Limit output history and reset test state

Author: buger

jest.config.js

@@ -25,4 +25,6 @@ module.exports = {
   detectOpenHandles: true,
   // Speed up test execution
   maxWorkers: process.env.CI ? 2 : '50%',
-};
+  // Recycle workers if they retain too much memory
+  workerIdleMemoryLimit: process.env.CI ? '1024MB' : undefined,
+};

src/state-machine/dispatch/history-snapshot.ts

@@ -1,13 +1,21 @@
 import type { EngineContext } from '../../types/engine';
 import { logger } from '../../logger';
 
+function getHistoryLimit(): number | undefined {
+  const raw = process.env.VISOR_TEST_HISTORY_LIMIT || process.env.VISOR_OUTPUT_HISTORY_LIMIT;
+  if (!raw) return undefined;
+  const n = parseInt(raw, 10);
+  return Number.isFinite(n) && n > 0 ? n : undefined;
+}
+
 /**
  * Build output history Map from journal for template rendering
  * This matches the format expected by AI providers.
  * Moved from LevelDispatch to reduce file size and improve reuse.
  */
 export function buildOutputHistoryFromJournal(context: EngineContext): Map<string, unknown[]> {
   const outputHistory = new Map<string, unknown[]>();
+  const limit = getHistoryLimit();
 
   try {
     const snapshot = context.journal.beginSnapshot();
@@ -41,7 +49,13 @@ export function buildOutputHistoryFromJournal(context: EngineContext): Map<strin
           continue;
         }
       } catch {}
-      if (payload !== undefined) outputHistory.get(checkId)!.push(payload);
+      if (payload !== undefined) {
+        const arr = outputHistory.get(checkId)!;
+        arr.push(payload);
+        if (limit && arr.length > limit) {
+          arr.splice(0, arr.length - limit);
+        }
+      }
     }
   } catch (error) {
     logger.debug(`[LevelDispatch] Error building output history: ${error}`);

src/state-machine/states/level-dispatch.ts

@@ -72,12 +72,20 @@ function recordOnFinishRoutingEvent(args: {
   addEvent('visor.routing', attrs);
 }
 
+function getHistoryLimit(): number | undefined {
+  const raw = process.env.VISOR_TEST_HISTORY_LIMIT || process.env.VISOR_OUTPUT_HISTORY_LIMIT;
+  if (!raw) return undefined;
+  const n = parseInt(raw, 10);
+  return Number.isFinite(n) && n > 0 ? n : undefined;
+}
+
 /**
  * Build output history Map from journal for template rendering
  * This matches the format expected by AI providers
  */
 function buildOutputHistoryFromJournal(context: EngineContext): Map<string, unknown[]> {
   const outputHistory = new Map<string, unknown[]>();
+  const limit = getHistoryLimit();
 
   try {
     const snapshot = context.journal.beginSnapshot();
@@ -94,7 +102,13 @@ function buildOutputHistoryFromJournal(context: EngineContext): Map<string, unkn
       // outputs_history['security'].last.issues[...] works in prompts and tests.
       const payload =
         entry.result.output !== undefined ? entry.result.output : (entry.result as unknown);
-      if (payload !== undefined) outputHistory.get(checkId)!.push(payload);
+      if (payload !== undefined) {
+        const arr = outputHistory.get(checkId)!;
+        arr.push(payload);
+        if (limit && arr.length > limit) {
+          arr.splice(0, arr.length - limit);
+        }
+      }
     }
   } catch (error) {
     // Silently fail - return empty map

src/state-machine/states/routing.ts

@@ -154,6 +154,13 @@ function createMemoryHelpers() {
   };
 }
 
+function getHistoryLimit(): number | undefined {
+  const raw = process.env.VISOR_TEST_HISTORY_LIMIT || process.env.VISOR_OUTPUT_HISTORY_LIMIT;
+  if (!raw) return undefined;
+  const n = parseInt(raw, 10);
+  return Number.isFinite(n) && n > 0 ? n : undefined;
+}
+
 type RoutingTrigger = 'on_success' | 'on_fail' | 'on_finish';
 type RoutingAction = 'run' | 'goto' | 'retry';
 type RoutingSource = 'run' | 'run_js' | 'goto' | 'goto_js' | 'transitions' | 'retry';
@@ -1252,6 +1259,7 @@ async function evaluateRunJs(
 ): Promise<string[]> {
   try {
     const sandbox = createSecureSandbox();
+    const historyLimit = getHistoryLimit();
 
     // Build outputs record and outputs_history
     const snapshotId = context.journal.beginSnapshot();
@@ -1287,8 +1295,12 @@ async function evaluateRunJs(
       try {
         const history = contextView.getHistory(checkIdFromJournal);
         if (history && history.length > 0) {
+          const trimmed =
+            historyLimit && history.length > historyLimit
+              ? history.slice(history.length - historyLimit)
+              : history;
           // Extract outputs from history (prefer output field if available)
-          outputsHistory[checkIdFromJournal] = history.map((r: any) =>
+          outputsHistory[checkIdFromJournal] = trimmed.map((r: any) =>
             r.output !== undefined ? r.output : r
           );
         }
@@ -1392,6 +1404,7 @@ export async function evaluateGoto(
   if (gotoJs) {
     try {
       const sandbox = createSecureSandbox();
+      const historyLimit = getHistoryLimit();
 
       // Build outputs record and outputs_history from the full session snapshot.
       // Do not filter by event here — on_finish (especially forEach post-children) may
@@ -1429,8 +1442,12 @@ export async function evaluateGoto(
         try {
           const history = contextView.getHistory(checkIdFromJournal);
           if (history && history.length > 0) {
+            const trimmed =
+              historyLimit && history.length > historyLimit
+                ? history.slice(history.length - historyLimit)
+                : history;
             // Extract outputs from history (prefer output field if available)
-            outputsHistory[checkIdFromJournal] = history.map((r: any) =>
+            outputsHistory[checkIdFromJournal] = trimmed.map((r: any) =>
               r.output !== undefined ? r.output : r
             );
           }
@@ -1572,6 +1589,7 @@ export async function evaluateTransitions(
   if (!transitions || transitions.length === 0) return undefined;
   try {
     const sandbox = createSecureSandbox();
+    const historyLimit = getHistoryLimit();
 
     // Build outputs record and outputs_history from the full session snapshot
     const snapshotId = context.journal.beginSnapshot();
@@ -1590,7 +1608,13 @@ export async function evaluateTransitions(
       try {
         const hist = view.getHistory(cid);
         if (hist && hist.length > 0) {
-          outputsHistory[cid] = hist.map((r: any) => (r.output !== undefined ? r.output : r));
+          const trimmed =
+            historyLimit && hist.length > historyLimit
+              ? hist.slice(hist.length - historyLimit)
+              : hist;
+          outputsHistory[cid] = trimmed.map((r: any) =>
+            r.output !== undefined ? r.output : r
+          );
         }
       } catch {}
     }

src/test-runner/index.ts

@@ -6,6 +6,7 @@ import { StateMachineExecutionEngine } from '../state-machine-execution-engine';
 import type { PRInfo } from '../pr-analyzer';
 import { RecordingOctokit } from './recorders/github-recorder';
 import { MemoryStore } from '../memory-store';
+import { SessionRegistry } from '../session-registry';
 import { setGlobalRecorder } from './recorders/global-recorder';
 // import { FixtureLoader } from './fixture-loader';
 import { type ExpectBlock } from './assertions';
@@ -35,6 +36,15 @@ export interface DiscoverOptions {
   testsPath?: string; // File, directory, or glob pattern
   cwd?: string;
 }
+function ensureTestEnvDefaults(): void {
+  if (!process.env.VISOR_TEST_MODE) process.env.VISOR_TEST_MODE = 'true';
+  if (!process.env.VISOR_TEST_PROMPT_MAX_CHARS) {
+    process.env.VISOR_TEST_PROMPT_MAX_CHARS = process.env.CI === 'true' ? '4000' : '8000';
+  }
+  if (!process.env.VISOR_TEST_HISTORY_LIMIT) {
+    process.env.VISOR_TEST_HISTORY_LIMIT = process.env.CI === 'true' ? '200' : '500';
+  }
+}
 /**
  * Very small glob-to-RegExp converter supporting **, *, and ?
  * - ** matches across path separators
@@ -180,6 +190,7 @@ export async function runSuites(
     }>;
   }>;
 }> {
+  ensureTestEnvDefaults();
   const perSuite: Array<{
     file: string;
     failures: number;
@@ -339,6 +350,10 @@ export class VisorTestRunner {
     try {
       MemoryStore.resetInstance();
     } catch {}
+    // Always clear AI sessions between cases to prevent cross-case leakage
+    try {
+      SessionRegistry.getInstance().clearAllSessions();
+    } catch {}
     // Always use StateMachineExecutionEngine
     const engine = new StateMachineExecutionEngine(undefined as any, recorder as unknown as any);
     try {
@@ -751,8 +766,11 @@ export class VisorTestRunner {
     const ghRec = defaultsAny?.github_recorder as
       | { error_code?: number; timeout_ms?: number }
       | undefined;
+    const envPromptCapRaw = process.env.VISOR_TEST_PROMPT_MAX_CHARS;
+    const envPromptCap = envPromptCapRaw ? parseInt(envPromptCapRaw, 10) : undefined;
     const defaultPromptCap: number | undefined =
-      options.promptMaxChars ||
+      options.promptMaxChars ??
+      (Number.isFinite(envPromptCap as number) ? (envPromptCap as number) : undefined) ??
       (typeof defaultsAny?.prompt_max_chars === 'number'
         ? defaultsAny.prompt_max_chars
         : undefined);
@@ -1224,6 +1242,10 @@ export class VisorTestRunner {
         try {
           MemoryStore.resetInstance();
         } catch {}
+        // Clear AI sessions before each stage to avoid leakage across stages
+        try {
+          SessionRegistry.getInstance().clearAllSessions();
+        } catch {}
         // Prepare default tag filters for this flow (inherit suite defaults)
         const parseTags = (v: unknown): string[] | undefined => {
           if (!v) return undefined;

tests/setup.ts

@@ -1,4 +1,14 @@
 // Test setup file to configure mocks and prevent external API calls
+import { MemoryStore } from '../src/memory-store';
+import { SessionRegistry } from '../src/session-registry';
+
+// Default test caps to reduce memory pressure (can be overridden by env)
+if (!process.env.VISOR_TEST_PROMPT_MAX_CHARS) {
+  process.env.VISOR_TEST_PROMPT_MAX_CHARS = process.env.CI === 'true' ? '4000' : '8000';
+}
+if (!process.env.VISOR_TEST_HISTORY_LIMIT) {
+  process.env.VISOR_TEST_HISTORY_LIMIT = process.env.CI === 'true' ? '200' : '500';
+}
 
 // Mock child_process.spawn globally to prevent real process spawns while leaving other methods intact
 jest.mock('child_process', () => {
@@ -62,6 +72,14 @@ afterEach(() => {
   // Ensure leaked git vars are cleared after each test
   const gitVars = ['GIT_DIR', 'GIT_WORK_TREE', 'GIT_INDEX_FILE', 'GIT_PREFIX', 'GIT_COMMON_DIR'];
   for (const k of gitVars) delete (process.env as NodeJS.ProcessEnv)[k];
+
+  // Clear global singletons between tests to avoid cross-test memory leaks
+  try {
+    MemoryStore.resetInstance();
+  } catch {}
+  try {
+    SessionRegistry.getInstance().clearAllSessions();
+  } catch {}
 });
 
 // Set global Jest timeout for all tests