@ai-agent-forge/agent-forge 0.85.0

package/dist/core/compaction/branch-summarization.js

@@ -0,0 +1,220 @@
+/**
+ * Branch summarization for tree navigation.
+ *
+ * When navigating to a different point in the session tree, this generates
+ * a summary of the branch being left so context isn't lost.
+ */
+import { BRANCH_SUMMARY_PROMPT, SUMMARIZATION_SYSTEM_PROMPT } from "@agent-forge/agent-core";
+import { SUMMARY_MODEL_INVOCATION_ERROR_CODES } from "../extensions/public-api.js";
+import { convertToLlm, createBranchSummaryMessage, createCompactionSummaryMessage, createCustomMessage, } from "../messages.js";
+import { estimateTokens, formatSummaryFailure } from "./compaction.js";
+import { computeFileLists, createFileOps, extractFileOpsFromMessage, formatFileOperations, serializeConversation, } from "./utils.js";
+// ============================================================================
+// Entry Collection
+// ============================================================================
+/**
+ * Collect entries that should be summarized when navigating from one position to another.
+ *
+ * Walks from oldLeafId back to the common ancestor with targetId, collecting entries
+ * along the way. Does NOT stop at compaction boundaries - those are included and their
+ * summaries become context.
+ *
+ * @param session - Session manager (read-only access)
+ * @param oldLeafId - Current position (where we're navigating from)
+ * @param targetId - Target position (where we're navigating to)
+ * @returns Entries to summarize and the common ancestor
+ */
+export function collectEntriesForBranchSummary(session, oldLeafId, targetId) {
+    // If no old position, nothing to summarize
+    if (!oldLeafId) {
+        return { entries: [], commonAncestorId: null };
+    }
+    // Find common ancestor (deepest node that's on both paths)
+    const oldPath = new Set(session.getBranch(oldLeafId).map((e) => e.id));
+    const targetPath = session.getBranch(targetId);
+    // targetPath is root-first, so iterate backwards to find deepest common ancestor
+    let commonAncestorId = null;
+    for (let i = targetPath.length - 1; i >= 0; i--) {
+        if (oldPath.has(targetPath[i].id)) {
+            commonAncestorId = targetPath[i].id;
+            break;
+        }
+    }
+    // Collect entries from old leaf back to common ancestor
+    const entries = [];
+    let current = oldLeafId;
+    while (current && current !== commonAncestorId) {
+        const entry = session.getEntry(current);
+        if (!entry)
+            break;
+        entries.push(entry);
+        current = entry.parentId;
+    }
+    // Reverse to get chronological order
+    entries.reverse();
+    return { entries, commonAncestorId };
+}
+// ============================================================================
+// Entry to Message Conversion
+// ============================================================================
+/**
+ * Extract AgentMessage from a session entry.
+ * Similar to getMessageFromEntry in compaction.ts but also handles compaction entries.
+ */
+function getMessageFromEntry(entry) {
+    switch (entry.type) {
+        case "message":
+            // Skip tool results - context is in assistant's tool call
+            if (entry.message.role === "toolResult")
+                return undefined;
+            return entry.message;
+        case "custom_message":
+            return createCustomMessage(entry.customType, entry.content, entry.display, entry.details, entry.timestamp);
+        case "branch_summary":
+            return createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);
+        case "compaction":
+            return createCompactionSummaryMessage(entry.summary, entry.tokensBefore, entry.timestamp);
+        // These don't contribute to conversation content
+        case "thinking_level_change":
+        case "model_change":
+        case "custom":
+        case "label":
+        case "session_info":
+            return undefined;
+    }
+}
+/**
+ * Prepare entries for summarization with token budget.
+ *
+ * Walks entries from NEWEST to OLDEST, adding messages until we hit the token budget.
+ * This ensures we keep the most recent context when the branch is too long.
+ *
+ * Also collects file operations from:
+ * - Tool calls in assistant messages
+ * - Existing branch_summary entries' details (for cumulative tracking)
+ *
+ * @param entries - Entries in chronological order
+ * @param tokenBudget - Maximum tokens to include (0 = no limit)
+ */
+export function prepareBranchEntries(entries, tokenBudget = 0) {
+    const messages = [];
+    const fileOps = createFileOps();
+    let totalTokens = 0;
+    // First pass: collect file ops from ALL entries (even if they don't fit in token budget)
+    // This ensures we capture cumulative file tracking from nested branch summaries
+    // Only extract from pi-generated summaries (fromHook !== true), not extension-generated ones
+    for (const entry of entries) {
+        if (entry.type === "branch_summary" && !entry.fromHook && entry.details) {
+            const details = entry.details;
+            if (Array.isArray(details.readFiles)) {
+                for (const f of details.readFiles)
+                    fileOps.read.add(f);
+            }
+            if (Array.isArray(details.modifiedFiles)) {
+                // Modified files go into both edited and written for proper deduplication
+                for (const f of details.modifiedFiles) {
+                    fileOps.edited.add(f);
+                }
+            }
+        }
+    }
+    // Second pass: walk from newest to oldest, adding messages until token budget
+    for (let i = entries.length - 1; i >= 0; i--) {
+        const entry = entries[i];
+        const message = getMessageFromEntry(entry);
+        if (!message)
+            continue;
+        // Extract file ops from assistant messages (tool calls)
+        extractFileOpsFromMessage(message, fileOps);
+        const tokens = estimateTokens(message);
+        // Check budget before adding
+        if (tokenBudget > 0 && totalTokens + tokens > tokenBudget) {
+            // If this is a summary entry, try to fit it anyway as it's important context
+            if (entry.type === "compaction" || entry.type === "branch_summary") {
+                if (totalTokens < tokenBudget * 0.9) {
+                    messages.unshift(message);
+                    totalTokens += tokens;
+                }
+            }
+            // Stop - we've hit the budget
+            break;
+        }
+        messages.unshift(message);
+        totalTokens += tokens;
+    }
+    return { messages, fileOps, totalTokens };
+}
+// ============================================================================
+// Summary Generation
+// ============================================================================
+const BRANCH_SUMMARY_PREAMBLE = `The user explored a different conversation branch before returning here.
+Summary of that exploration:
+
+`;
+/**
+ * Generate a summary of abandoned branch entries.
+ *
+ * @param entries - Session entries to summarize (chronological order)
+ * @param options - Generation options
+ */
+export async function generateBranchSummary(entries, options) {
+    const { invocation, model, signal, customInstructions, replaceInstructions, reserveTokens = 16384 } = options;
+    // Token budget = context window minus reserved space for prompt + response
+    const contextWindow = model.contextWindow || 128000;
+    const tokenBudget = contextWindow - reserveTokens;
+    const { messages, fileOps } = prepareBranchEntries(entries, tokenBudget);
+    if (messages.length === 0) {
+        return { summary: "No content to summarize" };
+    }
+    // Transform to LLM-compatible messages, then serialize to text
+    // Serialization prevents the model from treating it as a conversation to continue
+    const llmMessages = convertToLlm(messages);
+    const conversationText = serializeConversation(llmMessages);
+    // Build prompt
+    let instructions;
+    if (replaceInstructions && customInstructions) {
+        instructions = customInstructions;
+    }
+    else if (customInstructions) {
+        instructions = `${BRANCH_SUMMARY_PROMPT}\n\nAdditional focus: ${customInstructions}`;
+    }
+    else {
+        instructions = BRANCH_SUMMARY_PROMPT;
+    }
+    const promptText = `<conversation>\n${conversationText}\n</conversation>\n\n${instructions}`;
+    const maxTokens = Math.min(4096, model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY);
+    // One host-mediated completion: auth, transport selection, retry budget and
+    // cache routing stay host-owned, so SDK request behavior (timeouts, retries,
+    // attribution headers) stays consistent without running through agent state.
+    const result = await invocation.invoke({
+        version: 1,
+        systemPrompt: SUMMARIZATION_SYSTEM_PROMPT,
+        prompt: promptText,
+        maxTokens,
+    }, signal);
+    // Check if aborted or errored
+    if (result.status === "aborted") {
+        return { aborted: true };
+    }
+    if (result.status === "failed") {
+        // Transport exceptions historically surfaced raw; provider-side failures
+        // keep the labeled formatting.
+        return {
+            error: result.error.code === SUMMARY_MODEL_INVOCATION_ERROR_CODES.transportError
+                ? result.error.message
+                : formatSummaryFailure("Branch summarization", result.error),
+        };
+    }
+    // Prepend preamble to provide context about the branch summary
+    let summary = BRANCH_SUMMARY_PREAMBLE + result.text;
+    // Compute file lists and append to summary
+    const { readFiles, modifiedFiles } = computeFileLists(fileOps);
+    summary += formatFileOperations(readFiles, modifiedFiles);
+    return {
+        summary: summary || "No summary generated",
+        usage: result.usage,
+        readFiles,
+        modifiedFiles,
+    };
+}
+//# sourceMappingURL=branch-summarization.js.map

package/dist/core/compaction/branch-summarization.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"branch-summarization.js","sourceRoot":"","sources":["../../../src/core/compaction/branch-summarization.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAqB,qBAAqB,EAAE,2BAA2B,EAAE,MAAM,yBAAyB,CAAC;AAGhH,OAAO,EAAE,oCAAoC,EAAE,MAAM,6BAA6B,CAAC;AACnF,OAAO,EACN,YAAY,EACZ,0BAA0B,EAC1B,8BAA8B,EAC9B,mBAAmB,GACnB,MAAM,gBAAgB,CAAC;AAExB,OAAO,EAAE,cAAc,EAAE,oBAAoB,EAAE,MAAM,iBAAiB,CAAC;AACvE,OAAO,EACN,gBAAgB,EAChB,aAAa,EACb,yBAAyB,EAEzB,oBAAoB,EACpB,qBAAqB,GACrB,MAAM,YAAY,CAAC;AAsDpB,+EAA+E;AAC/E,mBAAmB;AACnB,+EAA+E;AAE/E;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,8BAA8B,CAC7C,OAA+B,EAC/B,SAAwB,EACxB,QAAgB,EACO;IACvB,2CAA2C;IAC3C,IAAI,CAAC,SAAS,EAAE,CAAC;QAChB,OAAO,EAAE,OAAO,EAAE,EAAE,EAAE,gBAAgB,EAAE,IAAI,EAAE,CAAC;IAChD,CAAC;IAED,2DAA2D;IAC3D,MAAM,OAAO,GAAG,IAAI,GAAG,CAAC,OAAO,CAAC,SAAS,CAAC,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACvE,MAAM,UAAU,GAAG,OAAO,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC;IAE/C,iFAAiF;IACjF,IAAI,gBAAgB,GAAkB,IAAI,CAAC;IAC3C,KAAK,IAAI,CAAC,GAAG,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QACjD,IAAI,OAAO,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC;YACnC,gBAAgB,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YACpC,MAAM;QACP,CAAC;IACF,CAAC;IAED,wDAAwD;IACxD,MAAM,OAAO,GAAmB,EAAE,CAAC;IACnC,IAAI,OAAO,GAAkB,SAAS,CAAC;IAEvC,OAAO,OAAO,IAAI,OAAO,KAAK,gBAAgB,EAAE,CAAC;QAChD,MAAM,KAAK,GAAG,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;QACxC,IAAI,CAAC,KAAK;YAAE,MAAM;QAClB,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QACpB,OAAO,GAAG,KAAK,CAAC,QAAQ,CAAC;IAC1B,CAAC;IAED,qCAAqC;IACrC,OAAO,CAAC,OAAO,EAAE,CAAC;IAElB,OAAO,EAAE,OAAO,EAAE,gBAAgB,EAAE,CAAC;AAAA,CACrC;AAED,+EAA+E;AAC/E,8BAA8B;AAC9B,+EAA+E;AAE/E;;;GAGG;AACH,SAAS,mBAAmB,CAAC,KAAmB,EAA4B;IAC3E,QAAQ,KAAK,CAAC,IAAI,EAAE,CAAC;QACpB,KAAK,SAAS;YACb,0DAA0D;YAC1D,IAAI,KAAK,CAAC,OAAO,CAAC,IAAI,KAAK,YAAY;gBAAE,OAAO,SAAS,CAAC;YAC1D,OAAO,KAAK,CAAC,OAAO,CAAC;QAEtB,KAAK,gBAAgB;YACpB,OAAO,mBAAmB,CAAC,KAAK,CAAC,UAAU,EAAE,KAAK,CAAC,OAAO,EAAE,KAAK,CAAC,OAAO,EAAE,KAAK,CAAC,OAAO,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC;QAE5G,KAAK,gBAAgB;YACpB,OAAO,0BAA0B,CAAC,KAAK,CAAC,OAAO,EAAE,KAAK,CAAC,MAAM,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC;QAEjF,KAAK,YAAY;YAChB,OAAO,8BAA8B,CAAC,KAAK,CAAC,OAAO,EAAE,KAAK,CAAC,YAAY,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC;QAE3F,iDAAiD;QACjD,KAAK,uBAAuB,CAAC;QAC7B,KAAK,cAAc,CAAC;QACpB,KAAK,QAAQ,CAAC;QACd,KAAK,OAAO,CAAC;QACb,KAAK,cAAc;YAClB,OAAO,SAAS,CAAC;IACnB,CAAC;AAAA,CACD;AAED;;;;;;;;;;;;GAYG;AACH,MAAM,UAAU,oBAAoB,CAAC,OAAuB,EAAE,WAAW,GAAW,CAAC,EAAqB;IACzG,MAAM,QAAQ,GAAmB,EAAE,CAAC;IACpC,MAAM,OAAO,GAAG,aAAa,EAAE,CAAC;IAChC,IAAI,WAAW,GAAG,CAAC,CAAC;IAEpB,yFAAyF;IACzF,gFAAgF;IAChF,6FAA6F;IAC7F,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC7B,IAAI,KAAK,CAAC,IAAI,KAAK,gBAAgB,IAAI,CAAC,KAAK,CAAC,QAAQ,IAAI,KAAK,CAAC,OAAO,EAAE,CAAC;YACzE,MAAM,OAAO,GAAG,KAAK,CAAC,OAA+B,CAAC;YACtD,IAAI,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,SAAS,CAAC,EAAE,CAAC;gBACtC,KAAK,MAAM,CAAC,IAAI,OAAO,CAAC,SAAS;oBAAE,OAAO,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;YACxD,CAAC;YACD,IAAI,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,aAAa,CAAC,EAAE,CAAC;gBAC1C,0EAA0E;gBAC1E,KAAK,MAAM,CAAC,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC;oBACvC,OAAO,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;gBACvB,CAAC;YACF,CAAC;QACF,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,KAAK,IAAI,CAAC,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC9C,MAAM,KAAK,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC;QACzB,MAAM,OAAO,GAAG,mBAAmB,CAAC,KAAK,CAAC,CAAC;QAC3C,IAAI,CAAC,OAAO;YAAE,SAAS;QAEvB,wDAAwD;QACxD,yBAAyB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;QAE5C,MAAM,MAAM,GAAG,cAAc,CAAC,OAAO,CAAC,CAAC;QAEvC,6BAA6B;QAC7B,IAAI,WAAW,GAAG,CAAC,IAAI,WAAW,GAAG,MAAM,GAAG,WAAW,EAAE,CAAC;YAC3D,6EAA6E;YAC7E,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY,IAAI,KAAK,CAAC,IAAI,KAAK,gBAAgB,EAAE,CAAC;gBACpE,IAAI,WAAW,GAAG,WAAW,GAAG,GAAG,EAAE,CAAC;oBACrC,QAAQ,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;oBAC1B,WAAW,IAAI,MAAM,CAAC;gBACvB,CAAC;YACF,CAAC;YACD,8BAA8B;YAC9B,MAAM;QACP,CAAC;QAED,QAAQ,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;QAC1B,WAAW,IAAI,MAAM,CAAC;IACvB,CAAC;IAED,OAAO,EAAE,QAAQ,EAAE,OAAO,EAAE,WAAW,EAAE,CAAC;AAAA,CAC1C;AAED,+EAA+E;AAC/E,qBAAqB;AACrB,+EAA+E;AAE/E,MAAM,uBAAuB,GAAG;;;CAG/B,CAAC;AAEF;;;;;GAKG;AACH,MAAM,CAAC,KAAK,UAAU,qBAAqB,CAC1C,OAAuB,EACvB,OAAqC,EACN;IAC/B,MAAM,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,EAAE,kBAAkB,EAAE,mBAAmB,EAAE,aAAa,GAAG,KAAK,EAAE,GAAG,OAAO,CAAC;IAE9G,2EAA2E;IAC3E,MAAM,aAAa,GAAG,KAAK,CAAC,aAAa,IAAI,MAAM,CAAC;IACpD,MAAM,WAAW,GAAG,aAAa,GAAG,aAAa,CAAC;IAElD,MAAM,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,oBAAoB,CAAC,OAAO,EAAE,WAAW,CAAC,CAAC;IAEzE,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC3B,OAAO,EAAE,OAAO,EAAE,yBAAyB,EAAE,CAAC;IAC/C,CAAC;IAED,+DAA+D;IAC/D,kFAAkF;IAClF,MAAM,WAAW,GAAG,YAAY,CAAC,QAAQ,CAAC,CAAC;IAC3C,MAAM,gBAAgB,GAAG,qBAAqB,CAAC,WAAW,CAAC,CAAC;IAE5D,eAAe;IACf,IAAI,YAAoB,CAAC;IACzB,IAAI,mBAAmB,IAAI,kBAAkB,EAAE,CAAC;QAC/C,YAAY,GAAG,kBAAkB,CAAC;IACnC,CAAC;SAAM,IAAI,kBAAkB,EAAE,CAAC;QAC/B,YAAY,GAAG,GAAG,qBAAqB,yBAAyB,kBAAkB,EAAE,CAAC;IACtF,CAAC;SAAM,CAAC;QACP,YAAY,GAAG,qBAAqB,CAAC;IACtC,CAAC;IACD,MAAM,UAAU,GAAG,mBAAmB,gBAAgB,wBAAwB,YAAY,EAAE,CAAC;IAE7F,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,KAAK,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,iBAAiB,CAAC,CAAC;IAEnG,4EAA4E;IAC5E,6EAA6E;IAC7E,6EAA6E;IAC7E,MAAM,MAAM,GAAG,MAAM,UAAU,CAAC,MAAM,CACrC;QACC,OAAO,EAAE,CAAC;QACV,YAAY,EAAE,2BAA2B;QACzC,MAAM,EAAE,UAAU;QAClB,SAAS;KACT,EACD,MAAM,CACN,CAAC;IAEF,8BAA8B;IAC9B,IAAI,MAAM,CAAC,MAAM,KAAK,SAAS,EAAE,CAAC;QACjC,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IAC1B,CAAC;IACD,IAAI,MAAM,CAAC,MAAM,KAAK,QAAQ,EAAE,CAAC;QAChC,yEAAyE;QACzE,+BAA+B;QAC/B,OAAO;YACN,KAAK,EACJ,MAAM,CAAC,KAAK,CAAC,IAAI,KAAK,oCAAoC,CAAC,cAAc;gBACxE,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,OAAO;gBACtB,CAAC,CAAC,oBAAoB,CAAC,sBAAsB,EAAE,MAAM,CAAC,KAAK,CAAC;SAC9D,CAAC;IACH,CAAC;IAED,+DAA+D;IAC/D,IAAI,OAAO,GAAG,uBAAuB,GAAG,MAAM,CAAC,IAAI,CAAC;IAEpD,2CAA2C;IAC3C,MAAM,EAAE,SAAS,EAAE,aAAa,EAAE,GAAG,gBAAgB,CAAC,OAAO,CAAC,CAAC;IAC/D,OAAO,IAAI,oBAAoB,CAAC,SAAS,EAAE,aAAa,CAAC,CAAC;IAE1D,OAAO;QACN,OAAO,EAAE,OAAO,IAAI,sBAAsB;QAC1C,KAAK,EAAE,MAAM,CAAC,KAA0B;QACxC,SAAS;QACT,aAAa;KACb,CAAC;AAAA,CACF","sourcesContent":["/**\n * Branch summarization for tree navigation.\n *\n * When navigating to a different point in the session tree, this generates\n * a summary of the branch being left so context isn't lost.\n */\n\nimport { type AgentMessage, BRANCH_SUMMARY_PROMPT, SUMMARIZATION_SYSTEM_PROMPT } from \"@agent-forge/agent-core\";\nimport type { Usage } from \"@agent-forge/ai/compat\";\nimport type { SummaryModelFactsV1, SummaryModelInvocationV1 } from \"../extensions/public-api.ts\";\nimport { SUMMARY_MODEL_INVOCATION_ERROR_CODES } from \"../extensions/public-api.ts\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport type { ReadonlySessionManager, SessionEntry } from \"../session-manager.ts\";\nimport { estimateTokens, formatSummaryFailure } from \"./compaction.ts\";\nimport {\n\tcomputeFileLists,\n\tcreateFileOps,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tserializeConversation,\n} from \"./utils.ts\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface BranchSummaryResult {\n\tsummary?: string;\n\tusage?: Usage;\n\treadFiles?: string[];\n\tmodifiedFiles?: string[];\n\taborted?: boolean;\n\terror?: string;\n}\n\n/** Details stored in BranchSummaryEntry.details for file tracking */\nexport interface BranchSummaryDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n}\n\nexport type { FileOperations } from \"./utils.ts\";\n\nexport interface BranchPreparation {\n\t/** Messages extracted for summarization, in chronological order */\n\tmessages: AgentMessage[];\n\t/** File operations extracted from tool calls */\n\tfileOps: FileOperations;\n\t/** Total estimated tokens in messages */\n\ttotalTokens: number;\n}\n\nexport interface CollectEntriesResult {\n\t/** Entries to summarize, in chronological order */\n\tentries: SessionEntry[];\n\t/** Common ancestor between old and new position, if any */\n\tcommonAncestorId: string | null;\n}\n\nexport interface GenerateBranchSummaryOptions {\n\t/** Host-owned single-shot completion boundary for the summary call. */\n\tinvocation: SummaryModelInvocationV1;\n\t/** Provider-agnostic facts about the resolved summary model. */\n\tmodel: SummaryModelFactsV1;\n\t/** Abort signal for cancellation */\n\tsignal: AbortSignal;\n\t/** Optional custom instructions for summarization */\n\tcustomInstructions?: string;\n\t/** If true, customInstructions replaces the default prompt instead of being appended */\n\treplaceInstructions?: boolean;\n\t/** Tokens reserved when selecting branch history (default 16384) */\n\treserveTokens?: number;\n}\n\n// ============================================================================\n// Entry Collection\n// ============================================================================\n\n/**\n * Collect entries that should be summarized when navigating from one position to another.\n *\n * Walks from oldLeafId back to the common ancestor with targetId, collecting entries\n * along the way. Does NOT stop at compaction boundaries - those are included and their\n * summaries become context.\n *\n * @param session - Session manager (read-only access)\n * @param oldLeafId - Current position (where we're navigating from)\n * @param targetId - Target position (where we're navigating to)\n * @returns Entries to summarize and the common ancestor\n */\nexport function collectEntriesForBranchSummary(\n\tsession: ReadonlySessionManager,\n\toldLeafId: string | null,\n\ttargetId: string,\n): CollectEntriesResult {\n\t// If no old position, nothing to summarize\n\tif (!oldLeafId) {\n\t\treturn { entries: [], commonAncestorId: null };\n\t}\n\n\t// Find common ancestor (deepest node that's on both paths)\n\tconst oldPath = new Set(session.getBranch(oldLeafId).map((e) => e.id));\n\tconst targetPath = session.getBranch(targetId);\n\n\t// targetPath is root-first, so iterate backwards to find deepest common ancestor\n\tlet commonAncestorId: string | null = null;\n\tfor (let i = targetPath.length - 1; i >= 0; i--) {\n\t\tif (oldPath.has(targetPath[i].id)) {\n\t\t\tcommonAncestorId = targetPath[i].id;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Collect entries from old leaf back to common ancestor\n\tconst entries: SessionEntry[] = [];\n\tlet current: string | null = oldLeafId;\n\n\twhile (current && current !== commonAncestorId) {\n\t\tconst entry = session.getEntry(current);\n\t\tif (!entry) break;\n\t\tentries.push(entry);\n\t\tcurrent = entry.parentId;\n\t}\n\n\t// Reverse to get chronological order\n\tentries.reverse();\n\n\treturn { entries, commonAncestorId };\n}\n\n// ============================================================================\n// Entry to Message Conversion\n// ============================================================================\n\n/**\n * Extract AgentMessage from a session entry.\n * Similar to getMessageFromEntry in compaction.ts but also handles compaction entries.\n */\nfunction getMessageFromEntry(entry: SessionEntry): AgentMessage | undefined {\n\tswitch (entry.type) {\n\t\tcase \"message\":\n\t\t\t// Skip tool results - context is in assistant's tool call\n\t\t\tif (entry.message.role === \"toolResult\") return undefined;\n\t\t\treturn entry.message;\n\n\t\tcase \"custom_message\":\n\t\t\treturn createCustomMessage(entry.customType, entry.content, entry.display, entry.details, entry.timestamp);\n\n\t\tcase \"branch_summary\":\n\t\t\treturn createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);\n\n\t\tcase \"compaction\":\n\t\t\treturn createCompactionSummaryMessage(entry.summary, entry.tokensBefore, entry.timestamp);\n\n\t\t// These don't contribute to conversation content\n\t\tcase \"thinking_level_change\":\n\t\tcase \"model_change\":\n\t\tcase \"custom\":\n\t\tcase \"label\":\n\t\tcase \"session_info\":\n\t\t\treturn undefined;\n\t}\n}\n\n/**\n * Prepare entries for summarization with token budget.\n *\n * Walks entries from NEWEST to OLDEST, adding messages until we hit the token budget.\n * This ensures we keep the most recent context when the branch is too long.\n *\n * Also collects file operations from:\n * - Tool calls in assistant messages\n * - Existing branch_summary entries' details (for cumulative tracking)\n *\n * @param entries - Entries in chronological order\n * @param tokenBudget - Maximum tokens to include (0 = no limit)\n */\nexport function prepareBranchEntries(entries: SessionEntry[], tokenBudget: number = 0): BranchPreparation {\n\tconst messages: AgentMessage[] = [];\n\tconst fileOps = createFileOps();\n\tlet totalTokens = 0;\n\n\t// First pass: collect file ops from ALL entries (even if they don't fit in token budget)\n\t// This ensures we capture cumulative file tracking from nested branch summaries\n\t// Only extract from pi-generated summaries (fromHook !== true), not extension-generated ones\n\tfor (const entry of entries) {\n\t\tif (entry.type === \"branch_summary\" && !entry.fromHook && entry.details) {\n\t\t\tconst details = entry.details as BranchSummaryDetails;\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\t// Modified files go into both edited and written for proper deduplication\n\t\t\t\tfor (const f of details.modifiedFiles) {\n\t\t\t\t\tfileOps.edited.add(f);\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t}\n\n\t// Second pass: walk from newest to oldest, adding messages until token budget\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tconst message = getMessageFromEntry(entry);\n\t\tif (!message) continue;\n\n\t\t// Extract file ops from assistant messages (tool calls)\n\t\textractFileOpsFromMessage(message, fileOps);\n\n\t\tconst tokens = estimateTokens(message);\n\n\t\t// Check budget before adding\n\t\tif (tokenBudget > 0 && totalTokens + tokens > tokenBudget) {\n\t\t\t// If this is a summary entry, try to fit it anyway as it's important context\n\t\t\tif (entry.type === \"compaction\" || entry.type === \"branch_summary\") {\n\t\t\t\tif (totalTokens < tokenBudget * 0.9) {\n\t\t\t\t\tmessages.unshift(message);\n\t\t\t\t\ttotalTokens += tokens;\n\t\t\t\t}\n\t\t\t}\n\t\t\t// Stop - we've hit the budget\n\t\t\tbreak;\n\t\t}\n\n\t\tmessages.unshift(message);\n\t\ttotalTokens += tokens;\n\t}\n\n\treturn { messages, fileOps, totalTokens };\n}\n\n// ============================================================================\n// Summary Generation\n// ============================================================================\n\nconst BRANCH_SUMMARY_PREAMBLE = `The user explored a different conversation branch before returning here.\nSummary of that exploration:\n\n`;\n\n/**\n * Generate a summary of abandoned branch entries.\n *\n * @param entries - Session entries to summarize (chronological order)\n * @param options - Generation options\n */\nexport async function generateBranchSummary(\n\tentries: SessionEntry[],\n\toptions: GenerateBranchSummaryOptions,\n): Promise<BranchSummaryResult> {\n\tconst { invocation, model, signal, customInstructions, replaceInstructions, reserveTokens = 16384 } = options;\n\n\t// Token budget = context window minus reserved space for prompt + response\n\tconst contextWindow = model.contextWindow || 128000;\n\tconst tokenBudget = contextWindow - reserveTokens;\n\n\tconst { messages, fileOps } = prepareBranchEntries(entries, tokenBudget);\n\n\tif (messages.length === 0) {\n\t\treturn { summary: \"No content to summarize\" };\n\t}\n\n\t// Transform to LLM-compatible messages, then serialize to text\n\t// Serialization prevents the model from treating it as a conversation to continue\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversation(llmMessages);\n\n\t// Build prompt\n\tlet instructions: string;\n\tif (replaceInstructions && customInstructions) {\n\t\tinstructions = customInstructions;\n\t} else if (customInstructions) {\n\t\tinstructions = `${BRANCH_SUMMARY_PROMPT}\\n\\nAdditional focus: ${customInstructions}`;\n\t} else {\n\t\tinstructions = BRANCH_SUMMARY_PROMPT;\n\t}\n\tconst promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n${instructions}`;\n\n\tconst maxTokens = Math.min(4096, model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY);\n\n\t// One host-mediated completion: auth, transport selection, retry budget and\n\t// cache routing stay host-owned, so SDK request behavior (timeouts, retries,\n\t// attribution headers) stays consistent without running through agent state.\n\tconst result = await invocation.invoke(\n\t\t{\n\t\t\tversion: 1,\n\t\t\tsystemPrompt: SUMMARIZATION_SYSTEM_PROMPT,\n\t\t\tprompt: promptText,\n\t\t\tmaxTokens,\n\t\t},\n\t\tsignal,\n\t);\n\n\t// Check if aborted or errored\n\tif (result.status === \"aborted\") {\n\t\treturn { aborted: true };\n\t}\n\tif (result.status === \"failed\") {\n\t\t// Transport exceptions historically surfaced raw; provider-side failures\n\t\t// keep the labeled formatting.\n\t\treturn {\n\t\t\terror:\n\t\t\t\tresult.error.code === SUMMARY_MODEL_INVOCATION_ERROR_CODES.transportError\n\t\t\t\t\t? result.error.message\n\t\t\t\t\t: formatSummaryFailure(\"Branch summarization\", result.error),\n\t\t};\n\t}\n\n\t// Prepend preamble to provide context about the branch summary\n\tlet summary = BRANCH_SUMMARY_PREAMBLE + result.text;\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\treturn {\n\t\tsummary: summary || \"No summary generated\",\n\t\tusage: result.usage as Usage | undefined,\n\t\treadFiles,\n\t\tmodifiedFiles,\n\t};\n}\n"]}

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

@@ -0,0 +1,178 @@
+/**
+ * Context compaction for long sessions.
+ *
+ * Pure functions for compaction logic. The session manager handles I/O,
+ * and after compaction the session is reloaded.
+ */
+import { type AgentMessage, type ThinkingLevel } from "@agent-forge/agent-core";
+import type { Usage } from "@agent-forge/ai/compat";
+import { type SummaryModelFactsV1, type SummaryModelInvocationV1 } from "../extensions/public-api.ts";
+import { type SessionEntry } from "../session-manager.ts";
+import { type FileOperations } from "./utils.ts";
+/** Details stored in CompactionEntry.details for file tracking */
+export interface CompactionDetails {
+    readFiles: string[];
+    modifiedFiles: string[];
+}
+/** Result from compact() - SessionManager adds uuid/parentUuid when saving */
+export interface CompactionResult<T = unknown> {
+    summary: string;
+    firstKeptEntryId: string;
+    tokensBefore: number;
+    estimatedTokensAfter?: number;
+    /** Usage from the LLM call(s) that generated this summary, if available */
+    usage?: Usage;
+    /** Extension-specific data (e.g., ArtifactIndex, version markers for structured compaction) */
+    details?: T;
+}
+export interface CompactionSettings {
+    enabled: boolean;
+    reserveTokens: number;
+    keepRecentTokens: number;
+}
+export declare const DEFAULT_COMPACTION_SETTINGS: CompactionSettings;
+/**
+ * Calculate total context tokens from usage.
+ * Uses the native totalTokens field when available, falls back to computing from components.
+ */
+export declare function calculateContextTokens(usage: Usage): number;
+/**
+ * Find the last valid assistant message usage from session entries.
+ */
+export declare function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined;
+export interface ContextUsageEstimate {
+    tokens: number;
+    usageTokens: number;
+    trailingTokens: number;
+    lastUsageIndex: number | null;
+}
+/**
+ * Estimate context tokens from messages, using the last assistant usage when available.
+ * If there are messages after the last usage, estimate their tokens with estimateTokens.
+ */
+export declare function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate;
+/**
+ * Check if compaction should trigger based on context usage.
+ */
+export declare function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean;
+/**
+ * Estimate token count for a message using chars/4 heuristic.
+ * This is conservative (overestimates tokens).
+ */
+export declare function estimateTokens(message: AgentMessage): number;
+/**
+ * Find the context-visible user-role message that starts the turn containing the given entry index.
+ * Returns -1 if no turn start found before the index.
+ */
+export declare function findTurnStartIndex(entries: SessionEntry[], entryIndex: number, startIndex: number): number;
+export interface CutPointResult {
+    /** Index of first entry to keep */
+    firstKeptEntryIndex: number;
+    /** Index of user message that starts the turn being split, or -1 if not splitting */
+    turnStartIndex: number;
+    /** Whether this cut splits a turn (cut point is not a user message) */
+    isSplitTurn: boolean;
+}
+/**
+ * Find the cut point in session entries that keeps approximately `keepRecentTokens`.
+ *
+ * Algorithm: Walk backwards from newest, accumulating estimated message sizes.
+ * Stop when we've accumulated >= keepRecentTokens. Cut at that point.
+ *
+ * Can cut at user OR assistant messages (never tool results). When cutting at an
+ * assistant message with tool calls, its tool results come after and will be kept.
+ *
+ * Returns CutPointResult with:
+ * - firstKeptEntryIndex: the entry index to start keeping from
+ * - turnStartIndex: if cutting mid-turn, the user message that started that turn
+ * - isSplitTurn: whether we're cutting in the middle of a turn
+ *
+ * Only considers entries between `startIndex` and `endIndex` (exclusive).
+ */
+export declare function findCutPoint(entries: SessionEntry[], startIndex: number, endIndex: number, keepRecentTokens: number): CutPointResult;
+/**
+ * Rebuild the historical per-call-site failure text from a structured
+ * invocation failure: tool-call violations keep the
+ * `${label} attempted to call a tool` wording, everything else formats as
+ * `${label} failed: ${message}`.
+ */
+export declare function formatSummaryFailure(label: string, error: {
+    code: string;
+    message: string;
+}): string;
+/**
+ * Semantic inputs for one summary generation. Model resolution, auth and
+ * transport stay inside the host-owned invocation boundary; these inputs only
+ * describe the semantic request and the output budget.
+ */
+export interface SummaryGenerationInputs {
+    /** Host-owned single-shot completion boundary for the summary call(s). */
+    invocation: SummaryModelInvocationV1;
+    /** Provider-agnostic facts about the resolved summary model. */
+    model: SummaryModelFactsV1;
+    /** Output token budget before the host's model-cap clamp. */
+    reserveTokens: number;
+    /** Optional custom focus appended to the summary prompt. */
+    customInstructions?: string;
+    /** Previous summary to update instead of writing an initial one. */
+    previousSummary?: string;
+    /** Reasoning effort forwarded when the summary model supports reasoning. */
+    thinkingLevel?: ThinkingLevel;
+    /** Cancellation boundary observed by the summary call. */
+    signal?: AbortSignal;
+    /** Error-message label for this call site, e.g. "Summarization". */
+    failureLabel: string;
+}
+/**
+ * Generate a summary of the conversation using the LLM.
+ * If previousSummary is provided, uses the update prompt to merge.
+ */
+export declare function generateSummary(currentMessages: AgentMessage[], inputs: SummaryGenerationInputs): Promise<string>;
+/** Generate or update a conversation summary and return its provider usage. */
+export declare function generateSummaryWithUsage(currentMessages: AgentMessage[], inputs: SummaryGenerationInputs): Promise<{
+    text: string;
+    usage: Usage;
+}>;
+export interface CompactionPreparation {
+    /** UUID of first entry to keep */
+    firstKeptEntryId: string;
+    /** Messages that will be summarized and discarded */
+    messagesToSummarize: AgentMessage[];
+    /** Messages that will be turned into turn prefix summary (if splitting) */
+    turnPrefixMessages: AgentMessage[];
+    /** Whether this is a split turn (cut point in middle of turn) */
+    isSplitTurn: boolean;
+    tokensBefore: number;
+    /** Summary from previous compaction, for iterative update */
+    previousSummary?: string;
+    /** File operations extracted from messagesToSummarize */
+    fileOps: FileOperations;
+    /** Compaction settions from settings.jsonl	*/
+    settings: CompactionSettings;
+}
+export declare function prepareCompaction(pathEntries: SessionEntry[], settings: CompactionSettings): CompactionPreparation | undefined;
+/**
+ * Semantic inputs for one whole compaction. Both split-turn summary calls
+ * (history and turn prefix) share the same invocation boundary and signal.
+ */
+export interface CompactionSummaryInputs {
+    /** Host-owned single-shot completion boundary for the summary call(s). */
+    invocation: SummaryModelInvocationV1;
+    /** Provider-agnostic facts about the resolved summary model. */
+    model: SummaryModelFactsV1;
+    /** Optional custom focus appended to the summary prompts. */
+    customInstructions?: string;
+    /** Reasoning effort forwarded when the summary model supports reasoning. */
+    thinkingLevel?: ThinkingLevel;
+    /** Cancellation boundary observed by the summary call(s). */
+    signal?: AbortSignal;
+}
+/**
+ * Generate summaries for compaction using prepared data.
+ * Returns CompactionResult - SessionManager adds uuid/parentUuid when saving.
+ *
+ * @param preparation - Pre-calculated preparation from prepareCompaction()
+ * @param inputs - Host-mediated invocation boundary plus semantic summary inputs
+ */
+export declare function compact(preparation: CompactionPreparation, inputs: CompactionSummaryInputs): Promise<CompactionResult>;
+//# sourceMappingURL=compaction.d.ts.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EACN,KAAK,YAAY,EAGjB,KAAK,aAAa,EAGlB,MAAM,yBAAyB,CAAC;AACjC,OAAO,KAAK,EAAoB,KAAK,EAAE,MAAM,wBAAwB,CAAC;AACtE,OAAO,EAEN,KAAK,mBAAmB,EAExB,KAAK,wBAAwB,EAC7B,MAAM,6BAA6B,CAAC;AAErC,OAAO,EAGN,KAAK,YAAY,EAEjB,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAIN,KAAK,cAAc,EAInB,MAAM,YAAY,CAAC;AAapB,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;AAMD;;;;;GAKG;AACH,wBAAgB,oBAAoB,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAAG,MAAM,CAKpG;AA+CD;;;;GAIG;AACH,MAAM,WAAW,uBAAuB;IACvC,0EAA0E;IAC1E,UAAU,EAAE,wBAAwB,CAAC;IACrC,gEAAgE;IAChE,KAAK,EAAE,mBAAmB,CAAC;IAC3B,6DAA6D;IAC7D,aAAa,EAAE,MAAM,CAAC;IACtB,4DAA4D;IAC5D,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,oEAAoE;IACpE,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,4EAA4E;IAC5E,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,0DAA0D;IAC1D,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,oEAAoE;IACpE,YAAY,EAAE,MAAM,CAAC;CACrB;AAED;;;GAGG;AACH,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,MAAM,EAAE,uBAAuB,GAC7B,OAAO,CAAC,MAAM,CAAC,CAEjB;AAED,+EAA+E;AAC/E,wBAAsB,wBAAwB,CAC7C,eAAe,EAAE,YAAY,EAAE,EAC/B,MAAM,EAAE,uBAAuB,GAC7B,OAAO,CAAC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,KAAK,CAAA;CAAE,CAAC,CAiDzC;AAsBD,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,wBAAgB,iBAAiB,CAChC,WAAW,EAAE,YAAY,EAAE,EAC3B,QAAQ,EAAE,kBAAkB,GAC1B,qBAAqB,GAAG,SAAS,CA4EnC;AAMD;;;GAGG;AACH,MAAM,WAAW,uBAAuB;IACvC,0EAA0E;IAC1E,UAAU,EAAE,wBAAwB,CAAC;IACrC,gEAAgE;IAChE,KAAK,EAAE,mBAAmB,CAAC;IAC3B,6DAA6D;IAC7D,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,4EAA4E;IAC5E,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,6DAA6D;IAC7D,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAED;;;;;;GAMG;AACH,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,MAAM,EAAE,uBAAuB,GAC7B,OAAO,CAAC,gBAAgB,CAAC,CA4E3B","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 {\n\ttype AgentMessage,\n\tSUMMARIZATION_PROMPT,\n\tSUMMARIZATION_SYSTEM_PROMPT,\n\ttype ThinkingLevel,\n\tTURN_PREFIX_SUMMARIZATION_PROMPT,\n\tUPDATE_SUMMARIZATION_PROMPT,\n} from \"@agent-forge/agent-core\";\nimport type { AssistantMessage, Usage } from \"@agent-forge/ai/compat\";\nimport {\n\tSUMMARY_MODEL_INVOCATION_ERROR_CODES,\n\ttype SummaryModelFactsV1,\n\ttype SummaryModelInvocationRequestV1,\n\ttype SummaryModelInvocationV1,\n} from \"../extensions/public-api.ts\";\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\tserializeConversation,\n\tserializeConversationTail,\n} from \"./utils.ts\";\n\n/**\n * Safety margin subtracted from the summary model's context window on top of\n * the output budget: covers the system prompt, the summarization instructions\n * and envelope tags around the conversation text.\n */\nconst SUMMARY_PROMPT_SAFETY_MARGIN_TOKENS = 4096;\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\n/**\n * Rebuild the historical per-call-site failure text from a structured\n * invocation failure: tool-call violations keep the\n * `${label} attempted to call a tool` wording, everything else formats as\n * `${label} failed: ${message}`.\n */\nexport function formatSummaryFailure(label: string, error: { code: string; message: string }): string {\n\tif (error.code === SUMMARY_MODEL_INVOCATION_ERROR_CODES.toolCall) {\n\t\treturn `${label} attempted to call a tool`;\n\t}\n\treturn `${label} failed: ${error.message}`;\n}\n\n/** Abort outcome of a summary invocation, normalized to the transport AbortError shape. */\nfunction summaryAbortError(): Error {\n\tconst error = new Error(\"This operation was aborted\");\n\terror.name = \"AbortError\";\n\treturn error;\n}\n\n/**\n * Run one summary completion through the host invocation boundary and map the\n * structured outcome back to the internal summary shape. The host owns auth,\n * transport selection, retry budget and cache routing; this helper only owns\n * the call-site failure label and the usage narrowing.\n */\nasync function invokeSummary(\n\tinvocation: SummaryModelInvocationV1,\n\trequest: Pick<SummaryModelInvocationRequestV1, \"systemPrompt\" | \"prompt\" | \"thinkingLevel\"> & { maxTokens: number },\n\tsignal: AbortSignal | undefined,\n\tfailureLabel: string,\n): Promise<{ text: string; usage: Usage }> {\n\tconst result = await invocation.invoke(\n\t\t{\n\t\t\tversion: 1,\n\t\t\tsystemPrompt: request.systemPrompt,\n\t\t\tprompt: request.prompt,\n\t\t\tmaxTokens: request.maxTokens,\n\t\t\tthinkingLevel: request.thinkingLevel,\n\t\t},\n\t\tsignal ?? new AbortController().signal,\n\t);\n\tif (result.status === \"aborted\") {\n\t\tthrow summaryAbortError();\n\t}\n\tif (result.status === \"failed\") {\n\t\t// Transport exceptions historically surfaced raw, without the call-site\n\t\t// label; provider-side failures keep their labeled formatting.\n\t\tthrow new Error(\n\t\t\tresult.error.code === SUMMARY_MODEL_INVOCATION_ERROR_CODES.transportError\n\t\t\t\t? result.error.message\n\t\t\t\t: formatSummaryFailure(failureLabel, result.error),\n\t\t);\n\t}\n\t// 包内收窄：invocation 结果原样透传 provider usage；压缩 usage 合并需要内部 Usage 形状。\n\treturn { text: result.text, usage: result.usage as Usage };\n}\n\n/**\n * Semantic inputs for one summary generation. Model resolution, auth and\n * transport stay inside the host-owned invocation boundary; these inputs only\n * describe the semantic request and the output budget.\n */\nexport interface SummaryGenerationInputs {\n\t/** Host-owned single-shot completion boundary for the summary call(s). */\n\tinvocation: SummaryModelInvocationV1;\n\t/** Provider-agnostic facts about the resolved summary model. */\n\tmodel: SummaryModelFactsV1;\n\t/** Output token budget before the host's model-cap clamp. */\n\treserveTokens: number;\n\t/** Optional custom focus appended to the summary prompt. */\n\tcustomInstructions?: string;\n\t/** Previous summary to update instead of writing an initial one. */\n\tpreviousSummary?: string;\n\t/** Reasoning effort forwarded when the summary model supports reasoning. */\n\tthinkingLevel?: ThinkingLevel;\n\t/** Cancellation boundary observed by the summary call. */\n\tsignal?: AbortSignal;\n\t/** Error-message label for this call site, e.g. \"Summarization\". */\n\tfailureLabel: string;\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\tinputs: SummaryGenerationInputs,\n): Promise<string> {\n\treturn (await generateSummaryWithUsage(currentMessages, inputs)).text;\n}\n\n/** Generate or update a conversation summary and return its provider usage. */\nexport async function generateSummaryWithUsage(\n\tcurrentMessages: AgentMessage[],\n\tinputs: SummaryGenerationInputs,\n): Promise<{ text: string; usage: Usage }> {\n\tconst {\n\t\tinvocation,\n\t\tmodel,\n\t\treserveTokens,\n\t\tcustomInstructions,\n\t\tpreviousSummary,\n\t\tthinkingLevel,\n\t\tsignal,\n\t\tfailureLabel,\n\t} = inputs;\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.8 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t);\n\n\t// Use update prompt if we have a previous summary, otherwise initial prompt\n\tlet basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tbasePrompt = `${basePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\n\t// Head-trim the serialized conversation to fit the summary model's input\n\t// window. Without this, the sessions that need compaction the most (context\n\t// near 100%) fail their summary request outright: the full history plus the\n\t// prompt exceeds the window. Trimming drops the OLDEST messages and keeps\n\t// the recent tail, matching the codex/ZCode discipline; the estimate is\n\t// chars/4, the same conservative heuristic as `estimateTokens`.\n\tconst fixedParts = basePrompt.length + (previousSummary?.length ?? 0) + SUMMARIZATION_SYSTEM_PROMPT.length;\n\tconst conversationBudgetTokens = summaryPromptBudgetTokens(model, maxTokens, fixedParts);\n\tconst llmMessages = convertToLlm(currentMessages);\n\tconst conversationText =\n\t\tconversationBudgetTokens === undefined\n\t\t\t? serializeConversation(llmMessages)\n\t\t\t: serializeConversationTail(llmMessages, conversationBudgetTokens).text;\n\n\t// Build the prompt with conversation wrapped in tags\n\tlet promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n`;\n\tif (previousSummary) {\n\t\tpromptText += `<previous-summary>\\n${previousSummary}\\n</previous-summary>\\n\\n`;\n\t}\n\tpromptText += basePrompt;\n\n\treturn invokeSummary(\n\t\tinvocation,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText, maxTokens, thinkingLevel },\n\t\tsignal,\n\t\tfailureLabel,\n\t);\n}\n\n/**\n * Input-token budget for the serialized conversation inside one summary\n * prompt: the summary model's context window minus the output budget, the\n * fixed prompt scaffolding, and a safety margin. Undefined when the model\n * declares no usable window (no trimming — behavior unchanged).\n */\nfunction summaryPromptBudgetTokens(\n\tmodel: SummaryModelFactsV1,\n\tmaxTokens: number,\n\tfixedChars: number,\n): number | undefined {\n\tif (!(model.contextWindow > 0)) return undefined;\n\tconst budget = model.contextWindow - maxTokens - SUMMARY_PROMPT_SAFETY_MARGIN_TOKENS - Math.ceil(fixedChars / 4);\n\treturn budget > 0 ? budget : undefined;\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 function prepareCompaction(\n\tpathEntries: SessionEntry[],\n\tsettings: CompactionSettings,\n): CompactionPreparation | undefined {\n\tif (pathEntries.length > 0 && pathEntries[pathEntries.length - 1].type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\n\tlet prevCompactionIndex = -1;\n\tfor (let i = pathEntries.length - 1; i >= 0; i--) {\n\t\tif (pathEntries[i].type === \"compaction\") {\n\t\t\tprevCompactionIndex = i;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\tlet previousSummary: string | undefined;\n\tlet boundaryStart = 0;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tconst firstKeptEntryIndex = pathEntries.findIndex((entry) => entry.id === prevCompaction.firstKeptEntryId);\n\t\tboundaryStart = firstKeptEntryIndex >= 0 ? firstKeptEntryIndex : prevCompactionIndex + 1;\n\t}\n\tconst boundaryEnd = pathEntries.length;\n\n\tconst tokensBefore = estimateContextTokens(buildSessionContext(pathEntries).messages).tokens;\n\n\tconst cutPoint = findCutPoint(pathEntries, boundaryStart, boundaryEnd, settings.keepRecentTokens);\n\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\n/**\n * Semantic inputs for one whole compaction. Both split-turn summary calls\n * (history and turn prefix) share the same invocation boundary and signal.\n */\nexport interface CompactionSummaryInputs {\n\t/** Host-owned single-shot completion boundary for the summary call(s). */\n\tinvocation: SummaryModelInvocationV1;\n\t/** Provider-agnostic facts about the resolved summary model. */\n\tmodel: SummaryModelFactsV1;\n\t/** Optional custom focus appended to the summary prompts. */\n\tcustomInstructions?: string;\n\t/** Reasoning effort forwarded when the summary model supports reasoning. */\n\tthinkingLevel?: ThinkingLevel;\n\t/** Cancellation boundary observed by the summary call(s). */\n\tsignal?: AbortSignal;\n}\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 inputs - Host-mediated invocation boundary plus semantic summary inputs\n */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tinputs: CompactionSummaryInputs,\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\tconst { invocation, model, customInstructions, thinkingLevel, signal } = inputs;\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(messagesToSummarize, {\n\t\t\t\tinvocation,\n\t\t\t\tmodel,\n\t\t\t\treserveTokens: settings.reserveTokens,\n\t\t\t\tcustomInstructions,\n\t\t\t\tpreviousSummary,\n\t\t\t\tthinkingLevel,\n\t\t\t\tsignal,\n\t\t\t\tfailureLabel: \"Summarization\",\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(turnPrefixMessages, {\n\t\t\tinvocation,\n\t\t\tmodel,\n\t\t\treserveTokens: settings.reserveTokens,\n\t\t\tthinkingLevel,\n\t\t\tsignal,\n\t\t\tfailureLabel: \"Turn prefix summarization\",\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(messagesToSummarize, {\n\t\t\tinvocation,\n\t\t\tmodel,\n\t\t\treserveTokens: settings.reserveTokens,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tsignal,\n\t\t\tfailureLabel: \"Summarization\",\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\tinputs: Omit<SummaryGenerationInputs, \"customInstructions\" | \"previousSummary\">,\n): Promise<{ text: string; usage: Usage }> {\n\tconst { invocation, model, reserveTokens, thinkingLevel, signal, failureLabel } = inputs;\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.5 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t); // Smaller budget for turn prefix\n\tconst fixedParts = TURN_PREFIX_SUMMARIZATION_PROMPT.length + SUMMARIZATION_SYSTEM_PROMPT.length;\n\tconst conversationBudgetTokens = summaryPromptBudgetTokens(model, maxTokens, fixedParts);\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText =\n\t\tconversationBudgetTokens === undefined\n\t\t\t? serializeConversation(llmMessages)\n\t\t\t: serializeConversationTail(llmMessages, conversationBudgetTokens).text;\n\tconst promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;\n\n\treturn invokeSummary(\n\t\tinvocation,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText, maxTokens, thinkingLevel },\n\t\tsignal,\n\t\tfailureLabel,\n\t);\n}\n"]}