@theokit/sdk 5.9.2 → 5.11.0

package/CHANGELOG.md

@@ -1,5 +1,109 @@
 # Changelog
 
+## 5.11.0
+
+### Minor Changes
+
+- [#734](https://github.com/usetheokit/theokit-sdk/pull/734) [`1f04344`](https://github.com/usetheokit/theokit-sdk/commit/1f0434491b706f7ea97f53d998fc33b300af1c33) Thanks [@usetheodev](https://github.com/usetheodev)! - Prompt caching now turns on for Anthropic models, and the cache counters they report reach the
+  turn's usage.
+  
+  For a model id that starts with `anthropic/` or `claude` (after an optional `openrouter/`,
+  `vertex/` or `bedrock/` prefix),
+  the agent loop sends the system prompt as text blocks: the configured prompt first, marked with
+  `cache_control: { type: "ephemeral" }`, then the `<active-memory>` block when memory adds one. The
+  Anthropic client sends the blocks as they are, and the OpenRouter client sends them as `content`
+  parts of the `system` message, with `cache_control` on the cacheable part. Every other provider
+  still receives one string. Bedrock ids (`anthropic.claude-…`) are not matched,
+  and their bodies do not change. Vertex Anthropic ids (`vertex/anthropic/claude-…`) are matched
+  and get the same blocks; no live Vertex response was measured for this change.
+  
+  Things a caller can observe:
+  
+  - **Billing.** Anthropic charges cache writes at 1.25 times the input price and cache reads at 0.1
+    times. A run with a single turn whose system prompt is above the model's caching minimum pays
+    the write premium and never reads the prefix back.
+  - **Order.** For these models the memory block now comes after the configured prompt instead of
+    before it. On a client that joins the blocks into one string, the text is the same and the order
+    is reversed.
+  - **What still misses the cache.** The cached prefix is the whole assembled prompt, so `<memory>`
+    facts, resolved skills or context that change between turns change block 1 and miss.
+  - **Counters.** The Anthropic streaming client reads `cache_read_input_tokens` and
+    `cache_creation_input_tokens` from `message_start`, where Anthropic reports them; a value in
+    `message_delta` still wins when it is above zero, and a zero there leaves the `message_start`
+    value in place. The non-streaming parser shared by the Bedrock and Vertex Anthropic clients now
+    returns both counters too.
+  - **Cost on the chat-completions client.** `prompt_tokens` there already includes the cached
+    slice, and the client used to report it whole as `inputTokens` beside `cacheReadTokens`, so the
+    estimated cost charged every cached token twice (about 8 times too high on a cached
+    OpenRouter Anthropic turn). `inputTokens` is now the uncached portion, as on the Anthropic
+    client, and `prompt_tokens_details.cache_write_tokens` is read into `cacheWriteTokens`. A
+    native OpenAI chat-completions run with a cache hit reports a lower `inputTokens` and a lower
+    cost than before. When a usage object carries both the nested `prompt_tokens_details` counts and the
+    Anthropic-style top-level ones, the nested count wins and a top-level count fills a nested
+    zero, the rule `normalizeUsage` already applied; the stream client used to let the top-level
+    field win.
+  - **Budget trackers.** A `BudgetTracker`'s `input` event now carries the whole prompt:
+    `inputTokens` plus `cacheReadTokens` and `cacheWriteTokens`. On the chat-completions client this
+    keeps the count a tracker saw before `inputTokens` became the uncached portion. On the Anthropic
+    clients, which always reported the uncached portion, a cached turn now counts its cached tokens
+    against `maxTokens`, and `createUsdBudgetTracker` prices them at the input rate, so a cost cap
+    trips earlier than the provider's bill would.
+  - **Input tokens on the Anthropic stream.** `input_tokens` is read from `message_start` too, so a
+    stream whose `message_delta` carries only `output_tokens` reports its uncached input instead of
+    none. A value above zero in `message_delta` still replaces it; a `null` or a zero there leaves
+    the `message_start` count.
+  - **Malformed Anthropic frames.** A `content_block_start`, `content_block_delta` or
+    `message_delta` frame that is not valid JSON now fails the turn with a retryable `NetworkError`
+    (`code: "stream_malformed_frame"`) instead of being skipped, which finished the turn with text
+    or tool input missing. Any other malformed frame is still skipped and reported on the
+    diagnostics channel. The error message and the diagnostic line name the event and the frame's
+    length, never its body, which can hold model output or tool input.
+
+## 5.10.0
+
+### Minor Changes
+
+- [#724](https://github.com/usetheokit/theokit-sdk/pull/724) [`bc2f665`](https://github.com/usetheokit/theokit-sdk/commit/bc2f665d4c5e03dda26858004ac60bd1f7d3f3bb) Thanks [@usetheodev](https://github.com/usetheodev)! - Memory can now be given a fact ceiling. Once the store is full, a new fact is refused with a
+  typed error. Before this, nothing stopped the store growing, and the one cap that did exist, 500
+  facts per dreaming sweep, left the rest out of clustering while the result still said
+  `status: "ok"`.
+  
+  Set `memory.maxFacts` on an agent to declare a ceiling. A new fact at or over it is refused with a
+  `ConfigurationError` whose code is `memory_ceiling_reached`, and nothing is written. The message
+  gives the directory, how many facts it holds, the ceiling and what to consolidate. Recording a fact
+  that is already stored counts as a second observation, not a new fact, so it is still admitted. A
+  value that is not a positive integer is refused with `invalid_memory_ceiling` by `Agent.create`
+  and `Agent.resume`, before the agent exists, and again by the writer before anything touches
+  disk. A prompt-driven memory write that hits the ceiling reports it on one diagnostic line and
+  stores nothing. With no ceiling declared, nothing changes and no extra read happens.
+  
+  The ceiling holds across processes. Every SDK process that writes one memory directory with a
+  ceiling takes a lock (`<directory>.lock`, beside the directory) around the count and the write, so
+  two processes one fact below the ceiling no longer both get in. The lock needs the optional
+  `proper-lockfile` peer. Without it the ceiling holds within one process only, and the SDK says so
+  once on its diagnostics channel. A writer that declares no ceiling, such as the Claude Code CLI,
+  takes no lock and can still add facts past it.
+  
+  A resumed agent keeps its ceiling and its `memory.directory`. The directory is now saved with the
+  agent, and a `memory` object passed to `Agent.resume` replaces only the fields it sets (a field
+  set to `undefined` keeps its persisted value). Before
+  this, `Agent.resume(id, { memory: { enabled: true } })` came back with no ceiling, and any resumed
+  agent wrote to the default store instead of the directory it was created with.
+  
+  `createInMemoryMarkdownProvider({ maxFacts })` in `@theokit/sdk-memory` applies the same ceiling
+  to the adapter's `write` and to the `memory_remember` tool. Each handle counts its own facts.
+  
+  `DreamingSweepResult` gains `factsLeftOut`, the number of deduplicated facts past the 500-fact
+  cap that a sweep did not cluster. It is reported on both dispatch paths, the `@theokit/sdk-memory`
+  peer and the SDK's own fallback. It is 0 on a skipped sweep and absent on a failed one. It is also
+  absent when an older `@theokit/sdk-memory` runs the sweep, because that version cannot count it,
+  so read absence as unknown, not as 0. `ClusterResult`, which `@theokit/sdk-memory` exports, now
+  requires `factsLeftOut` too. Code that builds its own `ClusterResult`, such as a mock of
+  `remPhase`, has to add it.
+  
+  `@theokit/sdk-memory` now needs `@theokit/sdk` `^5.10.0`, the first release that exports the
+  ceiling helpers it imports.
+
 ## 5.9.2
 
 ### Patch Changes

package/dist/{agent-WQ4LLTSY.cjs → agent-5DOR5T3E.cjs}

@@ -1,18 +1,18 @@
 'use strict';
 
-var chunkJU76WCW7_cjs = require('./chunk-JU76WCW7.cjs');
+var chunk5OHB5HJF_cjs = require('./chunk-5OHB5HJF.cjs');
 require('./chunk-J6GXS5A2.cjs');
-require('./chunk-FR6COQXZ.cjs');
+require('./chunk-N7ZIPW2N.cjs');
 require('./chunk-BUUUWQMB.cjs');
-require('./chunk-DLUCTZ56.cjs');
+require('./chunk-L3C6CCBG.cjs');
 require('./chunk-KIKKTW5O.cjs');
 require('./chunk-BV2MWEMV.cjs');
 require('./chunk-YEHQ3SIC.cjs');
 require('./chunk-GHX4P3V2.cjs');
 require('./chunk-TMKTSYDS.cjs');
-require('./chunk-VZ2JZ2EH.cjs');
-require('./chunk-RX3CFWQ2.cjs');
-require('./chunk-QY3CIMUV.cjs');
+require('./chunk-ZD66GFH5.cjs');
+require('./chunk-NQVWSSX2.cjs');
+require('./chunk-BZKF7TXX.cjs');
 require('./chunk-CQYBJOB2.cjs');
 require('./chunk-YJNYDJ2T.cjs');
 require('./chunk-WPO3VCY4.cjs');
@@ -40,7 +40,8 @@ require('./chunk-NKIQTLFA.cjs');
 require('./chunk-ZWACQM5U.cjs');
 require('./chunk-A4RAL2ER.cjs');
 require('./chunk-P22BZE7G.cjs');
-require('./chunk-OF2CLCTH.cjs');
+require('./chunk-R5YPDAKD.cjs');
+require('./chunk-ASK2OQM5.cjs');
 require('./chunk-ZF2LDKQQ.cjs');
 require('./chunk-GHHIJEFX.cjs');
 require('./chunk-JLRLCBJ4.cjs');
@@ -48,6 +49,7 @@ require('./chunk-C7GKRGKN.cjs');
 require('./chunk-TY56BKSK.cjs');
 require('./chunk-MPR2ISL2.cjs');
 require('./chunk-J5623RIV.cjs');
+require('./chunk-NWS2CS3Y.cjs');
 require('./chunk-3EE6LVWT.cjs');
 require('./chunk-SGC7LSHF.cjs');
 require('./chunk-HU666T3Q.cjs');
@@ -63,7 +65,7 @@ require('./chunk-6LHQPOMI.cjs');
 
 Object.defineProperty(exports, "Agent", {
   enumerable: true,
-  get: function () { return chunkJU76WCW7_cjs.Agent; }
+  get: function () { return chunk5OHB5HJF_cjs.Agent; }
 });
-//# sourceMappingURL=agent-WQ4LLTSY.cjs.map
-//# sourceMappingURL=agent-WQ4LLTSY.cjs.map
+//# sourceMappingURL=agent-5DOR5T3E.cjs.map
+//# sourceMappingURL=agent-5DOR5T3E.cjs.map

package/dist/{agent-WQ4LLTSY.cjs.map → agent-5DOR5T3E.cjs.map}

@@ -1 +1 @@
-{"version":3,"sources":[],"names":[],"mappings":"","file":"agent-WQ4LLTSY.cjs"}
+{"version":3,"sources":[],"names":[],"mappings":"","file":"agent-5DOR5T3E.cjs"}

package/dist/{agent-enXXu447.d.ts → agent-CmKwSyyD.d.ts}

@@ -1437,6 +1437,33 @@ interface MemorySettings {
      * share the writes.
      */
     directory?: string;
+    /**
+     * The most facts this agent's memory may hold. Default: no ceiling.
+     *
+     * The ceiling counts the write directory only: the per-memory fact files and any legacy
+     * `## Facts` bullets in that directory's `MEMORY.md`. A dreaming sweep's `factsBefore` counts
+     * every root it reads, so the two numbers can differ.
+     *
+     * At the ceiling a new fact is refused with a `ConfigurationError` whose code is
+     * `memory_ceiling_reached` and whose message names the directory, the count and what to
+     * consolidate; nothing is written. Recording a fact that is already stored is a second
+     * observation, not a new fact, and is still admitted. A value that is not a positive integer is
+     * refused with `invalid_memory_ceiling` by `Agent.create` and `Agent.resume`, before the agent
+     * exists, whether or not memory is enabled, and again by the writer before anything touches disk.
+     *
+     * The count and the write hold a lock that every SDK process writing this directory with a
+     * ceiling takes, so several processes on one directory cannot together pass it. That lock is
+     * the companion `<directory>.lock` beside the directory and needs the optional `proper-lockfile`
+     * peer; without it the ceiling holds within one process only, and the SDK says so once on its
+     * diagnostics channel. A writer that declares no ceiling, such as the Claude Code CLI, does not
+     * take the lock and can still add facts past it. When the lock stays busy past its retries the
+     * write is refused with `proper-lockfile`'s `ELOCKED` error, and nothing is written.
+     *
+     * A resumed agent keeps its ceiling and its `directory`: a `memory` object passed to
+     * `Agent.resume` replaces only the fields it sets, and a field set to `undefined` keeps its
+     * persisted value.
+     */
+    maxFacts?: number;
     /**
      * Whether the SDK auto-injects recalled facts as a `<memory>` block in the
      * LLM system prompt. Default `true`.

package/dist/{agent-GZVZXKME.js → agent-GWKRARAL.js}

@@ -1,16 +1,16 @@
-export { Agent } from './chunk-PEIV7LOU.js';
+export { Agent } from './chunk-4PHSVR2R.js';
 import './chunk-XKMVJ443.js';
-import './chunk-EXI7EHDN.js';
+import './chunk-7Y23B7XX.js';
 import './chunk-K2BQQ445.js';
-import './chunk-ZDF2IJKC.js';
+import './chunk-DPY2M74T.js';
 import './chunk-ZHWKT2NI.js';
 import './chunk-2SFBB54R.js';
 import './chunk-HI3W37FY.js';
 import './chunk-H73MEMQB.js';
 import './chunk-OR6XIWLB.js';
-import './chunk-47Z6LP3I.js';
-import './chunk-OZLQ6ZFM.js';
-import './chunk-EVH3OUMY.js';
+import './chunk-X3LOJJMO.js';
+import './chunk-IZDDSLL2.js';
+import './chunk-74OA5DIM.js';
 import './chunk-SUVCHX4L.js';
 import './chunk-2XC6KDJM.js';
 import './chunk-TE6XVTOA.js';
@@ -38,7 +38,8 @@ import './chunk-GTX6FEDP.js';
 import './chunk-FTZTZBCU.js';
 import './chunk-WE22OXQA.js';
 import './chunk-XZHJYSPB.js';
-import './chunk-BKMDRBJ7.js';
+import './chunk-7KIN3L2S.js';
+import './chunk-RGQICWFU.js';
 import './chunk-Q5EWJPRY.js';
 import './chunk-GGRPAQ3B.js';
 import './chunk-3JHIFQ4I.js';
@@ -46,6 +47,7 @@ import './chunk-ZDYDBXSD.js';
 import './chunk-X7EUUHXU.js';
 import './chunk-MQGSOKR3.js';
 import './chunk-GMRFAX4I.js';
+import './chunk-LZJXEPEQ.js';
 import './chunk-R7WIIPUR.js';
 import './chunk-I2ORT6QD.js';
 import './chunk-SVCPJRUB.js';
@@ -56,5 +58,5 @@ import './chunk-H3CVJSFC.js';
 import './chunk-Q7GYMXVC.js';
 import './chunk-B5MHRQJZ.js';
 import './chunk-CZJ6Q7CW.js';
-//# sourceMappingURL=agent-GZVZXKME.js.map
-//# sourceMappingURL=agent-GZVZXKME.js.map
+//# sourceMappingURL=agent-GWKRARAL.js.map
+//# sourceMappingURL=agent-GWKRARAL.js.map

package/dist/{agent-GZVZXKME.js.map → agent-GWKRARAL.js.map}

@@ -1 +1 @@
-{"version":3,"sources":[],"names":[],"mappings":"","file":"agent-GZVZXKME.js"}
+{"version":3,"sources":[],"names":[],"mappings":"","file":"agent-GWKRARAL.js"}

package/dist/{agent-Dh2YHSLj.d.cts → agent-XMkGzff_.d.cts}

@@ -1437,6 +1437,33 @@ interface MemorySettings {
      * share the writes.
      */
     directory?: string;
+    /**
+     * The most facts this agent's memory may hold. Default: no ceiling.
+     *
+     * The ceiling counts the write directory only: the per-memory fact files and any legacy
+     * `## Facts` bullets in that directory's `MEMORY.md`. A dreaming sweep's `factsBefore` counts
+     * every root it reads, so the two numbers can differ.
+     *
+     * At the ceiling a new fact is refused with a `ConfigurationError` whose code is
+     * `memory_ceiling_reached` and whose message names the directory, the count and what to
+     * consolidate; nothing is written. Recording a fact that is already stored is a second
+     * observation, not a new fact, and is still admitted. A value that is not a positive integer is
+     * refused with `invalid_memory_ceiling` by `Agent.create` and `Agent.resume`, before the agent
+     * exists, whether or not memory is enabled, and again by the writer before anything touches disk.
+     *
+     * The count and the write hold a lock that every SDK process writing this directory with a
+     * ceiling takes, so several processes on one directory cannot together pass it. That lock is
+     * the companion `<directory>.lock` beside the directory and needs the optional `proper-lockfile`
+     * peer; without it the ceiling holds within one process only, and the SDK says so once on its
+     * diagnostics channel. A writer that declares no ceiling, such as the Claude Code CLI, does not
+     * take the lock and can still add facts past it. When the lock stays busy past its retries the
+     * write is refused with `proper-lockfile`'s `ELOCKED` error, and nothing is written.
+     *
+     * A resumed agent keeps its ceiling and its `directory`: a `memory` object passed to
+     * `Agent.resume` replaces only the fields it sets, and a field set to `undefined` keeps its
+     * persisted value.
+     */
+    maxFacts?: number;
     /**
      * Whether the SDK auto-injects recalled facts as a `<memory>` block in the
      * LLM system prompt. Default `true`.

package/dist/{chunk-PEIV7LOU.js → chunk-4PHSVR2R.js}

@@ -1,13 +1,13 @@
 import { submit, emitRunEvent } from './chunk-XKMVJ443.js';
-import { enqueueSessionWrite, hydrateSession, persistTurnToTranscript, flushSessionWrites, appendSessionMessage, getSessionMessages, discardSession, clearSession } from './chunk-EXI7EHDN.js';
-import { IndexManager } from './chunk-ZDF2IJKC.js';
+import { enqueueSessionWrite, hydrateSession, persistTurnToTranscript, flushSessionWrites, appendSessionMessage, getSessionMessages, discardSession, clearSession } from './chunk-7Y23B7XX.js';
+import { IndexManager } from './chunk-DPY2M74T.js';
 import { currentPersonalityContext, warnPersonalitySwitchInsideFork } from './chunk-ZHWKT2NI.js';
 import { extractCodePlugins } from './chunk-2SFBB54R.js';
 import { createOpenAiCompatibleRuntime } from './chunk-HI3W37FY.js';
 import { persistActiveMemoryTranscript } from './chunk-OR6XIWLB.js';
-import { writeSessionSummary, resolveMemoryRoot, MEMORY_KINDS, appendFact, readFacts, legacyMemoryJsonPath, memoryMdPath, appendFactToMarkdown } from './chunk-47Z6LP3I.js';
-import { FsSessionStore } from './chunk-OZLQ6ZFM.js';
-import { discoverProviderPlugins, validateApiKeyShape, getConfiguredBaseUrl, httpRequest, isFixtureApiKey, presentProviderCredentialEnvVars, buildDefaultSummarizer, shouldUseRealLocalRuntime, parseSseStream, CircuitBreaker, resolveRunProvider, mergeExplicitApiKey, resolveProviderChain, abortError, tryFirstEvent, relayStream, renderToolResultContentText } from './chunk-EVH3OUMY.js';
+import { resolveMaxFacts, writeSessionSummary, resolveMemoryRoot, MEMORY_KINDS, appendFact, readFacts, legacyMemoryJsonPath, memoryMdPath, appendFactToMarkdown } from './chunk-X3LOJJMO.js';
+import { FsSessionStore } from './chunk-IZDDSLL2.js';
+import { discoverProviderPlugins, validateApiKeyShape, getConfiguredBaseUrl, httpRequest, isFixtureApiKey, presentProviderCredentialEnvVars, buildDefaultSummarizer, shouldUseRealLocalRuntime, parseSseStream, CircuitBreaker, resolveRunProvider, mergeExplicitApiKey, resolveProviderChain, abortError, tryFirstEvent, relayStream, renderToolResultContentText } from './chunk-74OA5DIM.js';
 import { resolveChildEnv } from './chunk-HKCKOAIO.js';
 import { DEFAULT_MAX_BYTES_PER_FILE, runDiscovery, truncateWithMarker } from './chunk-7QJYKVP6.js';
 import { warnFailureOnce, loadHookConfig, warnOnce } from './chunk-OPAKPEPG.js';
@@ -22,13 +22,14 @@ import { validateArtifactPath, PathTraversalError, safePathJoin } from './chunk-
 import { insideRoot, atOrInsideRoot } from './chunk-UKMBRMGT.js';
 import { mapWithConcurrency } from './chunk-7JAKO3NM.js';
 import { expandTilde, defaultBaseDir } from './chunk-FTZTZBCU.js';
-import { withFileLock } from './chunk-BKMDRBJ7.js';
+import { withFileLock } from './chunk-RGQICWFU.js';
 import { withCwdMutex } from './chunk-Q5EWJPRY.js';
 import { sweepStaleAtomicTemps, atomicWriteJson } from './chunk-3JHIFQ4I.js';
 import { loadSubagents, pluginBundleDirs } from './chunk-ZDYDBXSD.js';
 import { reportUndeclaredSources, theokitConfigRoot, undefinedVariablesIn, getTheokitHome, adapterForConfigPath, adaptersForSurface, projectConfigRoots, pluginBundleRoots } from './chunk-X7EUUHXU.js';
 import { readWorkspaceDir, parseSimpleYaml } from './chunk-MQGSOKR3.js';
 import { resolveEffectiveContextWindow, CONTEXT_WINDOW_FLOOR, CONTEXT_WINDOW_MARGIN } from './chunk-GMRFAX4I.js';
+import { isAnthropicFamilyModel } from './chunk-LZJXEPEQ.js';
 import { parseModelId } from './chunk-R7WIIPUR.js';
 import { getProviderProfile } from './chunk-I2ORT6QD.js';
 import { getCatalogModelInfo, isPatchedModelKey } from './chunk-SVCPJRUB.js';
@@ -335,6 +336,8 @@ function serializeMemory(memory) {
   if (memory.scope !== void 0) out.scope = memory.scope;
   if (memory.activeRecall !== void 0) out.activeRecall = { ...memory.activeRecall };
   if (memory.index !== void 0) out.index = serializeMemoryIndex(memory.index);
+  if (memory.maxFacts !== void 0) out.maxFacts = memory.maxFacts;
+  if (memory.directory !== void 0) out.directory = memory.directory;
   return out;
 }
 function serializeMemoryIndex(index) {
@@ -884,6 +887,7 @@ async function readMemoryFacts(cwd, config) {
 }
 async function appendMemoryFact(cwd, config, fact) {
   if (!config.enabled) return;
+  resolveMaxFacts(config.maxFacts);
   await migrateLegacyJson(cwd, config);
   const sanitized = {
     text: redactSecrets(fact.text),
@@ -3611,6 +3615,7 @@ function validateToolSchema(tool) {
 }
 function validateMemory(options) {
   const memory = options.memory;
+  resolveMaxFacts(memory?.maxFacts);
   if (memory?.storePath === void 0) return;
   const storePath = memory.storePath;
   if (storePath.includes("..") || storePath.startsWith("/") || /^[A-Z]:/i.test(storePath)) {
@@ -5800,14 +5805,27 @@ function stripThinkBlocks(content) {
 // src/internal/agent-loop/loop-llm-stream.ts
 function resolveSystemPromptWithMemoryAdditions(systemPrompt, additions) {
   if (additions === void 0 || additions.length === 0) return systemPrompt;
-  const block = `<active-memory>
-${escapeBlockBody(additions)}
-</active-memory>`;
+  const block = activeMemoryBlock(additions);
   if (systemPrompt === void 0 || systemPrompt.length === 0) return block;
   return `${block}
 
 ${systemPrompt}`;
 }
+function activeMemoryBlock(additions) {
+  return `<active-memory>
+${escapeBlockBody(additions)}
+</active-memory>`;
+}
+function buildLoopSystem(systemPrompt, additions, modelId) {
+  if (!isAnthropicFamilyModel(modelId) || systemPrompt === void 0 || systemPrompt.length === 0) {
+    return resolveSystemPromptWithMemoryAdditions(systemPrompt, additions);
+  }
+  const blocks = [{ text: systemPrompt, cacheable: true }];
+  if (additions !== void 0 && additions.length > 0) {
+    blocks.push({ text: activeMemoryBlock(additions) });
+  }
+  return blocks;
+}
 function toLlmTool(tool) {
   return { name: tool.name, description: tool.description, inputSchema: tool.inputSchema };
 }
@@ -5846,9 +5864,10 @@ async function streamLlmTurn(inputs, ctx) {
     {
       model: inputs.model.id ?? "auto",
       ...(() => {
-        const effective = resolveSystemPromptWithMemoryAdditions(
+        const effective = buildLoopSystem(
           inputs.systemPrompt,
-          ctx.memorySystemPromptAdditions
+          ctx.memorySystemPromptAdditions,
+          inputs.model.id
         );
         return effective !== void 0 ? { system: effective } : {};
       })(),
@@ -6835,29 +6854,32 @@ async function runIteration(inputs, ctx) {
   const llmOutput = await streamLlmTurn(inputs, ctx);
   await inputs.pluginManager?.runPostLlmCallHooks(hookCtx);
   accumulateUsage(ctx.usage, llmOutput);
-  if (inputs.budgetTracker !== void 0) {
-    const modelId = inputs.model.id ?? "auto";
-    const inputT = llmOutput.inputTokens ?? 0;
-    const outputT = llmOutput.outputTokens ?? 0;
-    try {
-      if (inputT > 0) {
-        inputs.budgetTracker.track({
-          tokens: inputT,
-          model: modelId,
-          type: "input"
-        });
-      }
-      if (outputT > 0) {
-        inputs.budgetTracker.track({
-          tokens: outputT,
-          model: modelId,
-          type: "output"
-        });
-      }
-    } catch {
+  trackTurnUsage(inputs, llmOutput);
+  return continueOrTerminate(inputs, ctx, llmOutput);
+}
+function trackTurnUsage(inputs, llmOutput) {
+  const tracker = inputs.budgetTracker;
+  if (tracker === void 0) return;
+  const modelId = inputs.model.id ?? "auto";
+  const inputT = (llmOutput.inputTokens ?? 0) + (llmOutput.cacheReadTokens ?? 0) + (llmOutput.cacheWriteTokens ?? 0);
+  const outputT = llmOutput.outputTokens ?? 0;
+  try {
+    if (inputT > 0) {
+      tracker.track({
+        tokens: inputT,
+        model: modelId,
+        type: "input"
+      });
     }
+    if (outputT > 0) {
+      tracker.track({
+        tokens: outputT,
+        model: modelId,
+        type: "output"
+      });
+    }
+  } catch {
   }
-  return continueOrTerminate(inputs, ctx, llmOutput);
 }
 async function transformLlmOutputText(inputs, text, ctx) {
   return inputs.pluginManager !== void 0 ? inputs.pluginManager.runTransformLlmOutputHooks(text, ctx) : text;
@@ -10607,10 +10629,12 @@ async function rehydrateExistingAgent(agentId, existing, options) {
   await discoverProviderPlugins();
   await validateRehydratedAgent(agentId, existing);
   const mergedLocal = options.local !== void 0 && existing.options.local !== void 0 ? { ...existing.options.local, ...options.local } : options.local ?? existing.options.local;
+  const mergedMemory = mergeResumedMemory(existing.options.memory, options.memory);
   const mergedOptions = {
     ...existing.options,
     ...options,
     ...mergedLocal !== void 0 ? { local: mergedLocal } : {},
+    ...mergedMemory !== void 0 ? { memory: mergedMemory } : {},
     // SE8 — normalize a caller-supplied bare-string model override at this resume
     // boundary (Agent.resume never flows through runCreateUnderSpan). Cloud resume
     // honors it; local resume forces the persisted model below.
@@ -10618,6 +10642,7 @@ async function rehydrateExistingAgent(agentId, existing, options) {
     mcpServers: void 0,
     agentId
   };
+  resolveMaxFacts(mergedOptions.memory?.maxFacts);
   if (existing.runtime === "cloud") {
     return new CloudAgent(mergedOptions, agentId);
   }
@@ -10625,6 +10650,12 @@ async function rehydrateExistingAgent(agentId, existing, options) {
   await agent.initialize();
   return agent;
 }
+function mergeResumedMemory(persisted, override) {
+  if (override === void 0) return persisted;
+  if (persisted === void 0) return override;
+  const named = Object.entries(override).filter(([, value]) => value !== void 0);
+  return { ...persisted, ...Object.fromEntries(named) };
+}
 async function validateRehydratedAgent(agentId, entry) {
   if (entry.runtime !== "local") return;
   const candidate = entry.options.local?.cwd ?? entry.cwd;
@@ -11053,7 +11084,7 @@ var Agent = class _Agent {
    */
   static async compact(agentId, options = {}) {
     const reg = await requireLocalAgent(agentId, "compact");
-    const { compactSessionTranscript, buildDefaultSummarizer: buildDefaultSummarizer2 } = await import('./compact-session-PVAWTKKL.js');
+    const { compactSessionTranscript, buildDefaultSummarizer: buildDefaultSummarizer2 } = await import('./compact-session-F2UDGZCN.js');
     const { cwd, model, store } = await openLocalStore(reg);
     return enqueueSessionWrite(
       cwd,
@@ -11083,7 +11114,7 @@ var Agent = class _Agent {
    */
   static async injectSessionTurn(agentId, turn) {
     const reg = await requireLocalAgent(agentId, "injectSessionTurn");
-    const { injectSessionTurn } = await import('./inject-session-O24XGRR6.js');
+    const { injectSessionTurn } = await import('./inject-session-E5UDFVLI.js');
     const { cwd, model, store } = await openLocalStore(reg);
     await injectSessionTurn({
       store,
@@ -11233,7 +11264,7 @@ async function openLocalStore(reg) {
   const cwd = reg.cwd ?? process.cwd();
   const optModel = reg.options.model;
   const model = reg.model?.id ?? (typeof optModel === "string" ? optModel : optModel?.id) ?? "unknown";
-  const { FsSessionStore: FsSessionStore2 } = await import('./fs-session-store-LCZBM3LT.js');
+  const { FsSessionStore: FsSessionStore2 } = await import('./fs-session-store-3VI4CPTX.js');
   const { defaultBaseDir: defaultBaseDir2, expandTilde: expandTilde2 } = await import('./session-transcript-AR7IBICT.js');
   const configuredSessionDir = reg.options.local?.sessionDir ?? reg.options.local?.baseDir;
   const baseDir = configuredSessionDir !== void 0 ? expandTilde2(configuredSessionDir) : defaultBaseDir2();
@@ -11241,5 +11272,5 @@ async function openLocalStore(reg) {
 }
 
 export { Agent, AgentBuilder, DEFAULT_AGENTIC_MODEL_ID, MEMORY_EMBEDDING_ADAPTERS, Tool, ToolError, UsageAccumulator, computeCost, generateCronId, getPricingEntry, managedSettingsPathFor, openRouterMemoryEmbeddingProviderAdapter, readManagedSettings, resolveSessionDir };
-//# sourceMappingURL=chunk-PEIV7LOU.js.map
-//# sourceMappingURL=chunk-PEIV7LOU.js.map
+//# sourceMappingURL=chunk-4PHSVR2R.js.map
+//# sourceMappingURL=chunk-4PHSVR2R.js.map