Merge ticket-T06 (T06 Storm Guard Isolation #23) into integration/v0.2

packages/ai-studio/AiGeneration/Studio/LoopGuard.js

@@ -6,6 +6,15 @@ import { emitStudioDebugEvent } from './DebugEvents';
 /**
  * The local loop guard: a circuit breaker for a BYOK agent.
  *
+ * Scope is strictly single-request (T06 #23): one guard instance per
+ * `aiRequest.id` (see `loopGuardsRef` in Utils.js), counting only that
+ * request's own turns — consecutive identical calls (repeat), consecutive
+ * failures (error storm), turns without user input, and turns without
+ * project progress (repeat-without-progress, sub-agents only). A child's
+ * startup failure never touches its parent's streak and never constrains
+ * the parent: `spawnSubAgent` failures fail only their own `spawn_agent`
+ * tool call (`success:false`) so the parent stays alive for serial fallback.
+ *
  * Ported from munder-difflin's `src/main/breaker.ts`, narrowed to what this
  * editor can actually observe. Under BYOK there is no server, so nothing counts
  * a runaway local agent: the hosted `repeated-tool-call-loop` error can never
@@ -44,6 +53,13 @@ export type LoopGuardDecision = {|
   changed: boolean,
 |};
 
+/**
+ * T06 (#23) isolation scope marker: guards are per-request only. A test
+ * asserts this constant exists so the single-request + repeat-without-
+ * progress contract cannot silently regress to a shared/global guard.
+ */
+export const LOOP_GUARD_SCOPE: string = 'single-request';
+
 export const DEFAULT_LOOP_GUARD_CONFIG = {
   enabled: true,
   /** Trips on the Nth consecutive identical call; N-1 are tolerated. */

packages/ai-studio/MANIFEST.json

@@ -26,9 +26,9 @@
       "sha256": "708de36210d000af56e795280eaa5dd450e32cedb4a54b7b031353a6cc4a679f"
     },
     "AiGeneration/Studio/LoopGuard.js": {
-      "bytes": 10420,
+      "bytes": 11296,
       "origin": "newIDE/app/src/AiGeneration/Studio/LoopGuard.js",
-      "sha256": "09bde299559edc5b9c5f74d7605c7c4ef8ba0eeb4f044168171183f51676b86d"
+      "sha256": "8fd346b75556051e7f03a15066ab562168f2028d5fca028583d9e7449be9b1ee"
     },
     "AiGeneration/Studio/NudgePolicy.js": {
       "bytes": 3821,

patches/14-ai-utils.patch

@@ -254,7 +254,7 @@
              !isCallApproved(functionCall)
          );
  
-@@ -554,13 +681,406 @@
+@@ -554,13 +681,409 @@
          }
        }
  
@@ -567,6 +567,9 @@
 +            });
 +          } catch (error) {
 +            console.error('Error while spawning a studio sub-agent:', error);
++            // T06 (#23): a child startup failure fails ONLY this tool call.
++            // It never touches the parent guard streak and never suspends or
++            // constrains the parent: the parent stays alive for serial fallback.
 +            spawnResult = { error: 'The sub-agent could not be started.' };
 +          }
 +
@@ -662,7 +665,7 @@
  
        // The "modified outside editor" callbacks each refresh the editor and can
        // trigger an in-game editor hot reload. Firing them once per function
-@@ -669,7 +1189,7 @@
+@@ -669,7 +1192,7 @@
            // Explorer sub-agent scripts are read-only: restrict their
            // `run_script` to non-mutating functions (defense in depth).
            runScriptReadOnly: isReadOnlyScriptContext,
@@ -671,15 +674,25 @@
              name: functionCall.name,
              arguments: functionCall.arguments,
              call_id: functionCall.call_id,
-@@ -761,13 +1281,55 @@
+@@ -761,13 +1284,65 @@
            return;
          }
  
 +        // Record outcomes for the loop guard's error-storm trip (a real success
 +        // clears it; a failure advances it).
++        // T06 (#23) isolation: ONLY editor-dispatched calls feed the parent
++        // guard. Spawn failures, offline-docs, local-plan, no-project, role
++        // denials, and loop-guard warnings already have terminal results and
++        // must never advance the parent error streak nor constrain it — a
++        // child that fails to start fails its own call; the parent retries
++        // serially instead of suspending.
 +        if (guard) {
++          const editorCallIds = new Set(
++            callsForEditor.map(functionCall => functionCall.call_id)
++          );
 +          results.forEach(result => {
 +            if (result.status !== 'finished') return;
++            if (!editorCallIds.has(result.call_id)) return;
 +            if (result.success) guard.recordSuccess();
 +            else guard.recordError();
 +          });
@@ -732,7 +745,7 @@
        } finally {
          // Flush the coalesced editor notifications for everything modified in
          // this batch (one hot reload instead of one per call). In `finally` so
-@@ -787,7 +1349,9 @@
+@@ -787,7 +1362,9 @@
          }
  
          // Release the lock so these calls can be retried if needed
@@ -743,7 +756,7 @@
          functionCallsToProcess.forEach(functionCall => {
            inFlightFunctionCallIdsRef.current.delete(
              `${aiRequest.id}:${functionCall.call_id}`
-@@ -824,9 +1388,12 @@
+@@ -824,9 +1401,12 @@
        getAssetStoreTagForNewObject,
        generateEvents,
        onSendEditorFunctionCallResults,
@@ -756,7 +769,7 @@
      ]
    );
  
-@@ -870,9 +1437,50 @@
+@@ -870,9 +1450,50 @@
      [onProcessFunctionCalls, allFunctionCallsToProcess]
    );
  
@@ -807,7 +820,7 @@
    };
  };
  
-@@ -977,7 +1585,6 @@
+@@ -977,7 +1598,6 @@
      editorFunctionCallResultsStorage,
      isFetchingSuggestions,
      setIsFetchingSuggestions,
@@ -815,7 +828,7 @@
      setSelectedAiRequestId,
      selectedAiRequest,
    } = React.useContext(AiRequestContext);
-@@ -1036,20 +1643,19 @@
+@@ -1036,20 +1656,19 @@
          // Then ask for some.
          if (
            !selectedAiRequest ||
@@ -847,7 +860,7 @@
  
          // Check if there are tools being run. If so, no suggestions at this time.
          const hasFunctionsCallsToProcess =
-@@ -1139,7 +1745,7 @@
+@@ -1139,7 +1758,7 @@
            : null;
          const preparedAiUserContent = await prepareAiUserContent({
            getAuthorizationHeader,
@@ -856,7 +869,7 @@
            simplifiedProjectJson,
            projectSpecificExtensionsSummaryJson,
            eventsJson: null,
-@@ -1153,7 +1759,7 @@
+@@ -1153,7 +1772,7 @@
            const aiRequestWorkingForSuggestions = await getAiRequestSuggestions(
              getAuthorizationHeader,
              {
@@ -865,7 +878,7 @@
                aiRequestId: selectedAiRequest.id,
                suggestionsType: isLastMessageFunctionCallOutputProjectInitialization
                  ? 'list-with-explanations'
-@@ -1525,39 +2131,12 @@
+@@ -1525,39 +2144,12 @@
      [profile, setSelectedAiRequestId]
    );
  

tests/integration/ai-byok-flow.test.mjs

@@ -690,3 +690,98 @@ describe('integration: T05 suspended UI and resume (terminal status card)', () =
     }
   });
 });
+
+describe('integration: T06 storm guard isolation (child failure never suspends parent)', () => {
+  it('LoopGuard is single-request with repeat-without-progress isolation', async () => {
+    const { readFileSync } = await import('node:fs');
+    const { fileURLToPath } = await import('node:url');
+    const { dirname, resolve } = await import('node:path');
+    const root = resolve(dirname(fileURLToPath(import.meta.url)), '..', '..');
+    const src = readFileSync(
+      resolve(root, 'packages/ai-studio/AiGeneration/Studio/LoopGuard.js'),
+      'utf8'
+    );
+    assert.ok(src.includes('LOOP_GUARD_SCOPE'), 'LOOP_GUARD_SCOPE marker missing');
+    assert.ok(src.includes("'single-request'"), 'guard scope must be single-request');
+    assert.ok(src.includes('maxTurnsWithoutProgress'), 'repeat-without-progress trip missing');
+    assert.ok(
+      src.includes("child") || src.includes('parent'),
+      'isolation note must name child/parent separation'
+    );
+  });
+
+  it('child spawn startup failure fails only its tool call (success:false, no parent suspend)', async () => {
+    const { readFileSync } = await import('node:fs');
+    const { fileURLToPath } = await import('node:url');
+    const { dirname, resolve } = await import('node:path');
+    const root = resolve(dirname(fileURLToPath(import.meta.url)), '..', '..');
+    const patch14 = readFileSync(resolve(root, 'patches/14-ai-utils.patch'), 'utf8');
+    // Failure shape: failed tool call the model sees.
+    assert.ok(
+      patch14.includes('The sub-agent could not be started.'),
+      'spawn startup failure message missing'
+    );
+    // Isolation markers.
+    for (const marker of ['T06 (#23)', 'editorCallIds', 'spawnFailureResults']) {
+      assert.ok(patch14.includes(marker), `patch 14 missing T06 marker: ${marker}`);
+    }
+    // Parent guard feeds ONLY editor-dispatched calls — never spawn/local failures.
+    assert.ok(
+      patch14.includes('if (!editorCallIds.has(result.call_id)) return;'),
+      'parent guard must skip non-editor (child/local) results'
+    );
+    // The spawn failure path never suspends or constrains the parent.
+    const spawnAt = patch14.indexOf('The sub-agent could not be started.');
+    assert.ok(spawnAt !== -1, 'spawn failure site missing');
+    const spawnSlice = patch14.slice(Math.max(0, spawnAt - 2000), spawnAt + 2000);
+    assert.ok(!spawnSlice.includes('suspendAiRequest'), 'spawn failure must not suspend the parent');
+    assert.ok(!spawnSlice.includes("status: 'error'"), 'spawn failure must not mark the parent error');
+    assert.ok(!spawnSlice.includes('guard.recordError'), 'spawn failure must not touch the parent streak');
+    // Parent stays alive: the failure is recorded AND sent as a normal tool error.
+    assert.ok(
+      patch14.includes('spawnFailureResults.push(failure)'),
+      'spawn failure must be pushed as a tool result'
+    );
+    assert.ok(
+      patch14.includes('...spawnFailureResults,'),
+      'spawn failures must be sent to the model for serial fallback'
+    );
+  });
+
+  it('patch 14 still dry-runs clean with isolation (real I/O, anti-revert intact)', async () => {
+    const { spawnSync } = await import('node:child_process');
+    const fs = await import('node:fs');
+    const os = await import('node:os');
+    const path = await import('node:path');
+    const { fileURLToPath } = await import('node:url');
+    const root = path.dirname(
+      path.dirname(path.dirname(fileURLToPath(import.meta.url)))
+    );
+    const upstream =
+      process.env.UPSTREAM_DIR || path.join(root, 'review', 'GDevelop');
+    const rel = 'newIDE/app/src/AiGeneration/Utils.js';
+    assert.ok(
+      fs.existsSync(path.join(upstream, rel)),
+      `missing fixture ${rel} (set UPSTREAM_DIR)`
+    );
+    const dir = fs.mkdtempSync(path.join(os.tmpdir(), 't06-p14-'));
+    try {
+      const dest = path.join(dir, rel);
+      fs.mkdirSync(path.dirname(dest), { recursive: true });
+      fs.copyFileSync(path.join(upstream, rel), dest);
+      const r = spawnSync(
+        'patch',
+        ['-p1', '-s', '--dry-run', '-i', path.join(root, 'patches/14-ai-utils.patch')],
+        { cwd: dir, encoding: 'utf8' }
+      );
+      assert.equal(r.status, 0, `patch 14 dry-run failed:\n${r.stdout}\n${r.stderr}`);
+      const patchText = fs.readFileSync(path.join(root, 'patches/14-ai-utils.patch'), 'utf8');
+      assert.ok(
+        !/^-[^-].*onSaveProjectAsWithStorageProvider/m.test(patchText),
+        'patch 14 deletes upstream-new symbol onSaveProjectAsWithStorageProvider'
+      );
+    } finally {
+      fs.rmSync(dir, { recursive: true, force: true });
+    }
+  });
+});