env, retry, callbacks, sessionId) {
     const { firstKeptEntryId, messagesToSummarize, turnPrefixMessages, isSplitTurn, tokensBefore, previousSummary, fileOps, settings, } = preparation;
     // Generate summaries and merge into one
     let summary;
@@ -691,18 +708,18 @@ export async function compact(preparation, model, apiKey, headers, customInstruc
         let historyText = "No prior history.";
         let historyUsage;
         if (messagesToSummarize.length > 0) {
-            const historyResult = await generateSummaryWithUsage(messagesToSummarize, model, settings.reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks);
+            const historyResult = await generateSummaryWithUsage(messagesToSummarize, model, settings.reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks, sessionId);
             historyText = historyResult.text;
             historyUsage = historyResult.usage;
         }
-        const turnPrefixResult = await generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, retry, callbacks);
+        const turnPrefixResult = await generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, retry, callbacks, sessionId);
         // Merge into single summary
         summary = `${historyText}\n\n---\n\n**Turn Context (split turn):**\n\n${turnPrefixResult.text}`;
         summaryUsage = historyUsage ? combineUsage(historyUsage, turnPrefixResult.usage) : turnPrefixResult.usage;
     }
     else {
         // Just generate history summary
-        const result = await generateSummaryWithUsage(messagesToSummarize, model, settings.reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks);
+        const result = await generateSummaryWithUsage(messagesToSummarize, model, settings.reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, env, retry, callbacks, sessionId);
         summary = result.text;
         summaryUsage = result.usage;
     }
@@ -723,13 +740,16 @@ export async function compact(preparation, model, apiKey, headers, customInstruc
 /**
  * Generate a summary for a turn prefix (when splitting a turn).
  */
-async function generateTurnPrefixSummary(messages, model, reserveTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, retry, callbacks) {
+async function generateTurnPrefixSummary(messages, model, reserveTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, retry, callbacks, sessionId) {
     const maxTokens = Math.min(Math.floor(0.5 * reserveTokens), model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY); // Smaller budget for turn prefix
     try {
-        return await summarizeMessages(messages, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, TURN_PREFIX_SUMMARIZATION_PROMPT, UPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT, undefined, retry, callbacks);
+        return await summarizeMessages(messages, model, maxTokens, apiKey, headers, env, signal, thinkingLevel, streamFn, TURN_PREFIX_SUMMARIZATION_PROMPT, UPDATE_TURN_PREFIX_SUMMARIZATION_PROMPT, undefined, retry, callbacks, sessionId);
     }
     catch (error) {
         const message = error instanceof Error ? error.message : String(error);
+        if (message === "Summarization attempted to call a tool") {
+            throw new Error("Turn prefix summarization attempted to call a tool");
+        }
         throw new Error(`Turn prefix summarization failed: ${message.replace(/^Summarization failed: /, "")}`);
     }
 }