@caupulican/pi-agent-core 0.81.41 → 0.81.43

package/dist/compaction/compaction.js

@@ -4,12 +4,13 @@
  * Pure functions for compaction logic. The session manager handles I/O,
  * and after compaction the session is reloaded.
  */
-import { completeSimple, uuidv7 } from "@caupulican/pi-ai";
+import { completeSimple } from "@caupulican/pi-ai/stream";
+import { uuidv7 } from "@caupulican/pi-ai/uuid";
 import { convertToLlm, createBranchSummaryMessage, createCompactionSummaryMessage, createCustomMessage, } from "../messages.js";
 import { buildSessionContext } from "../session/session-manager.js";
 import { addUsage, combineUsage, createEmptyUsage } from "../usage.js";
 import { extractCompactionFacts, renderFactsBlock } from "./extraction.js";
-import { computeFileLists, createFileOps, extractFileOpsFromMessage, SUMMARIZATION_SYSTEM_PROMPT, serializeConversation, } from "./utils.js";
+import { addPersistedFileOperations, computeFileLists, createFileOps, extractFileOpsFromMessage, isPlainRecord, SUMMARIZATION_SYSTEM_PROMPT, serializeConversation, } from "./utils.js";
 import { buildRetryPrompt, CompactionVerificationError, deterministicallyFillSummaryGaps, isCompactionSummaryStructurallyUsable, verifySummary, } from "./verification.js";
 /**
  * Extract file operations from messages and previous compaction entries.
@@ -21,15 +22,7 @@ function extractFileOperations(messages, entries, prevCompactionIndex) {
         const prevCompaction = entries[prevCompactionIndex];
         if (!prevCompaction.fromHook && prevCompaction.details) {
             // fromHook field kept for session file compatibility
-            const details = prevCompaction.details;
-            if (Array.isArray(details.readFiles)) {
-                for (const f of details.readFiles)
-                    fileOps.read.add(f);
-            }
-            if (Array.isArray(details.modifiedFiles)) {
-                for (const f of details.modifiedFiles)
-                    fileOps.edited.add(f);
-            }
+            addPersistedFileOperations(fileOps, prevCompaction.details);
         }
     }
     // Extract from tool calls in messages
@@ -94,12 +87,6 @@ function cloneVerificationReport(report) {
         failures: report.failures.map((failure) => ({ ...failure })),
     };
 }
-function isPlainRecord(value) {
-    if (!value || typeof value !== "object" || Array.isArray(value))
-        return false;
-    const prototype = Object.getPrototypeOf(value);
-    return prototype === Object.prototype || prototype === null;
-}
 function aggregateVerificationChecks(reports) {
     const checks = new Map();
     for (const report of reports) {
@@ -124,7 +111,7 @@ export const DEFAULT_COMPACTION_SETTINGS = {
     enabled: true,
     reserveTokens: 16384,
     keepRecentTokens: 20000,
-    triggerPercent: 0.7,
+    triggerPercent: 0.6,
 };
 // ============================================================================
 // Token calculation
@@ -501,13 +488,33 @@ function createSummarizationOptions(model, maxTokens, apiKey, headers, signal, t
     }
     return options;
 }
-async function completeSummarization(model, context, options, streamFn) {
+async function completeSummarization(model, context, options, streamFn, completion) {
+    if (completion)
+        return completion(model, context, options);
     if (!streamFn) {
         return completeSimple(model, context, options);
     }
     const stream = await streamFn(model, context, options);
     return stream.result();
 }
+async function completeSummarizationPrompt(promptText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, completion, usage, failureLabel = "Summarization") {
+    const response = await completeSummarization(model, {
+        systemPrompt: SUMMARIZATION_SYSTEM_PROMPT,
+        messages: [
+            {
+                role: "user",
+                content: [{ type: "text", text: promptText }],
+                timestamp: Date.now(),
+            },
+        ],
+    }, createSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel), streamFn, completion);
+    if (usage)
+        addUsage(usage, response.usage);
+    if (response.stopReason === "error") {
+        throw new Error(`${failureLabel} failed: ${response.errorMessage || "Unknown error"}`);
+    }
+    return response;
+}
 /**
  * Serialize messages to conversation text and, if a `preDigest` callback is supplied, run it
  * through that pass (a cheaper curation-model call that compresses older chunks — see
@@ -540,10 +547,10 @@ async function prepareSummarizationConversationText(currentMessages, preDigest,
  *   the same message span across multiple attempts (a verification-gate retry), compute this once
  *   via {@link prepareSummarizationConversationText} and pass it to every attempt.
  */
-export async function generateSummary(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, preDigest, factsBlock = "verification demands:\nfiles-modified-recall (must appear in ## Files):\nfiles-read-recall (must appear as exact paths in ## Files, path recall threshold applies):\nworking-set-recall (must appear in ## Working Set):\nopen-errors-recall (must appear in ## Open Problems):\nactions-recall (must appear in ## Done):\nmandatory-rules-recall (must appear in ### Mandatory Rules):\nactive-task-containment (must appear in ## Active Task):\ncancelled-work-dropped (must NOT appear outside ### Mandatory Rules):", chunked = false, precomputedConversationText) {
-    return (await generateSummaryWithUsage(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, preDigest, factsBlock, chunked, precomputedConversationText)).text;
+export async function generateSummary(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, preDigest, factsBlock = "verification demands:\nfiles-modified-recall (must appear in ## Files):\nfiles-read-recall (must appear as exact paths in ## Files, path recall threshold applies):\nworking-set-recall (must appear in ## Working Set):\nopen-errors-recall (must appear in ## Open Problems):\nactions-recall (must appear in ## Done):\nmandatory-rules-recall (must appear in ### Mandatory Rules):\nactive-task-containment (must appear in ## Active Task):\ncancelled-work-dropped (must NOT appear outside ### Mandatory Rules):", chunked = false, precomputedConversationText, completion) {
+    return (await generateSummaryWithUsage(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, preDigest, factsBlock, chunked, precomputedConversationText, completion)).text;
 }
-export async function generateSummaryWithUsage(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, preDigest, factsBlock = "verification demands:\nfiles-modified-recall (must appear in ## Files):\nfiles-read-recall (must appear as exact paths in ## Files, path recall threshold applies):\nworking-set-recall (must appear in ## Working Set):\nopen-errors-recall (must appear in ## Open Problems):\nactions-recall (must appear in ## Done):\nmandatory-rules-recall (must appear in ### Mandatory Rules):\nactive-task-containment (must appear in ## Active Task):\ncancelled-work-dropped (must NOT appear outside ### Mandatory Rules):", chunked = false, precomputedConversationText) {
+export async function generateSummaryWithUsage(currentMessages, model, reserveTokens, apiKey, headers, signal, customInstructions, previousSummary, thinkingLevel, streamFn, preDigest, factsBlock = "verification demands:\nfiles-modified-recall (must appear in ## Files):\nfiles-read-recall (must appear as exact paths in ## Files, path recall threshold applies):\nworking-set-recall (must appear in ## Working Set):\nopen-errors-recall (must appear in ## Open Problems):\nactions-recall (must appear in ## Done):\nmandatory-rules-recall (must appear in ### Mandatory Rules):\nactive-task-containment (must appear in ## Active Task):\ncancelled-work-dropped (must NOT appear outside ### Mandatory Rules):", chunked = false, precomputedConversationText, completion) {
     const usage = createEmptyUsage();
     const summaryBudget = getSummaryBudget(reserveTokens, model, factsBlock);
     const maxTokens = summaryBudget;
@@ -560,26 +567,13 @@ export async function generateSummaryWithUsage(currentMessages, model, reserveTo
         if (!chunked) {
             throw new Error("input-overflow: summarization request exceeds summarizer window");
         }
-        conversationText = await summarizeChunks(conversationText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, inputBound, previousSummary, promptSuffix, usage);
+        conversationText = await summarizeChunks(conversationText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, inputBound, previousSummary, promptSuffix, usage, completion);
     }
     const promptText = buildSummarizationPrompt(conversationText, previousSummary, promptSuffix);
     if (estimateStringTokens(promptText) > inputBound) {
         throw new Error("input-overflow: chunked summarization merge still exceeds summarizer window");
     }
-    const response = await completeSummarization(model, {
-        systemPrompt: SUMMARIZATION_SYSTEM_PROMPT,
-        messages: [
-            {
-                role: "user",
-                content: [{ type: "text", text: promptText }],
-                timestamp: Date.now(),
-            },
-        ],
-    }, createSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel), streamFn);
-    addUsage(usage, response.usage);
-    if (response.stopReason === "error") {
-        throw new Error(`Summarization failed: ${response.errorMessage || "Unknown error"}`);
-    }
+    const response = await completeSummarizationPrompt(promptText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, completion, usage);
     // A length-stopped checkpoint silently lost its tail sections — gating it as if complete
     // guarantees a verification failure. Fail loudly so the compaction ladder escalates instead.
     if (response.stopReason === "length") {
@@ -647,10 +641,10 @@ export function getChunkSummarizationTokenBudget(inputBound) {
 export function buildChunkSummarizationPrompt(chunk, index, total) {
     return `<conversation-chunk index="${index}" total="${total}">\n${chunk}\n</conversation-chunk>\n\nSummarize this chunk for a later checkpoint merge. Preserve exact file paths, commands, errors, user prohibitions, and active work. Output concise notes only.`;
 }
-async function summarizeChunks(conversationText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, inputBound, previousSummary, promptSuffix, usage) {
+async function summarizeChunks(conversationText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, inputBound, previousSummary, promptSuffix, usage, completion) {
     let reducedText = conversationText;
     for (let pass = 0; pass < 3; pass++) {
-        const summary = await summarizeChunkPass(reducedText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, inputBound, usage);
+        const summary = await summarizeChunkPass(reducedText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, inputBound, usage, completion);
         if (estimateStringTokens(buildSummarizationPrompt(summary, previousSummary, promptSuffix)) <= inputBound) {
             return summary;
         }
@@ -658,27 +652,14 @@ async function summarizeChunks(conversationText, model, maxTokens, apiKey, heade
     }
     throw new Error("input-overflow: chunked summarization merge still exceeds summarizer window");
 }
-async function summarizeChunkPass(conversationText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, inputBound, usage) {
+async function summarizeChunkPass(conversationText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, inputBound, usage, completion) {
     const maxChunkTokens = getChunkSummarizationTokenBudget(inputBound);
     const maxChunkChars = Math.max(1, maxChunkTokens * 4);
     const chunks = splitText(conversationText, maxChunkChars);
     const summaries = [];
     for (let i = 0; i < chunks.length; i++) {
         const promptText = buildChunkSummarizationPrompt(chunks[i], i + 1, chunks.length);
-        const response = await completeSummarization(model, {
-            systemPrompt: SUMMARIZATION_SYSTEM_PROMPT,
-            messages: [
-                {
-                    role: "user",
-                    content: [{ type: "text", text: promptText }],
-                    timestamp: Date.now(),
-                },
-            ],
-        }, createSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel), streamFn);
-        addUsage(usage, response.usage);
-        if (response.stopReason === "error") {
-            throw new Error(`Summarization failed: ${response.errorMessage || "Unknown error"}`);
-        }
+        const response = await completeSummarizationPrompt(promptText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, completion, usage);
         summaries.push(extractTextContent(response));
     }
     return summaries.join("\n\n");
@@ -829,7 +810,7 @@ async function generateVerifiedSummary(options) {
     const verificationGateFailures = [];
     const precomputedConversationText = await prepareSummarizationConversationText(options.messages, options.preDigest, options.signal);
     for (let attempt = 0; attempt < 2; attempt++) {
-        const generated = await generateSummaryWithUsage(options.messages, options.model, options.reserveTokens, options.apiKey, options.headers, options.signal, retryInstructions, options.previousSummary, options.thinkingLevel, options.streamFn, options.preDigest, options.factsBlock, options.chunked, precomputedConversationText);
+        const generated = await generateSummaryWithUsage(options.messages, options.model, options.reserveTokens, options.apiKey, options.headers, options.signal, retryInstructions, options.previousSummary, options.thinkingLevel, options.streamFn, options.preDigest, options.factsBlock, options.chunked, precomputedConversationText, options.completion);
         addUsage(usage, generated.usage);
         const summary = generated.text;
         const verification = verifySummary(summary, options.facts);
@@ -890,6 +871,7 @@ export async function compact(preparation, model, apiKey, headers, customInstruc
             previousSummary,
             thinkingLevel,
             streamFn,
+            completion: executionOptions?.completion,
             preDigest,
             facts,
             factsBlock,
@@ -898,7 +880,7 @@ export async function compact(preparation, model, apiKey, headers, customInstruc
     let summary = verified?.summary ?? "No prior history.";
     let summaryUsage = verified?.usage;
     if (isSplitTurn) {
-        const turnPrefix = await generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, apiKey, headers, signal, thinkingLevel, streamFn);
+        const turnPrefix = await generateTurnPrefixSummary(turnPrefixMessages, model, settings.reserveTokens, apiKey, headers, signal, thinkingLevel, streamFn, executionOptions?.completion);
         summary = `${summary}\n\n---\n\n**Turn Context (split turn):**\n\n${turnPrefix.text}`;
         summaryUsage = combineUsage(summaryUsage, turnPrefix.usage);
     }
@@ -997,22 +979,12 @@ export function createDeterministicCompaction(preparation) {
 /**
  * Generate a summary for a turn prefix (when splitting a turn).
  */
-async function generateTurnPrefixSummary(messages, model, reserveTokens, apiKey, headers, signal, thinkingLevel, streamFn) {
+async function generateTurnPrefixSummary(messages, model, reserveTokens, apiKey, headers, signal, thinkingLevel, streamFn, completion) {
     const maxTokens = Math.min(Math.floor(0.5 * reserveTokens), model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY); // Smaller budget for turn prefix
     const llmMessages = convertToLlm(messages);
     const conversationText = serializeConversation(llmMessages);
     const promptText = `<conversation>\n${conversationText}\n</conversation>\n\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;
-    const summarizationMessages = [
-        {
-            role: "user",
-            content: [{ type: "text", text: promptText }],
-            timestamp: Date.now(),
-        },
-    ];
-    const response = await completeSummarization(model, { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages }, createSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel), streamFn);
-    if (response.stopReason === "error") {
-        throw new Error(`Turn prefix summarization failed: ${response.errorMessage || "Unknown error"}`);
-    }
+    const response = await completeSummarizationPrompt(promptText, model, maxTokens, apiKey, headers, signal, thinkingLevel, streamFn, completion, undefined, "Turn prefix summarization");
     return {
         text: response.content
             .filter((c) => c.type === "text")