@illuma-ai/agents 1.1.2 → 1.1.4

package/dist/cjs/common/constants.cjs

@@ -64,6 +64,15 @@ const SUMMARIZATION_CONTEXT_THRESHOLD = 80;
  *  100% → graceful: use existing summary + recent messages, never block
  */
 const PROACTIVE_SUMMARY_THRESHOLD = 0.8;
+/**
+ * Number of recent conversation rounds (human+AI pairs) to keep in the
+ * windowed view when a summary is available. Everything older is covered
+ * by the summary. 2 rounds = last 2 user questions + 2 AI responses.
+ *
+ * This prevents wasting tokens on raw messages the summary already covers
+ * and keeps context tight for the LLM.
+ */
+const COMPACTION_RECENT_ROUNDS = 2;
 /**
  * Default reserve ratio (0-1) — fraction of context window to preserve as recent messages.
  * 0.3 means 30% of the context budget is reserved for the most recent messages,
@@ -101,6 +110,7 @@ const TOOL_DISCOVERY_CACHE_MAX_SIZE = 200;
  */
 const DEDUP_MAX_CONTENT_LENGTH = 10000;
 
+exports.COMPACTION_RECENT_ROUNDS = COMPACTION_RECENT_ROUNDS;
 exports.CONTEXT_SAFETY_BUFFER = CONTEXT_SAFETY_BUFFER;
 exports.DEDUP_MAX_CONTENT_LENGTH = DEDUP_MAX_CONTENT_LENGTH;
 exports.MIN_THINKING_BUDGET = MIN_THINKING_BUDGET;

package/dist/cjs/common/constants.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"constants.cjs","sources":["../../../src/common/constants.ts"],"sourcesContent":["// src/common/constants.ts\n\n/**\n * Minimum thinking budget allowed by the Anthropic API.\n * Extended thinking requires at least 1024 budget_tokens.\n * @see https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking\n */\nexport const MIN_THINKING_BUDGET = 1024;\n\n/**\n * Reduced thinking budget for subsequent ReAct iterations (tool-result turns).\n *\n * In a ReAct agent loop, the first LLM call processes the user's query and\n * may need deep reasoning. Subsequent iterations (after tool results return)\n * typically only need to decide \"call next tool\" or \"generate final response\"\n * — 1024 tokens is sufficient for this routing logic.\n *\n * This reduces wall-clock time per iteration from ~20-30s to ~5-10s,\n * compounding across multi-tool conversations (e.g., 10 tool calls).\n */\nexport const TOOL_TURN_THINKING_BUDGET = 1024;\n\n// ============================================================================\n// CONTEXT OVERFLOW MANAGEMENT\n//\n// Context overflow is handled mechanically — no token budget numbers are\n// exposed to the LLM. The system uses: pruning (Graph), summarization\n// (summarizeCallback), and auto-continuation (client.js max_tokens detection).\n//\n// See: docs/context-overflow-architecture.md\n// ============================================================================\n\n/**\n * Minimum number of attached documents before the multi-document delegation\n * hint is injected. Below this threshold, the agent processes documents\n * directly within its own context.\n */\nexport const MULTI_DOCUMENT_THRESHOLD = 3;\n\n/**\n * Context utilization safety buffer multiplier (0-1).\n * Applied as: effectiveMax = (maxContextTokens - maxOutputTokens) * CONTEXT_SAFETY_BUFFER\n *\n * Reserves headroom so the LLM doesn't hit hard token limits mid-generation.\n * 0.9 = 10% reserved for safety.\n */\nexport const CONTEXT_SAFETY_BUFFER = 0.9;\n\n// ============================================================================\n// SUMMARIZATION CONFIGURATION DEFAULTS\n//\n// These constants provide sensible defaults for the SummarizationConfig.\n// They can be overridden per-agent via AgentInputs.summarizationConfig.\n// ============================================================================\n\n/**\n * Default context utilization percentage (0-100) at which summarization triggers.\n * When the context window is ≥80% full, pruning + summarization activates.\n */\nexport const SUMMARIZATION_CONTEXT_THRESHOLD = 80;\n\n/**\n * Proactive summarization threshold (0-1 fraction of context window).\n * At this utilization level, background summarization fires BEFORE pruning is needed.\n * This gives the summary time to complete so it's ready when context actually fills up.\n *\n * Inspired by VS Code Copilot Chat's 3-tier strategy:\n *   80% → proactive background summary\n *   90% → pruning kicks in (with summary already cached)\n *  100% → graceful: use existing summary + recent messages, never block\n */\nexport const PROACTIVE_SUMMARY_THRESHOLD = 0.8;\n\n/**\n * Default reserve ratio (0-1) — fraction of context window to preserve as recent messages.\n * 0.3 means 30% of the context budget is reserved for the most recent messages,\n * ensuring the model always has immediate conversation history even after aggressive pruning.\n */\nexport const SUMMARIZATION_RESERVE_RATIO = 0.3;\n\n/**\n * Default EMA (Exponential Moving Average) alpha for pruning calibration.\n * Controls how quickly the calibration adapts to new token counts.\n * Higher α = faster adaptation (more responsive to recent changes).\n * Lower α = smoother adaptation (more stable across iterations).\n * 0.3 provides a balance between responsiveness and stability.\n */\nexport const PRUNING_EMA_ALPHA = 0.3;\n\n/**\n * Default initial calibration ratio for EMA pruning.\n * 1.0 means no adjustment on the first iteration (trust the raw token counts).\n * Subsequent iterations will adjust based on actual vs. estimated token usage.\n */\nexport const PRUNING_INITIAL_CALIBRATION = 1.0;\n\n// ============================================================================\n// TOOL DISCOVERY CACHING\n// ============================================================================\n\n/**\n * Maximum number of tool discovery entries to cache per conversation.\n * Prevents unbounded memory growth in very long conversations.\n */\nexport const TOOL_DISCOVERY_CACHE_MAX_SIZE = 200;\n\n// ============================================================================\n// MESSAGE DEDUPLICATION\n// ============================================================================\n\n/**\n * Maximum length of system message content to hash for deduplication.\n * Messages longer than this are always considered unique (hashing would be expensive).\n */\nexport const DEDUP_MAX_CONTENT_LENGTH = 10000;\n"],"names":[],"mappings":";;AAAA;AAEA;;;;AAIG;AACI,MAAM,mBAAmB,GAAG;AAEnC;;;;;;;;;;AAUG;AACI,MAAM,yBAAyB,GAAG;AAEzC;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AAEA;;;;AAIG;AACI,MAAM,wBAAwB,GAAG;AAExC;;;;;;AAMG;AACI,MAAM,qBAAqB,GAAG;AAErC;AACA;AACA;AACA;AACA;AACA;AAEA;;;AAGG;AACI,MAAM,+BAA+B,GAAG;AAE/C;;;;;;;;;AASG;AACI,MAAM,2BAA2B,GAAG;AAE3C;;;;AAIG;AACI,MAAM,2BAA2B,GAAG;AAE3C;;;;;;AAMG;AACI,MAAM,iBAAiB,GAAG;AAEjC;;;;AAIG;AACI,MAAM,2BAA2B,GAAG;AAE3C;AACA;AACA;AAEA;;;AAGG;AACI,MAAM,6BAA6B,GAAG;AAE7C;AACA;AACA;AAEA;;;AAGG;AACI,MAAM,wBAAwB,GAAG;;;;;;;;;;;;;;"}
+{"version":3,"file":"constants.cjs","sources":["../../../src/common/constants.ts"],"sourcesContent":["// src/common/constants.ts\n\n/**\n * Minimum thinking budget allowed by the Anthropic API.\n * Extended thinking requires at least 1024 budget_tokens.\n * @see https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking\n */\nexport const MIN_THINKING_BUDGET = 1024;\n\n/**\n * Reduced thinking budget for subsequent ReAct iterations (tool-result turns).\n *\n * In a ReAct agent loop, the first LLM call processes the user's query and\n * may need deep reasoning. Subsequent iterations (after tool results return)\n * typically only need to decide \"call next tool\" or \"generate final response\"\n * — 1024 tokens is sufficient for this routing logic.\n *\n * This reduces wall-clock time per iteration from ~20-30s to ~5-10s,\n * compounding across multi-tool conversations (e.g., 10 tool calls).\n */\nexport const TOOL_TURN_THINKING_BUDGET = 1024;\n\n// ============================================================================\n// CONTEXT OVERFLOW MANAGEMENT\n//\n// Context overflow is handled mechanically — no token budget numbers are\n// exposed to the LLM. The system uses: pruning (Graph), summarization\n// (summarizeCallback), and auto-continuation (client.js max_tokens detection).\n//\n// See: docs/context-overflow-architecture.md\n// ============================================================================\n\n/**\n * Minimum number of attached documents before the multi-document delegation\n * hint is injected. Below this threshold, the agent processes documents\n * directly within its own context.\n */\nexport const MULTI_DOCUMENT_THRESHOLD = 3;\n\n/**\n * Context utilization safety buffer multiplier (0-1).\n * Applied as: effectiveMax = (maxContextTokens - maxOutputTokens) * CONTEXT_SAFETY_BUFFER\n *\n * Reserves headroom so the LLM doesn't hit hard token limits mid-generation.\n * 0.9 = 10% reserved for safety.\n */\nexport const CONTEXT_SAFETY_BUFFER = 0.9;\n\n// ============================================================================\n// SUMMARIZATION CONFIGURATION DEFAULTS\n//\n// These constants provide sensible defaults for the SummarizationConfig.\n// They can be overridden per-agent via AgentInputs.summarizationConfig.\n// ============================================================================\n\n/**\n * Default context utilization percentage (0-100) at which summarization triggers.\n * When the context window is ≥80% full, pruning + summarization activates.\n */\nexport const SUMMARIZATION_CONTEXT_THRESHOLD = 80;\n\n/**\n * Proactive summarization threshold (0-1 fraction of context window).\n * At this utilization level, background summarization fires BEFORE pruning is needed.\n * This gives the summary time to complete so it's ready when context actually fills up.\n *\n * Inspired by VS Code Copilot Chat's 3-tier strategy:\n *   80% → proactive background summary\n *   90% → pruning kicks in (with summary already cached)\n *  100% → graceful: use existing summary + recent messages, never block\n */\nexport const PROACTIVE_SUMMARY_THRESHOLD = 0.8;\n\n/**\n * Number of recent conversation rounds (human+AI pairs) to keep in the\n * windowed view when a summary is available. Everything older is covered\n * by the summary. 2 rounds = last 2 user questions + 2 AI responses.\n *\n * This prevents wasting tokens on raw messages the summary already covers\n * and keeps context tight for the LLM.\n */\nexport const COMPACTION_RECENT_ROUNDS = 2;\n\n/**\n * Default reserve ratio (0-1) — fraction of context window to preserve as recent messages.\n * 0.3 means 30% of the context budget is reserved for the most recent messages,\n * ensuring the model always has immediate conversation history even after aggressive pruning.\n */\nexport const SUMMARIZATION_RESERVE_RATIO = 0.3;\n\n/**\n * Default EMA (Exponential Moving Average) alpha for pruning calibration.\n * Controls how quickly the calibration adapts to new token counts.\n * Higher α = faster adaptation (more responsive to recent changes).\n * Lower α = smoother adaptation (more stable across iterations).\n * 0.3 provides a balance between responsiveness and stability.\n */\nexport const PRUNING_EMA_ALPHA = 0.3;\n\n/**\n * Default initial calibration ratio for EMA pruning.\n * 1.0 means no adjustment on the first iteration (trust the raw token counts).\n * Subsequent iterations will adjust based on actual vs. estimated token usage.\n */\nexport const PRUNING_INITIAL_CALIBRATION = 1.0;\n\n// ============================================================================\n// TOOL DISCOVERY CACHING\n// ============================================================================\n\n/**\n * Maximum number of tool discovery entries to cache per conversation.\n * Prevents unbounded memory growth in very long conversations.\n */\nexport const TOOL_DISCOVERY_CACHE_MAX_SIZE = 200;\n\n// ============================================================================\n// MESSAGE DEDUPLICATION\n// ============================================================================\n\n/**\n * Maximum length of system message content to hash for deduplication.\n * Messages longer than this are always considered unique (hashing would be expensive).\n */\nexport const DEDUP_MAX_CONTENT_LENGTH = 10000;\n"],"names":[],"mappings":";;AAAA;AAEA;;;;AAIG;AACI,MAAM,mBAAmB,GAAG;AAEnC;;;;;;;;;;AAUG;AACI,MAAM,yBAAyB,GAAG;AAEzC;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AAEA;;;;AAIG;AACI,MAAM,wBAAwB,GAAG;AAExC;;;;;;AAMG;AACI,MAAM,qBAAqB,GAAG;AAErC;AACA;AACA;AACA;AACA;AACA;AAEA;;;AAGG;AACI,MAAM,+BAA+B,GAAG;AAE/C;;;;;;;;;AASG;AACI,MAAM,2BAA2B,GAAG;AAE3C;;;;;;;AAOG;AACI,MAAM,wBAAwB,GAAG;AAExC;;;;AAIG;AACI,MAAM,2BAA2B,GAAG;AAE3C;;;;;;AAMG;AACI,MAAM,iBAAiB,GAAG;AAEjC;;;;AAIG;AACI,MAAM,2BAA2B,GAAG;AAE3C;AACA;AACA;AAEA;;;AAGG;AACI,MAAM,6BAA6B,GAAG;AAE7C;AACA;AACA;AAEA;;;AAGG;AACI,MAAM,wBAAwB,GAAG;;;;;;;;;;;;;;;"}

package/dist/cjs/graphs/Graph.cjs

@@ -1124,68 +1124,167 @@ class StandardGraph extends Graph {
                 }
             }
             if (agentContext.pruneMessages) {
-                const { context, indexTokenCountMap, messagesToRefine } = agentContext.pruneMessages({
-                    messages: messages$1,
-                    usageMetadata: agentContext.currentUsage,
-                });
-                agentContext.indexTokenCountMap = indexTokenCountMap;
-                messagesToUse = context;
-                // ── Non-blocking summarization ──────────────────────────────────
-                // NEVER block the LLM call waiting for summarization. Instead:
-                //   1. If _cachedRunSummary exists → use it, fire async update
-                //   2. If persistedSummary exists → use it as fallback, fire async update
-                //   3. If NOTHING exists (first-ever prune) → skip summary, fire async generation
-                // The summary catches up asynchronously and is available for subsequent
-                // iterations (tool calls) and the next conversation turn.
+                // ── Context Compaction (Copilot-style: never delete messages) ─────
+                //
+                // DESIGN: Original messages are NEVER removed from the array.
+                // Instead, we build a "windowed view" for the LLM:
+                //   [system prompt] + [summary of older turns] + [recent turns that fit]
                 //
-                // SummarizationConfig integration:
-                //   - triggerType/triggerThreshold control WHEN summarization fires
-                //   - reserveRatio is enforced via calibrated maxTokens (above)
-                //   - initialSummary provides cross-run seeding as fallback before persistedSummary
-                let hasSummary = false;
+                // This ensures:
+                //   - No context is ever lost (summary covers older turns)
+                //   - We can always re-summarize from originals if summary is stale
+                //   - Conversation chaining works naturally across turns
+                //
+                // Flow:
+                //   1. Resolve best available summary (cached > persisted > seed)
+                //   2. Calculate token budget available for recent messages
+                //   3. Walk newest→oldest, build view of messages that fit
+                //   4. Assemble: [system] + [summary] + [recent window]
+                //   5. Fire background summary update for messages outside the window
                 const sumConfig = agentContext.summarizationConfig;
-                const shouldSummarize = this.shouldTriggerSummarization(messagesToRefine.length, agentContext.maxContextTokens ?? 0, agentContext.indexTokenCountMap, agentContext.instructionTokens, sumConfig);
-                if (messagesToRefine.length > 0 &&
-                    agentContext.summarizeCallback &&
-                    shouldSummarize) {
-                    try {
-                        let summary;
-                        let summarySource;
-                        if (this._cachedRunSummary != null) {
-                            summary = this._cachedRunSummary;
-                            summarySource = 'cached';
-                        }
-                        else if (agentContext.persistedSummary != null &&
-                            agentContext.persistedSummary !== '') {
-                            summary = agentContext.persistedSummary;
-                            this._cachedRunSummary = summary;
-                            summarySource = 'persisted';
+                const tokenCounter = agentContext.tokenCounter;
+                const maxTokens = agentContext.maxContextTokens ?? 0;
+                // Step 1: Resolve best available summary
+                let summary;
+                let summarySource;
+                if (this._cachedRunSummary != null) {
+                    summary = this._cachedRunSummary;
+                    summarySource = 'cached';
+                }
+                else if (agentContext.persistedSummary != null &&
+                    agentContext.persistedSummary !== '') {
+                    summary = agentContext.persistedSummary;
+                    this._cachedRunSummary = summary;
+                    summarySource = 'persisted';
+                }
+                else if (sumConfig?.initialSummary != null &&
+                    sumConfig.initialSummary !== '') {
+                    summary = sumConfig.initialSummary;
+                    this._cachedRunSummary = summary;
+                    summarySource = 'initial-seed';
+                }
+                else {
+                    summarySource = 'none';
+                }
+                // Step 2: Calculate token budget
+                // Apply EMA calibration for accuracy across iterations
+                const calibratedMax = pruneCalibration.applyCalibration(maxTokens, this._pruneCalibration);
+                const systemMsg = messages$1[0]?.getType() === 'system' ? messages$1[0] : null;
+                const systemTokens = systemMsg != null
+                    ? (agentContext.indexTokenCountMap[0] ?? 0)
+                    : 0;
+                const summaryMsg = summary != null && summary !== ''
+                    ? new messages.SystemMessage(`[Conversation Summary]\n${summary}`)
+                    : null;
+                const summaryTokens = summaryMsg != null && tokenCounter != null
+                    ? tokenCounter(summaryMsg)
+                    : 0;
+                // Budget for recent messages = total - system - summary - 3 (assistant priming)
+                const recentBudget = calibratedMax - systemTokens - summaryTokens - 3;
+                // Step 3: Determine window of recent messages to include.
+                //
+                // Two modes:
+                // A) No summary available → fill the budget (all messages that fit)
+                // B) Summary available → keep last 2 conversation rounds (H+A pairs)
+                //    + any trailing tool messages. The summary covers everything else.
+                //    This avoids wasting tokens on raw messages the summary already covers.
+                //
+                // A "round" = one human message + one AI response (+ any tool messages between).
+                const contentStart = systemMsg != null ? 1 : 0;
+                let usedTokens = 0;
+                let windowStart = messages$1.length; // index where the recent window begins
+                if (summary == null || summary === '') {
+                    // Mode A: No summary — include as many recent messages as fit in budget
+                    for (let i = messages$1.length - 1; i >= contentStart; i--) {
+                        const msgTokens = agentContext.indexTokenCountMap[i] ?? 0;
+                        if (usedTokens + msgTokens > recentBudget) {
+                            break;
                         }
-                        else if (sumConfig?.initialSummary != null &&
-                            sumConfig.initialSummary !== '') {
-                            // Cross-run seed: use initialSummary when no persisted summary exists
-                            summary = sumConfig.initialSummary;
-                            this._cachedRunSummary = summary;
-                            summarySource = 'initial-seed';
+                        usedTokens += msgTokens;
+                        windowStart = i;
+                    }
+                }
+                else {
+                    // Mode B: Summary exists — keep last 2 rounds (4 core messages: H+A+H+A)
+                    // Walk backward counting human messages as round boundaries.
+                    const MAX_RECENT_ROUNDS = constants.COMPACTION_RECENT_ROUNDS;
+                    let roundsSeen = 0;
+                    for (let i = messages$1.length - 1; i >= contentStart; i--) {
+                        const msgType = messages$1[i]?.getType();
+                        const msgTokens = agentContext.indexTokenCountMap[i] ?? 0;
+                        // Budget guard — even in round-limited mode, don't exceed budget
+                        if (usedTokens + msgTokens > recentBudget) {
+                            break;
                         }
-                        else {
-                            summarySource = 'none';
+                        usedTokens += msgTokens;
+                        windowStart = i;
+                        // Count a human message as a round boundary
+                        if (msgType === 'human') {
+                            roundsSeen++;
+                            if (roundsSeen >= MAX_RECENT_ROUNDS) {
+                                break;
+                            }
                         }
-                        // Single consolidated log for the entire prune+summarize decision
-                        console.debug(`[Graph:ContextMgmt] Pruned ${messages$1.length}→${context.length} msgs (${messagesToRefine.length} discarded) | summary=${summarySource}${summary ? ` (len=${summary.length})` : ''} | calibration=${this._pruneCalibration.ratio.toFixed(3)}(${this._pruneCalibration.iterations})`);
-                        // SCALE: Debounce background summarization — if a summary call is already
-                        // in-flight (from a prior tool iteration), accumulate messages instead of
-                        // firing another concurrent LLM call. At 2000 users with 3+ tool calls
-                        // per turn, this prevents 3x summary call volume.
+                    }
+                }
+                // Ensure we don't split tool-call / tool-result pairs.
+                // If windowStart lands on a ToolMessage, walk back to include its AI message.
+                while (windowStart > contentStart &&
+                    messages$1[windowStart]?.getType() === 'tool') {
+                    windowStart--;
+                    usedTokens += agentContext.indexTokenCountMap[windowStart] ?? 0;
+                }
+                const recentMessages = messages$1.slice(windowStart);
+                const compactedMessages = messages$1.slice(contentStart, windowStart);
+                const hasSummary = summaryMsg != null;
+                // Step 4: Assemble the windowed view
+                // [system] + [summary (covers compacted messages)] + [recent window]
+                const viewParts = [];
+                if (systemMsg != null) {
+                    viewParts.push(systemMsg);
+                }
+                if (summaryMsg != null) {
+                    viewParts.push(summaryMsg);
+                }
+                viewParts.push(...recentMessages);
+                messagesToUse = viewParts;
+                // Rebuild indexTokenCountMap for the windowed view so downstream
+                // analytics and summarization triggers see accurate token counts.
+                const viewTokenMap = {};
+                let viewIdx = 0;
+                if (systemMsg != null) {
+                    viewTokenMap[viewIdx] = systemTokens;
+                    viewIdx++;
+                }
+                if (summaryMsg != null) {
+                    viewTokenMap[viewIdx] = summaryTokens;
+                    viewIdx++;
+                }
+                for (let i = windowStart; i < messages$1.length; i++) {
+                    viewTokenMap[viewIdx] = agentContext.indexTokenCountMap[i];
+                    viewIdx++;
+                }
+                agentContext.indexTokenCountMap = viewTokenMap;
+                console.debug(`[Graph:Compaction] View: ${messages$1.length}→${viewParts.length} msgs ` +
+                    `(${compactedMessages.length} behind summary, ${recentMessages.length} in window) | ` +
+                    `summary=${summarySource}${summary ? ` (len=${summary.length})` : ''} | ` +
+                    `budget=${recentBudget}/${calibratedMax} used=${usedTokens}`);
+                // Step 5: Fire background summary update (non-blocking)
+                // Summarize messages outside the window so next iteration has a fresh summary.
+                // Only trigger if there are compacted messages worth summarizing.
+                if (compactedMessages.length > 0 &&
+                    agentContext.summarizeCallback) {
+                    const shouldSummarize = this.shouldTriggerSummarization(compactedMessages.length, maxTokens, agentContext.indexTokenCountMap, agentContext.instructionTokens, sumConfig);
+                    if (shouldSummarize) {
                         if (this._summaryInFlight) {
-                            this._pendingMessagesToRefine.push(...messagesToRefine);
-                            console.debug(`[Graph:ContextMgmt] Summary in-flight, queued ${messagesToRefine.length} msgs (pending=${this._pendingMessagesToRefine.length})`);
+                            this._pendingMessagesToRefine.push(...compactedMessages);
+                            console.debug(`[Graph:Compaction] Summary in-flight, queued ${compactedMessages.length} msgs (pending=${this._pendingMessagesToRefine.length})`);
                         }
                         else {
                             this._summaryInFlight = true;
                             const allMessages = this._pendingMessagesToRefine.length > 0
-                                ? [...this._pendingMessagesToRefine, ...messagesToRefine]
-                                : messagesToRefine;
+                                ? [...this._pendingMessagesToRefine, ...compactedMessages]
+                                : compactedMessages;
                             this._pendingMessagesToRefine = [];
                             agentContext
                                 .summarizeCallback(allMessages)
@@ -1195,40 +1294,17 @@ class StandardGraph extends Graph {
                                 }
                             })
                                 .catch((err) => {
-                                console.error('[Graph] Background summary failed (non-fatal):', err);
+                                console.error('[Graph:Compaction] Background summary update failed (non-fatal):', err);
                             })
                                 .finally(() => {
                                 this._summaryInFlight = false;
                             });
                         }
-                        if (summary != null && summary !== '') {
-                            hasSummary = true;
-                            const summaryMsg = new messages.SystemMessage(`[Conversation Summary]\n${summary}`);
-                            const systemIdx = messagesToUse[0]?.getType() === 'system' ? 1 : 0;
-                            messagesToUse = [
-                                ...messagesToUse.slice(0, systemIdx),
-                                summaryMsg,
-                                ...messagesToUse.slice(systemIdx),
-                            ];
-                        }
                     }
-                    catch (err) {
-                        console.error('[Graph] Summarization failed:', err);
-                    }
-                }
-                else if (messagesToRefine.length > 0) {
-                    // Log pruning even when no summarize callback (discard mode)
-                    console.debug(`[Graph:ContextMgmt] Pruned ${messages$1.length}→${context.length} msgs (${messagesToRefine.length} discarded, no summary callback) | calibration=${this._pruneCalibration.ratio.toFixed(3)}`);
                 }
-                // Deduplicate system messages that accumulate from repeated tool iterations
-                const { messages: dedupedMessages, removedCount } = dedup.deduplicateSystemMessages(messagesToUse);
-                if (removedCount > 0) {
-                    messagesToUse = dedupedMessages;
-                    console.debug(`[Graph:Dedup] Removed ${removedCount} duplicate system message(s)`);
-                }
-                // Post-prune context note for task-tool-enabled agents
-                if (messagesToRefine.length > 0 && contextPressure.hasTaskTool(agentContext.tools)) {
-                    const postPruneNote = contextPressure.buildPostPruneNote(messagesToRefine.length, hasSummary);
+                // Post-compaction context note for task-tool-enabled agents
+                if (compactedMessages.length > 0 && contextPressure.hasTaskTool(agentContext.tools)) {
+                    const postPruneNote = contextPressure.buildPostPruneNote(compactedMessages.length, hasSummary);
                     if (postPruneNote) {
                         messagesToUse = [
                             ...messagesToUse,
@@ -1237,6 +1313,14 @@ class StandardGraph extends Graph {
                     }
                 }
             }
+            // Deduplicate system messages — ALWAYS runs, not just during compaction.
+            // Duplicate system messages accumulate from repeated tool iterations,
+            // summary injections, and context notes across turns.
+            const { messages: dedupedMessages, removedCount } = dedup.deduplicateSystemMessages(messagesToUse);
+            if (removedCount > 0) {
+                messagesToUse = dedupedMessages;
+                console.debug(`[Graph:Dedup] Removed ${removedCount} duplicate system message(s)`);
+            }
             let finalMessages = messagesToUse;
             if (agentContext.useLegacyContent) {
                 finalMessages = content.formatContentStrings(finalMessages);