newmark-agent 0.3.7 → 0.3.8

package/dist/conversation-utility-host.bundle.cjs

@@ -327819,6 +327819,8 @@ function defaultConfig() {
       auto_compress: { _description: "Auto-compress history", _type: "boolean", value: true },
       compress_threshold_chars: { _description: "Compression threshold", _type: "integer", value: 8e4 },
       keep_recent_messages: { _description: "Keep recent messages", _type: "integer", value: 10 },
+      preserve_recent_messages: { _description: "dev-0.3.8 protected recent-message zone for context_history_manage (0 disables)", _type: "integer", value: 5 },
+      compression_cache_max: { _description: "dev-0.3.8 max folded-segment cache entries retained for restore/search", _type: "integer", value: 8 },
       structured_context_v2: { _description: "dev-0.3.0 structured context v2 (orchestrator + fixed order + snapshot)", _type: "boolean", value: true },
       build_history_persistence: { _description: "dev-0.3.0 append-only Build History persistence", _type: "boolean", value: true },
       branch_log_v2: { _description: "dev-0.3.0 branch long-log v2 (epoch summaries)", _type: "boolean", value: true },
@@ -335535,11 +335537,14 @@ var ToolExecutor = class {
       t3("linked_plan", "Read or update the current conversation linked Markdown plan. Update requires the current expected_revision.", { action: { type: "string", enum: ["get", "update"] }, markdown: { type: "string" }, expected_revision: { type: "number" } }, ["action"]),
       t3("build_history_query", "Read the concrete public work details of one historical Build Block. The prompt already exposes user input, final summary, and completion status; call this read-only tool only when the user asks what specifically happened. Select by newest-to-oldest history_index, or by run_id returned from an earlier query.", { history_index: { type: "number", minimum: 1, description: "1-based historical Build Block index from the request ledger; 1 is the newest previous task." }, run_id: { type: "string", description: "Exact run id returned by an earlier build_history_query result." }, max_events: { type: "number", minimum: 1, maximum: 200, description: "Maximum trailing public work events; defaults to 80." } }, []),
       t3("context_compress", "Actively compress the LLM context history for this conversation. This collapses older history entries into a concise summary while preserving the recent tail, which reduces context tokens and cost. IMPORTANT: it affects only the LLM context (what the model sees); the displayed conversation history shown to the user is never altered. Call this when the conversation is long, token pressure is high, or you judge that older turns are no longer needed in full. Idempotent and safe: repeated calls produce incremental summaries.", { keep_recent: { type: "number", minimum: 2, maximum: 60, description: "Recent message count to keep uncompressed at the tail. Defaults to the configured keep_recent_messages." }, force: { type: "boolean", description: "Compress even if the context is not yet over the automatic threshold. Defaults to false." } }, []),
-      t3("context_history_manage", "Manage the LLM context history for this conversation without affecting the displayed conversation history. This is the active context-management surface. Actions: list returns a bounded index of context entries (position, role, name, length, first-line preview); remove deletes the entry at a given position; summarize replaces a contiguous range of context entries with a concise local summary entry. The displayed conversation history (what the user sees) is never modified by any action.", {
-        action: { type: "string", enum: ["list", "remove", "summarize"], description: "list: index context entries. remove: delete the entry at position. summarize: fold entries [from, to] into one summary entry." },
+      t3("context_history_manage", "Manage the LLM context history for this conversation without affecting the displayed conversation history. This is the active context-management surface. Actions: list returns a bounded index of context entries (position, role, name, length, first-line preview); remove deletes the entry at a given position; summarize replaces a contiguous range of context entries with a concise local summary entry; restore reinserts the original messages of a folded segment from the compression cache using restore_id; search finds which cached folded segments contain a query (and the matching lines); status reports usage vs trigger/target budgets, the last compression, the compression cache, and the protected recent-message zone. The displayed conversation history (what the user sees) is never modified by any action. The recent context tail and the last user message are protected from remove/summarize unless dangerous is true.", {
+        action: { type: "string", enum: ["list", "remove", "summarize", "restore", "search", "status"], description: "list: index context entries. remove: delete the entry at position. summarize: fold entries [from, to] into one summary entry. restore: reinsert cached original messages by restore_id. search: find folded cache entries containing query. status: report context usage, budgets, cache, and protected zone." },
         position: { type: "number", minimum: 0, description: "0-based context entry index for remove, or the start of the range for summarize." },
         to: { type: "number", minimum: 0, description: "0-based inclusive end of the range for summarize. Defaults to position." },
-        limit: { type: "number", minimum: 5, maximum: 400, description: "Maximum context entries to list; defaults to 200." }
+        limit: { type: "number", minimum: 5, maximum: 400, description: "Maximum context entries to list (default 200), or search matches to return (default 20)." },
+        restore_id: { type: "string", description: "Cache id of a folded segment (from search or status) to restore into context." },
+        query: { type: "string", description: "Case-insensitive text to search for across cached folded segments and their summaries." },
+        dangerous: { type: "boolean", description: "Set true to override the protected recent-message zone and allow removing/summarizing entries that include the recent context tail or the last user message." }
       }, ["action"]),
       t3("question", "Ask user a multiple-choice question", { questions: { type: "array" } }, ["questions"]),
       t3("skill_download", "Download a skill", { name: { type: "string" }, source: { type: "string" } }, ["name", "source"]),
@@ -343418,6 +343423,8 @@ var Agent4 = class _Agent {
   continuations = [];
   activeConversationId = "default";
   lastCompression = null;
+  compressionCache = [];
+  nextCompressionCacheId = 1;
   workspaceConversations = /* @__PURE__ */ new Map();
   isSubagentRuntime = false;
   subagentName = "";
@@ -345879,6 +345886,7 @@ Review this persisted peer result and summarize or continue the parent task as n
     this.workspaceConversations.set(key3, {
       chatMessages: [...this.chatMessages],
       history: [...this.history],
+      compressionCache: [...this.compressionCache],
       plan: this.normalizeConversationPlan(this.conversationPlan),
       linkedPlan: this.normalizeLinkedPlan(this.linkedPlan),
       subagentState: this.subagents.serialize(),
@@ -345905,6 +345913,7 @@ Review this persisted peer result and summarize or continue the parent task as n
       title,
       chatMessages: [...this.chatMessages],
       history: [...this.history],
+      compressionCache: [...this.compressionCache],
       plan: this.normalizeConversationPlan(this.conversationPlan),
       linkedPlan: this.normalizeLinkedPlan(this.linkedPlan),
       subagentState: this.subagents.serialize(),
@@ -345932,6 +345941,8 @@ Review this persisted peer result and summarize or continue the parent task as n
     if (!key3) {
       this.chatMessages = [];
       this.history = [];
+      this.compressionCache = [];
+      this.nextCompressionCacheId = 1;
       this.conversationPlan = { items: [] };
       this.linkedPlan = { markdown: "", revision: 0 };
       this.workRuns = [];
@@ -345948,6 +345959,8 @@ Review this persisted peer result and summarize or continue the parent task as n
     const saved = this.workspaceConversations.get(key3);
     if (saved) {
       this.history = [...saved.history];
+      this.compressionCache = saved.compressionCache ? saved.compressionCache.map((entry) => ({ ...entry, messages: [...entry.messages] })) : [];
+      this.nextCompressionCacheId = Math.max(1, ...this.compressionCache.map((entry) => Number(entry.id.replace(/^ctx-cache-/, "")) || 0)) + 1;
       this.chatMessages = this.normalizeConversationChatMessages(saved.chatMessages, this.history);
       this.conversationPlan = this.normalizeConversationPlan(saved.plan);
       this.linkedPlan = this.normalizeLinkedPlan(saved.linkedPlan);
@@ -345967,6 +345980,8 @@ Review this persisted peer result and summarize or continue the parent task as n
     const stateKey = this.workspaceConversationStateKey();
     const persisted = stateKey && stored.conversations ? stored.conversations[stateKey] : null;
     this.history = persisted?.history ? [...persisted.history] : [];
+    this.compressionCache = persisted?.compressionCache ? persisted.compressionCache.map((entry) => ({ ...entry, messages: [...entry.messages] })) : [];
+    this.nextCompressionCacheId = Math.max(1, ...this.compressionCache.map((entry) => Number(entry.id.replace(/^ctx-cache-/, "")) || 0)) + 1;
     this.chatMessages = this.normalizeConversationChatMessages(persisted?.chatMessages || [], this.history);
     this.conversationPlan = this.normalizeConversationPlan(persisted?.plan);
     this.linkedPlan = this.normalizeLinkedPlan(persisted?.linkedPlan);
@@ -345984,6 +345999,7 @@ Review this persisted peer result and summarize or continue the parent task as n
     this.workspaceConversations.set(key3, {
       chatMessages: [...this.chatMessages],
       history: [...this.history],
+      compressionCache: [...this.compressionCache],
       plan: this.normalizeConversationPlan(this.conversationPlan),
       linkedPlan: this.normalizeLinkedPlan(this.linkedPlan),
       subagentState: this.subagents.serialize(),
@@ -346210,7 +346226,7 @@ Review this persisted peer result and summarize or continue the parent task as n
     } catch {
     }
     const action = String(input.action || "").trim();
-    if (!action) return { ok: false, output: "[context_history_manage] action is required (list|remove|summarize).", error: "action is required." };
+    if (!action) return { ok: false, output: "[context_history_manage] action is required (list|remove|summarize|restore|search|status).", error: "action is required." };
     if (action === "list") {
       const limit = Math.max(5, Math.min(400, Math.floor(Number(input.limit || 200))));
       const entries = this.history.slice(0, limit).map((message, index) => ({
@@ -346233,11 +346249,19 @@ Review this persisted peer result and summarize or continue the parent task as n
         metadata: { kind: "context-history-list" }
       };
     }
+    const protectedZone = this.contextHistoryProtectedZone();
     if (action === "remove") {
       const position = Math.floor(Number(input.position));
       if (!Number.isFinite(position) || position < 0 || position >= this.history.length) {
         return { ok: false, output: `[context_history_manage] remove position ${position} out of range (0..${this.history.length - 1}).`, error: "remove position out of range." };
       }
+      if (protectedZone.has(position) && !Boolean(input.dangerous)) {
+        return {
+          ok: false,
+          output: `[context_history_manage] remove position ${position} is protected (recent context tail or the last user message). Pass dangerous: true to override.`,
+          error: "remove position is in the protected context zone."
+        };
+      }
       const removed = this.history.splice(position, 1)[0];
       this.saveWorkspaceConversationState(true);
       return {
@@ -346259,6 +346283,14 @@ Review this persisted peer result and summarize or continue the parent task as n
       const to = Number.isFinite(toRaw) ? Math.min(this.history.length - 1, Math.max(from, toRaw)) : from;
       if (from >= this.history.length) return { ok: false, output: `[context_history_manage] summarize position ${from} out of range (0..${this.history.length - 1}).`, error: "summarize position out of range." };
       if (to - from < 1) return { ok: false, output: "[context_history_manage] summarize requires at least two entries in range.", error: "summarize requires a range of at least two entries." };
+      const protectedHit = this.history.slice(from, to + 1).some((_3, index) => protectedZone.has(from + index));
+      if (protectedHit && !Boolean(input.dangerous)) {
+        return {
+          ok: false,
+          output: "[context_history_manage] summarize range includes protected entries (recent context tail or the last user message). Pass dangerous: true to override.",
+          error: "summarize range overlaps the protected context zone."
+        };
+      }
       const segment = this.history.slice(from, to + 1);
       const chars = segment.reduce((sum, message) => sum + (typeof message.content === "string" ? message.content.length : JSON.stringify(message.content || "").length), 0);
       const summary = this.localCompressionSummary(
@@ -346272,6 +346304,8 @@ ${this.compressionHistoryContent(message.content || "")}`).join("\n\n").slice(0,
       const replacement = { role: "system", content: `[Context History Summary]
 ${summary}` };
       this.history.splice(from, to - from + 1, replacement);
+      this.pushCompressionCacheEntry(`[Context History Summary]
+${summary}`, segment, "local-summarize", true);
       this.saveWorkspaceConversationState(true);
       return {
         ok: true,
@@ -346289,6 +346323,111 @@ ${summary}` };
         metadata: { kind: "context-history-summarize" }
       };
     }
+    if (action === "restore") {
+      const restoreId = String(input.restore_id || "").trim();
+      const entry = this.compressionCache.find((item) => item.id === restoreId);
+      if (!entry) return { ok: false, output: `[context_history_manage] restore unknown restore_id: ${restoreId}.`, error: "restore_id not found." };
+      const summaryHeader = entry.summary.startsWith("[Context Compression") ? "[Context Compression" : "[Context History Summary]";
+      const markerIndex = this.history.findIndex((message) => String(message.role || "") === "system" && String(message.content || "").includes(summaryHeader) && String(message.content || "").includes(entry.summary.slice(0, 200)));
+      if (markerIndex < 0) {
+        return { ok: false, output: "[context_history_manage] restore failed: the folded summary is no longer present in context history (already re-folded or removed).", error: "restore target summary not found in history." };
+      }
+      this.history.splice(markerIndex, 1, ...entry.messages.map((message) => ({ ...message })));
+      this.compressionCache = this.compressionCache.filter((item) => item.id !== entry.id);
+      this.saveWorkspaceConversationState(true);
+      return {
+        ok: true,
+        output: JSON.stringify({
+          ok: true,
+          action: "restore",
+          restoreId: entry.id,
+          restoredEntries: entry.messages.length,
+          restoredChars: entry.foldedChars,
+          cacheRemaining: this.compressionCache.length,
+          displayHistory: { untouched: true, messageCount: this.chatMessages.length }
+        }, null, 2),
+        metadata: { kind: "context-history-restore" }
+      };
+    }
+    if (action === "search") {
+      const query = String(input.query || "").trim().toLowerCase();
+      const limit = Math.max(1, Math.min(200, Math.floor(Number(input.limit || 20))));
+      if (!query) return { ok: false, output: "[context_history_manage] search requires query.", error: "search requires query." };
+      const matches = [];
+      for (const entry of this.compressionCache) {
+        if (matches.length >= limit) break;
+        const hit = { cacheId: entry.id, at: entry.at, summary: entry.summary.slice(0, 500), matches: [] };
+        if (entry.summary.toLowerCase().includes(query)) {
+          hit.matches.push({ index: -1, snippet: this.snippetAround(entry.summary, query) });
+        }
+        entry.messages.forEach((message, index) => {
+          if (matches.length >= limit || hit.matches.length >= 40) return;
+          const content = this.compressionHistoryContent(message.content || message.reasoning_content || "");
+          if (content.toLowerCase().includes(query)) {
+            hit.matches.push({ index, snippet: this.snippetAround(content, query) });
+          }
+        });
+        if (hit.matches.length) matches.push(hit);
+      }
+      return {
+        ok: true,
+        output: JSON.stringify({
+          ok: true,
+          action: "search",
+          query: String(input.query || ""),
+          cacheEntries: this.compressionCache.length,
+          matches,
+          displayHistory: { untouched: true, messageCount: this.chatMessages.length }
+        }, null, 2),
+        metadata: { kind: "context-history-search" }
+      };
+    }
+    if (action === "status") {
+      const budget = this.compressionBudget(this.history);
+      const estimatedTokens = this.estimateContextTokens(this.history);
+      const maxTokens = this.contextMaxTokens();
+      const protectedStartIndex = this.contextHistoryProtectedStartIndex();
+      const lastUserIndex = this.history.map((message) => String(message.role || "")).lastIndexOf("user");
+      return {
+        ok: true,
+        output: JSON.stringify({
+          ok: true,
+          action: "status",
+          historyLength: this.history.length,
+          chatMessages: this.chatMessages.length,
+          estimatedTokens,
+          maxTokens,
+          triggerTokens: budget.triggerTokens,
+          targetTokens: budget.targetTokens,
+          summaryTokens: budget.summaryTokens,
+          usagePercent: maxTokens > 0 ? Math.round(estimatedTokens / maxTokens * 1e3) / 10 : 0,
+          thresholdReached: budget.triggerTokens > 0 && estimatedTokens >= budget.triggerTokens,
+          keepRecentMessages: this.config.getNum("context", "keep_recent_messages") || 10,
+          lastCompression: this.lastCompression ? {
+            at: this.lastCompression.at,
+            originalMessages: this.lastCompression.originalMessages,
+            compressedMessages: this.lastCompression.compressedMessages,
+            compressedTokens: this.lastCompression.compressedTokens,
+            model: this.lastCompression.model,
+            fallback: this.lastCompression.fallback
+          } : null,
+          cache: {
+            entries: this.compressionCache.length,
+            totalFoldedEntries: this.compressionCache.reduce((sum, item) => sum + item.foldedEntries, 0),
+            totalFoldedChars: this.compressionCache.reduce((sum, item) => sum + item.foldedChars, 0),
+            ids: this.compressionCache.map((item) => item.id)
+          },
+          protectedZone: {
+            preserveRecentMessages: this.config.getNum("context", "preserve_recent_messages") || 5,
+            protectedStartIndex,
+            lastUserMessageIndex: lastUserIndex,
+            protectedCount: protectedStartIndex >= 0 ? this.history.length - protectedStartIndex : 0
+          },
+          displayHistory: { untouched: true, messageCount: this.chatMessages.length }
+        }, null, 2),
+        metadata: { kind: "context-history-status" }
+      };
+    }
     return { ok: false, output: `[context_history_manage] Unknown action: ${action}`, error: `Unknown action: ${action}` };
   }
   recordContextCompressionStep() {
@@ -346689,6 +346828,7 @@ ${summary}` };
           model: fallbackUsed ? "model-switch-segmented-with-fallback" : modelName,
           fallback: fallbackUsed
         };
+        this.pushCompressionCacheEntry(summary2, originalMessages.slice(0, recentStart), "model-switch", fallbackUsed);
         this.persistCompressedHistory(summary2, recent.length, candidate2);
         this.saveWorkspaceConversationState(true);
         return { compressed: true, rounds, segments, droppedMessages: 0, estimatedTokens: this.estimateContextTokens(candidate2), maxTokens };
@@ -346722,6 +346862,7 @@ ${summary}` };
       model: fallbackUsed ? "model-switch-segmented-with-fallback" : modelName,
       fallback: fallbackUsed || droppedMessages > 0
     };
+    this.pushCompressionCacheEntry(summary, originalMessages.slice(0, recentStart), "model-switch", fallbackUsed || droppedMessages > 0);
     this.persistCompressedHistory(summary, recent.length, candidate);
     this.saveWorkspaceConversationState(true);
     return { compressed: true, rounds, segments, droppedMessages, estimatedTokens: this.estimateContextTokens(candidate), maxTokens };
@@ -349002,6 +349143,7 @@ Falling back to built-in engine.` }];
       model: compression.model,
       fallback: compression.fallback
     };
+    this.pushCompressionCacheEntry(compression.summary, middle, compression.model, compression.fallback);
     this.persistCompressedHistory(compression.summary, recent.length, msgs);
   }
   async buildCompressionSummary(middle, totalChars, budget, provider, signal, compressionModel, currentInstruction = "") {
@@ -349150,6 +349292,48 @@ ${text.slice(-tailChars).trimStart()}`;
     }
     if (this.isSubagentRuntime) this.subagentContextPersist?.(this.history.map((message) => ({ ...message })), this.lastCompression);
   }
+  pushCompressionCacheEntry(summary, messages, model, fallback) {
+    if (!messages.length) return;
+    const foldedChars = messages.reduce((sum, message) => sum + (typeof message.content === "string" ? message.content.length : JSON.stringify(message.content || "").length), 0);
+    this.compressionCache.push({
+      id: `ctx-cache-${this.nextCompressionCacheId}`,
+      at: (/* @__PURE__ */ new Date()).toISOString(),
+      summary,
+      messages: messages.map((message) => ({ ...message })),
+      foldedEntries: messages.length,
+      foldedChars,
+      model,
+      fallback
+    });
+    this.nextCompressionCacheId += 1;
+    const maxEntries = Math.max(0, Math.floor(this.config.getNum("context", "compression_cache_max") || 8));
+    if (this.compressionCache.length > maxEntries) {
+      this.compressionCache = this.compressionCache.slice(this.compressionCache.length - maxEntries);
+    }
+    this.saveWorkspaceConversationState(true);
+  }
+  contextHistoryProtectedStartIndex() {
+    const preserve = Math.max(0, Math.floor(this.config.getNum("context", "preserve_recent_messages") || 5));
+    const lastUserIndex = this.history.map((message) => String(message.role || "")).lastIndexOf("user");
+    const candidates = [];
+    if (preserve > 0 && this.history.length > 0) candidates.push(Math.max(0, this.history.length - preserve));
+    if (lastUserIndex >= 0) candidates.push(lastUserIndex);
+    return candidates.length ? Math.min(...candidates) : -1;
+  }
+  contextHistoryProtectedZone() {
+    const start = this.contextHistoryProtectedStartIndex();
+    const zone = /* @__PURE__ */ new Set();
+    if (start >= 0) for (let i4 = start; i4 < this.history.length; i4 += 1) zone.add(i4);
+    return zone;
+  }
+  snippetAround(content, query, radius = 150) {
+    const text = String(content || "");
+    const index = text.toLowerCase().indexOf(query.toLowerCase());
+    if (index < 0) return text.slice(0, radius * 2);
+    const from = Math.max(0, index - radius);
+    const to = Math.min(text.length, index + query.length + radius);
+    return `${from > 0 ? "\u2026" : ""}${text.slice(from, to).trim()}${to < text.length ? "\u2026" : ""}`;
+  }
   buildSystemPrompt() {
     const cwd = this.workspace.current?.path || this.rootPath;
     const enabledSkills = this.skills.active();

package/dist/core/agent.d.ts

@@ -58,6 +58,16 @@ export interface AutoRouteRatingResult {
 }
 export declare const ROOT_AGENT_ACTOR_ID = "00000000-0000-4000-8000-000000000001";
 export declare function normalizeIntelligenceTier(value: unknown): 'low' | 'medium' | 'high' | 'xhigh' | 'max' | 'ultra';
+export interface CompressionCacheEntry {
+    id: string;
+    at: string;
+    summary: string;
+    messages: Array<Record<string, unknown>>;
+    foldedEntries: number;
+    foldedChars: number;
+    model: string;
+    fallback: boolean;
+}
 type ConversationModelSelection = {
     kind: 'auto';
 } | {
@@ -286,6 +296,8 @@ export declare class Agent {
         model: string;
         fallback: boolean;
     } | null;
+    private compressionCache;
+    private nextCompressionCacheId;
     private workspaceConversations;
     isSubagentRuntime: boolean;
     private subagentName;
@@ -803,6 +815,10 @@ export declare class Agent {
     private compactSummaryBody;
     private formatCompressionSummary;
     private persistCompressedHistory;
+    private pushCompressionCacheEntry;
+    private contextHistoryProtectedStartIndex;
+    private contextHistoryProtectedZone;
+    private snippetAround;
     buildSystemPrompt(): string;
     /**
      * dev-0.3.0: assemble the model-request system prompt through the Context

package/dist/core/agent.js

@@ -191,6 +191,8 @@ class Agent {
     continuations = [];
     activeConversationId = 'default';
     lastCompression = null;
+    compressionCache = [];
+    nextCompressionCacheId = 1;
     workspaceConversations = new Map();
     isSubagentRuntime = false;
     subagentName = '';
@@ -3109,6 +3111,7 @@ class Agent {
         this.workspaceConversations.set(key, {
             chatMessages: [...this.chatMessages],
             history: [...this.history],
+            compressionCache: [...this.compressionCache],
             plan: this.normalizeConversationPlan(this.conversationPlan),
             linkedPlan: this.normalizeLinkedPlan(this.linkedPlan),
             subagentState: this.subagents.serialize(),
@@ -3138,6 +3141,7 @@ class Agent {
             title,
             chatMessages: [...this.chatMessages],
             history: [...this.history],
+            compressionCache: [...this.compressionCache],
             plan: this.normalizeConversationPlan(this.conversationPlan),
             linkedPlan: this.normalizeLinkedPlan(this.linkedPlan),
             subagentState: this.subagents.serialize(),
@@ -3167,6 +3171,8 @@ class Agent {
         if (!key) {
             this.chatMessages = [];
             this.history = [];
+            this.compressionCache = [];
+            this.nextCompressionCacheId = 1;
             this.conversationPlan = { items: [] };
             this.linkedPlan = { markdown: '', revision: 0 };
             this.workRuns = [];
@@ -3183,6 +3189,8 @@ class Agent {
         const saved = this.workspaceConversations.get(key);
         if (saved) {
             this.history = [...saved.history];
+            this.compressionCache = saved.compressionCache ? saved.compressionCache.map(entry => ({ ...entry, messages: [...entry.messages] })) : [];
+            this.nextCompressionCacheId = Math.max(1, ...this.compressionCache.map(entry => Number(entry.id.replace(/^ctx-cache-/, '')) || 0)) + 1;
             this.chatMessages = this.normalizeConversationChatMessages(saved.chatMessages, this.history);
             this.conversationPlan = this.normalizeConversationPlan(saved.plan);
             this.linkedPlan = this.normalizeLinkedPlan(saved.linkedPlan);
@@ -3202,6 +3210,8 @@ class Agent {
         const stateKey = this.workspaceConversationStateKey();
         const persisted = stateKey && stored.conversations ? stored.conversations[stateKey] : null;
         this.history = persisted?.history ? [...persisted.history] : [];
+        this.compressionCache = persisted?.compressionCache ? persisted.compressionCache.map(entry => ({ ...entry, messages: [...entry.messages] })) : [];
+        this.nextCompressionCacheId = Math.max(1, ...this.compressionCache.map(entry => Number(entry.id.replace(/^ctx-cache-/, '')) || 0)) + 1;
         this.chatMessages = this.normalizeConversationChatMessages(persisted?.chatMessages || [], this.history);
         this.conversationPlan = this.normalizeConversationPlan(persisted?.plan);
         this.linkedPlan = this.normalizeLinkedPlan(persisted?.linkedPlan);
@@ -3219,6 +3229,7 @@ class Agent {
         this.workspaceConversations.set(key, {
             chatMessages: [...this.chatMessages],
             history: [...this.history],
+            compressionCache: [...this.compressionCache],
             plan: this.normalizeConversationPlan(this.conversationPlan),
             linkedPlan: this.normalizeLinkedPlan(this.linkedPlan),
             subagentState: this.subagents.serialize(),
@@ -3466,7 +3477,7 @@ class Agent {
         catch { }
         const action = String(input.action || '').trim();
         if (!action)
-            return { ok: false, output: '[context_history_manage] action is required (list|remove|summarize).', error: 'action is required.' };
+            return { ok: false, output: '[context_history_manage] action is required (list|remove|summarize|restore|search|status).', error: 'action is required.' };
         if (action === 'list') {
             const limit = Math.max(5, Math.min(400, Math.floor(Number(input.limit || 200))));
             const entries = this.history.slice(0, limit).map((message, index) => ({
@@ -3489,11 +3500,19 @@ class Agent {
                 metadata: { kind: 'context-history-list' },
             };
         }
+        const protectedZone = this.contextHistoryProtectedZone();
         if (action === 'remove') {
             const position = Math.floor(Number(input.position));
             if (!Number.isFinite(position) || position < 0 || position >= this.history.length) {
                 return { ok: false, output: `[context_history_manage] remove position ${position} out of range (0..${this.history.length - 1}).`, error: 'remove position out of range.' };
             }
+            if (protectedZone.has(position) && !Boolean(input.dangerous)) {
+                return {
+                    ok: false,
+                    output: `[context_history_manage] remove position ${position} is protected (recent context tail or the last user message). Pass dangerous: true to override.`,
+                    error: 'remove position is in the protected context zone.',
+                };
+            }
             const removed = this.history.splice(position, 1)[0];
             this.saveWorkspaceConversationState(true);
             return {
@@ -3517,11 +3536,20 @@ class Agent {
                 return { ok: false, output: `[context_history_manage] summarize position ${from} out of range (0..${this.history.length - 1}).`, error: 'summarize position out of range.' };
             if (to - from < 1)
                 return { ok: false, output: '[context_history_manage] summarize requires at least two entries in range.', error: 'summarize requires a range of at least two entries.' };
+            const protectedHit = this.history.slice(from, to + 1).some((_, index) => protectedZone.has(from + index));
+            if (protectedHit && !Boolean(input.dangerous)) {
+                return {
+                    ok: false,
+                    output: '[context_history_manage] summarize range includes protected entries (recent context tail or the last user message). Pass dangerous: true to override.',
+                    error: 'summarize range overlaps the protected context zone.',
+                };
+            }
             const segment = this.history.slice(from, to + 1);
             const chars = segment.reduce((sum, message) => sum + (typeof message.content === 'string' ? message.content.length : JSON.stringify(message.content || '').length), 0);
             const summary = this.localCompressionSummary(`Workspace: ${this.workspace.current?.path || this.rootPath}\nMode: ${this.modeName()}`, segment.map((message, i) => `#${i + 1} [${String(message.role || 'unknown')}${message.name ? ` ${String(message.name)}` : ''}]\n${this.compressionHistoryContent(message.content || '')}`).join('\n\n').slice(0, 20000), segment.length, chars);
             const replacement = { role: 'system', content: `[Context History Summary]\n${summary}` };
             this.history.splice(from, to - from + 1, replacement);
+            this.pushCompressionCacheEntry(`[Context History Summary]\n${summary}`, segment, 'local-summarize', true);
             this.saveWorkspaceConversationState(true);
             return {
                 ok: true,
@@ -3539,6 +3567,116 @@ class Agent {
                 metadata: { kind: 'context-history-summarize' },
             };
         }
+        if (action === 'restore') {
+            const restoreId = String(input.restore_id || '').trim();
+            const entry = this.compressionCache.find(item => item.id === restoreId);
+            if (!entry)
+                return { ok: false, output: `[context_history_manage] restore unknown restore_id: ${restoreId}.`, error: 'restore_id not found.' };
+            const summaryHeader = entry.summary.startsWith('[Context Compression') ? '[Context Compression' : '[Context History Summary]';
+            const markerIndex = this.history.findIndex(message => String(message.role || '') === 'system' && String(message.content || '').includes(summaryHeader) && String(message.content || '').includes(entry.summary.slice(0, 200)));
+            if (markerIndex < 0) {
+                return { ok: false, output: '[context_history_manage] restore failed: the folded summary is no longer present in context history (already re-folded or removed).', error: 'restore target summary not found in history.' };
+            }
+            this.history.splice(markerIndex, 1, ...entry.messages.map(message => ({ ...message })));
+            this.compressionCache = this.compressionCache.filter(item => item.id !== entry.id);
+            this.saveWorkspaceConversationState(true);
+            return {
+                ok: true,
+                output: JSON.stringify({
+                    ok: true,
+                    action: 'restore',
+                    restoreId: entry.id,
+                    restoredEntries: entry.messages.length,
+                    restoredChars: entry.foldedChars,
+                    cacheRemaining: this.compressionCache.length,
+                    displayHistory: { untouched: true, messageCount: this.chatMessages.length },
+                }, null, 2),
+                metadata: { kind: 'context-history-restore' },
+            };
+        }
+        if (action === 'search') {
+            const query = String(input.query || '').trim().toLowerCase();
+            const limit = Math.max(1, Math.min(200, Math.floor(Number(input.limit || 20))));
+            if (!query)
+                return { ok: false, output: '[context_history_manage] search requires query.', error: 'search requires query.' };
+            const matches = [];
+            for (const entry of this.compressionCache) {
+                if (matches.length >= limit)
+                    break;
+                const hit = { cacheId: entry.id, at: entry.at, summary: entry.summary.slice(0, 500), matches: [] };
+                if (entry.summary.toLowerCase().includes(query)) {
+                    hit.matches.push({ index: -1, snippet: this.snippetAround(entry.summary, query) });
+                }
+                entry.messages.forEach((message, index) => {
+                    if (matches.length >= limit || hit.matches.length >= 40)
+                        return;
+                    const content = this.compressionHistoryContent(message.content || message.reasoning_content || '');
+                    if (content.toLowerCase().includes(query)) {
+                        hit.matches.push({ index, snippet: this.snippetAround(content, query) });
+                    }
+                });
+                if (hit.matches.length)
+                    matches.push(hit);
+            }
+            return {
+                ok: true,
+                output: JSON.stringify({
+                    ok: true,
+                    action: 'search',
+                    query: String(input.query || ''),
+                    cacheEntries: this.compressionCache.length,
+                    matches,
+                    displayHistory: { untouched: true, messageCount: this.chatMessages.length },
+                }, null, 2),
+                metadata: { kind: 'context-history-search' },
+            };
+        }
+        if (action === 'status') {
+            const budget = this.compressionBudget(this.history);
+            const estimatedTokens = this.estimateContextTokens(this.history);
+            const maxTokens = this.contextMaxTokens();
+            const protectedStartIndex = this.contextHistoryProtectedStartIndex();
+            const lastUserIndex = this.history.map(message => String(message.role || '')).lastIndexOf('user');
+            return {
+                ok: true,
+                output: JSON.stringify({
+                    ok: true,
+                    action: 'status',
+                    historyLength: this.history.length,
+                    chatMessages: this.chatMessages.length,
+                    estimatedTokens,
+                    maxTokens,
+                    triggerTokens: budget.triggerTokens,
+                    targetTokens: budget.targetTokens,
+                    summaryTokens: budget.summaryTokens,
+                    usagePercent: maxTokens > 0 ? Math.round((estimatedTokens / maxTokens) * 1000) / 10 : 0,
+                    thresholdReached: budget.triggerTokens > 0 && estimatedTokens >= budget.triggerTokens,
+                    keepRecentMessages: this.config.getNum('context', 'keep_recent_messages') || 10,
+                    lastCompression: this.lastCompression ? {
+                        at: this.lastCompression.at,
+                        originalMessages: this.lastCompression.originalMessages,
+                        compressedMessages: this.lastCompression.compressedMessages,
+                        compressedTokens: this.lastCompression.compressedTokens,
+                        model: this.lastCompression.model,
+                        fallback: this.lastCompression.fallback,
+                    } : null,
+                    cache: {
+                        entries: this.compressionCache.length,
+                        totalFoldedEntries: this.compressionCache.reduce((sum, item) => sum + item.foldedEntries, 0),
+                        totalFoldedChars: this.compressionCache.reduce((sum, item) => sum + item.foldedChars, 0),
+                        ids: this.compressionCache.map(item => item.id),
+                    },
+                    protectedZone: {
+                        preserveRecentMessages: this.config.getNum('context', 'preserve_recent_messages') || 5,
+                        protectedStartIndex,
+                        lastUserMessageIndex: lastUserIndex,
+                        protectedCount: protectedStartIndex >= 0 ? this.history.length - protectedStartIndex : 0,
+                    },
+                    displayHistory: { untouched: true, messageCount: this.chatMessages.length },
+                }, null, 2),
+                metadata: { kind: 'context-history-status' },
+            };
+        }
         return { ok: false, output: `[context_history_manage] Unknown action: ${action}`, error: `Unknown action: ${action}` };
     }
     recordContextCompressionStep() {
@@ -3980,6 +4118,7 @@ class Agent {
                     model: fallbackUsed ? 'model-switch-segmented-with-fallback' : modelName,
                     fallback: fallbackUsed,
                 };
+                this.pushCompressionCacheEntry(summary, originalMessages.slice(0, recentStart), 'model-switch', fallbackUsed);
                 this.persistCompressedHistory(summary, recent.length, candidate);
                 this.saveWorkspaceConversationState(true);
                 return { compressed: true, rounds, segments, droppedMessages: 0, estimatedTokens: this.estimateContextTokens(candidate), maxTokens };
@@ -4016,6 +4155,7 @@ class Agent {
             model: fallbackUsed ? 'model-switch-segmented-with-fallback' : modelName,
             fallback: fallbackUsed || droppedMessages > 0,
         };
+        this.pushCompressionCacheEntry(summary, originalMessages.slice(0, recentStart), 'model-switch', fallbackUsed || droppedMessages > 0);
         this.persistCompressedHistory(summary, recent.length, candidate);
         this.saveWorkspaceConversationState(true);
         return { compressed: true, rounds, segments, droppedMessages, estimatedTokens: this.estimateContextTokens(candidate), maxTokens };
@@ -6574,6 +6714,7 @@ class Agent {
             model: compression.model,
             fallback: compression.fallback,
         };
+        this.pushCompressionCacheEntry(compression.summary, middle, compression.model, compression.fallback);
         this.persistCompressedHistory(compression.summary, recent.length, msgs);
     }
     async buildCompressionSummary(middle, totalChars, budget, provider, signal, compressionModel, currentInstruction = '') {
@@ -6731,6 +6872,56 @@ class Agent {
         if (this.isSubagentRuntime)
             this.subagentContextPersist?.(this.history.map(message => ({ ...message })), this.lastCompression);
     }
+    pushCompressionCacheEntry(summary, messages, model, fallback) {
+        if (!messages.length)
+            return;
+        const foldedChars = messages.reduce((sum, message) => sum + (typeof message.content === 'string'
+            ? message.content.length
+            : JSON.stringify(message.content || '').length), 0);
+        this.compressionCache.push({
+            id: `ctx-cache-${this.nextCompressionCacheId}`,
+            at: new Date().toISOString(),
+            summary,
+            messages: messages.map(message => ({ ...message })),
+            foldedEntries: messages.length,
+            foldedChars,
+            model,
+            fallback,
+        });
+        this.nextCompressionCacheId += 1;
+        const maxEntries = Math.max(0, Math.floor(this.config.getNum('context', 'compression_cache_max') || 8));
+        if (this.compressionCache.length > maxEntries) {
+            this.compressionCache = this.compressionCache.slice(this.compressionCache.length - maxEntries);
+        }
+        this.saveWorkspaceConversationState(true);
+    }
+    contextHistoryProtectedStartIndex() {
+        const preserve = Math.max(0, Math.floor(this.config.getNum('context', 'preserve_recent_messages') || 5));
+        const lastUserIndex = this.history.map(message => String(message.role || '')).lastIndexOf('user');
+        const candidates = [];
+        if (preserve > 0 && this.history.length > 0)
+            candidates.push(Math.max(0, this.history.length - preserve));
+        if (lastUserIndex >= 0)
+            candidates.push(lastUserIndex);
+        return candidates.length ? Math.min(...candidates) : -1;
+    }
+    contextHistoryProtectedZone() {
+        const start = this.contextHistoryProtectedStartIndex();
+        const zone = new Set();
+        if (start >= 0)
+            for (let i = start; i < this.history.length; i += 1)
+                zone.add(i);
+        return zone;
+    }
+    snippetAround(content, query, radius = 150) {
+        const text = String(content || '');
+        const index = text.toLowerCase().indexOf(query.toLowerCase());
+        if (index < 0)
+            return text.slice(0, radius * 2);
+        const from = Math.max(0, index - radius);
+        const to = Math.min(text.length, index + query.length + radius);
+        return `${from > 0 ? '…' : ''}${text.slice(from, to).trim()}${to < text.length ? '…' : ''}`;
+    }
     buildSystemPrompt() {
         const cwd = this.workspace.current?.path || this.rootPath;
         const enabledSkills = this.skills.active();

package/dist/core/config.js

@@ -911,6 +911,8 @@ function defaultConfig() {
             auto_compress: { _description: "Auto-compress history", _type: "boolean", value: true },
             compress_threshold_chars: { _description: "Compression threshold", _type: "integer", value: 80000 },
             keep_recent_messages: { _description: "Keep recent messages", _type: "integer", value: 10 },
+            preserve_recent_messages: { _description: "dev-0.3.8 protected recent-message zone for context_history_manage (0 disables)", _type: "integer", value: 5 },
+            compression_cache_max: { _description: "dev-0.3.8 max folded-segment cache entries retained for restore/search", _type: "integer", value: 8 },
             structured_context_v2: { _description: "dev-0.3.0 structured context v2 (orchestrator + fixed order + snapshot)", _type: "boolean", value: true },
             build_history_persistence: { _description: "dev-0.3.0 append-only Build History persistence", _type: "boolean", value: true },
             branch_log_v2: { _description: "dev-0.3.0 branch long-log v2 (epoch summaries)", _type: "boolean", value: true },

package/dist/tools/index.js

@@ -374,11 +374,14 @@ class ToolExecutor {
             t('linked_plan', 'Read or update the current conversation linked Markdown plan. Update requires the current expected_revision.', { action: { type: 'string', enum: ['get', 'update'] }, markdown: { type: 'string' }, expected_revision: { type: 'number' } }, ['action']),
             t('build_history_query', 'Read the concrete public work details of one historical Build Block. The prompt already exposes user input, final summary, and completion status; call this read-only tool only when the user asks what specifically happened. Select by newest-to-oldest history_index, or by run_id returned from an earlier query.', { history_index: { type: 'number', minimum: 1, description: '1-based historical Build Block index from the request ledger; 1 is the newest previous task.' }, run_id: { type: 'string', description: 'Exact run id returned by an earlier build_history_query result.' }, max_events: { type: 'number', minimum: 1, maximum: 200, description: 'Maximum trailing public work events; defaults to 80.' } }, []),
             t('context_compress', 'Actively compress the LLM context history for this conversation. This collapses older history entries into a concise summary while preserving the recent tail, which reduces context tokens and cost. IMPORTANT: it affects only the LLM context (what the model sees); the displayed conversation history shown to the user is never altered. Call this when the conversation is long, token pressure is high, or you judge that older turns are no longer needed in full. Idempotent and safe: repeated calls produce incremental summaries.', { keep_recent: { type: 'number', minimum: 2, maximum: 60, description: 'Recent message count to keep uncompressed at the tail. Defaults to the configured keep_recent_messages.' }, force: { type: 'boolean', description: 'Compress even if the context is not yet over the automatic threshold. Defaults to false.' } }, []),
-            t('context_history_manage', 'Manage the LLM context history for this conversation without affecting the displayed conversation history. This is the active context-management surface. Actions: list returns a bounded index of context entries (position, role, name, length, first-line preview); remove deletes the entry at a given position; summarize replaces a contiguous range of context entries with a concise local summary entry. The displayed conversation history (what the user sees) is never modified by any action.', {
-                action: { type: 'string', enum: ['list', 'remove', 'summarize'], description: 'list: index context entries. remove: delete the entry at position. summarize: fold entries [from, to] into one summary entry.' },
+            t('context_history_manage', 'Manage the LLM context history for this conversation without affecting the displayed conversation history. This is the active context-management surface. Actions: list returns a bounded index of context entries (position, role, name, length, first-line preview); remove deletes the entry at a given position; summarize replaces a contiguous range of context entries with a concise local summary entry; restore reinserts the original messages of a folded segment from the compression cache using restore_id; search finds which cached folded segments contain a query (and the matching lines); status reports usage vs trigger/target budgets, the last compression, the compression cache, and the protected recent-message zone. The displayed conversation history (what the user sees) is never modified by any action. The recent context tail and the last user message are protected from remove/summarize unless dangerous is true.', {
+                action: { type: 'string', enum: ['list', 'remove', 'summarize', 'restore', 'search', 'status'], description: 'list: index context entries. remove: delete the entry at position. summarize: fold entries [from, to] into one summary entry. restore: reinsert cached original messages by restore_id. search: find folded cache entries containing query. status: report context usage, budgets, cache, and protected zone.' },
                 position: { type: 'number', minimum: 0, description: '0-based context entry index for remove, or the start of the range for summarize.' },
                 to: { type: 'number', minimum: 0, description: '0-based inclusive end of the range for summarize. Defaults to position.' },
-                limit: { type: 'number', minimum: 5, maximum: 400, description: 'Maximum context entries to list; defaults to 200.' },
+                limit: { type: 'number', minimum: 5, maximum: 400, description: 'Maximum context entries to list (default 200), or search matches to return (default 20).' },
+                restore_id: { type: 'string', description: 'Cache id of a folded segment (from search or status) to restore into context.' },
+                query: { type: 'string', description: 'Case-insensitive text to search for across cached folded segments and their summaries.' },
+                dangerous: { type: 'boolean', description: 'Set true to override the protected recent-message zone and allow removing/summarizing entries that include the recent context tail or the last user message.' },
             }, ['action']),
             t('question', 'Ask user a multiple-choice question', { questions: { type: 'array' } }, ['questions']),
             t('skill_download', 'Download a skill', { name: { type: 'string' }, source: { type: 'string' } }, ['name', 'source']),

package/dist/ui/index.html

@@ -2959,6 +2959,10 @@ button:focus, select:focus, input:focus, textarea:focus, [contenteditable]:focus
 .work-review-btn { height:28px; padding:0 10px; border:1px solid var(--glass-border-2); border-radius:6px; background:transparent; color:var(--text); cursor:pointer; font:600 11px var(--font); }
 .work-review-btn:hover { background:var(--control-hover-bg); border-color:var(--border-hover); }
 .work-review-list { border-top:1px solid var(--glass-border-1); }
+.work-review.collapsed .work-review-list { display: none; }
+.work-review-head { cursor: pointer; }
+.work-review-chevron { width:8px; height:8px; flex:0 0 auto; border-right:1px solid var(--text-dim); border-bottom:1px solid var(--text-dim); transform:rotate(-45deg); transition:transform 150ms ease; }
+.work-review.collapsed .work-review-chevron { transform:rotate(45deg); }
 .work-review-file { min-height:34px; display:grid; grid-template-columns:minmax(0,1fr) auto; align-items:center; gap:12px; padding:0 12px; color:var(--text); }
 .work-review-file:hover { background:var(--review-row-hover); }
 .work-review-path { min-width:0; overflow:hidden; text-overflow:ellipsis; white-space:nowrap; font:11px var(--font-mono); }
@@ -8208,6 +8212,12 @@ function normalizeWorkReviewDiffs(diffs) {
   return Object.keys(byPath).map(function(key) { return byPath[key]; });
 }
 
+window.toggleWorkReview = function(head) {
+  var review = head && head.closest ? head.closest('.work-review') : null;
+  if (!review) return;
+  review.classList.toggle('collapsed');
+};
+
 window.toggleWorkReviewFiles = function(button) {
   var review = button && button.closest ? button.closest('.work-review') : null;
   if (!review) return;
@@ -8301,15 +8311,16 @@ function addWorkReview(diffs) {
   var added = files.reduce(function(total, file) { return total + file.added; }, 0);
   var deleted = files.reduce(function(total, file) { return total + file.deleted; }, 0);
   var review = document.createElement('div');
-  review.className = 'work-review';
+  review.className = 'work-review collapsed';
   review.setAttribute('data-files', JSON.stringify(files));
   var rows = files.map(function(file, index) {
     return '<div class="work-review-file"' + (index >= 3 ? ' style="display:none"' : '') + '><span class="work-review-path">' + esc(file.path) + '</span><span><span class="work-review-add">+' + file.added + '</span><span class="work-review-del">-' + file.deleted + '</span></span></div>';
   }).join('');
   var editedLabel = files.length === 1 ? t('review.editedOne') : t('review.editedMany').replace('{count}', files.length);
-  review.innerHTML = '<div class="work-review-head"><div class="work-review-mark">' + iconSvg('file-diff', t('review.fileChanges'), 'small') + '</div>' +
+  review.innerHTML = '<div class="work-review-head" onclick="window.toggleWorkReview(this)"><div class="work-review-mark">' + iconSvg('file-diff', t('review.fileChanges'), 'small') + '</div>' +
     '<div><div class="work-review-title">' + esc(editedLabel) + '</div><div class="work-review-stats"><span class="work-review-add">+' + added + '</span><span class="work-review-del">-' + deleted + '</span></div></div>' +
-    '<div class="work-review-actions"><button class="work-review-btn" onclick="window.openWorkReview(this)">' + esc(t('review.open')) + '</button></div></div>' +
+    '<div class="work-review-actions"><button class="work-review-btn" onclick="window.openWorkReview(this);event.stopPropagation()">' + esc(t('review.open')) + '</button>' +
+    '<span class="work-review-chevron" aria-hidden="true"></span></div></div>' +
     '<div class="work-review-list">' + rows + (files.length > 3 ? '<button class="work-review-more" onclick="window.toggleWorkReviewFiles(this)">' + esc(t('review.showMore').replace('{count}', files.length - 3)) + '</button>' : '') + '</div>';
   els['chat-area'].appendChild(review);
   autoScrollIfAtBottom();

package/dist/wsl-agent-host.bundle.cjs

@@ -327823,6 +327823,8 @@ function defaultConfig() {
       auto_compress: { _description: "Auto-compress history", _type: "boolean", value: true },
       compress_threshold_chars: { _description: "Compression threshold", _type: "integer", value: 8e4 },
       keep_recent_messages: { _description: "Keep recent messages", _type: "integer", value: 10 },
+      preserve_recent_messages: { _description: "dev-0.3.8 protected recent-message zone for context_history_manage (0 disables)", _type: "integer", value: 5 },
+      compression_cache_max: { _description: "dev-0.3.8 max folded-segment cache entries retained for restore/search", _type: "integer", value: 8 },
       structured_context_v2: { _description: "dev-0.3.0 structured context v2 (orchestrator + fixed order + snapshot)", _type: "boolean", value: true },
       build_history_persistence: { _description: "dev-0.3.0 append-only Build History persistence", _type: "boolean", value: true },
       branch_log_v2: { _description: "dev-0.3.0 branch long-log v2 (epoch summaries)", _type: "boolean", value: true },
@@ -335539,11 +335541,14 @@ var ToolExecutor = class {
       t3("linked_plan", "Read or update the current conversation linked Markdown plan. Update requires the current expected_revision.", { action: { type: "string", enum: ["get", "update"] }, markdown: { type: "string" }, expected_revision: { type: "number" } }, ["action"]),
       t3("build_history_query", "Read the concrete public work details of one historical Build Block. The prompt already exposes user input, final summary, and completion status; call this read-only tool only when the user asks what specifically happened. Select by newest-to-oldest history_index, or by run_id returned from an earlier query.", { history_index: { type: "number", minimum: 1, description: "1-based historical Build Block index from the request ledger; 1 is the newest previous task." }, run_id: { type: "string", description: "Exact run id returned by an earlier build_history_query result." }, max_events: { type: "number", minimum: 1, maximum: 200, description: "Maximum trailing public work events; defaults to 80." } }, []),
       t3("context_compress", "Actively compress the LLM context history for this conversation. This collapses older history entries into a concise summary while preserving the recent tail, which reduces context tokens and cost. IMPORTANT: it affects only the LLM context (what the model sees); the displayed conversation history shown to the user is never altered. Call this when the conversation is long, token pressure is high, or you judge that older turns are no longer needed in full. Idempotent and safe: repeated calls produce incremental summaries.", { keep_recent: { type: "number", minimum: 2, maximum: 60, description: "Recent message count to keep uncompressed at the tail. Defaults to the configured keep_recent_messages." }, force: { type: "boolean", description: "Compress even if the context is not yet over the automatic threshold. Defaults to false." } }, []),
-      t3("context_history_manage", "Manage the LLM context history for this conversation without affecting the displayed conversation history. This is the active context-management surface. Actions: list returns a bounded index of context entries (position, role, name, length, first-line preview); remove deletes the entry at a given position; summarize replaces a contiguous range of context entries with a concise local summary entry. The displayed conversation history (what the user sees) is never modified by any action.", {
-        action: { type: "string", enum: ["list", "remove", "summarize"], description: "list: index context entries. remove: delete the entry at position. summarize: fold entries [from, to] into one summary entry." },
+      t3("context_history_manage", "Manage the LLM context history for this conversation without affecting the displayed conversation history. This is the active context-management surface. Actions: list returns a bounded index of context entries (position, role, name, length, first-line preview); remove deletes the entry at a given position; summarize replaces a contiguous range of context entries with a concise local summary entry; restore reinserts the original messages of a folded segment from the compression cache using restore_id; search finds which cached folded segments contain a query (and the matching lines); status reports usage vs trigger/target budgets, the last compression, the compression cache, and the protected recent-message zone. The displayed conversation history (what the user sees) is never modified by any action. The recent context tail and the last user message are protected from remove/summarize unless dangerous is true.", {
+        action: { type: "string", enum: ["list", "remove", "summarize", "restore", "search", "status"], description: "list: index context entries. remove: delete the entry at position. summarize: fold entries [from, to] into one summary entry. restore: reinsert cached original messages by restore_id. search: find folded cache entries containing query. status: report context usage, budgets, cache, and protected zone." },
         position: { type: "number", minimum: 0, description: "0-based context entry index for remove, or the start of the range for summarize." },
         to: { type: "number", minimum: 0, description: "0-based inclusive end of the range for summarize. Defaults to position." },
-        limit: { type: "number", minimum: 5, maximum: 400, description: "Maximum context entries to list; defaults to 200." }
+        limit: { type: "number", minimum: 5, maximum: 400, description: "Maximum context entries to list (default 200), or search matches to return (default 20)." },
+        restore_id: { type: "string", description: "Cache id of a folded segment (from search or status) to restore into context." },
+        query: { type: "string", description: "Case-insensitive text to search for across cached folded segments and their summaries." },
+        dangerous: { type: "boolean", description: "Set true to override the protected recent-message zone and allow removing/summarizing entries that include the recent context tail or the last user message." }
       }, ["action"]),
       t3("question", "Ask user a multiple-choice question", { questions: { type: "array" } }, ["questions"]),
       t3("skill_download", "Download a skill", { name: { type: "string" }, source: { type: "string" } }, ["name", "source"]),
@@ -343422,6 +343427,8 @@ var Agent4 = class _Agent {
   continuations = [];
   activeConversationId = "default";
   lastCompression = null;
+  compressionCache = [];
+  nextCompressionCacheId = 1;
   workspaceConversations = /* @__PURE__ */ new Map();
   isSubagentRuntime = false;
   subagentName = "";
@@ -345883,6 +345890,7 @@ Review this persisted peer result and summarize or continue the parent task as n
     this.workspaceConversations.set(key3, {
       chatMessages: [...this.chatMessages],
       history: [...this.history],
+      compressionCache: [...this.compressionCache],
       plan: this.normalizeConversationPlan(this.conversationPlan),
       linkedPlan: this.normalizeLinkedPlan(this.linkedPlan),
       subagentState: this.subagents.serialize(),
@@ -345909,6 +345917,7 @@ Review this persisted peer result and summarize or continue the parent task as n
       title,
       chatMessages: [...this.chatMessages],
       history: [...this.history],
+      compressionCache: [...this.compressionCache],
       plan: this.normalizeConversationPlan(this.conversationPlan),
       linkedPlan: this.normalizeLinkedPlan(this.linkedPlan),
       subagentState: this.subagents.serialize(),
@@ -345936,6 +345945,8 @@ Review this persisted peer result and summarize or continue the parent task as n
     if (!key3) {
       this.chatMessages = [];
       this.history = [];
+      this.compressionCache = [];
+      this.nextCompressionCacheId = 1;
       this.conversationPlan = { items: [] };
       this.linkedPlan = { markdown: "", revision: 0 };
       this.workRuns = [];
@@ -345952,6 +345963,8 @@ Review this persisted peer result and summarize or continue the parent task as n
     const saved = this.workspaceConversations.get(key3);
     if (saved) {
       this.history = [...saved.history];
+      this.compressionCache = saved.compressionCache ? saved.compressionCache.map((entry) => ({ ...entry, messages: [...entry.messages] })) : [];
+      this.nextCompressionCacheId = Math.max(1, ...this.compressionCache.map((entry) => Number(entry.id.replace(/^ctx-cache-/, "")) || 0)) + 1;
       this.chatMessages = this.normalizeConversationChatMessages(saved.chatMessages, this.history);
       this.conversationPlan = this.normalizeConversationPlan(saved.plan);
       this.linkedPlan = this.normalizeLinkedPlan(saved.linkedPlan);
@@ -345971,6 +345984,8 @@ Review this persisted peer result and summarize or continue the parent task as n
     const stateKey = this.workspaceConversationStateKey();
     const persisted = stateKey && stored.conversations ? stored.conversations[stateKey] : null;
     this.history = persisted?.history ? [...persisted.history] : [];
+    this.compressionCache = persisted?.compressionCache ? persisted.compressionCache.map((entry) => ({ ...entry, messages: [...entry.messages] })) : [];
+    this.nextCompressionCacheId = Math.max(1, ...this.compressionCache.map((entry) => Number(entry.id.replace(/^ctx-cache-/, "")) || 0)) + 1;
     this.chatMessages = this.normalizeConversationChatMessages(persisted?.chatMessages || [], this.history);
     this.conversationPlan = this.normalizeConversationPlan(persisted?.plan);
     this.linkedPlan = this.normalizeLinkedPlan(persisted?.linkedPlan);
@@ -345988,6 +346003,7 @@ Review this persisted peer result and summarize or continue the parent task as n
     this.workspaceConversations.set(key3, {
       chatMessages: [...this.chatMessages],
       history: [...this.history],
+      compressionCache: [...this.compressionCache],
       plan: this.normalizeConversationPlan(this.conversationPlan),
       linkedPlan: this.normalizeLinkedPlan(this.linkedPlan),
       subagentState: this.subagents.serialize(),
@@ -346214,7 +346230,7 @@ Review this persisted peer result and summarize or continue the parent task as n
     } catch {
     }
     const action = String(input2.action || "").trim();
-    if (!action) return { ok: false, output: "[context_history_manage] action is required (list|remove|summarize).", error: "action is required." };
+    if (!action) return { ok: false, output: "[context_history_manage] action is required (list|remove|summarize|restore|search|status).", error: "action is required." };
     if (action === "list") {
       const limit = Math.max(5, Math.min(400, Math.floor(Number(input2.limit || 200))));
       const entries = this.history.slice(0, limit).map((message, index) => ({
@@ -346237,11 +346253,19 @@ Review this persisted peer result and summarize or continue the parent task as n
         metadata: { kind: "context-history-list" }
       };
     }
+    const protectedZone = this.contextHistoryProtectedZone();
     if (action === "remove") {
       const position = Math.floor(Number(input2.position));
       if (!Number.isFinite(position) || position < 0 || position >= this.history.length) {
         return { ok: false, output: `[context_history_manage] remove position ${position} out of range (0..${this.history.length - 1}).`, error: "remove position out of range." };
       }
+      if (protectedZone.has(position) && !Boolean(input2.dangerous)) {
+        return {
+          ok: false,
+          output: `[context_history_manage] remove position ${position} is protected (recent context tail or the last user message). Pass dangerous: true to override.`,
+          error: "remove position is in the protected context zone."
+        };
+      }
       const removed = this.history.splice(position, 1)[0];
       this.saveWorkspaceConversationState(true);
       return {
@@ -346263,6 +346287,14 @@ Review this persisted peer result and summarize or continue the parent task as n
       const to = Number.isFinite(toRaw) ? Math.min(this.history.length - 1, Math.max(from, toRaw)) : from;
       if (from >= this.history.length) return { ok: false, output: `[context_history_manage] summarize position ${from} out of range (0..${this.history.length - 1}).`, error: "summarize position out of range." };
       if (to - from < 1) return { ok: false, output: "[context_history_manage] summarize requires at least two entries in range.", error: "summarize requires a range of at least two entries." };
+      const protectedHit = this.history.slice(from, to + 1).some((_3, index) => protectedZone.has(from + index));
+      if (protectedHit && !Boolean(input2.dangerous)) {
+        return {
+          ok: false,
+          output: "[context_history_manage] summarize range includes protected entries (recent context tail or the last user message). Pass dangerous: true to override.",
+          error: "summarize range overlaps the protected context zone."
+        };
+      }
       const segment = this.history.slice(from, to + 1);
       const chars = segment.reduce((sum, message) => sum + (typeof message.content === "string" ? message.content.length : JSON.stringify(message.content || "").length), 0);
       const summary = this.localCompressionSummary(
@@ -346276,6 +346308,8 @@ ${this.compressionHistoryContent(message.content || "")}`).join("\n\n").slice(0,
       const replacement = { role: "system", content: `[Context History Summary]
 ${summary}` };
       this.history.splice(from, to - from + 1, replacement);
+      this.pushCompressionCacheEntry(`[Context History Summary]
+${summary}`, segment, "local-summarize", true);
       this.saveWorkspaceConversationState(true);
       return {
         ok: true,
@@ -346293,6 +346327,111 @@ ${summary}` };
         metadata: { kind: "context-history-summarize" }
       };
     }
+    if (action === "restore") {
+      const restoreId = String(input2.restore_id || "").trim();
+      const entry = this.compressionCache.find((item) => item.id === restoreId);
+      if (!entry) return { ok: false, output: `[context_history_manage] restore unknown restore_id: ${restoreId}.`, error: "restore_id not found." };
+      const summaryHeader = entry.summary.startsWith("[Context Compression") ? "[Context Compression" : "[Context History Summary]";
+      const markerIndex = this.history.findIndex((message) => String(message.role || "") === "system" && String(message.content || "").includes(summaryHeader) && String(message.content || "").includes(entry.summary.slice(0, 200)));
+      if (markerIndex < 0) {
+        return { ok: false, output: "[context_history_manage] restore failed: the folded summary is no longer present in context history (already re-folded or removed).", error: "restore target summary not found in history." };
+      }
+      this.history.splice(markerIndex, 1, ...entry.messages.map((message) => ({ ...message })));
+      this.compressionCache = this.compressionCache.filter((item) => item.id !== entry.id);
+      this.saveWorkspaceConversationState(true);
+      return {
+        ok: true,
+        output: JSON.stringify({
+          ok: true,
+          action: "restore",
+          restoreId: entry.id,
+          restoredEntries: entry.messages.length,
+          restoredChars: entry.foldedChars,
+          cacheRemaining: this.compressionCache.length,
+          displayHistory: { untouched: true, messageCount: this.chatMessages.length }
+        }, null, 2),
+        metadata: { kind: "context-history-restore" }
+      };
+    }
+    if (action === "search") {
+      const query = String(input2.query || "").trim().toLowerCase();
+      const limit = Math.max(1, Math.min(200, Math.floor(Number(input2.limit || 20))));
+      if (!query) return { ok: false, output: "[context_history_manage] search requires query.", error: "search requires query." };
+      const matches = [];
+      for (const entry of this.compressionCache) {
+        if (matches.length >= limit) break;
+        const hit = { cacheId: entry.id, at: entry.at, summary: entry.summary.slice(0, 500), matches: [] };
+        if (entry.summary.toLowerCase().includes(query)) {
+          hit.matches.push({ index: -1, snippet: this.snippetAround(entry.summary, query) });
+        }
+        entry.messages.forEach((message, index) => {
+          if (matches.length >= limit || hit.matches.length >= 40) return;
+          const content = this.compressionHistoryContent(message.content || message.reasoning_content || "");
+          if (content.toLowerCase().includes(query)) {
+            hit.matches.push({ index, snippet: this.snippetAround(content, query) });
+          }
+        });
+        if (hit.matches.length) matches.push(hit);
+      }
+      return {
+        ok: true,
+        output: JSON.stringify({
+          ok: true,
+          action: "search",
+          query: String(input2.query || ""),
+          cacheEntries: this.compressionCache.length,
+          matches,
+          displayHistory: { untouched: true, messageCount: this.chatMessages.length }
+        }, null, 2),
+        metadata: { kind: "context-history-search" }
+      };
+    }
+    if (action === "status") {
+      const budget = this.compressionBudget(this.history);
+      const estimatedTokens = this.estimateContextTokens(this.history);
+      const maxTokens = this.contextMaxTokens();
+      const protectedStartIndex = this.contextHistoryProtectedStartIndex();
+      const lastUserIndex = this.history.map((message) => String(message.role || "")).lastIndexOf("user");
+      return {
+        ok: true,
+        output: JSON.stringify({
+          ok: true,
+          action: "status",
+          historyLength: this.history.length,
+          chatMessages: this.chatMessages.length,
+          estimatedTokens,
+          maxTokens,
+          triggerTokens: budget.triggerTokens,
+          targetTokens: budget.targetTokens,
+          summaryTokens: budget.summaryTokens,
+          usagePercent: maxTokens > 0 ? Math.round(estimatedTokens / maxTokens * 1e3) / 10 : 0,
+          thresholdReached: budget.triggerTokens > 0 && estimatedTokens >= budget.triggerTokens,
+          keepRecentMessages: this.config.getNum("context", "keep_recent_messages") || 10,
+          lastCompression: this.lastCompression ? {
+            at: this.lastCompression.at,
+            originalMessages: this.lastCompression.originalMessages,
+            compressedMessages: this.lastCompression.compressedMessages,
+            compressedTokens: this.lastCompression.compressedTokens,
+            model: this.lastCompression.model,
+            fallback: this.lastCompression.fallback
+          } : null,
+          cache: {
+            entries: this.compressionCache.length,
+            totalFoldedEntries: this.compressionCache.reduce((sum, item) => sum + item.foldedEntries, 0),
+            totalFoldedChars: this.compressionCache.reduce((sum, item) => sum + item.foldedChars, 0),
+            ids: this.compressionCache.map((item) => item.id)
+          },
+          protectedZone: {
+            preserveRecentMessages: this.config.getNum("context", "preserve_recent_messages") || 5,
+            protectedStartIndex,
+            lastUserMessageIndex: lastUserIndex,
+            protectedCount: protectedStartIndex >= 0 ? this.history.length - protectedStartIndex : 0
+          },
+          displayHistory: { untouched: true, messageCount: this.chatMessages.length }
+        }, null, 2),
+        metadata: { kind: "context-history-status" }
+      };
+    }
     return { ok: false, output: `[context_history_manage] Unknown action: ${action}`, error: `Unknown action: ${action}` };
   }
   recordContextCompressionStep() {
@@ -346693,6 +346832,7 @@ ${summary}` };
           model: fallbackUsed ? "model-switch-segmented-with-fallback" : modelName,
           fallback: fallbackUsed
         };
+        this.pushCompressionCacheEntry(summary2, originalMessages.slice(0, recentStart), "model-switch", fallbackUsed);
         this.persistCompressedHistory(summary2, recent.length, candidate2);
         this.saveWorkspaceConversationState(true);
         return { compressed: true, rounds, segments, droppedMessages: 0, estimatedTokens: this.estimateContextTokens(candidate2), maxTokens };
@@ -346726,6 +346866,7 @@ ${summary}` };
       model: fallbackUsed ? "model-switch-segmented-with-fallback" : modelName,
       fallback: fallbackUsed || droppedMessages > 0
     };
+    this.pushCompressionCacheEntry(summary, originalMessages.slice(0, recentStart), "model-switch", fallbackUsed || droppedMessages > 0);
     this.persistCompressedHistory(summary, recent.length, candidate);
     this.saveWorkspaceConversationState(true);
     return { compressed: true, rounds, segments, droppedMessages, estimatedTokens: this.estimateContextTokens(candidate), maxTokens };
@@ -349006,6 +349147,7 @@ Falling back to built-in engine.` }];
       model: compression.model,
       fallback: compression.fallback
     };
+    this.pushCompressionCacheEntry(compression.summary, middle, compression.model, compression.fallback);
     this.persistCompressedHistory(compression.summary, recent.length, msgs);
   }
   async buildCompressionSummary(middle, totalChars, budget, provider, signal, compressionModel, currentInstruction = "") {
@@ -349154,6 +349296,48 @@ ${text.slice(-tailChars).trimStart()}`;
     }
     if (this.isSubagentRuntime) this.subagentContextPersist?.(this.history.map((message) => ({ ...message })), this.lastCompression);
   }
+  pushCompressionCacheEntry(summary, messages, model, fallback) {
+    if (!messages.length) return;
+    const foldedChars = messages.reduce((sum, message) => sum + (typeof message.content === "string" ? message.content.length : JSON.stringify(message.content || "").length), 0);
+    this.compressionCache.push({
+      id: `ctx-cache-${this.nextCompressionCacheId}`,
+      at: (/* @__PURE__ */ new Date()).toISOString(),
+      summary,
+      messages: messages.map((message) => ({ ...message })),
+      foldedEntries: messages.length,
+      foldedChars,
+      model,
+      fallback
+    });
+    this.nextCompressionCacheId += 1;
+    const maxEntries = Math.max(0, Math.floor(this.config.getNum("context", "compression_cache_max") || 8));
+    if (this.compressionCache.length > maxEntries) {
+      this.compressionCache = this.compressionCache.slice(this.compressionCache.length - maxEntries);
+    }
+    this.saveWorkspaceConversationState(true);
+  }
+  contextHistoryProtectedStartIndex() {
+    const preserve = Math.max(0, Math.floor(this.config.getNum("context", "preserve_recent_messages") || 5));
+    const lastUserIndex = this.history.map((message) => String(message.role || "")).lastIndexOf("user");
+    const candidates = [];
+    if (preserve > 0 && this.history.length > 0) candidates.push(Math.max(0, this.history.length - preserve));
+    if (lastUserIndex >= 0) candidates.push(lastUserIndex);
+    return candidates.length ? Math.min(...candidates) : -1;
+  }
+  contextHistoryProtectedZone() {
+    const start = this.contextHistoryProtectedStartIndex();
+    const zone = /* @__PURE__ */ new Set();
+    if (start >= 0) for (let i4 = start; i4 < this.history.length; i4 += 1) zone.add(i4);
+    return zone;
+  }
+  snippetAround(content, query, radius = 150) {
+    const text = String(content || "");
+    const index = text.toLowerCase().indexOf(query.toLowerCase());
+    if (index < 0) return text.slice(0, radius * 2);
+    const from = Math.max(0, index - radius);
+    const to = Math.min(text.length, index + query.length + radius);
+    return `${from > 0 ? "\u2026" : ""}${text.slice(from, to).trim()}${to < text.length ? "\u2026" : ""}`;
+  }
   buildSystemPrompt() {
     const cwd = this.workspace.current?.path || this.rootPath;
     const enabledSkills = this.skills.active();

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "newmark-agent",
   "productName": "Newmark Agent",
-  "version": "0.3.7",
+  "version": "0.3.8",
   "description": "Newmark Agent — Portable AI coding agent with rich GUI and CLI (TypeScript)",
   "homepage": "https://github.com/positer/Newmark-Agent",
   "repository": {