@poncho-ai/harness 0.59.10 → 0.59.12

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.59.10 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.59.12 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
 > node scripts/embed-docs.js && tsup src/index.ts --format esm --dts
 
 [embed-docs] Generated poncho-docs.ts with 4 topics
@@ -9,8 +9,8 @@
 CLI Target: es2022
 ESM Build start
 ESM dist/isolate-F2PPSUL6.js 53.82 KB
-ESM dist/index.js            559.89 KB
-ESM ⚡️ Build success in 238ms
+ESM dist/index.js            560.73 KB
+ESM ⚡️ Build success in 254ms
 DTS Build start
-DTS ⚡️ Build success in 7368ms
-DTS dist/index.d.ts 102.06 KB
+DTS ⚡️ Build success in 8487ms
+DTS dist/index.d.ts 102.50 KB

package/CHANGELOG.md

@@ -1,5 +1,26 @@
 # @poncho-ai/harness
 
+## 0.59.12
+
+### Patch Changes
+
+- [`4dc94a7`](https://github.com/cesr/poncho-ai/commit/4dc94a7d7d1d411baa9ca4ac1e98053d6d2e66c3) Thanks [@cesr](https://github.com/cesr)! - `defaultAgentDefinition()` accepts an optional `tagline` to override the
+  opening line's descriptor ("You are **{name}**, {tagline}."). Default is
+  unchanged ("a helpful assistant built with Poncho"). Lets SDK consumers
+  shipping a differently-branded product keep the framework name out of the
+  agent's system prompt.
+
+## 0.59.11
+
+### Patch Changes
+
+- [`7464ad0`](https://github.com/cesr/poncho-ai/commit/7464ad04448095b34c4b1cbd52da559cf6bd6024) Thanks [@cesr](https://github.com/cesr)! - The user's memory file gets its own 1-hour Anthropic cache breakpoint.
+  It previously rode the uncached dynamic system tail (with todos + time),
+  which re-wrote the memory block — typically the bulk of a new
+  conversation's one-time cache cost — on every cold prefix, despite memory
+  only changing on explicit writes. System prompt is now three tiers:
+  static (1h), memory (1h), volatile todos+time (uncached).
+
 ## 0.59.10
 
 ### Patch Changes

package/dist/index.d.ts

@@ -853,9 +853,18 @@ interface DefaultAgentDefinitionOptions {
     maxSteps?: number;
     /** Hard timeout in seconds. Default: 300. */
     timeout?: number;
+    /**
+     * The descriptor that follows the name in the opening line
+     * ("You are **{name}**, {tagline}."). Default:
+     * "a helpful assistant built with Poncho". SDK consumers shipping a
+     * differently-branded product can override this so the framework name
+     * does not leak into the agent's system prompt.
+     */
+    tagline?: string;
 }
 declare const DEFAULT_AGENT_NAME = "agent";
 declare const DEFAULT_AGENT_DESCRIPTION = "A helpful Poncho assistant";
+declare const DEFAULT_TAGLINE = "a helpful assistant built with Poncho";
 declare const DEFAULT_MODEL_PROVIDER: "anthropic";
 declare const DEFAULT_MODEL_NAME = "claude-opus-4-5";
 declare const DEFAULT_TEMPERATURE = 0.2;
@@ -2383,4 +2392,4 @@ type NewEntryNoId = NewConversationEntry extends infer T ? T extends NewConversa
  */
 declare const appendEntriesSafe: (store: ConversationStore, conversation: Pick<Conversation, "conversationId" | "ownerId" | "tenantId">, entries: NewEntryNoId[], log: Logger) => Promise<ConversationEntry[]>;
 
-export { type ActiveConversationRun, type ActiveSubagentRun, type AgentFrontmatter, AgentHarness, type AgentIdentity, type AgentLimitsConfig, type AgentModelConfig, AgentOrchestrator, type ApprovalEventItem, type ArchivedToolResult$1 as ArchivedToolResult, type BashConfig, BashEnvironmentManager, type BashExecutionLimits, type BuiltInToolToggles, CALLBACK_LOCK_STALE_MS, type CallbackStartedEntry, type CompactMessagesOptions, type CompactResult, type CompactionConfig, type ContinuationHooks, type Conversation, type ConversationCreateInit, type ConversationEntry, type ConversationState, type ConversationStatusSnapshot, type ConversationStore, type ConversationSummary, type CreateSkillToolsOptions, type CronJobConfig, DEFAULT_AGENT_DESCRIPTION, DEFAULT_AGENT_NAME, DEFAULT_MAX_STEPS, DEFAULT_MODEL_NAME, DEFAULT_MODEL_PROVIDER, DEFAULT_TEMPERATURE, DEFAULT_TIMEOUT, type DefaultAgentDefinitionOptions, type EventSink, type ExecuteTurnResult, type HarnessOptions, type HarnessRunOutput, type HistorySource, InMemoryConversationStore, InMemoryEngine, InMemoryStateStore, type IsolateBinding, type IsolateConfig, LocalMcpBridge, LocalUploadStore, MAX_CONCURRENT_SUBAGENTS, MAX_CONTINUATION_COUNT, MAX_SUBAGENT_CALLBACK_COUNT, MAX_SUBAGENT_NESTING, type MainMemory, type McpConfig, type MemoryConfig, type MemoryStore, type MessagingChannelConfig, type ModelProviderFactory, type MountProvider, type NetworkConfig, type NewConversationEntry, OPENAI_CODEX_CLIENT_ID, type OpenAICodexAuthConfig, type OpenAICodexDeviceAuthRequest, type OpenAICodexSession, type OrchestratorHooks, type OrchestratorOptions, type OtlpConfig, type OtlpOption, PONCHO_UPLOAD_SCHEME, type ParsedAgent, type PendingSubagentApproval, type PendingSubagentResult, type PendingToolCall, type PonchoConfig, PonchoFsAdapter, PostgresEngine, type ProviderConfig, type Recurrence, type RecurrenceType, type Reminder, type ReminderCreateInput, type ReminderStatus, type ReminderStore, type RemoteMcpServerConfig, type RunConversationTurnOpts, type RunConversationTurnResult, type RunOutcome, type RunRequest, type RuntimeRenderContext, S3UploadStore, STALE_SUBAGENT_THRESHOLD_MS, STORAGE_SCHEMA_VERSION, type SecretsStore, type SkillContextEntry, type SkillMetadata, type SkillSource, SqliteEngine, type StateConfig, type StateProviderName, type StateStore, type StorageConfig, type StorageEngine, type StorageFactoryOptions, type StorageProvider, type StoredApproval, type SubagentManager, type SubagentResult, type SubagentResultEntry, type SubagentSpawnResult, type SubagentSummary, type SubagentTranscript, type SubagentTranscriptMode, TOOL_RESULT_ARCHIVE_PARAM, type TelemetryConfig, TelemetryEmitter, type TenantTokenPayload, type ToolAccess, type ToolCall, ToolDispatcher, type ToolExecutionResult, type TurnDraftState, type TurnResultMetadata, type TurnSection, type UploadStore, type UploadsConfig, VFS_SCHEME, VercelBlobUploadStore, type VfsDirEntry, type VfsStat, type VirtualMount, abnormalEndResponse, appendEntriesSafe, applyTurnMetadata, buildAgentDirectoryName, buildApprovalCheckpoints, buildAssistantMetadata, buildSkillContextWindow, buildToolCompletedText, cloneSections, compactMessages, completeOpenAICodexDeviceAuth, computeNextOccurrence, createBashTool, createConversationStore, createConversationStoreFromEngine, createDefaultTools, createDeleteDirectoryTool, createDeleteTool, createEditTool, createMemoryStore, createMemoryStoreFromEngine, createMemoryTools, createModelProvider, createReminderStore, createReminderStoreFromEngine, createReminderTools, createSearchTools, createSecretsStore, createSkillTools, createStateStore, createStorageEngine, createSubagentTools, createTodoStoreFromEngine, createTurnDraftState, createUploadStore, createWriteTool, decodeFileInputData, defaultAgentDefinition, deleteOpenAICodexSession, deriveUploadKey, ensureAgentIdentity, estimateTokens, estimateTotalTokens, executeConversationTurn, findSafeSplitPoint, flushTurnDraft, generateAgentId, getAgentStoreDirectory, getModelContextWindow, getOpenAICodexAccessToken, getOpenAICodexAuthFilePath, getOpenAICodexRequiredScopes, getPendingSubagentResults, getPonchoStoreRoot, isMessageArray, jsonSchemaToZod, lastAssistantText, loadCanonicalHistory, loadPonchoConfig, loadRunHistory, loadSkillContext, loadSkillInstructions, loadSkillMetadata, loadSkillMetadataFromDirs, loadVfsSkillMetadata, mergeSkills, normalizeApprovalCheckpoint, normalizeOtlp, normalizeScriptPolicyPath, normalizeToolAccess, parseAgentFile, parseAgentMarkdown, parseSkillFrontmatter, ponchoDocsTool, readOpenAICodexSession, readSkillResource, realResponseText, recordStandardTurnEvent, renderAgentPrompt, resolveAgentIdentity, resolveCompactionConfig, resolveEnv, resolveMemoryConfig, resolveRunRequest, resolveSkillDirs, resolveStateConfig, runConversationTurn, slugifyStorageComponent, startOpenAICodexDeviceAuth, verifyTenantToken, withToolResultArchiveParam, writeOpenAICodexSession };
+export { type ActiveConversationRun, type ActiveSubagentRun, type AgentFrontmatter, AgentHarness, type AgentIdentity, type AgentLimitsConfig, type AgentModelConfig, AgentOrchestrator, type ApprovalEventItem, type ArchivedToolResult$1 as ArchivedToolResult, type BashConfig, BashEnvironmentManager, type BashExecutionLimits, type BuiltInToolToggles, CALLBACK_LOCK_STALE_MS, type CallbackStartedEntry, type CompactMessagesOptions, type CompactResult, type CompactionConfig, type ContinuationHooks, type Conversation, type ConversationCreateInit, type ConversationEntry, type ConversationState, type ConversationStatusSnapshot, type ConversationStore, type ConversationSummary, type CreateSkillToolsOptions, type CronJobConfig, DEFAULT_AGENT_DESCRIPTION, DEFAULT_AGENT_NAME, DEFAULT_MAX_STEPS, DEFAULT_MODEL_NAME, DEFAULT_MODEL_PROVIDER, DEFAULT_TAGLINE, DEFAULT_TEMPERATURE, DEFAULT_TIMEOUT, type DefaultAgentDefinitionOptions, type EventSink, type ExecuteTurnResult, type HarnessOptions, type HarnessRunOutput, type HistorySource, InMemoryConversationStore, InMemoryEngine, InMemoryStateStore, type IsolateBinding, type IsolateConfig, LocalMcpBridge, LocalUploadStore, MAX_CONCURRENT_SUBAGENTS, MAX_CONTINUATION_COUNT, MAX_SUBAGENT_CALLBACK_COUNT, MAX_SUBAGENT_NESTING, type MainMemory, type McpConfig, type MemoryConfig, type MemoryStore, type MessagingChannelConfig, type ModelProviderFactory, type MountProvider, type NetworkConfig, type NewConversationEntry, OPENAI_CODEX_CLIENT_ID, type OpenAICodexAuthConfig, type OpenAICodexDeviceAuthRequest, type OpenAICodexSession, type OrchestratorHooks, type OrchestratorOptions, type OtlpConfig, type OtlpOption, PONCHO_UPLOAD_SCHEME, type ParsedAgent, type PendingSubagentApproval, type PendingSubagentResult, type PendingToolCall, type PonchoConfig, PonchoFsAdapter, PostgresEngine, type ProviderConfig, type Recurrence, type RecurrenceType, type Reminder, type ReminderCreateInput, type ReminderStatus, type ReminderStore, type RemoteMcpServerConfig, type RunConversationTurnOpts, type RunConversationTurnResult, type RunOutcome, type RunRequest, type RuntimeRenderContext, S3UploadStore, STALE_SUBAGENT_THRESHOLD_MS, STORAGE_SCHEMA_VERSION, type SecretsStore, type SkillContextEntry, type SkillMetadata, type SkillSource, SqliteEngine, type StateConfig, type StateProviderName, type StateStore, type StorageConfig, type StorageEngine, type StorageFactoryOptions, type StorageProvider, type StoredApproval, type SubagentManager, type SubagentResult, type SubagentResultEntry, type SubagentSpawnResult, type SubagentSummary, type SubagentTranscript, type SubagentTranscriptMode, TOOL_RESULT_ARCHIVE_PARAM, type TelemetryConfig, TelemetryEmitter, type TenantTokenPayload, type ToolAccess, type ToolCall, ToolDispatcher, type ToolExecutionResult, type TurnDraftState, type TurnResultMetadata, type TurnSection, type UploadStore, type UploadsConfig, VFS_SCHEME, VercelBlobUploadStore, type VfsDirEntry, type VfsStat, type VirtualMount, abnormalEndResponse, appendEntriesSafe, applyTurnMetadata, buildAgentDirectoryName, buildApprovalCheckpoints, buildAssistantMetadata, buildSkillContextWindow, buildToolCompletedText, cloneSections, compactMessages, completeOpenAICodexDeviceAuth, computeNextOccurrence, createBashTool, createConversationStore, createConversationStoreFromEngine, createDefaultTools, createDeleteDirectoryTool, createDeleteTool, createEditTool, createMemoryStore, createMemoryStoreFromEngine, createMemoryTools, createModelProvider, createReminderStore, createReminderStoreFromEngine, createReminderTools, createSearchTools, createSecretsStore, createSkillTools, createStateStore, createStorageEngine, createSubagentTools, createTodoStoreFromEngine, createTurnDraftState, createUploadStore, createWriteTool, decodeFileInputData, defaultAgentDefinition, deleteOpenAICodexSession, deriveUploadKey, ensureAgentIdentity, estimateTokens, estimateTotalTokens, executeConversationTurn, findSafeSplitPoint, flushTurnDraft, generateAgentId, getAgentStoreDirectory, getModelContextWindow, getOpenAICodexAccessToken, getOpenAICodexAuthFilePath, getOpenAICodexRequiredScopes, getPendingSubagentResults, getPonchoStoreRoot, isMessageArray, jsonSchemaToZod, lastAssistantText, loadCanonicalHistory, loadPonchoConfig, loadRunHistory, loadSkillContext, loadSkillInstructions, loadSkillMetadata, loadSkillMetadataFromDirs, loadVfsSkillMetadata, mergeSkills, normalizeApprovalCheckpoint, normalizeOtlp, normalizeScriptPolicyPath, normalizeToolAccess, parseAgentFile, parseAgentMarkdown, parseSkillFrontmatter, ponchoDocsTool, readOpenAICodexSession, readSkillResource, realResponseText, recordStandardTurnEvent, renderAgentPrompt, resolveAgentIdentity, resolveCompactionConfig, resolveEnv, resolveMemoryConfig, resolveRunRequest, resolveSkillDirs, resolveStateConfig, runConversationTurn, slugifyStorageComponent, startOpenAICodexDeviceAuth, verifyTenantToken, withToolResultArchiveParam, writeOpenAICodexSession };

package/dist/index.js

@@ -670,6 +670,7 @@ var loadPonchoConfig = async (workingDir) => {
 import { randomBytes } from "crypto";
 var DEFAULT_AGENT_NAME = "agent";
 var DEFAULT_AGENT_DESCRIPTION = "A helpful Poncho assistant";
+var DEFAULT_TAGLINE = "a helpful assistant built with Poncho";
 var DEFAULT_MODEL_PROVIDER = "anthropic";
 var DEFAULT_MODEL_NAME = "claude-opus-4-5";
 var DEFAULT_TEMPERATURE = 0.2;
@@ -685,6 +686,7 @@ var defaultAgentDefinition = (opts = {}) => {
   temperature: ${opts.temperature}` : "";
   const maxSteps = opts.maxSteps ?? DEFAULT_MAX_STEPS;
   const timeout = opts.timeout ?? DEFAULT_TIMEOUT;
+  const tagline = opts.tagline ?? DEFAULT_TAGLINE;
   return `---
 name: ${name}
 id: ${id}
@@ -699,7 +701,7 @@ limits:
 
 # {{name}}
 
-You are **{{name}}**, a helpful assistant built with Poncho.
+You are **{{name}}**, ${tagline}.
 
 Working directory: {{runtime.workingDir}}
 Environment: {{runtime.environment}}
@@ -10720,11 +10722,11 @@ ${skillContextWindow}${browserContext}${fsContext}${isolateContext}` : `${agentP
       const timeContext = `
 
 Current UTC time (hour precision): ${hourlyTime}`;
-      const dynamicPart = `${memoryContext}${todoContext}${timeContext}`;
-      return { staticPart, dynamicPart };
+      const dynamicPart = `${todoContext}${timeContext}`;
+      return { staticPart, memoryPart: memoryContext, dynamicPart };
     };
-    let { staticPart: staticSystemPart, dynamicPart: dynamicSystemPart } = await buildSystemPromptParts();
-    let systemPrompt = `${staticSystemPart}${dynamicSystemPart}`;
+    let { staticPart: staticSystemPart, memoryPart: memorySystemPart, dynamicPart: dynamicSystemPart } = await buildSystemPromptParts();
+    let systemPrompt = `${staticSystemPart}${memorySystemPart}${dynamicSystemPart}`;
     let lastPromptFingerprint = `${this.agentFileFingerprint}
 ${this.skillFingerprint}`;
     const pushEvent = (event) => {
@@ -11203,6 +11205,18 @@ ${textContent}` };
                 anthropic: { cacheControl: { type: "ephemeral", ttl: "1h" } }
               }
             },
+            // Memory: per-user, conversation-independent, changes only on
+            // an explicit memory write — its own 1h breakpoint means a
+            // memory edit busts THIS block forward but a normal turn reads
+            // it (plus everything before it) from cache. Breakpoint budget:
+            // Anthropic allows 4; this is #2 of 3 (static, memory, tail).
+            ...memorySystemPart.length > 0 ? [{
+              role: "system",
+              content: memorySystemPart,
+              providerOptions: {
+                anthropic: { cacheControl: { type: "ephemeral", ttl: "1h" } }
+              }
+            }] : [],
             ...dynamicSystemPart.length > 0 ? [{ role: "system", content: dynamicSystemPart }] : [],
             ...cachedMessages
           ] : cachedMessages;
@@ -11844,8 +11858,8 @@ ${textContent}` };
               const currentFingerprint = `${this.agentFileFingerprint}
 ${this.skillFingerprint}`;
               if (currentFingerprint !== lastPromptFingerprint) {
-                ({ staticPart: staticSystemPart, dynamicPart: dynamicSystemPart } = await buildSystemPromptParts());
-                systemPrompt = `${staticSystemPart}${dynamicSystemPart}`;
+                ({ staticPart: staticSystemPart, memoryPart: memorySystemPart, dynamicPart: dynamicSystemPart } = await buildSystemPromptParts());
+                systemPrompt = `${staticSystemPart}${memorySystemPart}${dynamicSystemPart}`;
                 lastPromptFingerprint = currentFingerprint;
               }
             }
@@ -14644,6 +14658,7 @@ export {
   DEFAULT_MAX_STEPS,
   DEFAULT_MODEL_NAME,
   DEFAULT_MODEL_PROVIDER,
+  DEFAULT_TAGLINE,
   DEFAULT_TEMPERATURE,
   DEFAULT_TIMEOUT,
   InMemoryConversationStore,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.59.10",
+  "version": "0.59.12",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",

package/src/default-agent.ts

@@ -37,10 +37,19 @@ export interface DefaultAgentDefinitionOptions {
   maxSteps?: number;
   /** Hard timeout in seconds. Default: 300. */
   timeout?: number;
+  /**
+   * The descriptor that follows the name in the opening line
+   * ("You are **{name}**, {tagline}."). Default:
+   * "a helpful assistant built with Poncho". SDK consumers shipping a
+   * differently-branded product can override this so the framework name
+   * does not leak into the agent's system prompt.
+   */
+  tagline?: string;
 }
 
 export const DEFAULT_AGENT_NAME = "agent";
 export const DEFAULT_AGENT_DESCRIPTION = "A helpful Poncho assistant";
+export const DEFAULT_TAGLINE = "a helpful assistant built with Poncho";
 export const DEFAULT_MODEL_PROVIDER = "anthropic" as const;
 export const DEFAULT_MODEL_NAME = "claude-opus-4-5";
 export const DEFAULT_TEMPERATURE = 0.2;
@@ -66,6 +75,7 @@ export const defaultAgentDefinition = (
     opts.temperature !== undefined ? `\n  temperature: ${opts.temperature}` : "";
   const maxSteps = opts.maxSteps ?? DEFAULT_MAX_STEPS;
   const timeout = opts.timeout ?? DEFAULT_TIMEOUT;
+  const tagline = opts.tagline ?? DEFAULT_TAGLINE;
 
   return `---
 name: ${name}
@@ -81,7 +91,7 @@ limits:
 
 # {{name}}
 
-You are **{{name}}**, a helpful assistant built with Poncho.
+You are **{{name}}**, ${tagline}.
 
 Working directory: {{runtime.workingDir}}
 Environment: {{runtime.environment}}

package/src/harness.ts

@@ -2331,13 +2331,17 @@ ${typeStubs}
 Code is wrapped in an async IIFE — use \`return\` to return a value to the tool result.`;
     }
 
-    // Split the system prompt into a static portion (stable across turns
-    // and jobs within an hour, modulo MCP connect/skill author/memory edit)
-    // and a dynamic tail (memory, todos, time). The static portion gets a
-    // 1-hour Anthropic cache breakpoint downstream; the tail rides the
-    // existing 5-min message-level breakpoint. See the streamText site for
-    // the breakpoint wiring.
-    const buildSystemPromptParts = async (): Promise<{ staticPart: string; dynamicPart: string }> => {
+    // Split the system prompt into THREE cacheability tiers (see the
+    // streamText site for the breakpoint wiring):
+    //   1. staticPart — agent body + skills + runtime context. Stable across
+    //      turns, conversations, and jobs within an hour. 1h breakpoint.
+    //   2. memoryPart — the user's memory file. Per-user, shared by every
+    //      conversation, and only changes on an explicit memory write — so
+    //      it gets its own 1h breakpoint instead of riding the volatile
+    //      tail (where it busted the message-history cache for no reason).
+    //   3. dynamicPart — todos + hour-quantized time. Genuinely volatile
+    //      within a conversation; uncached, kept as small as possible.
+    const buildSystemPromptParts = async (): Promise<{ staticPart: string; memoryPart: string; dynamicPart: string }> => {
       const agentPrompt = renderCurrentAgentPrompt();
       const tenantSkills = await this.getSkillsForTenant(input.tenantId);
       const skillContextWindow = buildSkillContextWindow(tenantSkills);
@@ -2358,13 +2362,13 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
         return `${weekday} ${d.toISOString().slice(0, 13)}Z`;
       })();
       const timeContext = `\n\nCurrent UTC time (hour precision): ${hourlyTime}`;
-      const dynamicPart = `${memoryContext}${todoContext}${timeContext}`;
-      return { staticPart, dynamicPart };
+      const dynamicPart = `${todoContext}${timeContext}`;
+      return { staticPart, memoryPart: memoryContext, dynamicPart };
     };
-    let { staticPart: staticSystemPart, dynamicPart: dynamicSystemPart } =
+    let { staticPart: staticSystemPart, memoryPart: memorySystemPart, dynamicPart: dynamicSystemPart } =
       await buildSystemPromptParts();
     // Concatenated form for legacy consumers (token estimation, telemetry).
-    let systemPrompt = `${staticSystemPart}${dynamicSystemPart}`;
+    let systemPrompt = `${staticSystemPart}${memorySystemPart}${dynamicSystemPart}`;
     let lastPromptFingerprint = `${this.agentFileFingerprint}\n${this.skillFingerprint}`;
 
     const pushEvent = (event: AgentEvent): AgentEvent => {
@@ -3008,6 +3012,20 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
                   anthropic: { cacheControl: { type: "ephemeral", ttl: "1h" } },
                 },
               },
+              // Memory: per-user, conversation-independent, changes only on
+              // an explicit memory write — its own 1h breakpoint means a
+              // memory edit busts THIS block forward but a normal turn reads
+              // it (plus everything before it) from cache. Breakpoint budget:
+              // Anthropic allows 4; this is #2 of 3 (static, memory, tail).
+              ...(memorySystemPart.length > 0
+                ? [{
+                    role: "system" as const,
+                    content: memorySystemPart,
+                    providerOptions: {
+                      anthropic: { cacheControl: { type: "ephemeral", ttl: "1h" } },
+                    },
+                  }]
+                : []),
               ...(dynamicSystemPart.length > 0
                 ? [{ role: "system" as const, content: dynamicSystemPart }]
                 : []),
@@ -3794,9 +3812,9 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
             agent = this.parsedAgent as ParsedAgent;
             const currentFingerprint = `${this.agentFileFingerprint}\n${this.skillFingerprint}`;
             if (currentFingerprint !== lastPromptFingerprint) {
-              ({ staticPart: staticSystemPart, dynamicPart: dynamicSystemPart } =
+              ({ staticPart: staticSystemPart, memoryPart: memorySystemPart, dynamicPart: dynamicSystemPart } =
                 await buildSystemPromptParts());
-              systemPrompt = `${staticSystemPart}${dynamicSystemPart}`;
+              systemPrompt = `${staticSystemPart}${memorySystemPart}${dynamicSystemPart}`;
               lastPromptFingerprint = currentFingerprint;
             }
           }