Durable mcp

[pedrofrxncx]

Summary by cubic

Introduces a durable workflow engine with persisted executions, event-driven scheduling, and a sandboxed TypeScript runner. Adds workflow/execution tools, a Connection binding and gateway transport for MCPs, and an Event Bus handler for workflow events.

• New Features

  • Durable workflow runtime with checkpoints, retries/timeouts, cancellation, and unified events/signals.
  • DAG-based executor for parallel step execution and @ref resolution.
  • QuickJS sandbox for code steps (console, fetch) with TypeScript transpilation via sucrase.
  • Workflow collections/tools: CRUD for workflows and executions, start/resume/cancel, list/get step results.
  • PostgreSQL setup with idempotent table/index creation and locking.
  • Event Bus binding with a runtime handler for "workflow.execution.created".
  • Connection Binding and MCP gateway transport for tool steps; executions track gateway_id.
  • Workflow validation on create: step type checks, @ref validation, and TypeScript validation for code steps.

• Migration

  • Configure DATABASE, EVENT_BUS, and CONNECTION bindings; tables are created on startup.
  • Replace Meta Ads secrets with META_ACCESS_TOKEN in GitHub workflows and env; docs updated to use direct token auth.
  • Install new deps (QuickJS, sucrase, @ai-sdk/mcp, @modelcontextprotocol/sdk).

^{Written for commit bb91b6c. Summary will update on new commits.}

[cubic-dev-ai]

25 issues found across 42 files

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="mcp-studio/server/sandbox/index.ts">

<violation number="1" location="mcp-studio/server/sandbox/index.ts:117">
P1: Setting the interrupt handler on the shared runtime inside `createContext` causes a race condition. Since the runtime is cached and shared (line 73), each new context overwrites the previous interrupt handler, meaning only the last context&#39;s deadline will be enforced. Earlier contexts will not have their timeouts respected, which could allow runaway code execution in a sandbox environment.</violation>
</file>

<file name="mcp-studio/server/tools/execution.ts">

<violation number="1" location="mcp-studio/server/tools/execution.ts:150">
P2: Missing `outputSchema` in `createGetTool`. All other tools in this file define an outputSchema, and `GET_BINDING.outputSchema` is validated to exist on line 28. Consider adding `outputSchema: GET_BINDING.outputSchema` for consistency.</violation>
</file>

<file name="connection-binding/app.json">

<violation number="1" location="connection-binding/app.json:14">
P2: The `inputSchema` for `COLLECTION_CONNECTIONS_GET` is missing `required` and `additionalProperties` fields. For a GET operation, the `id` parameter should be required. Without this, the schema allows calling GET without an id, which would be invalid.</violation>
</file>

<file name="mcp-studio/server/db/schemas/workflow.ts">

<violation number="1" location="mcp-studio/server/db/schemas/workflow.ts:31">
P2: Missing FOREIGN KEY constraint on `workflow_collection_id`. Other tables in this schema (like `workflow_execution_step_result` and `workflow_event`) properly define foreign key constraints. Consider adding `FOREIGN KEY (workflow_collection_id) REFERENCES workflow_collection(id)` for data integrity.</violation>

<violation number="2" location="mcp-studio/server/db/schemas/workflow.ts:49">
P2: Missing FOREIGN KEY constraint on `workflow_id`. Other tables in this schema properly define foreign key constraints. Consider adding `FOREIGN KEY (workflow_id) REFERENCES workflow(id)` for data integrity, or use inline syntax.</violation>
</file>

<file name="mcp-studio/server/tools/workflow.ts">

<violation number="1" location="mcp-studio/server/tools/workflow.ts:246">
P2: Malformed JSON example in the description. The step definition is truncated and starts with `true }` which is invalid JSON. This will confuse users trying to understand the workflow syntax. The step should include the full definition with `name` and `action` properties.</violation>
</file>

<file name="mcp-studio/server/db/transformers.ts">

<violation number="1" location="mcp-studio/server/db/transformers.ts:115">
P2: Using `Number(row.visible_at)` directly can produce `NaN` (if undefined) or `0` (if null) without validation. Consider using `toNumberOrNull` or adding explicit null handling like the `consumed_at` field does.</violation>

<violation number="2" location="mcp-studio/server/db/transformers.ts:117">
P2: This duplicates the `safeJsonParse` logic defined earlier in the file. Using `payload: safeJsonParse(row.payload)` would be more consistent and includes error handling that this inline version lacks.</violation>
</file>

<file name="mcp-studio/server/engine/steps/code-step.ts">

<violation number="1" location="mcp-studio/server/engine/steps/code-step.ts:173">
P1: Memory leak: `runtime` is created with a unique ID but never disposed. Only `ctx` is disposed in the finally block, leaving the runtime cached indefinitely. Use `using` syntax or manually call `runtime[Symbol.dispose]()` in the finally block.</violation>
</file>

<file name="mcp-studio/server/tools/_helpers.ts">

<violation number="1" location="mcp-studio/server/tools/_helpers.ts:72">
P0: SQL injection vulnerability: `fieldName` is interpolated directly into SQL without validation. While values use parameterized queries (`?`), identifiers are not protected. Consider validating field names against an allowlist of permitted columns.</violation>

<violation number="2" location="mcp-studio/server/tools/_helpers.ts:151">
P0: SQL injection vulnerability: `direction` is directly interpolated into SQL without validation. It should be restricted to only &#39;ASC&#39; or &#39;DESC&#39;.</violation>

<violation number="3" location="mcp-studio/server/tools/_helpers.ts:152">
P1: SQL injection vulnerability: `nulls` is directly interpolated into SQL without validation. It should be restricted to only &#39;FIRST&#39; or &#39;LAST&#39;.</violation>
</file>

<file name="mcp-studio/server/sandbox/builtins/fetch.ts">

<violation number="1" location="mcp-studio/server/sandbox/builtins/fetch.ts:45">
P2: Memory leak: handles created by `ctx.newNumber()` are not tracked or disposed. In QuickJS-emscripten, handles passed to `setProp` should be disposed after use, or tracked for later cleanup.</violation>
</file>

<file name="mcp-studio/server/engine/context.ts">

<violation number="1" location="mcp-studio/server/engine/context.ts:34">
P2: If `getExecution` returns `null` (execution not found), this method silently continues without error. A missing execution during cancellation check could indicate data corruption or an invalid `executionId`. Consider throwing an error when the execution doesn&#39;t exist.</violation>
</file>

<file name="mcp-studio/server/sandbox/utils/call-function.ts">

<violation number="1" location="mcp-studio/server/sandbox/utils/call-function.ts:11">
P1: Memory leak: QuickJS handles (`thisArgHandle`, `argHandles`, `result`, `resultPromise`) are never disposed. Each call leaks memory since QuickJS requires manual handle cleanup.</violation>
</file>

<file name="mcp-studio/server/types/env.ts">

<violation number="1" location="mcp-studio/server/types/env.ts:45">
P2: The optional parameter contradicts the comment stating MCP rejects undefined inputs. If MCP validation truly rejects undefined, this parameter should be required, not optional.</violation>
</file>

<file name="bun.lock">

<violation number="1" location="bun.lock:2693">
P0: Absolute path from a developer&#39;s local machine was accidentally committed to the lock file (`file:/Users/pedrofranca/Documents/dev/decocms/mesh/packages/bindings`). This will cause build failures for anyone else and leaks local filesystem information.</violation>
</file>

<file name="mcp-studio/server/engine/executor.ts">

<violation number="1" location="mcp-studio/server/engine/executor.ts:182">
P1: The `updateStepResult` call is not awaited. This async function may not complete before the error is re-thrown, potentially leaving step failures unrecorded in the database and causing an unhandled promise rejection.</violation>
</file>

<file name="mcp-studio/server/events/handler.ts">

<violation number="1" location="mcp-studio/server/events/handler.ts:22">
P1: Fire-and-forget async calls may cause lost workflow executions. The function spawns async `executeWorkflow` calls without awaiting them or returning the promises. If the execution context terminates (common in serverless/workers), pending workflows will be dropped. Consider making the function async and using `Promise.allSettled()` to await all executions, or use `waitUntil()` if in a Cloudflare Worker context.</violation>
</file>

<file name="mcp-studio/server/db/queries/executions.ts">

<violation number="1" location="mcp-studio/server/db/queries/executions.ts:310">
P1: Return type mismatch: When `ON CONFLICT DO NOTHING` triggers (race condition), `result[0]` will be `undefined` but the return type promises `WorkflowExecutionStepResult`. Either change return type to `WorkflowExecutionStepResult | null` or use `ON CONFLICT DO UPDATE ... RETURNING *` to always return a result.</violation>

<violation number="2" location="mcp-studio/server/db/queries/executions.ts:419">
P1: Pagination is broken: `totalCount` is set to the returned result count (after LIMIT), not the actual total in the database. This makes `hasMore` always false. You need a separate `COUNT(*)` query to get the true total count.</violation>
</file>

<file name="mcp-studio/server/sandbox/utils/to-quickjs.ts">

<violation number="1" location="mcp-studio/server/sandbox/utils/to-quickjs.ts:23">
P1: Memory leak: QuickJS handle `hv` is not disposed after `ctx.setProp()`. Each object property conversion leaks memory since handles are not freed after use.</violation>
</file>

<file name="mcp-studio/server/utils/ref-resolver.ts">

<violation number="1" location="mcp-studio/server/utils/ref-resolver.ts:60">
P1: The condition `refStr.startsWith(&quot;input&quot;)` incorrectly matches step names starting with &#39;input&#39; (e.g., `@inputName`, `@inputData`) as input references. Should use the same pattern as &#39;item&#39;: check for exact match OR prefix with dot.</violation>
</file>

<file name="mcp-studio/server/engine/steps/tool-step.ts">

<violation number="1" location="mcp-studio/server/engine/steps/tool-step.ts:188">
P1: MCP client connection is never closed, causing a resource leak. The client should be closed in a `finally` block to ensure cleanup regardless of success or error.</violation>
</file>

<file name="mcp-studio/package.json">

<violation number="1" location="mcp-studio/package.json:18">
P1: Local file references (`file:../../mesh/packages/...`) point to paths outside the repository that don&#39;t exist. This will break package installation in CI/CD and for other developers. Consider using the original npm package versions, workspace references, or ensuring the path exists in the monorepo structure.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P1: Setting the interrupt handler on the shared runtime inside createContext causes a race condition. Since the runtime is cached and shared (line 73), each new context overwrites the previous interrupt handler, meaning only the last context's deadline will be enforced. Earlier contexts will not have their timeouts respected, which could allow runaway code execution in a sandbox environment.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/sandbox/index.ts, line 117:

<comment>Setting the interrupt handler on the shared runtime inside `createContext` causes a race condition. Since the runtime is cached and shared (line 73), each new context overwrites the previous interrupt handler, meaning only the last context&#39;s deadline will be enforced. Earlier contexts will not have their timeouts respected, which could allow runaway code execution in a sandbox environment.</comment>

<file context>
@@ -0,0 +1,146 @@
+    };
+
+    // Set up interrupt handler for this context
+    runtime.setInterruptHandler(() =&gt; {
+      const shouldInterrupt = deadline &gt; 0 &amp;&amp; Date.now() &gt; deadline;
+      if (shouldInterrupt) {
</file context>

[cubic-dev-ai]

P2: Missing outputSchema in createGetTool. All other tools in this file define an outputSchema, and GET_BINDING.outputSchema is validated to exist on line 28. Consider adding outputSchema: GET_BINDING.outputSchema for consistency.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/tools/execution.ts, line 150:

<comment>Missing `outputSchema` in `createGetTool`. All other tools in this file define an outputSchema, and `GET_BINDING.outputSchema` is validated to exist on line 28. Consider adding `outputSchema: GET_BINDING.outputSchema` for consistency.</comment>

<file context>
@@ -0,0 +1,205 @@
+  createPrivateTool({
+    id: &quot;COLLECTION_WORKFLOW_EXECUTION_GET&quot;,
+    description: &quot;Get a single workflow execution by ID with step results&quot;,
+    inputSchema: GET_BINDING.inputSchema,
+    execute: async ({
+      context,
</file context>

[cubic-dev-ai]

P2: The inputSchema for COLLECTION_CONNECTIONS_GET is missing required and additionalProperties fields. For a GET operation, the id parameter should be required. Without this, the schema allows calling GET without an id, which would be invalid.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At connection-binding/app.json, line 14:

<comment>The `inputSchema` for `COLLECTION_CONNECTIONS_GET` is missing `required` and `additionalProperties` fields. For a GET operation, the `id` parameter should be required. Without this, the schema allows calling GET without an id, which would be invalid.</comment>

<file context>
@@ -0,0 +1,617 @@
+    {
+      &quot;name&quot;: &quot;COLLECTION_CONNECTIONS_UPDATE&quot;,
+      &quot;description&quot;: &quot;Update an existing MCP connection in the organization&quot;,
+      &quot;inputSchema&quot;: {
+        &quot;type&quot;: &quot;object&quot;,
+        &quot;properties&quot;: {
</file context>

[cubic-dev-ai]

P2: Missing FOREIGN KEY constraint on workflow_id. Other tables in this schema properly define foreign key constraints. Consider adding FOREIGN KEY (workflow_id) REFERENCES workflow(id) for data integrity, or use inline syntax.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/db/schemas/workflow.ts, line 49:

<comment>Missing FOREIGN KEY constraint on `workflow_id`. Other tables in this schema properly define foreign key constraints. Consider adding `FOREIGN KEY (workflow_id) REFERENCES workflow(id)` for data integrity, or use inline syntax.</comment>

<file context>
@@ -0,0 +1,165 @@
+const postgresWorkflowExecutionTableIdempotentQuery = `
+CREATE TABLE IF NOT EXISTS workflow_execution (
+  id TEXT PRIMARY KEY,
+  workflow_id TEXT NOT NULL,
+  status TEXT NOT NULL CHECK(status IN (&#39;enqueued&#39;, &#39;cancelled&#39;, &#39;success&#39;, &#39;error&#39;, &#39;running&#39;)),
+  input JSONB,
</file context>

[cubic-dev-ai]

P2: Missing FOREIGN KEY constraint on workflow_collection_id. Other tables in this schema (like workflow_execution_step_result and workflow_event) properly define foreign key constraints. Consider adding FOREIGN KEY (workflow_collection_id) REFERENCES workflow_collection(id) for data integrity.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/db/schemas/workflow.ts, line 31:

<comment>Missing FOREIGN KEY constraint on `workflow_collection_id`. Other tables in this schema (like `workflow_execution_step_result` and `workflow_event`) properly define foreign key constraints. Consider adding `FOREIGN KEY (workflow_collection_id) REFERENCES workflow_collection(id)` for data integrity.</comment>

<file context>
@@ -0,0 +1,165 @@
+const postgresWorkflowTableIdempotentQuery = `
+  CREATE TABLE IF NOT EXISTS workflow (
+    id TEXT PRIMARY KEY,
+    workflow_collection_id TEXT,
+    steps JSONB NOT NULL DEFAULT &#39;{}&#39;,
+    input JSONB,
</file context>

[cubic-dev-ai]

P1: Pagination is broken: totalCount is set to the returned result count (after LIMIT), not the actual total in the database. This makes hasMore always false. You need a separate COUNT(*) query to get the true total count.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/db/queries/executions.ts, line 419:

<comment>Pagination is broken: `totalCount` is set to the returned result count (after LIMIT), not the actual total in the database. This makes `hasMore` always false. You need a separate `COUNT(*)` query to get the true total count.</comment>

<file context>
@@ -0,0 +1,555 @@
+    [...params, limit, offset],
+  );
+
+  const totalCount = result.length;
+
+  return {
</file context>

[cubic-dev-ai]

P1: Memory leak: QuickJS handle hv is not disposed after ctx.setProp(). Each object property conversion leaks memory since handles are not freed after use.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/sandbox/utils/to-quickjs.ts, line 23:

<comment>Memory leak: QuickJS handle `hv` is not disposed after `ctx.setProp()`. Each object property conversion leaks memory since handles are not freed after use.</comment>

<file context>
@@ -0,0 +1,126 @@
+      if (Array.isArray(value)) {
+        const arr = ctx.newArray();
+        value.forEach((v, i) =&gt; {
+          const hv = toQuickJS(ctx, v);
+          ctx.setProp(arr, String(i), hv);
+        });
</file context>

[cubic-dev-ai]

P1: The condition refStr.startsWith("input") incorrectly matches step names starting with 'input' (e.g., @inputName, @inputData) as input references. Should use the same pattern as 'item': check for exact match OR prefix with dot.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/utils/ref-resolver.ts, line 60:

<comment>The condition `refStr.startsWith(&quot;input&quot;)` incorrectly matches step names starting with &#39;input&#39; (e.g., `@inputName`, `@inputData`) as input references. Should use the same pattern as &#39;item&#39;: check for exact match OR prefix with dot.</comment>

<file context>
@@ -0,0 +1,299 @@
+  }
+
+  // Input reference: @input.path.to.value
+  if (refStr.startsWith(&quot;input&quot;)) {
+    const path = refStr.length &gt; 5 ? refStr.substring(6) : &quot;&quot;; // Remove &#39;input.&#39; or &#39;input&#39;
+    return { type: &quot;input&quot;, path };
</file context>

[cubic-dev-ai]

P1: MCP client connection is never closed, causing a resource leak. The client should be closed in a finally block to ensure cleanup regardless of success or error.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/engine/steps/tool-step.ts, line 188:

<comment>MCP client connection is never closed, causing a resource leak. The client should be closed in a `finally` block to ensure cleanup regardless of success or error.</comment>

<file context>
@@ -0,0 +1,258 @@
+      },
+    ],
+  });
+  await client.connect(transport);
+
+  // Fetch tool schema for type coercion
</file context>

[cubic-dev-ai]

P1: Local file references (file:../../mesh/packages/...) point to paths outside the repository that don't exist. This will break package installation in CI/CD and for other developers. Consider using the original npm package versions, workspace references, or ensuring the path exists in the monorepo structure.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/package.json, line 18:

<comment>Local file references (`file:../../mesh/packages/...`) point to paths outside the repository that don&#39;t exist. This will break package installation in CI/CD and for other developers. Consider using the original npm package versions, workspace references, or ensuring the path exists in the monorepo structure.</comment>

<file context>
@@ -14,40 +14,40 @@
-    &quot;@decocms/bindings&quot;: &quot;1.0.1-alpha.26&quot;,
-    &quot;@decocms/runtime&quot;: &quot;1.0.0-alpha.31&quot;,
+    &quot;@ai-sdk/mcp&quot;: &quot;^1.0.1&quot;,
+    &quot;@decocms/bindings&quot;: &quot;file:../../mesh/packages/bindings&quot;,
+    &quot;@decocms/runtime&quot;: &quot;file:../../mesh/packages/runtime&quot;,
+    &quot;@jitl/quickjs-wasmfile-release-sync&quot;: &quot;^0.31.0&quot;,
</file context>

✅ Addressed in b34caa5

[pedrofrxncx]

@cubic-ai

[cubic-dev-ai]

@cubic-ai

@pedrofrxncx I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

23 issues found across 54 files

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="mcp-studio/server/sandbox/builtins/fetch.ts">

<violation number="1" location="mcp-studio/server/sandbox/builtins/fetch.ts:29">
P2: The `_headers` variable is parsed from init options but never used. Headers passed to the Response constructor are silently ignored. Either implement the headers functionality or remove the dead code.</violation>
</file>

<file name="mcp-studio/server/types/env.ts">

<violation number="1" location="mcp-studio/server/types/env.ts:45">
P2: Optional parameter contradicts the comment. If MCP tool validation rejects `undefined` inputs, then `params` should be required (not optional) to ensure callers always pass an empty object `{}`.</violation>
</file>

<file name="mcp-studio/server/tools/workflow.ts">

<violation number="1" location="mcp-studio/server/tools/workflow.ts:246">
P2: Malformed JSON example in documentation. Line 229 starts with `true },` which is invalid - the first step definition is truncated. It should likely be something like `{ &quot;name&quot;: &quot;fetch_users&quot;, &quot;action&quot;: { &quot;toolName&quot;: &quot;GET_USERS&quot; }, &quot;transformCode&quot;: &quot;export default async (i) =&gt; i[0]&quot; },`</violation>
</file>

<file name="mcp-studio/server/engine/executor.ts">

<violation number="1" location="mcp-studio/server/engine/executor.ts:182">
P1: Missing `await` on async `updateStepResult` call. The database update may not complete before the error is thrown, and any database errors will be silently lost as unhandled promise rejections.</violation>
</file>

<file name="mcp-studio/server/engine/context.ts">

<violation number="1" location="mcp-studio/server/engine/context.ts:32">
P1: `checkCancelled()` silently succeeds if the execution doesn&#39;t exist. When `getExecution` returns null/undefined, the optional chaining causes the status check to evaluate to false, and no error is thrown. Consider throwing an error when the execution is not found to prevent operations on non-existent executions.</violation>
</file>

<file name="mcp-studio/server/tools/_helpers.ts">

<violation number="1" location="mcp-studio/server/tools/_helpers.ts:77">
P0: SQL injection vulnerability: `fieldName` is directly interpolated into SQL without sanitization. While values are properly parameterized, field names are not. Consider validating field names against an allowlist or using a safe identifier escaping function.</violation>

<violation number="2" location="mcp-studio/server/tools/_helpers.ts:151">
P0: SQL injection vulnerability: `direction` is not validated before interpolation. It should be restricted to &#39;ASC&#39; or &#39;DESC&#39; only.</violation>

<violation number="3" location="mcp-studio/server/tools/_helpers.ts:152">
P0: SQL injection vulnerability: `nulls` is not validated before interpolation. It should be restricted to &#39;FIRST&#39; or &#39;LAST&#39; only.</violation>
</file>

<file name="mcp-studio/server/db/schemas/workflow.ts">

<violation number="1" location="mcp-studio/server/db/schemas/workflow.ts:49">
P2: Missing FOREIGN KEY constraint on `workflow_id`. This column references the `workflow` table but has no FK constraint, unlike similar relationships elsewhere in this schema (e.g., `workflow_execution_step_result.execution_id` and `workflow_event.execution_id` both have FK constraints). This could lead to orphaned executions referencing non-existent workflows.</violation>
</file>

<file name="mcp-studio/server/utils/ref-resolver.ts">

<violation number="1" location="mcp-studio/server/utils/ref-resolver.ts:60">
P1: Bug: `startsWith(&quot;input&quot;)` incorrectly matches step names like `@inputs` or `@inputData`. Should use the same pattern as the &quot;item&quot; case: exact match OR dot prefix.</violation>
</file>

<file name="mcp-studio/server/engine/steps/code-step.ts">

<violation number="1" location="mcp-studio/server/engine/steps/code-step.ts:37">
P2: Regex `[^}]*` will fail for interfaces with nested object types. The pattern stops at the first `}`, so `interface Input { user: { name: string }; }` would extract incomplete content. Consider using a brace-counting approach or a proper parser, or document this as a known limitation for simple interfaces only.</violation>
</file>

<file name=".github/workflows/SECRETS.md">

<violation number="1" location=".github/workflows/SECRETS.md:38">
P2: Missing warning about short-lived tokens. Tokens from Graph API Explorer typically expire in 1-2 hours. For production deployments via GitHub Actions, users should be warned to exchange for a long-lived token (60 days) or the deployment will break. Consider adding a note like:
  - ⚠️ **Important**: Tokens from Graph API Explorer are short-lived. For production, exchange for a long-lived token (see meta-ads/README.md)</violation>
</file>

<file name="connection-binding/app.json">

<violation number="1" location="connection-binding/app.json:14">
P1: The `inputSchema` for `COLLECTION_CONNECTIONS_GET` is missing `required: [&quot;id&quot;]` and `additionalProperties: false`. Without marking `id` as required, the schema allows GET requests without specifying which connection to retrieve, which would fail at runtime.</violation>
</file>

<file name="mcp-studio/server/db/queries/executions.ts">

<violation number="1" location="mcp-studio/server/db/queries/executions.ts:426">
P1: Pagination logic is broken: `totalCount` is set to the returned items count instead of actual total count from database. This causes `hasMore` to always be `false`. A separate `COUNT(*)` query is needed for accurate pagination.</violation>
</file>

<file name="mcp-studio/server/sandbox/utils/call-function.ts">

<violation number="1" location="mcp-studio/server/sandbox/utils/call-function.ts:11">
P1: Memory leak: QuickJS handles are created but never disposed. Following the pattern in other files (e.g., `to-quickjs.ts`, `console.ts`, `fetch.ts`), handles should be disposed after use. Consider wrapping in try/finally to dispose `thisArgHandle`, `argHandles`, `result`, and `resultPromise.handle`.</violation>
</file>

<file name="mcp-studio/server/sandbox/index.ts">

<violation number="1" location="mcp-studio/server/sandbox/index.ts:112">
P2: The `setDeadline` function is defined but never exposed to callers. It&#39;s not returned or attached to the context, making it inaccessible and effectively dead code. If dynamic deadline updates are intended, consider exposing this function; otherwise, remove it to reduce confusion.</violation>

<violation number="2" location="mcp-studio/server/sandbox/index.ts:117">
P1: Interrupt handler is set per-context but QuickJS only supports one handler per runtime. When multiple contexts are created, only the last context&#39;s deadline will be checked. Earlier contexts will never be interrupted, potentially causing runaway executions.

Consider moving deadline tracking to a shared structure at the runtime level (e.g., a Map of context IDs to deadlines), and have a single interrupt handler check all active deadlines.</violation>
</file>

<file name="mcp-studio/server/tools/execution.ts">

<violation number="1" location="mcp-studio/server/tools/execution.ts:30">
P2: Copy-paste error: Error message references wrong binding name. This checks `CREATE_BINDING` but the error message says `COLLECTION_WORKFLOW_EXECUTION_GET` instead of `COLLECTION_WORKFLOW_EXECUTION_CREATE`.</violation>
</file>

<file name="mcp-studio/server/engine/steps/tool-step.ts">

<violation number="1" location="mcp-studio/server/engine/steps/tool-step.ts:136">
P3: Comment says &quot;force http if not local&quot; but code sets `https:`. The comment should say &quot;force https if not local&quot;.</violation>
</file>

<file name="mcp-studio/server/sandbox/utils/to-quickjs.ts">

<violation number="1" location="mcp-studio/server/sandbox/utils/to-quickjs.ts:23">
P1: Memory leak: QuickJS handles created for object properties are not disposed after `setProp`. Add `hv.dispose()` after the `setProp` call.</violation>
</file>

<file name="mcp-studio/server/events/handler.ts">

<violation number="1" location="mcp-studio/server/events/handler.ts:22">
P1: Unwaited async operation in event handler. `executeWorkflow` returns a Promise but is not awaited, so the function returns before workflow execution completes. In event bus handlers, this can cause the runtime to terminate early, dropping workflow executions. Consider making the handler async and using `Promise.allSettled()` to wait for all executions.</violation>
</file>

<file name=".github/workflows/deploy.yml">

<violation number="1" location=".github/workflows/deploy.yml:77">
P1: The `META_ACCESS_TOKEN` secret won&#39;t be deployed. The `scripts/deploy.ts` file still references `META_APP_ID` and `META_APP_SECRET` in the `envVarsToPass` array (lines 149-150) but doesn&#39;t include the new `META_ACCESS_TOKEN`. This means the secret is available in the runner but won&#39;t be propagated to the deployed MCPs.</violation>
</file>

<file name="meta-ads/package.json">

<violation number="1" location="meta-ads/package.json:12">
P2: Replacing `tsc --noEmit` with `oxlint` removes type checking from the `check` script. These tools serve different purposes—`tsc` catches type errors while `oxlint` catches linting/style issues. Consider keeping both: `&quot;check&quot;: &quot;tsc --noEmit &amp;&amp; bun run oxlint server/&quot;`.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P2: The _headers variable is parsed from init options but never used. Headers passed to the Response constructor are silently ignored. Either implement the headers functionality or remove the dead code.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/sandbox/builtins/fetch.ts, line 29:

<comment>The `_headers` variable is parsed from init options but never used. Headers passed to the Response constructor are silently ignored. Either implement the headers functionality or remove the dead code.</comment>

<file context>
@@ -0,0 +1,186 @@
+        let status = 200;
+        let statusText = &quot;OK&quot;;
+        // @ts-ignore
+        let _headers: Record&lt;string, string&gt; = {};
+
+        if (init &amp;&amp; init !== ctx.null &amp;&amp; init !== ctx.undefined) {
</file context>

✅ Addressed in 2a62a6c

[cubic-dev-ai]

P2: Optional parameter contradicts the comment. If MCP tool validation rejects undefined inputs, then params should be required (not optional) to ensure callers always pass an empty object {}.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/types/env.ts, line 45:

<comment>Optional parameter contradicts the comment. If MCP tool validation rejects `undefined` inputs, then `params` should be required (not optional) to ensure callers always pass an empty object `{}`.</comment>

<file context>
@@ -0,0 +1,66 @@
+    };
+  }&gt;;
+  // Accepts an (empty) object because MCP tool validation rejects `undefined` inputs.
+  COLLECTION_CONNECTIONS_LIST: (params?: Record&lt;string, never&gt;) =&gt; Promise&lt;{
+    items: {
+      id: string;
</file context>

✅ Addressed in e5155fe

[cubic-dev-ai]

P2: Malformed JSON example in documentation. Line 229 starts with true }, which is invalid - the first step definition is truncated. It should likely be something like { "name": "fetch_users", "action": { "toolName": "GET_USERS" }, "transformCode": "export default async (i) => i[0]" },

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/tools/workflow.ts, line 246:

<comment>Malformed JSON example in documentation. Line 229 starts with `true },` which is invalid - the first step definition is truncated. It should likely be something like `{ &quot;name&quot;: &quot;fetch_users&quot;, &quot;action&quot;: { &quot;toolName&quot;: &quot;GET_USERS&quot; }, &quot;transformCode&quot;: &quot;export default async (i) =&gt; i[0]&quot; },`</comment>

<file context>
@@ -0,0 +1,379 @@
+
+Example workflow with a step that references the output of another step:
+{ &quot;title&quot;: &quot;Get first user and then fetch orders&quot;, &quot;steps&quot;: [
+  true }, &quot;transformCode&quot;: &quot;export default async (i) =&gt; i[0]&quot; },
+  { &quot;name&quot;: &quot;fetch_orders&quot;, &quot;action&quot;: { &quot;toolName&quot;: &quot;GET_ORDERS&quot; }, &quot;input&quot;: { &quot;user&quot;: &quot;@fetch_users.user&quot; } },
+]}
</file context>

[cubic-dev-ai]

P1: Missing await on async updateStepResult call. The database update may not complete before the error is thrown, and any database errors will be silently lost as unhandled promise rejections.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/engine/executor.ts, line 182:

<comment>Missing `await` on async `updateStepResult` call. The database update may not complete before the error is thrown, and any database errors will be silently lost as unhandled promise rejections.</comment>

<file context>
@@ -0,0 +1,219 @@
+        return { step, result };
+      } catch (err) {
+        console.error(&quot;Step failed&quot;, step.name, err);
+        updateStepResult(env, executionId, step.name, {
+          error: err instanceof Error ? err.message : String(err),
+          completed_at_epoch_ms: Date.now(),
</file context>

[cubic-dev-ai]

P1: checkCancelled() silently succeeds if the execution doesn't exist. When getExecution returns null/undefined, the optional chaining causes the status check to evaluate to false, and no error is thrown. Consider throwing an error when the execution is not found to prevent operations on non-existent executions.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/engine/context.ts, line 32:

<comment>`checkCancelled()` silently succeeds if the execution doesn&#39;t exist. When `getExecution` returns null/undefined, the optional chaining causes the status check to evaluate to false, and no error is thrown. Consider throwing an error when the execution is not found to prevent operations on non-existent executions.</comment>

<file context>
@@ -0,0 +1,73 @@
+    return this.env.MESH_REQUEST_CONTEXT?.meshUrl ?? &quot;&quot;;
+  }
+
+  async checkCancelled(): Promise&lt;void&gt; {
+    const execution = await getExecution(this.env, this.executionId);
+    if (execution?.status === &quot;cancelled&quot;) {
</file context>

[cubic-dev-ai]

P3: Comment says "force http if not local" but code sets https:. The comment should say "force https if not local".

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/engine/steps/tool-step.ts, line 136:

<comment>Comment says &quot;force http if not local&quot; but code sets `https:`. The comment should say &quot;force https if not local&quot;.</comment>

<file context>
@@ -0,0 +1,258 @@
+const fixProtocol = (url: URL) =&gt; {
+  const isLocal = url.hostname === &quot;localhost&quot; || url.hostname === &quot;127.0.0.1&quot;;
+  if (!isLocal) {
+    // force http if not local
+    url.protocol = &quot;https:&quot;;
+  }
</file context>

[cubic-dev-ai]

P1: Memory leak: QuickJS handles created for object properties are not disposed after setProp. Add hv.dispose() after the setProp call.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/sandbox/utils/to-quickjs.ts, line 23:

<comment>Memory leak: QuickJS handles created for object properties are not disposed after `setProp`. Add `hv.dispose()` after the `setProp` call.</comment>

<file context>
@@ -0,0 +1,126 @@
+      if (Array.isArray(value)) {
+        const arr = ctx.newArray();
+        value.forEach((v, i) =&gt; {
+          const hv = toQuickJS(ctx, v);
+          ctx.setProp(arr, String(i), hv);
+        });
</file context>

[cubic-dev-ai]

P1: Unwaited async operation in event handler. executeWorkflow returns a Promise but is not awaited, so the function returns before workflow execution completes. In event bus handlers, this can cause the runtime to terminate early, dropping workflow executions. Consider making the handler async and using Promise.allSettled() to wait for all executions.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/events/handler.ts, line 22:

<comment>Unwaited async operation in event handler. `executeWorkflow` returns a Promise but is not awaited, so the function returns before workflow execution completes. In event bus handlers, this can cause the runtime to terminate early, dropping workflow executions. Consider making the handler async and using `Promise.allSettled()` to wait for all executions.</comment>

<file context>
@@ -0,0 +1,34 @@
+/**
+ * Handle a batch of workflow events.
+ */
+export function handleWorkflowEvents(events: WorkflowEvent[], env: Env): void {
+  for (const event of events) {
+    if (!event.subject) continue;
</file context>

[cubic-dev-ai]

P1: The META_ACCESS_TOKEN secret won't be deployed. The scripts/deploy.ts file still references META_APP_ID and META_APP_SECRET in the envVarsToPass array (lines 149-150) but doesn't include the new META_ACCESS_TOKEN. This means the secret is available in the runner but won't be propagated to the deployed MCPs.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At .github/workflows/deploy.yml, line 77:

<comment>The `META_ACCESS_TOKEN` secret won&#39;t be deployed. The `scripts/deploy.ts` file still references `META_APP_ID` and `META_APP_SECRET` in the `envVarsToPass` array (lines 149-150) but doesn&#39;t include the new `META_ACCESS_TOKEN`. This means the secret is available in the runner but won&#39;t be propagated to the deployed MCPs.</comment>

<file context>
@@ -74,8 +74,7 @@ jobs:
           PERPLEXITY_API_KEY: ${{ secrets.PERPLEXITY_API_KEY }}
-          META_APP_ID: ${{ secrets.META_APP_ID }}
-          META_APP_SECRET: ${{ secrets.META_APP_SECRET }}
+          META_ACCESS_TOKEN: ${{ secrets.META_ACCESS_TOKEN }}
 
       - name: Notify success
</file context>

[cubic-dev-ai]

P2: Replacing tsc --noEmit with oxlint removes type checking from the check script. These tools serve different purposes—tsc catches type errors while oxlint catches linting/style issues. Consider keeping both: "check": "tsc --noEmit && bun run oxlint server/".

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At meta-ads/package.json, line 12:

<comment>Replacing `tsc --noEmit` with `oxlint` removes type checking from the `check` script. These tools serve different purposes—`tsc` catches type errors while `oxlint` catches linting/style issues. Consider keeping both: `&quot;check&quot;: &quot;tsc --noEmit &amp;&amp; bun run oxlint server/&quot;`.</comment>

<file context>
@@ -9,7 +9,7 @@
     &quot;build&quot;: &quot;bun run build:server&quot;,
     &quot;publish&quot;: &quot;cat app.json | deco registry publish -w /shared/deco -y&quot;,
-    &quot;check&quot;: &quot;tsc --noEmit&quot;,
+    &quot;check&quot;: &quot;bun run oxlint server/&quot;,
     &quot;dev:tunnel&quot;: &quot;deco link -p 3003 -- PORT=3003 bun run dev&quot;
   },
</file context>

✅ Addressed in 27da5de

[cubic-dev-ai]

1 issue found across 6 files (changes from recent commits).

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="mcp-studio/server/tools/workflow.ts">

<violation number="1" location="mcp-studio/server/tools/workflow.ts:305">
P0: Typo introduces SQL syntax error. The stray `y` character at the beginning of the `UPDATE` statement will cause all workflow update operations to fail with a syntax error.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P0: Typo introduces SQL syntax error. The stray y character at the beginning of the UPDATE statement will cause all workflow update operations to fail with a syntax error.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At mcp-studio/server/tools/workflow.ts, line 305:

<comment>Typo introduces SQL syntax error. The stray `y` character at the beginning of the `UPDATE` statement will cause all workflow update operations to fail with a syntax error.</comment>

<file context>
@@ -302,7 +302,7 @@ async function updateWorkflowCollection(
 
   const sql = `
-        UPDATE workflow_collection
+y        UPDATE workflow_collection
         SET ${setClauses.join(&quot;, &quot;)}
         WHERE id = ?
</file context>

Suggested change

	y UPDATE workflow_collection
	UPDATE workflow_collection