@graphorin/memory 0.6.1 → 0.7.0

package/src/context-engine/compaction/compactor.ts

@@ -0,0 +1,403 @@
+/**
+ * Auto-compaction subsystem core (RB-46 / suggested DEC-162 /
+ * ADR-050). Trims the in-flight ContextEngine message buffer when
+ * the assembled token count crosses a per-provider threshold; the
+ * trim is summarized via the configured section template; the
+ * trim is followed by synchronous-await firing of registered
+ * post-compaction hooks (lifecycle owned by Phase 12; this module
+ * exposes the trim primitive).
+ *
+ * Disjoint from the Phase 10c consolidator: never writes to
+ * `semantic_facts` / `episodic_episodes` / `procedural_rules`;
+ * reads only `memory.metadata(...)` / `memory.semantic.search(...)`
+ * / `memory.procedural.activate(...)` / `memory.working.read(...)`
+ * (non-mutating).
+ *
+ * @packageDocumentation
+ */
+
+import type { Message, ModelSpec, SystemMessage } from '@graphorin/core';
+import {
+  type ImperativeScanResult as ScanResult,
+  scanImperativePatterns,
+} from '@graphorin/observability/redaction';
+import { UNTRUSTED_CONTENT_OPEN_PREFIX, wrapUntrusted } from '../../internal/envelope.js';
+import type { ContextLocalePack } from '../locale-packs/index.js';
+import {
+  type ContextTokenCounter,
+  countMessageTokens,
+  HEURISTIC_TOKEN_COUNTER,
+  renderMessageText,
+} from '../token-counter.js';
+import { clearOldToolResults } from './clear-tool-results.js';
+import {
+  buildSummarizerPrompt,
+  type CompactionMetadataPayload,
+  renderFinalSummary,
+  SUMMARY_TEMPLATE_NAME,
+  SUMMARY_TEMPLATE_VERSION,
+} from './templates/summary-9-section.js';
+import type {
+  CompactionContext,
+  CompactionResult,
+  CompactionSource,
+  CompactionStrategy,
+  CompactionSummarizer,
+} from './types.js';
+
+/**
+ * Default count of recent turns preserved verbatim by the
+ * `summarize-old-preserve-recent` strategy.
+ *
+ * @stable
+ */
+export const DEFAULT_PRESERVE_RECENT_TURNS = 6;
+
+/** Opening marker of the inbound-untrusted envelope (tools sanitize layer). */
+const UNTRUSTED_MARKER = UNTRUSTED_CONTENT_OPEN_PREFIX;
+
+/**
+ * Wall-clock budget for the CE-15 injection scan of the summarizer
+ * output. The scanner's 5ms default exists for the per-tool-result
+ * hot path; a compaction already paid for an LLM call, the summary
+ * body is a single bounded text, and the pattern catalogue is fixed,
+ * so the scan here runs UNBUDGETED - a wall-clock budget only turned
+ * scheduler noise into a `null` verdict (a 50ms budget was observed
+ * expiring on a degraded CI runner, silently failing the check open).
+ */
+const COMPACTION_SCAN_BUDGET_MS = Number.POSITIVE_INFINITY;
+
+/**
+ * CE-15 trust decision for a freshly produced summary. Exported for
+ * unit tests because the `scan === null` branch (scanner budget
+ * exceeded) MUST fail closed: the scanner's contract is "null = the
+ * caller applies its own best-effort degradation", and treating null
+ * as "no hits" would commit a poisoned summary as trusted exactly on
+ * the contended hosts where the budget expires.
+ *
+ * @internal
+ */
+export function resolveSummaryTrust(
+  windowUntrusted: boolean,
+  scan: ScanResult | null,
+): 'trusted' | 'untrusted-derived' {
+  if (windowUntrusted) return 'untrusted-derived';
+  if (scan === null) return 'untrusted-derived';
+  return scan.hits.length > 0 ? 'untrusted-derived' : 'trusted';
+}
+
+/** CE-15: does the compacted window carry inbound-untrusted envelopes? */
+function windowContainsUntrusted(messages: ReadonlyArray<Message>): boolean {
+  return messages.some((message) => renderMessageText(message).includes(UNTRUSTED_MARKER));
+}
+
+/**
+ * CE-15: wrap the LLM-authored summary body in a derived-trust
+ * envelope. Envelope marker sequences inside the body are neutralized
+ * first so summarizer output influenced by injected text cannot break
+ * out of the envelope and masquerade as authoritative system text.
+ * Delegates to the shared `internal/envelope.ts` helper (W-083) so the
+ * whole memory package neutralizes markers with one scheme; the
+ * output is byte-identical to the historical inline implementation on
+ * literal-marker inputs.
+ */
+function wrapSummaryAsDerived(body: string): string {
+  return wrapUntrusted(body, { trust: 'derived', tool: 'compaction-summarizer' });
+}
+
+/**
+ * Trim the in-flight buffer using the
+ * `summarize-old-preserve-recent` strategy. Returns the trimmed
+ * messages + summary metadata; the caller is responsible for
+ * appending the summary to the message buffer (Phase 12 owns the
+ * lifecycle; this module exposes the primitive).
+ *
+ * @stable
+ */
+export interface ExecuteCompactionInput {
+  readonly messages: ReadonlyArray<Message>;
+  readonly source: CompactionSource;
+  readonly strategy: CompactionStrategy;
+  readonly localePack: ContextLocalePack;
+  readonly summarizer: CompactionSummarizer;
+  readonly tokenCounter?: ContextTokenCounter;
+  readonly thresholdTokens: number;
+  readonly runId: string;
+  readonly sessionId: string;
+  readonly agentId: string;
+  readonly scope: import('@graphorin/core').SessionScope;
+  readonly providerTrust?: import('@graphorin/core').LocalProviderTrust;
+  readonly now?: () => number;
+  readonly signal?: AbortSignal;
+}
+
+/**
+ * Perform a compaction call. Returns the result envelope containing
+ * the produced summary, the dropped/preserved message slices, and
+ * the per-event metadata. Phase 12 / `agent.compact()` is the
+ * lifecycle owner; this function is the trim primitive.
+ *
+ * @stable
+ */
+export async function executeCompaction(input: ExecuteCompactionInput): Promise<CompactionResult> {
+  const counter = input.tokenCounter ?? HEURISTIC_TOKEN_COUNTER;
+  const now = input.now ?? Date.now;
+  const startTs = now();
+  const beforeTokens = await countMessageTokens(input.messages, counter);
+
+  if (input.strategy.kind === 'custom') {
+    const ctx: CompactionContext = {
+      runId: input.runId,
+      sessionId: input.sessionId,
+      agentId: input.agentId,
+      scope: input.scope,
+      source: input.source,
+      messages: input.messages,
+      beforeTokens,
+      thresholdTokens: input.thresholdTokens,
+      preserveRecentTurns: DEFAULT_PRESERVE_RECENT_TURNS,
+      providerTrust: input.providerTrust ?? 'public-tls',
+    };
+    const result = await input.strategy.compact(ctx);
+    return result;
+  }
+
+  if (input.strategy.kind === 'clear-old-tool-results') {
+    return executeClearStrategy(input, input.strategy, counter, now, startTs, beforeTokens);
+  }
+
+  const preserveRecentTurns = input.strategy.preserveRecentTurns ?? DEFAULT_PRESERVE_RECENT_TURNS;
+  const summarizerModelInput = input.strategy.summarizerModel;
+  const summarizerTimeoutMs = input.strategy.summarizerTimeoutMs;
+
+  // Slice the older portion + preserve the recent N turns. We treat
+  // a "turn" as a single message for the v0.1 surface; the spec
+  // notes more nuanced segmentation is post-MVP per Q-121.
+  let olderCount = Math.max(0, input.messages.length - preserveRecentTurns);
+  // context-engine-01: never split an assistant/tool pair. A purely
+  // positional slice can strand a `role:'tool'` message at the head of
+  // the preserved window while its `tool_calls` assistant partner is
+  // summarized away - OpenAI-compatible servers 400 on the orphan and
+  // `invalid-request` is fallback-ineligible, so the run would die right
+  // after an otherwise successful compaction. Snap the boundary backward
+  // until the preserved slice starts on a non-`tool` message: preserving
+  // more than requested is always transcript-safe. (The
+  // clear-tool-results strategy is immune by construction - it replaces
+  // content in place and never removes messages.)
+  while (olderCount > 0 && input.messages[olderCount]?.role === 'tool') {
+    olderCount -= 1;
+  }
+  const olderRaw = input.messages.slice(0, olderCount);
+  const preservedTail = input.messages.slice(olderCount);
+
+  // C4: keep the most recent USER messages verbatim across compaction -
+  // only assistant/tool content gets summarized away. User words are the
+  // task statement; a paraphrase in the summary routinely drops the
+  // constraint that mattered. The carried messages are removed from the
+  // summarizer's input (no double-counting) and re-inserted between the
+  // summary and the preserved tail, keeping their relative order.
+  const preserveUserMessages = Math.max(0, input.strategy.preserveUserMessages ?? 2);
+  const carriedUserIndices = new Set<number>();
+  if (preserveUserMessages > 0) {
+    for (let i = olderRaw.length - 1; i >= 0 && carriedUserIndices.size < preserveUserMessages; ) {
+      if (olderRaw[i]?.role === 'user') carriedUserIndices.add(i);
+      i -= 1;
+    }
+  }
+  // Degenerate guard: on a tiny (user-dominated) window, carrying users
+  // could empty the summarizable slice and turn the whole compaction into
+  // a no-op - there, positional summarization wins.
+  if (carriedUserIndices.size > 0 && carriedUserIndices.size === olderRaw.length) {
+    carriedUserIndices.clear();
+  }
+  const olderMessages = olderRaw.filter((_, idx) => !carriedUserIndices.has(idx));
+  const carriedUserMessages = olderRaw.filter((_, idx) => carriedUserIndices.has(idx));
+  const preservedMessages = [...carriedUserMessages, ...preservedTail];
+
+  if (olderMessages.length === 0) {
+    return Object.freeze({
+      summary: '',
+      summaryTokens: 0,
+      beforeTokens,
+      afterTokens: beforeTokens,
+      droppedMessageIds: Object.freeze([]),
+      droppedMessageIndices: Object.freeze([]),
+      preservedMessages: Object.freeze([...preservedMessages]),
+      trimmedMessages: Object.freeze([...preservedMessages]),
+      source: input.source,
+      durationMs: now() - startTs,
+      hooksFiredCount: 0,
+      summaryTrust: 'trusted',
+    });
+  }
+
+  const template = {
+    preamble: input.localePack.compactionSummaryTemplate.preamble,
+    sections: input.localePack.compactionSummaryTemplate.sections,
+  };
+  const prompt = buildSummarizerPrompt({
+    template,
+    olderMessages,
+    ...(input.strategy.summarizerInputCharBudget !== undefined
+      ? { maxDumpChars: input.strategy.summarizerInputCharBudget }
+      : {}),
+  });
+
+  const summarizerInput: {
+    prompt: string;
+    model?: ModelSpec | string;
+    signal?: AbortSignal;
+    timeoutMs?: number;
+  } = { prompt };
+  if (summarizerModelInput !== undefined) summarizerInput.model = summarizerModelInput;
+  if (input.signal !== undefined) summarizerInput.signal = input.signal;
+  if (summarizerTimeoutMs !== undefined) summarizerInput.timeoutMs = summarizerTimeoutMs;
+  const summarized = await input.summarizer.summarize(summarizerInput);
+
+  // Section 9 metadata payload - stable shape so consumers can
+  // deserialize and reason about a compaction event.
+  const metadata: CompactionMetadataPayload = {
+    compactedAtIso: new Date(startTs).toISOString(),
+    // CE-14: positional labels, prefixed with the compaction instant so
+    // they no longer collide across compaction events (core Message has
+    // no id - these are indices into THIS compaction's dropped slice).
+    compactedFromMessageIds: Object.freeze(olderMessages.map((_, idx) => `c${startTs}_msg_${idx}`)),
+    compactedFromMessageIndices: Object.freeze(olderMessages.map((_, idx) => idx)),
+    compactedFromTokens: await countMessageTokens(olderMessages, counter),
+    summaryTokens: summarized.usageTokens ?? (await counter.countText(summarized.text)),
+    summarizerModel: resolveSummarizerModelLabel(summarizerModelInput, input.summarizer.id),
+    templateName: SUMMARY_TEMPLATE_NAME,
+    templateVersion: SUMMARY_TEMPLATE_VERSION,
+    preserveRecentTurns,
+  };
+
+  // CE-15: the summary must not launder untrusted content into a
+  // trusted system message. When the compacted window carried
+  // untrusted envelopes - or the injection heuristics flag the
+  // summarizer output itself - the LLM-authored body is committed
+  // inside a derived-trust envelope (sticky across re-compactions:
+  // the envelope re-triggers this detection next time around).
+  const summaryScan = scanImperativePatterns(summarized.text, undefined, COMPACTION_SCAN_BUDGET_MS);
+  const summaryTrust = resolveSummaryTrust(windowContainsUntrusted(olderMessages), summaryScan);
+  const finalSummary = renderFinalSummary({
+    template,
+    summaryFromLlm:
+      summaryTrust === 'untrusted-derived'
+        ? wrapSummaryAsDerived(summarized.text)
+        : summarized.text,
+    preservedMessages,
+    metadata,
+  });
+  const summaryTokens = await counter.countText(finalSummary);
+
+  const summaryMessage: SystemMessage = {
+    role: 'system',
+    content: finalSummary,
+  };
+  const trimmedMessages: ReadonlyArray<Message> = Object.freeze([
+    summaryMessage,
+    ...preservedMessages,
+  ]);
+  const afterTokens = await countMessageTokens(trimmedMessages, counter);
+  const durationMs = now() - startTs;
+
+  return Object.freeze({
+    summary: finalSummary,
+    summaryTokens,
+    beforeTokens,
+    afterTokens,
+    droppedMessageIds: metadata.compactedFromMessageIds,
+    droppedMessageIndices: metadata.compactedFromMessageIndices,
+    preservedMessages: Object.freeze([...preservedMessages]),
+    trimmedMessages,
+    source: input.source,
+    durationMs,
+    hooksFiredCount: 0,
+    summaryTrust,
+  });
+}
+
+/**
+ * SOTA-1: clear the oldest tool results (zero-LLM), then summarize only if the
+ * cleared buffer is still over the threshold. The summarizer runs on the
+ * already-reduced window, so a few-tool-result buffer compacts for free.
+ */
+async function executeClearStrategy(
+  input: ExecuteCompactionInput,
+  strategy: Extract<CompactionStrategy, { kind: 'clear-old-tool-results' }>,
+  counter: ContextTokenCounter,
+  now: () => number,
+  startTs: number,
+  beforeTokens: number,
+): Promise<CompactionResult> {
+  const outcome = await clearOldToolResults(
+    input.messages,
+    {
+      ...(strategy.keepToolUses !== undefined ? { keepToolUses: strategy.keepToolUses } : {}),
+      ...(strategy.clearAtLeast !== undefined ? { clearAtLeast: strategy.clearAtLeast } : {}),
+      ...(strategy.excludeTools !== undefined ? { excludeTools: strategy.excludeTools } : {}),
+      ...(strategy.externalize !== undefined ? { externalize: strategy.externalize } : {}),
+      ...(strategy.clearToolInputs !== undefined
+        ? { clearToolInputs: strategy.clearToolInputs }
+        : {}),
+      ...(strategy.readResultToolName !== undefined
+        ? { readResultToolName: strategy.readResultToolName }
+        : {}),
+    },
+    counter,
+  );
+  const afterClearTokens = await countMessageTokens(outcome.messages, counter);
+
+  const fallback = strategy.summarizeFallback;
+  if (fallback !== false && afterClearTokens > input.thresholdTokens) {
+    // Clearing did not reclaim enough - summarize the ALREADY-cleared buffer.
+    const summarized = await executeCompaction({
+      ...input,
+      messages: outcome.messages,
+      strategy: { kind: 'summarize-old-preserve-recent', ...(fallback ?? {}) },
+    });
+    return Object.freeze({
+      ...summarized,
+      // Report against the original buffer + carry the cleared indices.
+      beforeTokens,
+      droppedMessageIndices: Object.freeze([
+        ...outcome.clearedIndices,
+        ...summarized.droppedMessageIndices.filter((i) => !outcome.clearedIndices.includes(i)),
+      ]),
+      durationMs: now() - startTs,
+    });
+  }
+
+  // Zero-LLM clear-only outcome: content replaced in place, nothing summarized.
+  const clearedSet = new Set(outcome.clearedIndices);
+  const preservedMessages = outcome.messages.filter((_, i) => !clearedSet.has(i));
+  return Object.freeze({
+    summary: '',
+    summaryTokens: 0,
+    beforeTokens,
+    afterTokens: afterClearTokens,
+    droppedMessageIds: Object.freeze(outcome.clearedIndices.map((i) => `c${startTs}_cleared_${i}`)),
+    droppedMessageIndices: Object.freeze([...outcome.clearedIndices]),
+    preservedMessages: Object.freeze(preservedMessages),
+    trimmedMessages: Object.freeze([...outcome.messages]),
+    source: input.source,
+    durationMs: now() - startTs,
+    hooksFiredCount: 0,
+    summaryTrust: 'trusted',
+  });
+}
+
+function resolveSummarizerModelLabel(
+  modelInput: ModelSpec | string | undefined,
+  summarizerId: string | undefined,
+): string | null {
+  if (typeof modelInput === 'string') return modelInput;
+  if (modelInput !== undefined) {
+    const provider =
+      'provider' in modelInput
+        ? `${modelInput.provider.name}:${modelInput.model}`
+        : `${modelInput.name}:${modelInput.modelId}`;
+    return provider;
+  }
+  return summarizerId ?? null;
+}

package/src/context-engine/compaction/hooks/reanchor-persona-block.ts

@@ -0,0 +1,55 @@
+/**
+ * `reanchorPersonaBlock({ blockLabel = 'persona' })` - built-in
+ * post-compaction hook. Queries `memory.working.read(...)` for the
+ * persona block and returns it as a system-content fragment for
+ * re-injection.
+ *
+ * @packageDocumentation
+ */
+
+import type { MessageContent } from '@graphorin/core';
+import type { HookDeps, NamedPostCompactionHook } from './types.js';
+
+/**
+ * Build a `reanchorPersonaBlock` hook.
+ *
+ * @stable
+ */
+export function reanchorPersonaBlock(
+  options: { readonly blockLabel?: string } = {},
+): NamedPostCompactionHook {
+  const blockLabel = options.blockLabel ?? 'persona';
+  return {
+    id: 'reanchorPersonaBlock',
+    async resolveContent(deps: HookDeps): Promise<ReadonlyArray<MessageContent>> {
+      const value = await deps.memory.working.read(deps.scope, blockLabel);
+      if (value === null || value.length === 0) return [];
+      // context-engine-02: the D2 privacy filter gates re-anchored content
+      // exactly as it gates the assembled prompt - a secret-tier persona
+      // block the assembly withheld must not ship to the provider via the
+      // post-compaction splice. `read()` returns only the value, so the
+      // sensitivity comes from the block listing; a block absent from the
+      // listing (unwritten, default-value fallback) is evaluated at the
+      // engine's default sensitivity.
+      if (deps.allowSensitivity !== undefined) {
+        const blocks = await deps.memory.working.list(deps.scope);
+        const block = blocks.find((b) => b.label === blockLabel);
+        if (!deps.allowSensitivity(block?.sensitivity)) return [];
+      }
+      return [
+        {
+          type: 'text',
+          text: `<memory_blocks anchor="post-compaction">\n  <block label="${escapeXml(blockLabel)}">\n    ${escapeXml(value)}\n  </block>\n</memory_blocks>`,
+        },
+      ];
+    },
+  };
+}
+
+function escapeXml(value: string): string {
+  return value
+    .replace(/&/g, '&amp;')
+    .replace(/</g, '&lt;')
+    .replace(/>/g, '&gt;')
+    .replace(/"/g, '&quot;');
+}

package/src/context-engine/compaction/hooks/reanchor-pinned-facts.ts

@@ -0,0 +1,66 @@
+/**
+ * `reanchorPinnedFacts({ pinnedFactIds, maxTokens })` - built-in
+ * post-compaction hook. Looks up the explicitly-pinned fact IDs
+ * via `memory.semantic.get(...)` and returns them as a
+ * system-content fragment with their tags.
+ *
+ * @packageDocumentation
+ */
+
+import type { MessageContent } from '@graphorin/core';
+import { type ContextTokenCounter, HEURISTIC_TOKEN_COUNTER } from '../../token-counter.js';
+import type { HookDeps, NamedPostCompactionHook } from './types.js';
+
+/**
+ * Build a `reanchorPinnedFacts` hook.
+ *
+ * @stable
+ */
+export function reanchorPinnedFacts(options: {
+  readonly pinnedFactIds: ReadonlyArray<string>;
+  readonly maxTokens?: number;
+  readonly tokenCounter?: ContextTokenCounter;
+}): NamedPostCompactionHook {
+  const ids = options.pinnedFactIds;
+  const maxTokens = Math.max(0, options.maxTokens ?? 2000);
+  const counter = options.tokenCounter ?? HEURISTIC_TOKEN_COUNTER;
+  return {
+    id: 'reanchorPinnedFacts',
+    async resolveContent(deps: HookDeps): Promise<ReadonlyArray<MessageContent>> {
+      if (ids.length === 0) return [];
+      const lines = ['<memory_pinned_facts anchor="post-compaction">'];
+      let runningTokens = 0;
+      for (const id of ids) {
+        const fact = await deps.memory.semantic.get(deps.scope, id).catch(() => null);
+        if (fact === null || fact.deletedAt !== undefined) continue;
+        // context-engine-02: pinning is not a privacy override - a
+        // secret-tier fact the D2 filter withholds from the assembled
+        // prompt must not re-enter via the post-compaction splice.
+        if (deps.allowSensitivity !== undefined && !deps.allowSensitivity(fact.sensitivity)) {
+          continue;
+        }
+        const tagBlob =
+          fact.tags !== undefined ? ` tags="${escapeXml((fact.tags ?? []).join(','))}"` : '';
+        const fragment = `  <fact id="${escapeXml(fact.id)}"${tagBlob}>${escapeXml(fact.text)}</fact>`;
+        const fragmentTokens = await counter.countText(fragment);
+        if (maxTokens > 0 && runningTokens + fragmentTokens > maxTokens) {
+          lines.push('  <!-- additional pinned facts truncated to fit budget -->');
+          break;
+        }
+        runningTokens += fragmentTokens;
+        lines.push(fragment);
+      }
+      if (lines.length === 1) return [];
+      lines.push('</memory_pinned_facts>');
+      return [{ type: 'text', text: lines.join('\n') }];
+    },
+  };
+}
+
+function escapeXml(value: string): string {
+  return value
+    .replace(/&/g, '&amp;')
+    .replace(/</g, '&lt;')
+    .replace(/>/g, '&gt;')
+    .replace(/"/g, '&quot;');
+}

package/src/context-engine/compaction/hooks/reanchor-project-rules.ts

@@ -0,0 +1,57 @@
+/**
+ * `reanchorProjectRules({ ruleTagsAllowlist? })` - built-in
+ * post-compaction hook. Queries `memory.procedural.activate(...)`
+ * and returns the active rule set as a system-content fragment for
+ * re-injection into the trimmed buffer.
+ *
+ * @packageDocumentation
+ */
+
+import type { MessageContent } from '@graphorin/core';
+import type { HookDeps, NamedPostCompactionHook } from './types.js';
+
+/**
+ * Build a `reanchorProjectRules` hook.
+ *
+ * @stable
+ */
+export function reanchorProjectRules(
+  options: { readonly ruleTagsAllowlist?: ReadonlyArray<string> } = {},
+): NamedPostCompactionHook {
+  const allowlist =
+    options.ruleTagsAllowlist !== undefined ? new Set(options.ruleTagsAllowlist) : null;
+  return {
+    id: 'reanchorProjectRules',
+    async resolveContent(deps: HookDeps): Promise<ReadonlyArray<MessageContent>> {
+      const procedural = deps.procedural ?? {};
+      const rules = await deps.memory.procedural.activate(deps.scope, procedural);
+      // context-engine-02: apply the same D2 privacy decision `assemble()`
+      // applies to active rules - the hook's output ships to the provider.
+      const visible =
+        deps.allowSensitivity === undefined
+          ? rules
+          : rules.filter((rule) => deps.allowSensitivity?.(rule.sensitivity) === true);
+      const filtered =
+        allowlist === null
+          ? visible
+          : visible.filter((rule) => (rule.tags ?? []).some((tag) => allowlist.has(tag)));
+      if (filtered.length === 0) return [];
+      const lines = ['<memory_rules anchor="post-compaction">'];
+      for (const rule of filtered) {
+        lines.push(
+          `  <rule priority="${rule.priority}" tags="${escapeXml((rule.tags ?? []).join(','))}">${escapeXml(rule.text)}</rule>`,
+        );
+      }
+      lines.push('</memory_rules>');
+      return [{ type: 'text', text: lines.join('\n') }];
+    },
+  };
+}
+
+function escapeXml(value: string): string {
+  return value
+    .replace(/&/g, '&amp;')
+    .replace(/</g, '&lt;')
+    .replace(/>/g, '&gt;')
+    .replace(/"/g, '&quot;');
+}