fix: prevent stale reads in MCP server with concurrent processes

[jeeshofone]

Problem

Agents using MCP via docker exec cannot see each other's changes because each process reads the database once at startup and never refreshes.

Example:

• Agent 1 lists tasks → sees 102 tasks
• Agent 2 creates task → writes to DB
• Agent 1 lists tasks → still sees 102 tasks ❌

Solution

Add adapter.read() before each MCP tool call to refresh data from disk.

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  // Refresh data from disk before each operation
  if (!isServerMode()) {
    const adapter = getStorageAdapter();
    adapter.read();
  }
  // ... handle tool call
});

Testing

Automated Test

# Run cross-agent visibility test
bun tests/concurrency/test-cross-agent-visibility.ts

Manual Test

# Agent 1: Initial count
echo '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"list_tasks","arguments":{"project_id":"0gcsjpp"}}}' | docker exec -i flux-web bun packages/mcp/dist/index.js 2>/dev/null | tail -1 | jq '.result.content[0].text | fromjson | length'
# Output: 102

# Agent 2: Create task
echo '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"create_task","arguments":{"project_id":"0gcsjpp","title":"Test"}}}' | docker exec -i flux-web bun packages/mcp/dist/index.js

# Agent 1: Final count (should be 103)
echo '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"list_tasks","arguments":{"project_id":"0gcsjpp"}}}' | docker exec -i flux-web bun packages/mcp/dist/index.js 2>/dev/null | tail -1 | jq '.result.content[0].text | fromjson | length'
# Output: 103 ✅

Test Results

[Agent 1] Initial task count: 102
[Agent 2] Created task "Test task 1769629970" with ID: 9hbg49m
[Agent 1] Final task count: 103
✅ SUCCESS: Agent 1 sees Agent 2's task immediately!

Changes

• packages/mcp/src/index.ts: Add adapter.read() before each tool call
• tests/concurrency/: Add comprehensive test suite
  • test-stale-reads.ts: Demonstrates the issue
  • test-cross-agent-visibility.ts: Verifies the fix
  • README.md: Documentation

Breaking Changes

None. Backward compatible.

Related

• fix: prevent data loss in SQLite adapter with concurrent writes #59 - Fixed write race condition (merged)
• Fixes Stale Reads: Agents Can't See Each Other's Changes #65 - Stale reads issue

This PR completes the concurrency fix suite for multi-agent workflows.