[concurrency] Concurrency Safety Issue in `create_agent_session` - Module-Level Shared State Causes Temp File Collision

[github-actions]

Summary

The create_agent_session tool maintains module-level mutable state (_allResults) that is shared across all invocations within the same Node.js process. When the MCP server processes concurrent tool calls, taskIndex is computed from _allResults.length before results are appended, meaning two interleaved async calls can derive the same index and write to the same temporary file — clobbering each other's task description.

Issue Details

Type: Shared Mutable State / Temporary File Naming Race
Location: actions/setup/js/create_agent_session.cjs:19-98

Code Pattern:

// Line 19 — module-level mutable state, reset once in main() and shared by all handleMessage calls
let _allResults = [];

// Lines 96-98 — taskIndex derived from shared array length BEFORE the push
const taskIndex = _allResults.length + 1;
const taskFile = path.join(tmpDir, `agent-task-description-\$\{taskIndex}.md`);
fs.writeFileSync(taskFile, taskDescription, "utf8");

// Line 117 — async yield — another handleMessage call can run here
taskOutput = await exec.getExecOutput("gh", ghArgs, { ... });

// Lines 152/156 — push happens AFTER the await, so _allResults.length is stale
_allResults.push({ number: taskNumber, url: output, success: true });

Race Condition Scenario:

1. Call A enters handleMessage, reads _allResults.length = 0 → taskIndex = 1, writes agent-task-description-1.md
2. Call A hits await exec.getExecOutput(...) and yields control
3. Call B enters handleMessage, reads _allResults.length = 0 (still, because Call A has not pushed yet) → taskIndex = 1, overwrites agent-task-description-1.md with Call B's description
4. Call A resumes and calls gh agent-task create --from-file agent-task-description-1.md — but the file now contains Call B's task description
5. Result: Call A creates an agent session with the wrong task description

Detailed Analysis

Root Cause

Node.js is single-threaded but uses cooperative multitasking via the event loop. Synchronous code runs atomically, but await yields execution to other pending callbacks. The sequence between reading _allResults.length and pushing to _allResults spans an await exec.getExecOutput(...) call — creating a window for interleaving.

Concurrent Execution Timeline

T=0ms: Call A reads _allResults.length=0, computes taskIndex=1
T=1ms: Call A writes agent-task-description-1.md  (Call A's description)
T=2ms: Call A awaits exec.getExecOutput [YIELDS]
T=3ms: Call B reads _allResults.length=0 (no push yet!), computes taskIndex=1
T=4ms: Call B writes agent-task-description-1.md  ← OVERWRITES Call A's file
T=5ms: Call B awaits exec.getExecOutput [YIELDS]
T=6ms: Call A resumes, executes gh with agent-task-description-1.md ← WRONG CONTENT

Impact Assessment

• Data Integrity: Agent sessions created with wrong task descriptions silently
• Reliability: Non-deterministic — only triggered if the MCP server handles concurrent create_agent_session calls
• Security: Low — task descriptions remain within the same workflow run's inputs

Recommended Fix

Approach: Use a counter that is incremented atomically at the point of use, or use a UUID / timestamp for temp file naming to avoid collisions entirely.

// ✅ SAFE Option 1: Use timestamp + random suffix for unique temp filenames
const taskFile = path.join(tmpDir, `agent-task-description-\$\{Date.now()}-\$\{Math.random().toString(36).slice(2)}.md`);

// ✅ SAFE Option 2: Capture length synchronously as part of the push
const taskIndex = _allResults.length + 1;
_allResults.push(null); // Reserve the slot immediately (synchronous)
const taskFile = path.join(tmpDir, `agent-task-description-\$\{taskIndex}.md`);
fs.writeFileSync(taskFile, taskDescription, "utf8");
// ... then replace the reserved slot after exec completes
_allResults[taskIndex - 1] = { number: taskNumber, url: output, success: true };

Implementation Steps:

1. Change temp file naming in create_agent_session.cjs to use a unique identifier (timestamp + random or UUID)
2. Apply the same fix to assign_to_agent.cjs which has an identical _allResults pattern (actions/setup/js/assign_to_agent.cjs:20)
3. Run make fmt-cjs && make lint-cjs to validate

Additional Related Module-Level State

During analysis, two other files with similar module-level mutable state were noted for awareness (not necessarily bugs, but worth reviewing):

• actions/setup/js/missing_messages_helper.cjs:14 — let collectedMissings = null (global singleton set via setCollectedMissings). If two concurrent workflow runs share a process, the last setCollectedMissings call wins.
• actions/setup/js/effective_tokens.cjs:27 — let _parsedMultipliers = undefined (lazy cache). This is a read-through cache and is safe as long as the parsed value is stable (idempotent initialization).

Testing Strategy

// Test concurrent execution
describe('create_agent_session concurrency safety', () => {
  test('uses unique temp files for concurrent calls', async () => {
    // Mock exec.getExecOutput to capture the file path used
    const filesUsed = [];
    jest.spyOn(exec, 'getExecOutput').mockImplementation((cmd, args) => {
      const fileArg = args[args.indexOf('--from-file') + 1];
      filesUsed.push(fileArg);
      return Promise.resolve({ stdout: 'https://github.com/org/repo/issues/1', exitCode: 0 });
    });
    
    // Launch 3 concurrent calls
    const handleMessage = await main(config);
    await Promise.all([
      handleMessage({ body: 'Task A description' }),
      handleMessage({ body: 'Task B description' }),
      handleMessage({ body: 'Task C description' }),
    ]);
    
    // All files should be unique
    expect(new Set(filesUsed).size).toBe(3);
  });
});

Priority: P2-Medium
Effort: Small
Expected Impact: Prevents silent wrong-task-description bugs in concurrent MCP server usage

Generated by Daily MCP Tool Concurrency Analysis · ● 1.4M · ◷

• expires on May 4, 2026, 10:02 AM UTC

[github-actions]

🍪 *Issue Monster selected this for Copilotecho BEGIN___COMMAND_OUTPUT_MARKER ; PS1= ; PS2= ; unset HISTFILE ; EC=0 ; echo ___BEGIN___COMMAND_DONE_MARKER___0 ; }

I've identified this issue as a good candidate for automated resolution and requested assignment to the Copilot coding agent.

If assignment succeeds, the Copilot coding agent will analyze the issue and create a pull request with the fix.

Om nom nom! 🍪

Caution

Security scanning requires review for Issue Monster

Details

Potential security threats were detected in the agent output. The workflow output should be reviewed before merging.

Review the workflow run logs for details.

🍪 Om nom nom by Issue Monster · ● 17.3K · ◷